Escolar Documentos
Profissional Documentos
Cultura Documentos
Investigação Operacional
Volume 22
Número 2
Apdio Dezembro 2002
CESUR - Instituto Superior Técnico
Av. Rovisco Pais - 1049 - 001 LISBOA
Telef. 21 840 74 55 - Fax. 21 840 98 84
http://www.apdio.pt
Propriedade:
APDIO — Associação Portuguesa de Investigação Operacional
ESTATUTO EDITORIAL
Patrocinadores
Fundação Calouste Gulbenkian
ISSN nº 0874-5161
Dep. Legal nº 130 761 / 98
Execução Gráfica: J. F. Macedo - Astrografe
700 Ex.
2002/12
Comissão Editorial
Esta Revista é distribuı́da gratuitamente aos sócios da APDIO. As informações sobre inscrições
na Associação, assim como a correspondência para a Revista devem ser enviadas para a sede
da APDIO - Associação Portuguesa de Investigação Operacional - CESUR, Instituto Superior
Técnico, Av. Rovisco Pais, 1049-001 Lisboa.
Assinatura: 25 Euros
L. Grilo, I. Themido / Investigação Operacional, 22 (2002) 147-166 147
∗
Luı́s Miguel Grilo Isabel Hall Themido∗†
∗
Escola Superior de Tecnologia, Instituto Politécnico de Tomar
lgrilo@ipt.pt
†
IST/CESUR, Universidade Técnica de Lisboa
ithemido@cesur.civil.ist.utl.pt
Abstract
This paper presents the development of regression models to explain the sales of two
consumer groups of products that present remarkable variability and seasonality (personal
hygiene paper pulp based products). Models for the market share of two groups of products
and for the sales of the total market were also developed in order to evaluate alternative
approaches to sales forecasting.
Using data from Renova, the Portuguese leader in this market, several linear models
were developed. In these models price is an important explanatory variable. Another
variable that is also significant in some models is the market share of other brands, where
own brands have the major contribution. For the sales models, the lagged sales variable has
a significant explanatory power, which reflects the importance of the brand under analysis
and consumer habits.
Resumo
Keywords: Personal Hygiene Paper, Sales Modelling, Forecasting, Market Share, Regression, Marke-
ting
Title: Sales modelling of consumer products: an application to the personal hygiene paper market
1 Introdução
A Renova é uma empresa nacional que se dedica ao marketing e venda de produtos de grande
consumo na área da higiene e limpeza, sendo que os produtos comercializados ultrapassam
uma centena e encontram-se distribuı́dos por mais de 15 categorias. Esta dinâmica potencia
a sua integração numa rede internacional de comércio, deixando de ser um mero actor a
nı́vel nacional. A sua nova estratégia consiste, simultaneamente, numa aposta na crescente
satisfação do consumidor, nas preocupações ambientais e no reforço da sua posição de liderança
no mercado.
Não se sabe, ao certo, quando teve inı́cio a produção de papel em Portugal, mas a crer
na tradição, terá sido, já há mais de dois séculos, durante a época pombalina. O que sabe-
mos com segurança é que, actualmente, os produtos de papel são muito importantes na vida
quotidiana, a ponto de serem mesmo considerados produtos de primeira necessidade. Deste
modo, consideramos interessante realizar este trabalho cujo objectivo é desenvolver modelos
econométricos, que permitam perceber (conhecer) o impacto das variáveis de marketing sobre
as vendas e, eventualmente, prever, no curto e médio prazo, as vendas de, somente, dois grupos
de produtos da megamarca Renova, os quais existem, desde há vários anos, no mercado por-
tuguês. Por razões estratégicas e de sigilo, passaremos a designá-los por Grupo X (somatório
de cinco produtos) e Grupo Y (somatório de 7 produtos).
Como nos cingimos, apenas, ao mercado interno, mais concretamente a Portugal Continen-
tal (não considerando, portanto, uma análise às vendas nos mercados internacionais 1 , nem nos
mercados das regiões autónomas dos Açores e Madeira), julgamos importante analisar, breve-
mente, em que condições se realiza a distribuição em Portugal, dado estarmos conscientes da
multiplicidade de operações que, actualmente, ligam a produção ao consumo, bem como do
facto do mercado português estar cada vez mais dependente das regras de distribuição. Deste
modo, para melhor entendermos o processo de vendas num contexto integrado, passamos a
caracterizar a cadeia logı́stica que serve de base ao presente estudo.
1
Actualmente, a empresa exporta os seus produtos, apenas, para o mercado espanhol. No entanto, conscien-
tes de que a aposta noutros mercados implica uma estratégia bem delineada a nı́vel de produtos (nomeadamente
no que respeita ao mercado europeu), os seus responsáveis afirmaram, em entrevista concedida à revista Exame,
que uma estratégia a médio prazo poderá passar pela presença dos seus produtos nos palop ou na Europa de
Leste (bastante atractivos, em termos de custos).
L. Grilo, I. Themido / Investigação Operacional, 22 (2002) 147-166 149
Consumidor Final
De acordo com informação Nielsen2 , Portugal é o paı́s da Europa que possui maior número
de mercearias (comércio tradicional) por habitante. No entanto, as mesmas têm vindo a dar
lugar a formas mais recentes de comércio, como os supermercados, hipermercados e os discount.
Enquanto que os primeiros estão já muito próximos da maturidade, os hipermercados estão
em plena fase de crescimento (aumentando a sua importância, através da abertura de novos
pontos de venda) e os últimos estão, ainda, em fase de introdução3 .
O que nos propomos fazer neste ponto do trabalho, consiste em analisar, por um lado,
a posição da Renova no mercado, verificando (por exemplo) se as suas vendas reflectiram o
panorama económico do perı́odo em análise e, por outro lado, tentando, sempre que possı́vel,
compará-la com as marcas concorrentes que operam na mesma área de negócio. Para proce-
dermos a tal análise, recorremos, sobretudo, a dados bimestrais Nielsen, para o perı́odo Janeiro
de 1992 a Dezembro de 1996 (dimensão da amostra: 30 observações).
Antes, porém, de avançarmos, convém salientar que existem diferenças entre as sucessões
das vendas da Renova, quando comparadas com as compras dos retalhistas captadas pela
Nielsen (mesmo que se trate do mesmo produto). As vendas da Renova são rigorosamente
conhecidas, enquanto que as compras do consumidor final só podem ser estimadas a partir de
dados recolhidos pela Nielsen. Infelizmente, estes últimos dados representam, em média, cerca
de 77%4 do mercado total, ou seja, a Nielsen não cobre a totalidade dos circuitos de distri-
buição para os grupos de produtos estudados. Para além disso, também não recolhe informação
dos grossistas – as suas compras à Renova só aparecem como compras Nielsen na leitura dos
pequenos retalhistas. De salientar, ainda, que, se compararmos as sucessões das vendas ao
consumidor final e compras dos retalhistas (ambas em volume) recolhidas pela Nielsen (Figura
2), verificamos que o seu andamento é bastante semelhante, facto que, em nosso entender se
deve ao conceito adoptado recentemente pelos retalhistas no mercado português (principal-
mente pelas grandes superfı́cies), quando se trata de novas aquisições: filosofia just-in-time.
É sabido, que os retalhistas conhecem, com mais ou menos precisão, as compras do consumi-
dor final, de modo que compram, em cada bimestre5 , o suficiente para satisfazer a respectiva
2
Multinacional criada nos eua e instalada em Portugal desde 1968; presta serviços na área do consumo, for-
necendo informação sobre o comportamento dos produtos e marcas nos pontos de venda do mercado Continental
ou total (designação da Nielsen).
3
Para um maior desenvolvimento deste assunto, veja-se [6].
4
Valor estimado por nós, com base no perı́odo em análise [4].
5
Note-se que nos referimos às compras, assim como a outras variáveis, em termos bimestrais, pelo facto
150 L. Grilo, I. Themido / Investigação Operacional, 22 (2002) 147-166
18000 1e5
16000 90000
14000 80000
12000 70000
10000 60000
8000 50000
6000 40000
4000 30000
Vendas Vendas
2000 20000
Compras Compras
Stock Stock
29 DN
39 DN
49 DN
59 DN
69 DN
29 FJ
39 FJ
49 FJ
59 FJ
69 FJ
29 AJ
39 AJ
49 AJ
59 AJ
69 AJ
29 OS
39 OS
49 OS
59 OS
69 OS
29J M
39J M
49J M
59J M
69J M
29 A M
39 A M
49 A M
59 A M
69 A M
29 DN
39 DN
49 DN
59 DN
69 DN
29 FJ
39 FJ
49 FJ
59 FJ
69 FJ
29 AJ
39 AJ
49 AJ
59 AJ
69 AJ
29 OS
39 OS
49 OS
59 OS
69 OS
29J M
39J M
49J M
59J M
69J M
29 A M
39 A M
49 A M
59 A M
Figura 2: Representação das vendas, compras e stocks em volume dos retalhistas, para os Grupos X
e Y.
42000 1,4e5
16000
40000
1,2e5
) s e da di nu 0001( l at o T o dacr e Ms a dne V
15000
38000
1e5
14000 36000
80000
34000
13000
60000
32000
12000
30000 40000
39 DN
49 DN
59 DN
69 DN
29 FJ
39 FJ
49 FJ
59 FJ
69 FJ
29 AJ
39 AJ
49 AJ
59 AJ
69 AJ
29 OS
39 OS
49 OS
59 OS
69 OS
29J M
39J M
49J M
59J M
69J M
29 A M
39 A M
49 A M
59 A M
69 A M
29 DN
39 DN
49 DN
59 DN
69 DN
29 FJ
39 FJ
49 FJ
59 FJ
69 FJ
29 AJ
39 AJ
49 AJ
59 AJ
69 AJ
29 OS
39 OS
49 OS
59 OS
69 OS
29J M
39J M
49J M
59J M
69J M
29 A M
39 A M
49 A M
59 A M
69 A M
Tempo (Bimestres) Tempo (Bimestres)
Figura 3: Representação das vendas em volume dos Grupos X e Y, para o mercado total e para a
Renova.
procura, mantendo, obviamente, uma margem de stocks apenas para evitar rupturas.
desta ser a maior desagregação do tempo que nos foi possı́vel fazer. Estamos, no entanto, conscientes, de que
as compras dos retalhistas não se fazem uma única vez no bimestre, assim como as vendas ao consumidor final,
que são diárias.
L. Grilo, I. Themido / Investigação Operacional, 22 (2002) 147-166 151
De acordo com alguns especialistas, a razão para que o mercado total registe, actualmente,
tal tendência tem a ver com facto de o PIB também se encontrar a crescer. Outro motivo
apontado para este crescimento, deve-se ao facto de termos assistido, nos últimos tempos, a
uma redução na quantidade de papel por unidade, o que induz, obviamente, a um aumento
do consumo de unidades dos produtos deste grupo, por forma a manter o consumo de papel,
por indivı́duo. De notar, que o preço tem vindo a decrescer ao longo do perı́odo em análise,
havendo, ainda, quem considere que este produto está, também, a ser utilizado para fins
diferentes do habitual.
Também neste bimestre, as vendas da Renova são superiores às verificadas nos outros
bimestres. No entanto, a partir do inı́cio de 1995 nota-se uma ligeira tendência decrescente
desta sucessão, pois, tal como havı́amos verificado no caso do Grupo X, também no Grupo
Y, as marcas da distribuição começaram a ter um maior peso no mercado total (embora aqui
o peso seja ligeiramente menor – Figura 4). De notar, todavia, que a Renova não perdeu
significativa quota de mercado para as outras marcas, no perı́odo em análise, e que, também,
as vendas do mercado total são relativamente inferiores nos últimos bimestres.
No que diz respeito às vendas e preços, no Grupo X, é possı́vel descortinar uma relação
inversa8 , em certa medida já esperada, entre estas variáveis (Figura 5). Uma particularidade
6
As marcas da distribuição (também, conhecidas por marcas próprias) entraram pela primeira vez em Por-
tugal em 1984. Dadas as poucas lojas que as vendiam, a sua quota de mercado era inferior a 1%, donde a sua
importância foi-se reduzindo, acabando por se extinguirem no final da década de 80. Ressurgiram no mercado
devido ao crescente número e importância dos hipermercados e à evolução das principais organizações retalhistas
[2].
7
De salientar que, a variável designada por outras marcas neste estudo, não representa as principais marcas
concorrentes da Renova, como a Scott, Kleenex ou Colhogar, mas antes marcas menos conhecidas ou com
menos importância no mercado. O peso das marcas da distribuição nas outras marcas tem vindo a crescer,
representando, no final do perı́odo em estudo, mais de 70%, quando nos estamos a referir ao Grupo X, e 90%, no
caso do Grupo Y. Actualmente, a Nielsen já considera a separação, entre outras marcas e marcas da distribuição
mas, lamentavelmente, para o perı́odo considerado, não disponhamos de tal desagregação.
8
Em [8], encontramos um gráfico, referente a um estudo sobre produtos de papel, onde é notória uma relação
inversa entre os preços de uma marca de papel higiénico e as vendas correspondentes.
152 L. Grilo, I. Themido / Investigação Operacional, 22 (2002) 147-166
50 70
46 60
42 50
38 40
34 30
30 20
26 10
22 0
18 -10
Renova Renova
29 DN
39 DN
49 DN
59 DN
69 DN
29 FJ
39 FJ
49 FJ
59 FJ
69 FJ
29 AJ
39 AJ
49 AJ
59 AJ
69 AJ
29 OS
39 OS
49 OS
59 OS
69 OS
29J M
39J M
49J M
59J M
69J M
29 A M
39 A M
49 A M
59 A M
69 A M
29 DN
39 DN
49 DN
59 DN
69 DN
29 FJ
39 FJ
49 FJ
59 FJ
69 FJ
29 AJ
39 AJ
49 AJ
59 AJ
69 AJ
29 OS
39 OS
49 OS
59 OS
69 OS
29J M
39J M
49J M
59J M
69J M
29 A M
39 A M
49 A M
59 A M
69 A M
Figura 4: Quota de mercado Renova e quota das outras marcas para os Grupos X e Y.
17000 52 95000 23
16000 22
85000
48
) e da di nu/s o ducs e( o da noi c al f e d oç er P
21
40 19
13000 55000
18
36
12000 45000
17
11000 32 35000 16
Vendas Vendas
Preços Preços
29 DN
39 DN
49 DN
59 DN
69 DN
29 FJ
39 FJ
49 FJ
59 FJ
69 FJ
29 AJ
39 AJ
49 AJ
59 AJ
69 AJ
29 OS
39 OS
49 OS
59 OS
69 OS
29J M
39J M
49J M
59J M
69J M
29 A M
39 A M
49 A M
59 A M
69 A M
29 DN
39 DN
49 DN
59 DN
69 DN
29 FJ
39 FJ
49 FJ
59 FJ
69 FJ
29 AJ
39 AJ
49 AJ
59 AJ
69 AJ
29 OS
39 OS
49 OS
59 OS
69 OS
29J M
39J M
49J M
59J M
69J M
29 A M
39 A M
49 A M
59 A M
69 A M
Tempo (Bimestres) Tempo (Bimestres)
Os dados recolhidos são provenientes de três fontes de informação. Como a fonte externa a que
a Renova recorre é a Nielsen, foram-nos cedidos (pela fonte interna - Renova), entre outros,
dados sobre: as vendas ao consumidor final9 , a quota de mercado, o preço ao consumidor,
stock nos retalhistas e as distribuições numérica e ponderada. Por outro lado, recorremos ao
INE e ao Banco de Portugal, como fontes documentais; na primeira recolhemos informações
9
As vendas a retalho de muitos bens de consumo podem ser estimadas para um determinado perı́odo, a partir
de uma amostra estratificada do bem em análise. É feito um inventário do montante de produto existente na
prateleira e na área de armazém temporário, ao mesmo tempo que é recolhida informação sobre as compras
feitas pelo retalhista desde a última auditoria. Assim, Vendas = Stock inicial + Compras – Stock final.
L. Grilo, I. Themido / Investigação Operacional, 22 (2002) 147-166 153
sobre o Produto Interno Bruto a preços correntes (PIBpc ) e o Índice de Preços no Consumidor
(IPC) e, na última, o Índice de Preços implı́citos no PIB.
Vendas (t) – representa as vendas dos retalhistas para os grupos de produtos Renova
(compras do consumidor final), no perı́odo t (em 1000 unidades).
Quota (t) – variável dependente (em percentagem) que representa a quota de mercado
das vendas em volume da Renova, no perı́odo t.
Vendas totais (t) – variável dependente que representa as vendas dos retalhistas para o
mercado total, no perı́odo t (em 1000 unidades).
a) Variáveis de Marketing :
Preço deflacionado (t)– variável que representa o preço médio deflacionado dos produtos
Renova (em escudos por unidade), ao consumidor final. O processo de deflação dos preços teve
por base o IPC mensal, para a classe de artigos “outros bens e serviços cuidados pessoais não
duráveis”.
Preço relativo (t)– esta variável (em percentagem), foi obtida pelo rácio entre o preço
médio Renova e o preço médio do mercado total.
Preço total deflacionado (t)– representa o preço médio deflacionado do mercado total
(em escudos por unidade), ao consumidor final. O processo de deflação foi o utilizado na
variável preço deflacionado.
Distribuição ponderada (t) – percentagem calculada com base nas vendas em valor da
classe de produtos das lojas “negociantes” verificadas no perı́odo, sobre o total das vendas
em valor realizadas durante o mesmo perı́odo em todas as lojas que negoceiam a classe de
produtos.
Stock (t) e sotck total (t)– stock em volume, existente no retalho, da marca Renova e
do mercado total, respectivamente.
10
Algumas destas variáveis também foram consideradas desfasadas.
154 L. Grilo, I. Themido / Investigação Operacional, 22 (2002) 147-166
Perı́odo (t) – esta variável irá permitir que o modelo ajuste uma tendência, caso se
verifique a sua existência. Foi construı́da da seguinte forma: atribuı́mos o valor 1 ao pri-
meiro perı́odo de vendas disponı́vel, o bimestre Janeiro/Fevereiro 92, o valor 2 ao bimestre,
Março/Abril 92 e assim sucessivamente.
Seno(t) e coseno(t) – como muitas funções periódicas podem representar-se por uma com-
binação linear de senos e cosenos, podemos estimar a componente sazonal, numa perspectiva
determinı́stica, através de uma análise de regressão em termos de polinómios trigonométricos 11 .
JF, MA, MJ, JA, SO, ND – estas variáveis representam, respectivamente, os bimes-
tres Janeiro/Fevereiro, Março/Abril, Maio/Junho, Julho/Agosto, Setembro/Outubro e No-
vembro/Dezembro. São variáveis binárias (do tipo dummy), cujos coeficientes representam a
sazonalidade, com o seguinte significado:
(
1, se nos encontramos no bimestre correspondente
Bimestre = .
0, caso contrário
PIB (t) – variável externa que representa o Produto Interno Bruto a preços constantes de
1992 (deflacionado com recurso ao Índice de Preços implı́citos no PIB ou deflactor do PIB)12 .
Vendas (t-1) e vendas totais (t-1) – variáveis dependentes desfasadas (vendas da Renova
e do mercado total, respectivamente, no bimestre t-1), que representam os hábitos de consumo
ou a própria notoriedade da marca (apenas no caso das vendas Renova) em cada perı́odo.
d) Acções da Concorrência:
Este é, certamente, um conjunto de variáveis que influenciam as vendas da Renova, contudo,
algumas delas, dificilmente as podemos prever e quantificar no passado e no presente.
11
De acordo com [10], como os dados são bimestrais e a sazonalidade é anual, terı́amos: sen(2πt/6) e
cos(2πt/6), contudo, estas variáveis não se mostraram estatisticamente significativas.
12
Note-se que, embora os valores desta variável fossem efectivos até 1993, para evitarmos quebras de estrutura
(pois, a partir deste ano a recolha do indicador em causa passou a ser feita de forma manifestamente diferente),
estimámos esses mesmos valores, tendo por base a taxa de variação verificada na sucessão recolhida até esta
data. Posteriormente, tivemos ainda que proceder a uma desagregação da variável, pois a mesma apresenta-se
disponı́vel apenas trimestralmente, quando necessitamos de dados bimestrais.
L. Grilo, I. Themido / Investigação Operacional, 22 (2002) 147-166 155
Quota outras marcas (t) – dá-nos a quota de mercado, em percentagem, das outras
marcas.
Obviamente que outras variáveis, desde que disponı́veis13 , poderiam ter sido consideradas,
por exemplo: publicidade, acções promocionais, topos de gôndola, folhetos, etc..
O orçamento gasto em publicidade, para ambos os grupos de produtos, não está disponı́vel
para todo o perı́odo em estudo, variável que poderia ser considerada desfasada um ou dois
bimestres, para que o modelo implementasse a influência da publicidade um e/ou dois bimestres
depois. Seria, ainda, interessante se tivéssemos disponı́vel o valor da publicidade da Renova
em termos relativos, face à despesa total do mercado em publicidade. Saliente-se, no entanto,
que de acordo com um responsável da Renova, aliás, em consonância com alguns estudos
empı́ricos, a publicidade é uma variável que, para o tipo de produtos em estudo, não deverá
ser significativa14 .
k
X
Vt =α0 + αi Xit + εt
i=1
onde,
εt = termo aleatório.
Quando são estabelecidas relações mais complexas entre variáveis, ao reflectir a interacção
provável entre elas, o modelo de regressão multiplicativo considera-se um melhor ajustamento:
k
Xitαi .εt ,
Y
Vt = α 0 .
i=1
k
X
lnVt = lnα0 + (αi .lnXit ) + ln εt .
i=1
Nos modelos citados, os coeficientes αi são determinados por ajuste da sucessão histórica
e representam a importância de cada variável independente na explicação da variável vendas,
enquanto que ε é a variável residual.
m
X
Qit = Vit / Vjt
j=1
m
X
Vit = Qit . Vjt .
j=1
15
Os coeficientes dos modelos multiplicativos (elasticidades) possuem grande vantagem, quando comparados
com os do modelo aditivo, pois adimensionam a contribuição das variáveis independentes na regressão, já que
reflectem a variação percentual das vendas, dada uma variação de 1% na variável independente.
16
Uma propriedade desejável para qualquer modelo de quota de mercado é que este possua consistência lógica,
isto é, as estimativas produzidas pelo modelo de quota de mercado variam entre 0 e 1, devendo a soma dessas
quotas de mercado estimadas para todas as marcas, num determinado perı́odo, ser igual a 1 [5].
L. Grilo, I. Themido / Investigação Operacional, 22 (2002) 147-166 157
Dos estudos realizados em Portugal, importa destacar um trabalho onde, recorrendo a da-
dos bimestrais Nielsen, se usaram modelos de regressão para previsão de vendas bimestrais ao
retalho e ao consumidor final de um produto de higiene pessoal de grande consumo. Nos mo-
delos obtidos, o stock no mercado (funcionando como barómetro do canal logı́stico) e algumas
variáveis de marketing, surgem como principais variáveis explicativas [12].
2.4 Modelação
- modelar as quotas de mercado e as vendas do mercado total, para ambos os grupos, com
o intuito de proporcionar modelos alternativos de previsão de vendas.
A selecção de variáveis a serem incluı́das num modelo de regressão é uma das tarefas mais
difı́ceis, pois se, por um lado, temos que ser cautelosos no sentido de não eliminarmos variáveis
explicativas importantes, dado que isso iria prejudicar o poder explicativo do modelo e logo
conduzir a estimativas enviesadas dos coeficientes de regressão e das previsões efectuadas, por
outro, são vários os autores a afirmar que, na prática, os modelos parcimoniosos conduzem, em
158 L. Grilo, I. Themido / Investigação Operacional, 22 (2002) 147-166
Quadro 1: Valores estimados para o modelo de vendas Renova ao consumidor final: Grupo X.
Vendas (t) = 60794.25 – 240.73 Preço relativo (t) – 243.14 Distribuição ponderada (t-1) (1)
2
Variáveis Coeficiente Desvio t p – level Ord. Ent. R Aj. (%)
Padrão
Constante 60794.25 8932.36 6.81 0.00 – –
Preço relativo (t) -240.73 51.37 -4.69 0.00 1 26.3
Distribuição
ponderada (t-1) -243.14 68.73 -3.54 0.00 2 47.8
R2 (%) R2 Ajustado D. Padrão Erro Absoluto F da Graus
(%) Regressão Médio (%) Regressão Liberdade
Stepwise 51.4 47.8 750.8 3.9 14.3 27
Entre outros aspectos, como o de validar o sinal de cada coeficiente, tivemos sempre em
atenção o seu nı́vel de significância17 . Para além de assegurar que os parâmetros estimados
tenham as propriedades estatı́sticas desejáveis, é imprescindı́vel verificar se o modelo obtido
respeita as premissas básicas. Assim, para testar a hipótese de linearidade usámos o teste
reset (pois o que está em causa é a análise da forma funcional), para a heterocedasticidade
o teste Breusch-Pagan (baseado nos multiplicadores de Lagrange) e, para a autocorrelação o
teste Breusch-Durbin-Godfrey [3]. Verificámos, ainda, a normalidade dos resı́duos por simples
observação do gráfico que reflecte a probabilidade destes se aproximarem de uma curva com
distribuição Normal.
Os modelos aditivos, que analisaremos a seguir para os Grupos X e Y, foram os que melhor
se adaptaram às sucessões disponı́veis e que “passaram” nos testes às hipóteses básicas.
Modelos das Vendas Renova Na equação (1) o preço relativo surge com coeficiente ne-
gativo, pois quanto maior o seu valor menor será a predisposição dos indivı́duos para adquirir
produtos do Grupo X, o que se justifica pelo seguinte motivo: está implı́cito que um maior
17
Sendo desejável que este valor seja inferior a 5%, considerámos, excepcionalmente, e desde que a sua
introdução não causasse instabilidade no modelo, estimativas com nı́veis de significância não superior a 15%.
Na área de marketing, esta prática verifica-se com alguma frequência, com o intuito de dar visibilidade a variáveis
conceptualmente importantes, embora estatisticamente pouco relevantes face aos limitados dados disponı́veis
[12 ].
L. Grilo, I. Themido / Investigação Operacional, 22 (2002) 147-166 159
Quadro 2: Valores estimados para o modelo de vendas Renova ao consumidor final: Grupo X.
Vendas (t) = 23970.32 – 237.94 Preço relativo (t) + 6.00 PIB (t) + 0.25 Vendas (t-1) (2)
2
Variáveis Coeficiente Desvio t p – level Ord. Ent. R Aj. (%)
Padrão
Constante 23970.32 7960.41 3.01 0.00 – –
Preço relativo (t) -237.94 70.75 -3.36 0.00 1 26.3
PIB (t) 6.00 2.71 2.22 0.04 2 35.1
Vendas (t-1) 0.25 0.17 1.53 0.14 3 38.2
R2 (%) R2 Ajustado D. Padrão Erro Absoluto F da Graus
(%) Regressão Médio (%) Regressão Liberdade
Stepwise 44.6 38.2 816.8 4.4 7.00 26
valor desta variável pressupõe a existência de um preço mais elevado (dos produtos Renova,
relativamente às restantes marcas), logo mais próximo do preço de reserva de cada consumidor
(o que implica que o valor do Excedente Lı́quido do Consumidor, será menor) – neste sentido,
verificamos uma rigidez à subida de preço.
Em relação à variável distribuição ponderada, numa primeira análise, dirı́amos ter sinal
simétrico do teoricamente esperado. No entanto, esta variável apresenta uma correlação ne-
gativa com a variável dependente, ao contrário, do que a teoria deixa prever, dado que esta
instância, em particular, tem associado o fenómeno seguinte: embora a marca Renova esteja
presente nas superfı́cies comerciais com maior volume de vendas (no que respeita, obviamente,
à comercialização da classe de produtos onde a marca se insere), é também nestas lojas (hi-
permercados) que o consumidor se depara com a existência de marcas próprias, com preços
menores (pois, nos supermercados e comércio tradicional, tais produtos não existem). Na
verdade, numa análise que fizemos às quotas de mercado nestes três tipos diferentes de lojas
(para os últimos bimestres do perı́odo em análise), constatámos que as vendas Renova deste
grupo de produtos se reparte da seguinte forma: cerca de 40%, para o comércio tradicional (o
qual tem registado uma ligeira quebra, dado o encerramento de parte destas lojas), enquanto
os supermercados e hipermercados são responsáveis por cerca de 30% para cada loja. Assim,
muito provavelmente as vendas da Renova serão superiores sempre que esta marca estiver pre-
sente em lojas de menor volume de negócios, pois aqui não estão sujeitas à concorrência das
marcas da distribuição. Neste contexto, é razoável admitir que o sinal não estará incorrecto.
Um outro modelo de vendas, que consideramos interessante apresentar, tem a equação (2).
Conforme se pode ver no Quadro 2, este modelo apresenta um erro percentual absoluto
médio ligeiramente mais elevado que o anterior. Em (2) o PIB e as vendas desfasadas são as
variáveis explicativas, em vez da distribuição ponderada de (1). O coeficiente positivo do PIB
poderá indiciar que o crescimento das vendas da Renova, que se verificou, principalmente,
a partir do bimestre Maio/Junho de 1995 ( ver Figura 3), terá sido, pelo menos em parte,
influenciado pelo crescimento do PIB, registado em quase todo o perı́odo em análise. Tais
resultados estão em conformidade com o que já havı́amos dito e permitem-nos classificar os
bens em causa como normais. A variável dependente desfasada que surge como explicativa,
embora seja a última a ser incluı́da no modelo, com p-level de 0.14, contribui com 3.1 pontos
percentuais para o R2 ajustado, em que o valor final obtido é de 38.2% (de salientar que, de
acordo com alguns autores, em modelos microeconómicos, é possı́vel que o valor deste indicador
seja relativamente baixo, permitindo, ainda assim, que tenhamos um bom modelo). O sinal
160 L. Grilo, I. Themido / Investigação Operacional, 22 (2002) 147-166
M od elo do Qu ad ro 2 .
Mo del o d o Quad ro 1.
20000
2000 0
16000
1600 0
12000
1200 0
8000
8000
4000
4000
Valores
Valores 0
0 Observados
Observados
Valores
Valores -4000
-4000 Previstos
Previstos
Resíduos
Resíduos
JF92
JF93
JF94
JF95
JF9 6
JA9 2
JA93
JA94
JA95
JA9 6
MJ92
N D92
MJ93
N D93
MJ94
ND9 4
MJ9 5
ND 95
M J9 6
ND 96
MA92
SO92
MA93
SO93
MA94
SO9 4
MA9 5
SO95
MA96
SO96
JF92
JA92
JF9 3
JA9 3
JF94
JA94
JF95
JA95
JF96
JA96
MJ92
ND 92
MJ9 3
ND 93
MJ94
N D94
MJ95
ND9 5
MJ9 6
ND 96
SO92
SO93
SO94
SO9 5
SO96
MA9 2
MA93
MA94
MA95
MA9 6
Tempo (Bimestres)
T empo (Bimestres)
Figura 6: Qualidade do ajuste dos modelos de vendas Renova ao consumidor final: Grupo X.
positivo significa que os hábitos de consumo (ou, a notoriedade da marca) tendem a manter-
se, muito provavelmente, no perı́odo actual. A qualidade dos ajustes a que os anteriores
modelos conduzem, bem como o comportamento dos respectivos resı́duos (estes aparentam
estacionaridade na média, que é zero, e na variância), pode observar-se na Figura 6.
Na equação (3), o preço relativo desfasado tem sinal teoricamente esperado. A quota das
outras marcas desfasada (onde as marcas próprias, com preço inferior, possuem o maior peso),
é a primeira variável a entrar no modelo e logo com um R2 ajustado de 64.8% (embora, com
p-level de .
0.13), o que mostra bem a importância desta variável neste modelo (Quadro 3). A
p
variável 1 3 P eriodo(t), permite ajustar uma tendência não linear (decaimento exponencial
amortecido), tal como esperávamos, pela observação do comportamento da quota de mercado
Renova18 (ver Figura 4).
18
De referir, a tı́tulo de curiosidade, que, em mercados onde a distribuição é já bastante sofisticada (como o
francês), se verifica a existência de um limite condicionado pelos fabricantes, ao crescimento das marcas próprias,
combatendo a distribuição (com recurso ao aumento das acções de promoção, por exemplo). De acordo, com
L. Grilo, I. Themido / Investigação Operacional, 22 (2002) 147-166 161
Modelos das Vendas do Mercado Total Para as vendas do mercado total apresentamos
as equações (4) e (5), com variáveis explicativas diferentes, mas onde os valores do R 2 e do
erro médio são sensivelmente iguais (Quadros 4 e 5).
Quando a concorrência é maior os preços dos bens têm tendência a baixar o que, associado
ao crescimento do PIB, faz aumentar o consumo, mesmo para estes bens, até porque os
consumidores tendem a diversificar o uso destes produtos. Daqui se compreende que em
(4) o sinal do coeficiente associado ao preço deflacionado seja negativo, revelando que estes
bens são ordinários. As variáveis D1 e D2 permitem que o modelo ajuste duas tendências
lineares crescentes, relativamente diferentes, confirmando que existe uma tendência crescente,
ligeiramente mais acentuada a partir do bimestre Setembro/Outubro95.
François Glémet, consultor da McKinsey, há um topo de quota de mercado para as marcas próprias que, para
a média de todas as categorias de produtos, situar-se-á entre 20 e 25% do mercado. Contudo Kotler prevê
que à medida que as marcas do retalho se tornam melhores e com maior aceitação, as marcas dos fabricantes
terão de descer os seus preços para um nı́vel próximo daquelas. Ainda segundo alguns especialistas, o valor da
marca sai reforçado nesta guerra e, mais do que nunca, construir e deter uma marca lı́der constitui um seguro
de vida nos tempos que correm. Numa amostra de 75 grandes superfı́cies retalhistas alimentares, recolhida em
Portugal, 53 venderam marcas próprias e o seu volume de vendas (503.6 milhões de contos) representou 97.2%
da facturação total da amostra (518.1 milhões de contos)]. Dado o peso actual das marcas da distribuição, as
marcas que actuam no mercado nacional começam a reagir, primeiramente com recurso à descida dos preços
[13,14].
162 L. Grilo, I. Themido / Investigação Operacional, 22 (2002) 147-166
Quadro 6: Valores estimados para o modelo de vendas Renova ao consumidor final: Grupo Y.
Vendas (t) = 354364.70 – 3589.10 Preço deflacionado (t-1) – 2154.89 Preço relativo (t-1) +
+ 10433.22 JF – 10307.80 MJ – 15220.30 JA – 10975.40 SO (6)
Variáveis Coeficiente Desvio t p – level Ord. Ent. R2 Aj. (%)
Padrão
Constante 354364.70 63659.39 5.57 0.00 – –
Preço deflacionado (t-1) -3589.10 913.15 -3.93 0.00 6 78.3
Preço relativo (t-1) -2154.89 471.52 -4.57 0.00 5 65.2
JF 10433.22 2830.28 3.69 0.00 1 33.2
MJ -10307.80 2824.24 -3.65 0.00 3 51.6
JA -15220.30 2810.73 -5.42 0.00 2 45.0
SO -10975.40 2806.21 -3.91 0.00 4 61.6
R2 (%) R2 Ajustado D. Padrão Erro Absoluto F da Graus
(%) Regressão Médio (%) Regressão Liberdade
Stepwise 82.8 78.3 5111.8 6.9 18.4 23
De modo análogo ao Grupo X, apresentamos, neste ponto, os modelos para as vendas e quota
de mercado da Renova e, ainda, das vendas do mercado total, do Grupo Y.
Modelos das Vendas Renova Embora, na equação deste modelo (6), os dois preços, defla-
cionado e relativo (desfasados), apareçam significativos, pela análise das contribuições directas
verificamos que é o último que tem maior contribuição na explicação das vendas Renova.
Relativamente ao sinal de todos os coeficientes estimados (Quadro 6), estes não contrariam
o que a teoria económica deixa prever, tendo em conta os dados e a informação a priori
que possuı́amos. Ainda assim, achámos por bem tecer alguns comentários, os quais podem,
eventualmente, adaptar-se a modelos seguintes. O sinal negativo dos coeficientes associados
aos preços indica-nos que estamos em presença de bens ordinários (contrariando a polı́tica
de preços que previmos ser praticada pelos retalhistas). Por outro lado, as variáveis dummy,
com coeficientes estimados significativos a 5%, permitem-nos adiantar que no bimestre JF,
em média, os indivı́duos estariam dispostos a comprar mais, relativamente aos bimestres ND
(omitido) e MA (o consumo médio destes produtos no bimestre MA é indiferente ao que se
efectua no bimestre omitido, visto o coeficiente respectivo não ser significativamente diferente
L. Grilo, I. Themido / Investigação Operacional, 22 (2002) 147-166 163
Quadro 7: Valores estimados para o modelo de vendas Renova ao consumidor final: Grupo Y.
Vendas (t) = 21596.79 + 9964.78 Coseno (t) – 4859.02 Seno (t) + 0.63 Vendas (t-1) (7)
Variáveis Coeficiente Desvio t p – level Ord. Ent. R2 Aj. (%)
Padrão
Constante 21596.79 8764.75 2.46 0.02 – –
Coseno (t) 9964.78 1500.33 6.64 0.00 1 62.7
Seno (t) -4859.02 2209.59 -2.20 0.05 3 77.4
Vendas (t-1) 0.63 0.15 4.10 0.00 2 74.1
R2 (%) R2 Ajustado D. Padrão Erro Absoluto F da Graus
(%) Regressão Médio (%) Regressão Liberdade
Stepwise 79.7 77.4 5522.6 7.1 34.0 26
Modelo do Quadro 6.
Modelo do Quadro 7.
1e 5
1e5
80000
80000
60000
60000
40000
40000
20000
20000
Valores
0 Valores
Observados 0
Observados
Valores
-2 0000 Valores
Previstos -20000
Previstos
Resíduos
Resíduos
J F9 2
JF93
JF9 4
J F95
JF96
JA9 2
J A93
JA9 4
JA 95
J A9 6
MJ9 2
N D 92
MJ 93
ND 9 3
M J9 4
N D9 4
MJ9 5
ND 95
M J96
ND 9 6
SO9 2
SO93
SO9 4
S O95
SO9 6
MA9 2
MA 93
M A9 4
MA9 5
MA 96
JF92
JF93
JF94
JF9 5
J F9 6
JA92
JA9 3
JA9 4
JA 95
J A96
MJ92
ND 92
MJ93
ND 93
MJ9 4
N D94
MJ9 5
N D9 5
M J96
ND9 6
MA92
SO92
MA93
SO93
MA9 4
SO9 4
MA9 5
S O9 5
M A96
SO96
Tempo (Bimestres)
Tempo (Bimestres)
Figura 7: Qualidade do ajuste dos modelos de vendas Renova ao consumidor final: Grupo Y.
de zero), dado o sinal positivo do coeficiente associado à variável em causa; nos bimestres
MJ, JA e SO, os indivı́duos predispõem-se a consumir (logo comprar), em média, uma menor
quantidade de produtos inseridos no Grupo Y (relativamente aos bimestres ND e MA), já que
todos possuem sinal negativo. Este comportamento, em nosso entender, muito provavelmente,
resulta do facto das condições climatéricas influenciarem, em larga medida, o consumo deste
tipo de bens (tal como já havı́amos referido, nomeadamente com recurso à análise da Figura
3).
Pela observação do andamento da sucessão das vendas Renova (Figura 3), podemos veri-
ficar, na fase final do perı́odo em análise, que os picos (“vértices”) de sazonalidade tendem a
arredondar-se (graficamente, mais semelhante com as funções seno e coseno), comportamento
que se espera manter no futuro com a nova polı́tica de diferenciação do produto (por exem-
plo, é cada vez mais frequente ver o produto em causa, junto dos produtos de cosmética para
senhora). Mais uma vez, as vendas desfasadas, além de reflectirem a sazonalidade, também
podem representar a própria notoriedade da marca. Na Figura 7, apresentamos os gráficos
que mostram a qualidade do ajuste dos modelos.
164 L. Grilo, I. Themido / Investigação Operacional, 22 (2002) 147-166
De salientar, em relação à qualidade do ajuste dos modelos, que se é verdade que um modelo
que não consegue explicar/modelar o passado (de forma aceitável e, no caso, economicamente
sensı́vel), não deverá servir para prever (a não ser por acaso!), também não podemos esquecer
que nem sempre os melhores ajustamentos conduzem às melhores previsões.
19
Embora as marcas da distribuição tenham uma quota de mercado de cerca de 30% para ambos os grupos
no último bimestre (ND96), a Renova detém 55% de quota de mercado no Grupo Y (contra 60%, no primeiro
bimestre – JF92), enquanto que, no caso do Grupo X, passou de 50% para 34%.
L. Grilo, I. Themido / Investigação Operacional, 22 (2002) 147-166 165
Modelos das Vendas do Mercado Total Tal como no modelo de vendas Renova (6),
também, na equação (9) as dummies representam a sazonalidade (até porque o comportamento
é semelhante em ambas as sucessões, ver Figura 3), indicando-nos, porém, que é nos bimestres
JF e ND que se vende mais deste produto, em média. As vendas totais desfasadas (primeira
variável a entrar no modelo – Quadro 9) reflectem não só a sazonalidade como, também, os
hábitos de consumo.
3 Conclusões
Dos vários modelos de regressão aditivos desenvolvidos, seleccionámos aqueles que minimi-
zavam os valores dos indicadores que medem o desvio entre as previsões e os valores observados,
sendo que os erros absolutos médios dos modelos que apresentamos encontram-se entre 2.0 e
7.1%. Como antevı́amos, tivemos mais dificuldade em modelar o Grupo X, donde a capacidade
explicativa alcançada pelo modelo de vendas Renova deste grupo é inferior ao do Y. Embora
conscientes que um grande número de condicionalismos não foram contemplados nos modelos,
relativamente às variáveis, destaca-se o seguinte: foi possı́vel confirmar que a variável stock não
é significativa em nenhum dos modelos; das variáveis explicativas que aparecem significativas,
o preço, a quota das outras marcas e, ainda, a variável vendas desfasada, assumiram particular
importância, pelo que devem ser motivo de alguma atenção.
Gostarı́amos de referir, ainda, que durante a elaboração deste trabalho nos ocorreram
possı́veis e interessantes áreas de estudo, a desenvolver, eventualmente, no futuro. Estas, são,
entre outras:
- A modelação das vendas de cada um dos produtos pertencentes a cada grupo, permitindo,
posteriormente (pela soma as previsões parciais), verificar até que ponto o resultado está
próximo da previsão total.
- A modelação das vendas a cada um dos grupos económicos clientes da Renova, possibi-
litando conhecer o valor das vendas ao consumidor, por grupo de produtos ou por produto,
166 L. Grilo, I. Themido / Investigação Operacional, 22 (2002) 147-166
sendo então possı́vel entrar com outras variáveis explicativas, que estão inerentes a cada cliente
e produto, como é o caso da publicidade, dos topos de gôndola, etc..
- Desenvolvimento de modelos, por regiões (dado que a Nielsen divide o Paı́s em 5 regiões),
com vista a explicar e a prever as vendas em cada uma delas.
- Utilização de outras técnicas que, tendo em conta, por exemplo, o tipo de produto mais
vendido, possibilitem a segmentação do mercado, com vista a implementar estratégias direc-
cionadas de marketing.
Por fim, pensamos que nesta era da globalização, em que obter e trocar informação é cada
vez mais fácil, permitindo um maior conhecimento a priori que, quando conjugado com este
tipo de análises quantitativas e até, porque não, com alguma intuição sobre o fenómeno em
causa, poderá facilitar a tomada de decisões, principalmente nesta área de marketing. Neste
sentido, os gestores poderão mais facilmente gerir a mudança, estando melhor preparados para
reagir à incerteza do meio envolvente.
4 Referências
[1] Cooper, L. and Nakanishi, M., Market-Share Analysis – Evaluating Competitive Marketing Ef-
fectiveness, Kluwer Academic Publishers, Boston (1993).
[2] Farhangmeher, M. and Veiga, P., The changing consumer in Portugal, Research in Marketing 5
(1995) 485-502.
[3] Greene, W. H., Econometric Analysis, Macmillan,(2nd Edition), New York (1993).
[4] Grilo, L. M., Contribuições para a modelação de vendas de produtos de grande consumo: Uma
aplicação ao mercado dos produtos transformados de papel, Tese de Mestrado em Matemática
aplicada à Economia e Gestão, Instituto Superior de Economia e Gestão - Universidade Técnica
de Lisboa (1997).
[5] Hanssens, D. M., Parsons, L. J. and Schultz, R. L., Market Response Models: Econometric and
Time Series Analysis, Kluwer Academic Publishers, Pennsylvania (1992).
[6] Helfer, J. P. and Orsoni, J., Marketing, Edições Sı́labo, Lda., Lisboa (1996).
[7] Lilien, G. L., Kotler, P., Marketing Decision Making – A Model Building Approach, Harper &
Row Publishers, New York (1983).
[8] Lilien, G. L., Kotler, P. and Moorthy, K. S., Marketing Models, Pretince-Hall International
Editions, New Jersey (1992).
[9] Mendes, A. and Themido, I., Modelação de elasticidades de quotas de mercado para produtos de
grande consumo, Revista Estudos de Economia 4 (1998).
[10] Murteira, B., Müller, D. A. and Turkman, K. F., Análise de Sucessões Cronológicas, McGraw-
Hill, Lisboa (1993).
[11] Penny, N. J. and Broom, D., The Tesco Approach to Store Location, in N. Wrigley (ed.), Store
Choice Location and Market Analysis, London Routledge (1988).
[12] Silva, F., Themido, I. H., Um modelo causal de previsão de um produto de grande consumo,
Revista de Investigação Operacional 1 (1998) 33-48.
[13] Vasco, Rute S., Acabaram-se as vantagens competitivas, Revista Exame, Agosto (1997a) 76-79.
[14] Vasco, Rute S., Marcas da distribuição: o dilema dos produtores, Revista Exame, Outubro
(1997b) 104-107.
P. Infante, J. Rodrigues Dias / Investigação Operacional, 22 (2002) 167-179 167
∗ ∗
Paulo Infante J. Rodrigues Dias
∗
Departamento de Matemática, Universidade de Évora
{pinfante, jrd}@dmat.uevora.pt
Abstract
In this paper we analyse the influence of some lifetime models, with different failure
rates, in the expected total cost of the productive system per cycle and, on the other hand,
in the optimal solution for the inspection period. Particularly, we analyse the practical
consequence of one approximate solution for the inspection period. For this study, using
X̄ and R charts, we consider different working costs and, as an optimization method, we
use simulation techniques.
Resumo
Keywords: Statistical quality control, control charts, failure rate, costs, inspection period, simulation
Title: Analysis of the importance of the lifetime distribution on the inspection period in statistical
quality control
168 P. Infante, J. Rodrigues Dias / Investigação Operacional, 22 (2002) 167-179
1 Introdução
Todo o sistema produtivo, seja qual for a sua natureza e complexidade, está sujeito ao apa-
recimento de perı́odos de mau funcionamento. Como se desconhece o instante em que estes
perı́odos podem começar, já que o tempo de vida do sistema é uma variável aleatória, é ne-
cessário, durante o seu funcionamento, retirar da linha de produção amostras, que então se
analisam.
De um modo geral, a imensa bibliografia sobre controlo de qualidade não tem, usualmente,
uma abordagem económica, apesar de ela ser extremamente importante. Duncan terá sido o
primeiro autor a considerá-la em [4]. Em [1], [6], [11] e [21 - 23] podem encontrar-se extensas
sı́nteses e análises bibliográficas sobre este e outros aspectos. Em [13] encontra-se o modelo
que serviu de base ao desenvolvimento do presente trabalho. Em [14 - 16], [18 - 20], bem como
em [7 - 9] outras abordagens foram feitas.
Neste trabalho pretende-se analisar até que ponto é que o tipo e a forma da distribuição do
tempo de vida do sistema influenciam o intervalo de amostragem que minimiza o respectivo
custo total médio de funcionamento por ciclo. Assim, para além da tradicional distribuição
exponencial, à qual corresponde uma taxa de risco constante, considera-se a distribuição nor-
mal, com diferentes coeficientes de variação, a distribuição de Weibull com diferentes valores
do parâmetro de forma, a distribuição lognormal e a distribuição de Hjorth apresentada em
[5], as quais cobrem taxas de risco crescentes, decrescentes, crescentes e depois decrescentes
e, finalmente, em forma de banheira. Refira-se, a propósito, que Collani em [2] escreveu: “In
general the investigations are restricted to the case of exponentially distributed times between
the ocurrence of assignable causes”.
Vamos supor que a produção, estando sob controlo estatı́stico, segue uma distribuição normal
com média µ0 e desvio padrão σ0 . Admite-se, por outro lado, que, após uma falha do sistema, a
qualidade, embora ainda com distribuição normal, passa a ter uma nova média µ 1 = µ0 + λσ0 ,
P. Infante, J. Rodrigues Dias / Investigação Operacional, 22 (2002) 167-179 169
com λ 6=0.
C1 - custo de obtenção e análise de uma amostra, bem como da marcação dos pontos nas
cartas;
De acordo com o modelo adoptado, considerando que não há paragem do processo produ-
tivo quando um ponto cai fora dos limites de controlo, os custos envolvidos podem-se decompor
em 3 parcelas:
a) uma referente aos falsos alarmes: verifica-se um custo C2 por cada procura de uma causa
que não existe.
b) outra referente às falhas do sistema: verificam-se os custos de detecção C 3 , de reparação
C4 , e ainda o custo de mau funcionamento do sistema durante (T5 +T3 ) dado por (G-
G’)×(T5 +T3 ) e o custo G×T4 porque supomos não haver produção enquanto o sistema
é reparado.
c) outra parcela, por fim, referente à amostragem e análise de produtos dada por C 1 ×(Na +Nd ),
onde Na representa o número de amostras inspeccionadas antes do sistema falhar e N d
o número de amostras inspeccionadas depois do sistema falhar.
Representando por C o custo total por ciclo, por T o tempo de vida do sistema, por I k
uma variável indicatriz que assume o valor 1 com probabilidade η (probabilidade de ocorrer
um falso alarme), tomando G-G’=Gm e representando por P o perı́odo de amostragem, tem-se
Int(T /P )
T
X µ µ ¶ ¶
C = C2 Ik + C3 + C4 + Gm (T3 + T5 ) + GT4 + Int + N d C1 (1)
k=1
P
Designando por H o intervalo de tempo entre o instante em que o sistema falha e o instante
em que a primeira amostra, após a falha, é analisada, suponha-se que a esperança matemática
170 P. Infante, J. Rodrigues Dias / Investigação Operacional, 22 (2002) 167-179
Int(T /P )
T
X µ µ ¶ ¶
C = C2 Ik + C3 + C4 + Gm (T3 + T5 ) + GT4 + Int + N d C1 (2)
k=1
P
Com base neste pressuposto, que do ponto de vista intuitivo é facilmente aceitável e que
iremos analisar ao longo deste artigo, tem-se
P
E [T5 |λ] = kP − (3)
2
sendo k o número médio de amostras inspeccionadas, desde o instante em que o sistema falha,
até à sua detecção (média da distribuição geométrica). Assim, tem-se
T P
µ · µ ¶¸ ¶
E Int + k P − E(T ) ∼
= kP − (4)
P 2
T E(T ) 1
· µ ¶¸
E Int ∼
= − (5)
P P 2
pelo que o custo total médio por ciclo vem dado por
E (T ) 1 P E (T ) 1
· µ ¶ ¸ · µ ¶ ¸ ·µ ¶ ¸
E (C) ∼
= η − C2 + C3 + C4 + Gm T3 + kP − + GT4 + − + k C1
P 2 2 P 2
(6)
E (T ) 1 P E (T ) 1
· µ ¶ ¸ · µ ¶ ¸ ·µ ¶ ¸
E (C) ∼
= η − C2 + C3 + C4 + Gm T3 + kP − + GT4 + − + k C1
P 2 2 P 2
(7)
Analogamente, quando se opta pela paragem do processo no caso de um ponto sair fora
dos limites de controlo, a aproximação do perı́odo de amostragem que minimiza o custo total
médio por ciclo é dada por ([7]):
E (T ) 1 P E (T ) 1
· µ ¶ ¸ · µ ¶ ¸ ·µ ¶ ¸
E (C) ∼
= η − C2 + C3 + C4 + Gm T3 + kP − + GT4 + − + k C1
P 2 2 P 2
(8)
E (T ) 1 P E (T ) 1
· µ ¶ ¸ · µ ¶ ¸ ·µ ¶ ¸
E (C) ∼
= η − C2 + C3 + C4 + Gm T3 + kP − + GT4 + − + k C1
P 2 2 P 2
(9)
E (T ) 1 P E (T ) 1
· µ ¶ ¸ · µ ¶ ¸ ·µ ¶ ¸
E (C) ∼
= η − C2 + C3 + C4 + Gm T3 + kP − + GT4 + − + k C1
P 2 2 P 2
(10)
onde em vez dos valores individuais dos custos C1 e Gm , aparece o seu cociente designado por
r. Pode afirmar-se que as expressões de P anteriormente obtidas constituem uma generalização
desta, facto este que nos parece de realçar.
Com base no modelo antes apresentado, fez-se o estudo de um conjunto de situações, quer de
natureza teórica, quer de natureza fundamentalmente prática, com o objectivo fundamental
de ver até que ponto é que a distribuição do tempo de vida era relevante. Com este intuito,
obteve-se todo um conjunto de resultados, alguns dos quais aqui se apresentam e analisam.
Para este estudo iremos considerar C2 =20, T2 =2, C3 =10, T3 =1, C4 =100 e T4 =3, valores
que nos parecem razoáveis. Refira-se que, em [4], o custo C1 é proporcional ao tamanho da
√
amostra e em [7] considera-se C1 =5 n, reflectindo, assim, uma certa economia de escala. Neste
trabalho também consideramos este valor para C1 , apesar de apenas se utilizarem amostras
de tamanho 8.
Iremos utilizar uma carta de controlo para a média e outra para a amplitude, adoptando
limites de controlo “3-sigma”, o que é perfeitamente aceitável e usual. No entanto, em [7] e
em [9], consideram-se outros limites de controlo e outras dimensões amostrais.
Refira-se, ainda, que se considera apenas o caso em que não há paragem do processo
produtivo quando se detecta um ponto fora dos limites de controlo das cartas. No entanto,
havendo paragem, resultados perfeitamente idênticos foram obtidos, como se pode ver em [7].
Por outro lado, os casos em que são acentuadamente diferentes os custos de mau funcio-
namento (Gm =100, em que se toma G=100 e G’=0, e Gm =1100, em que se toma G=1000 e
172 P. Infante, J. Rodrigues Dias / Investigação Operacional, 22 (2002) 167-179
G’=-100).
Finalmente, dois casos associados a diferentes probabilidades de detectar a falha após esta
ter ocorrido (uma alteração suave e uma alteração acentuada da média).
Muitas vezes, devido ao facto de modelar situações concretas (o que pode acontecer, por
exemplo, em sistemas electrónicos e em sistemas em que a complexidade é grande), e também
devido à sua simplicidade (em termos de tratamento analı́tico), tem-se considerado que o
tempo de vida de um sistema segue a distribuição exponencial. Como se pode ver em [16], no
caso da distribuição do tempo de vida ser exponencial, uma interpretação geométrica mostra
facilmente que E(H) é sempre maior que P/2. Neste trabalho apenas consideramos o caso em
que λ=1000, pois para outros valores do parâmetro λ (em particular, para λ=100 e λ=10000)
resultados análogos foram obtidos.
A distribuição de tempo de vida Weibull, pelo facto de cobrir diferentes taxas de risco,
é cada vez mais utilizada. Neste trabalho consideramos o mesmo parâmetro de escala e 3
valores diferentes para o parâmetro de forma. Assim, ao considerarmos β=0.5, estamos em
presença de um sistema com taxa de risco decrescente, enquanto que ao considerarmos β=3 e
β=4 estamos em presença de um sistema com taxa de risco crescente.
No caso em que o tempo de vida segue uma distribuição normal começa por considerar-se
uma média µ=1000 e um desvio padrão σ=100, considerando-se posteriormente outros valores
do desvio padrão. Recorde-se que a taxa de risco associada a esta distribuição é crescente.
Considerámos também a distribuição lognormal, para a qual o sistema começa por ter uma
taxa de risco crescente, para depois ser decrescente.
P. Infante, J. Rodrigues Dias / Investigação Operacional, 22 (2002) 167-179 173
Tabela 1: Regressão linear da forma E(H)=aP, para diferentes distribuições do tempo de vida de um
sistema.
Distribuição do Valor do Coeficiente de
Tempo de Vida Declive (a) Determinação
Exponencial (λ=0.001) 0.5031 0.999992
(α=1000; β=0.5) 0.5376 0.999700
Weibull (α=1000; β=3) 0.5001 0.999998
(α=1000; β=4) 0.4999 0.999999
Normal (µ=1000; σ=100) 0.5000 0.999999
Lognormal (µ=5; σ=1) 0.4985 0.999975
Hjorth (θ = β=1; δ=0.01) 0.6091 0.994408
Por fim, considerámos a distribuição de Hjorth, tomando valores para os seus parâmetros
por forma a que a taxa de risco do sistema considerado se apresente em forma de banheira,
como se pode ver em [16] .
(1) Para todos os casos considerados, os ajustamentos obtidos foram muito bons, a avaliar
pelo valor do respectivo coeficiente de determinação.
(2) No caso da distribuição exponencial, o valor do declive é bastante próximo de 0.5, estando
de acordo com a interpretação geométrica antes referida. Também no caso da distribuição
lognormal o declive é muito próximo de 0.5, embora, neste caso, inferior.
(3) Nos casos da distribuição de Weibull, quando a taxa de risco é crescente, e também no
caso da distribuição normal, o valor do declive obtido é igual a 0.5, o que está de acordo
com o pressuposto por nós admitido. Nestes casos, a aproximação por nós obtida pode
ser considerada como solução “exacta”.
(5) No caso da distribuição de Hjorth, o declive é bastante superior a 0.5. Neste caso,
pudemos constatar que quanto maior o perı́odo de inspecção considerado, também maior
o valor do desvio de E(H) obtido.
30
20
E(H)
y = 0,5032x
10
0
0 10 20 30 40 50
P
Figura 1: Regressão linear de E(H) em função de P para um tempo de vida normal com µ=1000 e
σ=10 (r2 =0.960428).
(2) Quando o tempo de vida do sistema é quase constante (coeficiente de variação igual
a 0.01%) não conseguimos nenhum ajustamento, como se ilustra no diagrama de dis-
persão representado na Fig. 2. Refira-se que, neste caso, em que o tempo de vida é
quase constante, uma constatação análoga se verificou em [7] quando se considerou uma
distribuição lognormal com µ=5 e σ=0.01.
Neste ponto iremos analisar a influência da distribuição do tempo de vida sobre o custo to-
tal médio por ciclo, considerando as diferentes situações anteriormente referidas, procurando
estudar o seu comportamento, bem como verificar até que ponto as aproximações podem ser
consideradas como óptimas ou quase.
A Fig. 3 ilustra o comportamento do custo total médio por ciclo em função do perı́odo
de inspecção considerando um tempo de vida com distribuição exponencial. Como se pode
P. Infante, J. Rodrigues Dias / Investigação Operacional, 22 (2002) 167-179 175
40
30
E(H)
20
10
0
0 10 20 30 40 50
P
Figura 2: Diagrama de dispersão de E(H) em função de P para um tempo de vida normal com µ=1000
e σ=0.1.
observar, apenas existe um único mı́nimo. Para todas as alterações referidas anteriormente e
para os dois valores do custo de mau funcionamento considerados, as curvas dos custos totais
médios por ciclo, para as diferentes distribuições indicadas na tabela 1, são análogas à que se
apresenta na Fig. 3.
Na Fig. 4, que corresponde a uma das situações em que o custo total é mais sensı́vel a
alterações do perı́odo de inspecção, representam-se alguns resultados obtidos neste contexto.
Podemos verificar que o custo total médio por ciclo é pouco sensı́vel a uma variação, ainda
que elevada, do perı́odo de amostragem P. Na realidade, basta constatar que, por exemplo,
erros relativos no cálculo de P na ordem dos 20% provocam apenas erros relativos no cálculo
de E(C) na ordem dos 2.5%.
Esta fraca sensibilidade do custo total médio em relação ao perı́odo de inspecção havia já
sido constatada, noutras situações, por Rodrigues Dias (1983). Este facto permite reforçar a
importância que na prática assumem as expressões (7) e (8) obtidas para calcular aproximada-
mente o perı́odo P, as quais, para todas as distribuições da Tabela 1, podem ser consideradas
como óptimas ou quase.
Quando se considera o caso de um tempo de vida com distribuição normal, mas com pe-
quenos coeficientes de variação, a situação anterior altera-se. Contudo, como se pode observar
na Fig. 5, no caso em que o coeficiente de variação é igual a 1%, podem observar-se vários
mı́nimos relativos, mas para valores bastante elevados de P e distantes da zona do mı́nimo
absoluto.
Para um coeficiente de variação igual a 0.1%, o custo total médio por ciclo apresenta
grandes oscilações com muitos mı́nimos relativos (Fig. 6).
No entanto, considerando uma pequena alteração da média, E(C) não apresenta oscilações
relevantes, para quaisquer dos dois custos de mau funcionamento considerados. Pensamos que
176 P. Infante, J. Rodrigues Dias / Investigação Operacional, 22 (2002) 167-179
200000
150000
E(C)
100000
50000
0
0 2 4 6 8 10
Figura 3: Custo total médio por ciclo E(C) em função de P, no caso de uma alteração da média
λ = 0.5,com Gm =1100.
30
%var relat. E(C)
25
20
15
10
5
0
-30 -20 -10 0 10 20 30
%var. relat. P
Figura 4: Variação do custo total médio por ciclo em função da variação do perı́odo de inspecção P,
no caso de uma alteração da média λ=0.5, com Gm =1100.
P. Infante, J. Rodrigues Dias / Investigação Operacional, 22 (2002) 167-179 177
45000
E(C)
25000
5000
0 10 20 30 40 50
P
Figura 5: Custo total médio por ciclo E(C) em função de P, no caso T∩N(1000, 10) e de uma alteração
da média λ=2, com Gm =1100.
50000
E(C)
25000
0
0 10 20 30 40 50
Figura 6: Custo total médio por ciclo em função de P, no caso T∩N(1000, 0.1) e de uma alteração da
média λ=2, com Gm =1100.
178 P. Infante, J. Rodrigues Dias / Investigação Operacional, 22 (2002) 167-179
a explicação para tal reside no facto de pequenas alterações necessitarem de um grande número
de amostras até serem detectadas, pelo que o valor de E(H) tem uma fraca contribuição para
o custo médio associado à falha e, consequentemente, para E(C).
4 Considerações Finais
Neste trabalho havia dois objectivos fundamentais a atingir: por um lado, sendo o tempo de
vida uma variável aleatória, analisar a importância da sua distribuição, quer sobre o custo
total médio de funcionamento de um sistema por ciclo, quer sobre o perı́odo de inspecção; por
outro lado, analisar o nı́vel de precisão de soluções aproximadas obtidas para o perı́odo de
inspecção.
(a) De um modo geral, com excepção do caso em que o tempo de vida pode ser considerado
como uma constante, com reduzido interesse prático, pode-se concluir que, apesar dos
resultados obtidos não serem independentes da distribuição do tempo de vida, esta é
pouco relevante na perspectiva de determinar o perı́odo de inspecção que minimiza o
custo total médio por ciclo.
(b) As soluções aproximadas obtidas para o perı́odo de inspecção, que generalizam uma
aproximação obtida por Nakagawa e Yasui (1979), que, por sua vez, é generalizada por
uma outra obtida por Rodrigues Dias (1990a), num contexto algo diferente, podem ser
consideradas “óptimas” ou “quase-óptimas”. Refira-se que, no que concerne à distri-
buição do tempo de vida, a única grandeza que nelas aparece é a respectiva esperança
matemática, o que está de acordo com a conclusão atrás referida.
5 Referências
[1] Collani, E., An Updated Bibliography of Economic Quality Control Procedures, Economic Qua-
lity Control, No 3, 1988, pp. 48-62.
[2] Collani, E., A Note on Optimal Inspections and Sampling Procedures, Economic Quality Control,
No 12, 1988, pp. 1-15.
[3] Dhillon, B. S., Life Distributions, IEEE Transactions on Reliability, Vol. R-30, N. 5, 1981, pp.
457-460.
[4] Duncan, A. J., The Economic Design of X̄- Charts Used to Maintain Current Control of a
Process, Journal of the American Statistical Association, Vol. 51, N. 274, 1956, pp. 228-242.
[5] Hjorth, U., A Reliability Distribution with Increasing, Decreasing, Constant and Bathub-Shaped
Failure Rates, Technometrics, Vol. 22, No 1, 1980, pp. 99-107.
[6] Ho, C.; Case, K. E., Economic Design of Control Charts: A Literature Review 1981-1991, Journal
of Quality Technology, Vol. 26, No 1, 1994, pp. 39-53.
[1] Infante, P., Optimização em Controlo Estatı́stico de Qualidade, Provas de Aptidão Pedagógica e
Capacidade Cientı́fica (Trabalho de Sı́ntese), Universidade de Évora, 1997.
P. Infante, J. Rodrigues Dias / Investigação Operacional, 22 (2002) 167-179 179
[7] Infante, P.; Rodrigues Dias, J., Comparação de resultados em controlo de qualidade usando
intervalos constantes e diferentes, Estatı́stica: a diversidade na unidade. Editores: Manuela
Souto e Isabel Pereira, 1998, pp. 237-246.
[8] Infante, P.; Rodrigues Dias, J., Influência dos Limites de Controlo na Minimização de Custos
em Controlo Estatı́stico de Qualidade, 2000. Submetido a publicação.
[9] Kleijnen, J. ; Groenendal, W. V. , Simulation A Statistical Perspective, John Wiley, 1992.
[10] Montgomery, D.C., The Economic Design of Control Charts: A Review and Literature Survey,
Journal of Quality Technology, Vol. 12, N. 2, 1980, pp. 75-87.
[11] Nakagawa, T.; Yasui, K., Approximate Calculation of Inspeccion Policy with Weibull Failure
Times, IEEE Transactions on Reliability, Vol. R-28, N. 5, 1979, pp. 403-404.
[12] Rodrigues Dias, J., Cartas de Controlo de Qualidade nas Indústrias Agro-Alimentares, Jornadas
de Informação Agrária / 81, 1981, Évora.
[13] Rodrigues Dias, J., Optimização do Intervalo de Amostragem em Controlo Estatı́stico de Quali-
dade, Congresso 81, Ordem dos Engenheiros, Tema 1, Comunicação 16, 1981, Lisboa.
[14] Rodrigues Dias, J., Influence de la Période d’Inspection sur les Coûts dans l’Inspection Périodique
de Systémes, Revue de Statistique Appliquée, vol. XXXI, N.4, 1983, pp. 5-15.
[15] Rodrigues Dias, J., Polı́ticas de Inspecção de Sistemas, Tese de Doutoramento, Universidade de
Évora, 1987.
[16] Rodrigues Dias, J., Validade de uma solução Aproximada para o Perı́odo de Inspecção de Sistemas
com Diferentes Tipos de Taxas de Risco, Investigação Operacional, Vol. 8, N. 1, 1988, pp. 107-
116.
[17] Rodrigues Dias, J., A New Approximation for the Inspection Period of Systems with Different
Failure Rates - European Journal of Operational Research, N.45, 1990, 219-223.
[18] Rodrigues Dias, J., Some Approximate Inspection Policies for a System with Imperfect Inspecti-
ons, Recherche Opérationnelle / Operations Research Vol. 24, N o 2, 1990, pp. 191-199.
[19] Rodrigues Dias, J.; Infante, P., Influência da Distribuição do Tempo de Vida do Sistema em Con-
trolo Estatı́stico de Qualidade - II Congreso Galego de Estatistica e Investigacion de Operacions,
Vigo, 1995, 11 pág’s.
[20] Tagaras, G., A Survey of Recent Developments in the Design of Adaptive Control Charts, Jounal
of Quality Technology, Vol. 30, No 3, 1998, pp. 212-231.
[21] Woodall, W. H.; Montgomery, D. C., Research Issues and Ideas in Statistical Process Control,
Journal of Quality Technology, Vol. 31, No 4, 1999, pp. 376-386.
[22] Vance, L. C., A Bibliography of Statistical Quality control Chart Techniques, 1970-1980, Journal
of Quality Technology, Vol. 15, No 2, 1983, pp. 59-62.
H. Alvelos, J.A. Sarsfield Cabral / Investigação Operacional, 22 (2002) 181-198 181
∗ †
Helena Alvelos J.A. Sarsfield Cabral
∗
Secção Autónoma de Gestão e Engenharia Industrial, Universidade de Aveiro
helena@egi.ua.pt
†
Departamento de Engenharia Mecânica e Gestão Industrial, Faculdade de Engenharia da
Universidade do Porto
jacabral@fe.up.pt
Abstract
This paper discusses the use of control charts in monitoring customer satisfaction survey
data over time. Using simulated multinomial data, the performance of the X S chart,
the modified p chart (or X P chart) and the χ2 chart are compared, assuming both in-
control and out-of-control situations. It was concluded that the first two charts are almost
equivalent, but for practical reasons, it is suggested that the X P chart should be preferred.
It is clearly shown that, in many circumstances, both the X P and the χ2 charts are not
able to detect not only statistically significant changes in the satisfaction profile but also
sizeable ones. Taking into account the poor individual performance of each of these two
charts it is recommended that they should be operated together. However, when an out-of-
control point occurs, the direction of the statistically significant change in the underlying
satisfaction level can be very difficult to interpret. Hence, a new method based on the value
concepts is proposed as a tool to help decision-makers in interpreting the out-of-control
signals.
Keywords: customer satisfaction, categorical data, control charts, goodness-of-fit, multinomial distri-
bution, value theory
1 Introduction
Sound quality management requires measuring consumers’ perceptions of the quality of goods
and services delivered. Accordingly, the use of questionnaires to appraise customer satisfaction
has been growing in recent years. In fact, many institutions (not only hotels and hospitals)
use now to ask their customers to fulfil satisfaction questionnaires. It is also common that
182 H. Alvelos, J.A. Sarsfield Cabral / Investigação Operacional, 22 (2002) 181-198
customers are surveyed frequently, on a regular basis, because the goal is not only to estimate
the level of customer satisfaction in a given moment but also to evaluate its evolution over
time. In these circumstances, the questionnaires are repeated periodically (once a month,
twice a year, etc.).
Analysing and interpreting customer satisfaction survey data over time is a difficult task.
When a Likert scale is adopted to assess a quality dimension, a typical procedure consists
of calculating the average value (or the variance) of all the responses and comparing the
scores obtained. A graphic representation of these values can also be developed. Based on
this procedure, it is often puzzling to verify if there is a statistically significant change in the
respondents’ satisfaction profile over the time period studied, particularly when the judgement
is exclusively based on the evolution of the data.
The ability of detecting significant changes in the satisfaction profile is a crucial requirement
in understanding the evolution of customer preferences and it is certainly necessary in deciding
when and how to implement corrective actions. The main question lies in distinguishing when
the discrepancy in the data indicates a significant change in customers’ opinion or when it is
simply attributable to random factors. This distinction turns out to be decisive, allowing for
action strictly when justified.
The use of control charts in processing results from questionnaires can help in this task.
Hayes (1992) recommends Shewart control charts (widely employed in industrial processes)
for monitoring the process that generates the customer satisfaction data. On the one hand,
those charts can help in identifying changes attributable to assignable causes of variation that
should be isolated and removed. On the other hand, control charts prevent decision-makers in
making corrections based on common-cause variation.
Wardell and Candia (1996) suggest the use of the modified p chart (an extension of the
p chart, also called X P chart) and the χ2 chart in monitoring satisfaction survey data. The
same authors also suggest the superiority of the χ2 chart when compared both to the X S chart
(whose limits are based on subgroup standard deviations) and to the X P chart.
The main objective of this paper is to evaluate and compare the performance of those charts
using simulated satisfaction survey data. Each simulated process assumes a stable consumer
satisfaction profile. It is also assumed that the customer satisfaction level is accessed using a
questionnaire involving a unique question concerning a service or a product. The respondents,
which number of was arbitrarily fixed at N = 70, express their opinions using a Likert-type
scale with five categories, 0 representing the most dissatisfied possible and 4 the most satisfied.
Hence, a multinomial distribution (with 5 parameters) is used to generate the response data.
The X S chart is used to control the expected value of a normally distributed random variable.
Considering the situation addressed in this paper, the data (consisting on the number of
“zeros”, “ones”, “twos”, “threes” and “fours”) is multinomial and not normal. Nevertheless,
according to the central limit theorem, the sample size used (N = 70) is large enough for the
distribution of the sample mean,X, to be considered approximately normal. If the parameters
of the multinomial distribution are unknown, the central line and the control limits of the X S
H. Alvelos, J.A. Sarsfield Cabral / Investigação Operacional, 22 (2002) 181-198 183
CL = X
U CL = X + A3 · S x
LCL = X − A3 · S x
where:
CL - central line
Notice that, given the large sample size assumed, the process standard deviation is com-
puted using the subgroup standard deviations instead of the subgroup ranges. This procedure
leads to control limits more robust to normality than those obtained from the subgroup ranges
used in the classic X R chart (see Wardell and Candia, 1996).
The modified p chart (or X P chart) proposed by Wardell and Candia (1996) is an extension
of the p chart to the general case of more than two categories. The average of the sample results
is again used to control the expected value, so the central line of both the X P and the X S
charts is the same. It is also assumed that the central limit theorem applies in order to establish
the limits as ±3 standard deviations from the mean. Consequently, the probability of the type
I error (denoted by α) should be about 0.27%.
The difference between the X S chart and the modified p control chart is that, in the latter,
the estimate of the process standard deviation is based on the estimates of the parameters of
the underlying multinomial distribution. The multinomial distributions considered here are
characterised by the parameters N , p0 , p1 , p2 and p3 , where pi represents expected proportion
of observations in the ith category and N represents the sample size.
Under these circumstances, the control limits of the modified p chart are:
3
U LC = X + √ · σ̂x
n
3
LCL = X − √ · σ̂x
n
where
xi = 0, 1, 2, 3 or 4,
m
P
yij
j=1
p̂i =
n·m
with
Extensive simulations showed very small differences between the control limits of these two
charts (the points outside the control limits were always coincident in both charts). Considering
sample sizes similar to N = 70, it was concluded that the two charts are almost equivalent.
However, Wardell and Candia (1996) suggest that the X P (or modified p) chart should be
preferred because it does not require constants that are difficult to find (remember that N is
large) and because the X P chart handles more easily varying sample sizes.
When the parameters of the multinomial distributions are known, the limits of the two
charts are the same: the standard deviation of X is no longer estimated. Therefore, the upper
√
and lower limits of the X S chart should be computed replacing S x and A3 by σx and 3/ n
respectively. In the X P chart the estimated standard deviation,σ̂x , should be replaced by the
theoretical value, σx .
3 The χ2 chart
The χ2 chart is based on the χ2 goodness-of-fit statistic and test. Considering the situation
analysed in this paper, the hypothesis set to be tested, H0 , states that there is a certain
probability pi that an observation of the underlying distribution is classified in category i. The
values of pi represent the probability function of the theoretical (or hypothesised) distribution.
The test, particularly suitable for multinomial distributions, uses the Q statistic defined as
follows:
k−1
X (Yi − Ei )2
Q=
i=0
Ei
where
When H0 is true, and the assumptions mentioned later in this section are met, the sampling
distribution of Q follows approximately the χ2 distribution with k − 1 degrees of freedom.
The value of Q represents the overall lack of fit. If the null hypothesis is true, the observed
Yi should be close to the corresponding Ei and hence Q should be small. If the null hypothesis
is not true, the difference between the real and the expected values (or the lack of agreement
between Yi and Ei ) should be reflected on a large value of Q. Consequently, the χ2 test is
one-sided and the chart only has the upper control limit.
Since the χ2k distribution is right skewed, the usual procedure (used when the normality is
assumed) of setting the upper control limit at µ+ 3σ would not lead to a significance level of
α = 0.27%. UCL values for the χ2 chart complying with the standard value of α = 0.27% can
be found in Duncan (1950) and in Wardell and Candia (1996). Those UCL values are specified
as a function of the number of degrees of freedom.
The approximation of the χ2k − 1 distribution by the Q statistic is only valid when the
sample size is large (usually n ≥ 30 is required) and the expected value of the number of
elements in each category is reasonable high. It is usually accepted (Hildbrand and Ott, 1991)
that the minimum expected value for all categories should be 5 (Ei ≥ 5, for all i). This rule
is a simplification of the one owed to Cochran (1954) and mentioned by Marcucci (1985) – no
more than 20% of the expected frequencies should be less than 5 and none of them should be
less than 1 – which becomes more flexible as the number of categories grow.
Another rule presented by Yarnold (1970) suggests that, being k the number of categories
(k ≥ 3) and r the number of categories with expected frequencies of less than 5, the minimum
expected frequency should be at least 5 · r/k.
Those rules are particularly important when customer satisfaction questionnaires are con-
cerned. In fact, it is very common that the respondents tend to concentrate their answers in
few categories, living in blank the others. Wardell and Candia (1996) overcome this problem
combining adjacent categories until the rule of 5 is met. This procedure seems to be arguable:
the information concerning the distribution of the answers over the merged categories is lost,
the control limit value is no longer fixed (UCL is a function of the number of categories), and
the pooled categories may not even make sense.
In order to illustrate the Q statistic behaviour vis-à-vis the chi-squared probability density
function (p.d.f.), Figure 1 shows the histograms of the Q statistic values produced using
simulated data coming out of four multinomial distributions with five categories (0, 1, 2, 3
and 4). Each histogram was obtained by simulating 5 000 independent samples with N = 70.
The χ24 p.d.f. was superimposed to all histograms.
In the cases of Figure 1A and 1B, the histograms are fairly close to the χ 24 p.d.f., while, as
expected, in Figure 1C and in Figure 1D the histograms are very different from the χ 24 p.d.f.:
the underlying distributions does not match the assumptions of the χ2 test.
186 H. Alvelos, J.A. Sarsfield Cabral / Investigação Operacional, 22 (2002) 181-198
0.20 0.20
A - Underlying distribution: B - Underlying distribution:
0.15 14 .14 .14 14
. 14
. 0.15 02 .02 .10 18
. 38
.
0.10 0.10
0.05 0.05
0.00 0.00
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
Q, c 2 Q, c2
4 4
0.30 0.30
0.20 0.20
0.15 0.15
0.10 0.10
0.05 0.05
0.00 0.00
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20 22 38
2 c2
Q, c Q,
4 4
Figure 1: Comparison between the χ24 p.d.f. and the Q statistic histograms obtained from four
underlying multinomial distributions.
H. Alvelos, J.A. Sarsfield Cabral / Investigação Operacional, 22 (2002) 181-198 187
The objective of this section is to present a comparison between the performance of the X P
chart and the χ2 chart when the underlying process is in the state of control (i. e. when H0 is
true). The study is based on simulated data, assuming that all the distributions parameters
are known (in this situation the X S chart and the X P chart are equivalent).
The performance of the charts is measured by the Average Run Length (ARL), which
represents the number of samples that, on average, can be observed between two consecutive
out-of-control signals. When H0 holds (or, in other words, when the process is in-control), the
reciprocal of the ARL (denoted by α) represents the false alarm rate.
A number of 500 000 independent samples of 70 simulated answers were drawn using each
in-control underlying distribution. Based on the estimated α (α̂), a 95% confidence interval
was calculated for α. The ARL considered here is the reverse of α̂.
Table I presents α̂ and ARL values produced using 10 underlying distributions. The first
seven distributions are symmetrical, the other three being skewed. The last distribution is
approximately the same as the one presented by Wardell and Candia (1996). The 95% con-
fidence intervals of α are shown in Figure 2. All the results refer the modified p and the χ 2
charts. It is important to notice that the scale of the XX axis is not numerical since it does
not translate any quantitative measure. It is a qualitative scale where different distributions
are represented.
A general conclusion that can be drawn from the results presented in Table I is that
the in-control ARL values are inconstant, heavily depending on the shape of the underlying
distribution. However, the α risk associated to the modified p chart is apparently not too
sensitive to the distribution shape, staying close to the theoretical value of 0.27% (ARL ≈ 370).
Significant departures from those values can be expected when the χ2 chart is used. Moreover,
the χ2 chart usually shows a lower ARL (or a higher α) than the modified p chart (excluding
distributions number 7 and 9, exhibiting overlapping confidence intervals).
188 H. Alvelos, J.A. Sarsfield Cabral / Investigação Operacional, 22 (2002) 181-198
a [%] c 2 chart
0.55%
0.50% modified p chart
0.45%
0.40%
0.35%
0.30%
0.25%
0.20%
1 2 3 4 5 6 7 8 9 10
Distribution No.
The minimum ARL value of the modified p chart occurs when the population distribution is
extremely skewed (distribution number 10). However, it can not be asserted that symmetrical
distributions tend to have higher values of ARL than non-symmetrical ones. The variation of
α depends on how the distribution of the test statistic, X, matches the tails of the normal
distribution.
The performance of the χ2 chart is clearly associated with the validity conditions of the
χ2 test: the worst results are observed when the expected values of some categories of the
underlying distributions are small (between 2 and 5).
All the results obtained with other simulations using multinomial distributions (not men-
tioned in Table I) are in line with those just described. It can be concluded that, in general,
when the process is in-control the modified p chart induces larger and more stable ARLs than
the χ2 chart.
Throughout this section, the out-of-control performance of both the X P and χ2 charts will be
illustrated using four examples selected from an extensive set of simulated scenarios. Now, the
ARL measures the chart ability in detecting out-of-control situations. The lower the ARL, the
better will be the sensitivity of the control chart on signalling departures from the in-control
state. When the process is out-of-control, the reciprocal of the ARL is denoted by 1 - β, β
representing the probability of not detecting that situation on the first trial.
In the first two examples the in-control situation is characterised by a common symmetric
multinomial distribution: the distribution 05. 12. 36. 12. 05, represented in the first row of Tables
II and III (this distribution has the number 5 on Table I). Two scenarios were conceived in
simulating the evolution from the in-control to the out-of-control state:
(i) the initial symmetric distribution (distribution number 1) evolved toward other sym-
H. Alvelos, J.A. Sarsfield Cabral / Investigação Operacional, 22 (2002) 181-198 189
Table II: Estimated β and ARL for some out-of-control situations (symmetric underlying distributions
progressively concentrated on their mean).
Underlying distribution Modified p chart χ2 chart
_ _
No. β(%) ARL β(%) ARL
1 05. 12. 36. 12. 05 0.99784 463 0.99635 274
2 05. 11. 38. 11. 05 0.99852 678 0.99680 313
3 05. 10. 40. 10. 05 0.99803 508 0.99360 156
4 05. 09. 42. 09. 05 0.99902 1016 0.98277 58
5 05. 08. 44. 08. 05 0.99877 813 0.95349 22
6 05. 07. 46. 07. 05 0.99926 1355 0.87894 8
7 05. 06. 48. 06. 05 0.99951 2033 0.73427 4
8 05. 05. 50. 05. 05 0.99975 4065 0.51328 2
1.0
b^ [%]
0.9
0.8
0.7
0.6
0.5 c 2 chart
0.4 modified p chart
0.3
1 2 3 4 5 6 7 8
Distribution No.
metric distributions progressively concentrated on their (constant) mean (see Table II);
(ii) the initial symmetric distribution (distribution number 1) evolved toward other sym-
metric distributions progressively scattered about their (constant) mean (see Table III);
Tables II and III present the values of β and ARL of the modified p and χ2 charts obtained
considering seven out-of-control underlying distributions selected according the two scenarios.
All the results refers to 500 000 simulations with N = 70, and they are graphically represented
in Figure 3 and in Figure 4 (notice that the XX axis scale has no numerical meaning).
Comparing the results showed on Figures 3 and 4, it is clear that the modified p chart is
not able to detect departures from the in-control state that are of the nature considered in
both scenarios. This conclusion was certainly expected because all the envisaged underlying
distributions have the same expected value. Under these circumstances, the sample mean
(which is the statistic used in the modified p chart) is not able to detect changes in the shape
of the distribution, even when the dispersion about the mean increases substantially.
On the contrary, the χ2 chart responds considerably well to those situations: the higher
the concentration (or the greater the dispersion), the more the distribution moves away from
the control situation, the quicker the change is detected. It can then be concluded that when
190 H. Alvelos, J.A. Sarsfield Cabral / Investigação Operacional, 22 (2002) 181-198
Table III: Estimated β and ARL for some out-of-control situations (symmetric underlying distributions
progressively scattered about their mean).
Underlying distribution Modified p chart χ2 chart
_ _
No. β(%) ARL β(%) ARL
1 05. 12. 36. 12. 05 0.99784 463 0.99635 274
2 06. 12. 34. 12. 06 0.99729 369 0.98573 70
3 07. 12. 32. 12. 07 0.99508 203 0.95104 20
4 08. 12. 30. 12. 08 0.99139 116 0.85335 7
5 09. 12. 28. 12. 09 0.98795 83 0.68873 3
6 10. 12. 26. 12. 10 0.98425 63 0.46775 2
7 11. 12. 24. 12. 11 0.97883 47 0.26353 1
8 12. 12. 22. 12. 12 0.97416 39 0.13262 1
1.0
b^ [%] 0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2 c 2 chart
0.1 modified p chart
0.0
1 2 3 4 5 6 7 8
Distribution No.
Table IV: Estimated β and ARL for some out-of-control situations (left skewed underlying distribu-
tions, Ei ≥ 5).
Underlying distribution Modified p chart χ2 chart
_ _
No. β(%) ARL β(%) ARL
1 20. 17. 14. 11. 08 0.99716 352 0.99718 354
2 21. 18. 14. 10. 07 0.99466 187 0.99713 348
3 21. 19. 14. 10. 06 0.99003 100 0.99508 203
4 21. 20. 14. 09. 06 0.98428 64 0.99121 114
5 22. 20. 13. 09. 06 0.97372 38 0.98883 89
6 22. 21. 13. 09. 05 0.95601 23 0.97699 43
7 23. 21. 12. 09. 05 0.93221 15 0.96993 33
8 24. 22. 11. 08. 05 0.86420 7 0.93230 15
9 26. 21. 13. 05. 05 0.69969 3 0.86741 8
10 26. 23. 11. 05. 05 0.62390 3 0.77523 4
11 26. 25. 09. 05. 05 0.55016 2 0.60725 3
12 27. 26. 07. 05. 05 0.42763 2 0.41057 2
the transformations on the shape of the underlying distribution do not affect considerably its
average value, the χ2 chart seems to be an appropriate tool for detecting significant shifts
(considering that the prerequisites of the χ2 test are met).
The in-control situations of the two other scenarios are based on asymmetric distributions.
The first one, presented on Table IV, evolves from a left-skewed distribution (which has the
number 1 on Table IV) to other distributions progressively more skewed to the left. Across
that gradual evolution there is a quite relevant change in the distributions average. Remark
that the prerequisites of the χ2 test hold in all the situations (Ei ≥ 5).
The results displayed on Table IV and Figure 5 leads to the conclusion that the modified p
chart is faster at detecting the changes than the χ2 chart (regarding the situation hypothesised
in third scenario). This behaviour is a consequence of the relative small changes on the expected
values of each category compared to the relevant shifts in the distribution averages (remember
that the XX axis scale has no numerical meaning).
Finally, Table V and Figure 6 represent the results obtained considering a right skewed
in-control distribution having two categories with expected values equal to 2. This distribution
is quite similar to the one described by Wardell and Candia (1996).
The out-of-control distributions assume that the respondents are progressively more satis-
fied with the product or service. Accordingly, as one progresses down Table V, the number of
answers on the higher-score categories increases.
At first glance, the results of this simulation are surprising: although reasonable changes
on the average and on the shape of the distributions are produced, both charts perform poorly.
Figure 7 shows the histograms of the statistic X (used on X P chart) derived from distribution
number 1 (representing the in-control situation) and from distribution number 2 (the worst
result presented on Table V). Taking into account that the control limits of theX P chart are
LCL = 2.9 and UCL = 3.6, it is easy to identify what has occurred. The distribution of
the test statistic, X, produced with the underlying distribution number 2 has larger average
192 H. Alvelos, J.A. Sarsfield Cabral / Investigação Operacional, 22 (2002) 181-198
1.0
b^ [%]
0.9
0.8
0.7
0.6
0.5 c 2 chart
0.4 modified p chart
0.3
1 2 3 4 5 6 7 8 9 10 11 12
Distribution No.
Figure 5: Evolution of the β risk. Third scenario (left skewed distributions, Ei ≥ 5).
Table V: Estimated β and ARL for some out-of-control situations (right skewed underlying distribu-
tions, E1,2,3 ≤ 2).
Underlying distribution Modified p chart χ2 chart
_ _
No. β(%) ARL β(%) ARL
1 02. 02. 10. 18. 38 0.99655 290 0.99466 187
2 00. 02. 12. 18. 38 0.99870 769 0.99001 100
3 00. 00. 14. 18. 38 0.99839 620 0.96053 25
4 00. 02. 10. 20. 38 0.99750 401 0.99371 159
5 00. 00. 12. 20. 38 0.99641 279 0.98835 86
6 00. 00. 10. 22. 38 0.99355 155 0.98716 78
7 00. 00. 10. 20. 40 0.98273 58 0.99574 235
8 00. 00. 10. 18. 42 0.95907 24 0.99672 305
9 00. 00. 08. 18. 44 0.86989 8 0.99199 125
10 00. 00. 06. 18. 46 0.66659 3 0.96817 31
11 00. 00. 06. 16. 48 0.52359 2 0.92476 13
12 00. 00. 06. 14. 50 0.38144 2 0.83680 6
H. Alvelos, J.A. Sarsfield Cabral / Investigação Operacional, 22 (2002) 181-198 193
1.0
b^ [%]
0.9
0.8
0.7
0.6
0.5 c 2 chart
0.4 modified p chart
0.3
1 2 3 4 5 6 7 8 9 10 11 12
Distribution No.
Figure 6: Evolution of the β risk. Fourth scenario (right skewed distributions, E 1,2,3 ≤ 2).
but smaller standard deviation than the corresponding distribution obtained with the process
in-control (see histograms of Figure 7). Therefore, when distribution number 2 holds the
probability of having a point outside the control limits is very small.
With the χ2 chart another phenomenon takes place: when the assumptions of the χ2 test
do not hold, the distribution of the Q statistic no longer remains close to the χ 2k distribution.
Therefore, some Q distributions have no favourable locations and shapes to perform with the
χ2 test. This fact is illustrated in Figure 8, showing the histograms of the Q statistic obtained
with the underlying distributions 3 and 8, the χ24 p.d.f. and the control limit UCL = 16.2.
The histogram area beyond the UCL is very small, regardless of the fact that the underlying
out-of-control distribution does not agree with the corresponding in-control distribution. The
ARL of distribution number 8 is 305, while the ARL of distribution number 3 is 25, although
the latter is closer to the in-control distribution (the distribution number 1 in Table V).
Taking into account all the results presented on this section, it is recommended that,
whenever it is possible, the two charts should be used together (in spite of the inevitable
increase in the false alarm rate). An observation that lies out of the control limits of any one
of the two charts is a strong indication that a significant change in the process has occurred.
The change is attributable to assignable causes deserving further investigation.
If the circumstances are similar to those described in Table V (that are very common,
in practice) the χ2 chart will not work properly (neither the modified p chart). The better
solution for this problem is to reformulate the questionnaires in order to obtain less categories
having higher expected values, and to proceed as in the previous paragraph. This procedure
can avoid the problems induced when categories are merged a posteriori.
When an out-of-control signal occurs that means the underlying customer satisfaction level
(or profile) departure from the one considered initially. A decision-maker may have difficulties
194 H. Alvelos, J.A. Sarsfield Cabral / Investigação Operacional, 22 (2002) 181-198
0.10 0.10
0.05 0.05
0.00 0.00
2.8 2.9 3.0 3.1 3.2 3.3 3.4 3.5 3.6 3.7 2.8 2.9 3.0 3.1 3.2 3.3 3.4 3.5 3.6 3.7
X X
Figure 7: Histograms of the test statistic X (distributions numbers 1 and 2, Table V).
0.30 0.30
Underlying distribution: Underlying distribution:
00.00. 10. 18. 42 00. 00. 14. 18. 38
0.25 (distribution No. 8 - Table V) 0.25 (distribution No. 3 - Table V)
0.20 0.20
0.15 0.15
0.05 0.05
0.00 0.00
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
2 2
Q, c Q, c
4 4
Figure 8: Comparison between the χ24 p.d.f. and the histograms of the Q statistic obtained with
distributions number 3 and 8 (Table V).
H. Alvelos, J.A. Sarsfield Cabral / Investigação Operacional, 22 (2002) 181-198 195
20
V
15
10
V1
V2
5 V3
V4
V5
0
1 2 3 4 5
Questionnaire Category
in interpreting the direction of the change. From his point of view, there might be doubts
whether the new (different) situation is worse or better than the former. The transition from
one state to the other may even be indifferent to the decision-maker. For example, using the
underlying distributions considered in Table III, is not easy to decide if distribution 8 reveals
a poor or better situation than the one depicted by distribution 1 (the in-control scenario).
The problem can be stated in other words: how to identify situations that actually deserve
a corrective action? When the control chart detects a statistical significant change in the
underlying distribution it does not necessarily imply that the decision-maker assigns less value
to the new situation. If it is possible to capture the decision-maker value function, the problem
has a solution. Construction of value functions must follow some rules (see Clemen, 1996)
that guarantee the consistency of the decision model. We will not address here that issue, too
technical for the scope of this paper. On the other hand, we will suppose that the overall value
function satisfies the requirements for additivity.
Having in mind the quality dimension under analysis, five conjectured value functions were
constructed. Those functions are presented in Table VI and in Figure 9. Note that the value
scale has only a relative meaning (useful for comparisons between alternatives) so no special
meaning shall be assigned to a particular value.
The first function, V1 , shows that the decision-maker has the same increase (decrease) in
satisfaction when changing from any category to the next one. For instance, a change from
category 0 to 1 fully compensates a change from category 4 to 3. On the other hand, V 2 and
196 H. Alvelos, J.A. Sarsfield Cabral / Investigação Operacional, 22 (2002) 181-198
11,0
OV
10,0
9,0 V1
8,0 V2
7,0 V3
6,0 V4
5,0 V5
4,0
3,0
1 2 3 4 5 6 7 8
Distribution No. (Table III)
Figure 10: Evolution of the value assigned by five decision-makers (data from Table III).
V4 are typical of a decision-maker that avoids low categories, while a decision-maker with V 3
or V5 seeks more for the high performances, with the sole difference that, in V3 , category 4 is
less important than in the other case.
Using the underlying distributions presented on Table III, Figure 10 shows the overall
value, OV j , of each satisfaction profile that would be assigned by decision-makers having
value functions V1 , V2 , V3, V4 and V5 , respectively. The values OV j were calculated using the
following expression:
4
Vj (i) · ni
P
i=0
OVj = ,
n
where:
We stress again that OV j has no absolute significance, so its use must be restricted to
comparing two different situations for the same decision-maker (using the same value function).
In particular, any cross-comparison between values coming from different value functions is
meaningless.
Note that the value evolution corresponding to the decision-makers number 4 and 2 are
opposition to the value evolution of decision-makers number 3 and 5: the first are decreasing
while the second are increasing. The value allocated by decision-maker number 1 remain
stable.
This simulation shows that when the control chart signals an out-of-control situation there
H. Alvelos, J.A. Sarsfield Cabral / Investigação Operacional, 22 (2002) 181-198 197
are circumstances in which some sort of corrective action is necessary and there are other cases
in which it isn’t: it depends on the value attributed, as a whole, to the new situation by the
decision-maker.
The following procedure is then advisable when assessing customer satisfaction over time:
1. Monitor the satisfaction level (or profile) using X P and χ2 charts together.
3. If there is a negative change in the overall value, a detailed examination of the process
is required in order to identify and remove assignable causes of variation. If the change in the
underlying distribution leads to a substantial increase of the overall value, the process should
be investigated in order to determine what actually happened. Finally, if the value change is
insignificant no action is required.
7 Conclusions
It is not possible to establish which of the two charts proposed by Wardell and Candia (1996)
(χ2 chart or the modified p chart) is the best for monitoring customer satisfaction over time.
The modified p chart seems to perform slightly better when customer satisfaction level is in-
control, but unfortunately, both charts respond very poorly to many out-of-control situations.
In general, if the assumptions of the χ2 statistic are met, the χ2 chart can be preferred.
On average, it detects more quickly shifts in the underlying satisfaction level that are likely
to occur in practice. However, it is very common that some categories of the questionnaires
have low expected values: in this situation the χ2 statistic is not suitable, being advisable to
reformulate the questionnaires in order to obtain less categories having higher expected values.
This seams to be a better procedure than the one recommended by Wardell and Candia (1996)
consisting of merging adjacent classes a posteriori so that the expected value of each category
is ≥ 5.
As remarked by Wardell and Candia (1996), the signals of the charts are often difficult to
interpret (particularly when the χ2 chart is used because it does not indicate the direction of
change). Value analysis can be a positive contribution to overcome this problem, allowing to
distinguish between relevant alarm signals and those corresponding to insignificant shifts in
the decision-maker value - not deserving a great of concern and investigation - or even gains.
Varying sample sizes is a normal situation when dealing with satisfaction survey data. This
issue is not considered in this paper, although unequal sample sizes can be easily handled by
both the X P and the χ2 charts. Varying sample sizes and inconstant time intervals between
198 H. Alvelos, J.A. Sarsfield Cabral / Investigação Operacional, 22 (2002) 181-198
Another difficulty not considered in this paper, is the instability of the decision-maker
over time. If the surveys cover a long period of time it is likely that the decision-maker
value function change significantly during that period. This is other area requiring additional
analysis.
8 Acknowledgements
The authors acknowledge, with thanks, the comments and suggestions of Professor M. Matos,
of the Faculdade de Engenharia of Porto University, on an earlier version of this paper.
9 References
Clemen, R. T. (1996), Making Hard Decisions: An Introduction to Decision Analysis, Duxbury
Press, Pacific Grove.
Cochran, W.G. (1954), “The χ2 Goodness-of-Fit Test”, Annals of Mathematical Statistics, 23:
315-345.
Duncan, A. J. (1950), “A Chi-Square Chart for Controlling a Set of Percentages”, Industrial
Quality Control, 7: 11-15.
Hayes, B. E. (1992), Measuring Customer Satisfaction: Development and Use of Questionnaires,
ASQC Quality Press, Milwaukee.
Hildbrand, D. and L. Ott. (1991), Statistical Thinking for Managers, Duxbury Press, Belmont.
Marcucci, M. (1985), “Monitoring Multinomial Processes”, Journal of Quality Technology, 17:
86-91.
Montgomery, D. C. (1985), Introduction to Statistical Quality Control, John Wiley & Sons, New
York.
Wardell, D. G. and M. R. Candia. (1996), “Statistical Process Monitoring of Customer Satisfac-
tion Survey Data”, Quality Management Journal, 3, no. 4: 36-50.
Yarnold, J.K. (1970), “The Minimum Expectation in χ2 Goodness-of-Fit Tests and the Accuracy
of Approximations for the Null Distributions”, Journal of the American Statistical Association,
65: 864-886.
P. Borges, L. Sørensen, V. Vidal / Investigação Operacional, 22 (2002) 199-212 199
∗ † ‡
Pedro C. Borges Lene Sørensen René Victor Valqui Vidal
∗
Faculdade de Economia e Gestão, Universidade Católica Portuguesa, Porto
pmb@etc.pt
†
Center for Tele-Information, Technical University of Denmark, Denmark
lene@cti.dtu.dk
‡
Dept. of Informatics and Mathematical Modelling, Technical University of Denmark, Denmark
vvv@imm.dtu.dk
Abstract
Strategy development, planning, and multicriteria are important concepts within or-
ganisations, which almost every day have to comply with the threats and opportunities
of the surrounding environment and have to deal with them. This paper gives an intro-
ductory description to why, how and for what, organisations develop strategies, plan and
can use multicriteria. On the methodological side focus is on soft Operational Research
methodologies and multicriteria decision aid. The paper ends up with short presentations
on situations in which the methodologies have been used in practice.
1 Introduction
As OR practitioners, consultants, teachers and private persons, we are often asked the ques-
tions of why, how and for what when it comes to strategy development, planning and ac-
ceptance and dealing with multicriteria. There are no simple and unambiguous answers to
the questions but a broad range of views and ideas to form some kind of answer. The main
purpose of this paper is to give an overall presentation of the concepts and contents of organi-
sational strategy development, planning and multicriteria as well as some popular approaches
to support such processes.
The outline of the paper is as follows. Section 2 defines and links the concepts and provides
the background and motivation for the paper. Section 3 focuses on strategy development.
200 P. Borges, L. Sørensen, V. Vidal / Investigação Operacional, 22 (2002) 199-212
Here different types of strategies are presented along with two approaches for supporting
the development of such strategies. The approaches are the SWOT analysis and the Future
Workshop. In section 4 the concept of planning is dealt with. Planning is seen as an interactive
process where organisational members are involved. Within this perspective, the scenario
methodology and the Strategic Options Development and Analysis (SODA) are presented.
Other so-called soft approaches can be found in [9]. The task of making personal choices is
closely linked with the existence of multicriteria. Section 5 presents the fundamental concepts
of the multicriteria approach and its different opportunities for methodological support. In
section 6, three examples are shortly presented. They represent different situations to which
the concepts and methodologies have been applied. The examples deal with formulating
information technology strategies in Danish schools, with energy planning on a national level
and lastly with national climate change mitigation strategies. Finally, the conclusions are
presented in section 7.
All individuals in Society are directly or indirectly members (belong to) or related to or-
ganisations (a family, a working place, schools, business organisations, a trade union, local
communities, sport clubs, etc.). Indeed the whole society can be considered as a large-scale
organisation.
Organisations develop usually from day to day in a smooth evolutionary process. Some-
times, it can be foreseen that the organisation will not function as usual or it is not desirable
that it functions as usual. This might occur due to radical changes in the environment (external
factors) or/and major alterations within the organisation itself (internal factors).
External factors could be: the appearance of a new strong competitor, demands that
the school uses more information technology (IT), a public organisation getting privatised,
pollution problems in a local community, environmental legislation imposed in the organisation,
public demand that a football team always wins, a war situation in a country, etc.
Internal factors could be: the modernisation of the organisation, the re-engineering of
a working process, highly qualified people leaving the organisation, the introduction of IT
technology, change of the geographical location of the organisation, the best player is not
available for the next match of a football team, a revolutionary situation in a country, etc.
What can be done? In a situation where radical changes are needed, we believe that
it is important that a group of persons from the organisation, maybe assisted by a facilita-
tor (often from a consulting firm), develops a strategy to be able to cope with the changes
that the future brings about. Strategy development usually involves explicit formulation or
formation of reachable objectives (goals, visions) for the future state of the organisation in
question. Reachable objectives means that although strategy development focuses primarily
on objectives, account is taken of means and resources available.
Examples of simple strategy forms are: kill the organisation, do nothing and pray, conquer
the competitors, formulate visions for the future of the organisation, hire experts to strength
the organisation, etc. These are not unrealistic strategies. In practice, a real-life strategy will
be a complex mixture of these forms. The development, implementation and eventual re-design
P. Borges, L. Sørensen, V. Vidal / Investigação Operacional, 22 (2002) 199-212 201
of a strategy are complex and cumbersome tasks that usually involve all the individuals of the
organisation. This could be described as anticipatory decision-making in a complex system of
decisions. By anticipatory decision-making is meant that one decides what to do and how to
do it, before passing to action. When developing or reformulating strategies, it is advisable to
perceive strategy formulation as a never ending process and not just an isolated act.
How to develop strategies? The answer is planning. Planning is a set of co-ordinated ac-
tivities seeking the fulfilment of desired objectives for the future and actions for bringing them
about. Planning is a process that involves making and assessing each of a set of interrelated
decisions before action is required, in a situation in which it is believed that unless action is
taken a desired future state is not likely to be reached. On the other hand, if suitable action
is taken, the likelihood of a desirable outcome can be increased. What is important in any
planning activity is not only the result, the plan, but also the process by which such a plan
has been formatted because usually plans have to be reformulated, or redesigned as time goes
by.
Planning in practice can by carried out in different forms: a top-down form where the top
managers, eventually assisted by experts hired from a consulting firm, formulate the plan to
realise the formulated strategy; a bottom-up form where different groups, assisted by facili-
tators, through dialogue participate in the formation of a plan; a decorative form where the
plan will never be implemented; an oppressive form where planning is a way to show who has
power; etc.
During the planning process, decision-makers will often be confronted with situations where
they want to select an alternative, an action or a project from a given set of possibilities where
several points of view and often contradictory criteria must be taken into account.
What can be done? These kinds of decision-making problems are the area of Multicriteria
Decision Aid that provides some methods and tools in order to help and support decision
situations where several criteria are taken into consideration. Multicriteria decisions involve
tasks requiring comparison of different alternatives (actions, solutions, options, etc.) that are
described by means of their characteristics or attributes that can be objective or subjective.
Many organisations deal with all three areas at once; strategy development, planning and
deciding which decisions. It is not always easy to distinguish between the three areas, since
dealing with organisational issues somewhat involves all three.
3 Strategy Development
In real-life, strategy development is conditioned by the way the organisation works while solving
problems and taking decisions. Any organisation has a history and will have a tendency to
develop strategies in a way similar to that other big problems have been solved. Changing this
routine usually demands a complete re-design of the organisation. This means that a system
for strategy development has to be suitably tailored for the particular organisation.
what they wish to achieve, meaning that visions and goals have been identified and explicitly
formulated. Secondly, the strategies are made in advance of the action to which they apply.
Thirdly, the strategies are made consciously and purposefully. Fourthly, once the strategy has
been formulated what is left is the problem of implementation; this is a rather complex and
uncertain top-down process demanding a lot of planning.
There are several approaches and tools that can be used to support the process of strategy
development. Approaches to enhance group discussion, dialogue, creativity, collective problem
solving and participation are a must in such a process. Furthermore, it is advisable to use a
facilitator that is a person that primarily focuses on the process of strategy development itself
assuring that it will be accomplished with fruitful results.
A popular method within the business sector is the so-called SWOT (or TOWS) analysis.
During the recent past, other sectors and local communities have also used this method. Its
popularity resides in its simplicity. SWOT is an acronym formed from threats, opportunities,
weakness and strengths. SWOT gives some guidelines for the systematic analyses of the inter-
nal and external environments of an organisation. It involves the assessment and appreciation
of the external factors and from those identifies opportunities and threats posed to the organ-
isation by the external environment. Similarly, the internal factors are used to list strengths
and weakness inherent to the current status of the organisation.
The representation of strengths, weakness, opportunities and threats in tabular form, gives
origin to the SWOT matrix. This matrix suggests four different ways of generating strategies
by combining the minimisation of threats and weakness, and maximisation of strengths and
opportunities. SWOT has shown to be a fruitful approach to support a group of persons in the
first stage of a creative process while developing strategies where emphasis is given to dialogue
and participation. See for example [5] or [6] for more detail.
Another popular approach within community work is the so-called Future Workshop that was
developed in Europe by groups of individuals in local communities and grass-root movements
P. Borges, L. Sørensen, V. Vidal / Investigação Operacional, 22 (2002) 199-212 203
in their work to obtain influence in important political decision processes. Later, this approach
has also been used by private firms especially in relation to the development of strategies for
product development.
The Future Workshop is composed of five stages that are organised and carried out sys-
tematically and sequentially from stage to stage. These stages are:
– Preparation, where the theme for the workshop is formulated and the environment and
other facilities are found.
– Critique, where the problem (or the actual situation) is presented through criticisms and
a kind of summary is achieved.
– Fantasy, where possible solutions (or strategies) are created based on wishes and expec-
tations; these have the character of utopias or visions.
– Realisation, where stage two and three are compared with reality and work is done to
the realisation of the solution (or strategy) selected.
– Follow up, that analyses and summarises the earlier processes and the processes to be
carried out in connection with the implementation process.
Obviously, at each stage of the Future Workshop different techniques and methods (fantasy
sessions, brainstorming, etc.) are used under the advisorship of a facilitator. See [7] for more
detail on the workshop.
Compared to the SWOT method, the Future Workshop demands much more commitment
and ability to be creative from the part of the participants. The Future Workshop is focusing
more in the group process than in the problem itself, quite the opposite to the SWOT method.
4 Planning
– Implementation: design of procedures and organisation so that the plan can be carried
out
The order in which these phases are given does not represent the order in which they should
be completed. Recall that the set of decisions involved in planning are systemic, hence the
parts of a plan and the phases of a planning process must interact.
– Participative: all the members of the organisation have to plan for themselves
– Integrative: both at the different levels of the organisation and at the strategic and
tactical levels integrating planning and implementation
In general there are two approaches to planning, those that specify how good decisions
ought to be made (normative), and those that take departure on the actual way how decisions
are made and that try to support the positive elements and reduce the negative elements of the
process (supportive). The first are called the hard planning approach and the second the soft
planning approach. The hard planning approach is often based on an application of a method
where a method is a tool that usually demands the support of an expert for its application.
The soft planning approach is often based on application of a methodology, which is a set
of recommendations of how the planning process can be carried out usually supported by a
facilitator. A methodology can use several methods.
In supportive planning, the so-called soft methodologies have been developed (see [9]).
These try to support the creative, intuitive and motivating elements in the planning process.
The main guidelines are identification, dialogue, participation and action as fundamental el-
ements in an interactive planning process. Negotiations among the members of the planning
P. Borges, L. Sørensen, V. Vidal / Investigação Operacional, 22 (2002) 199-212 205
A popular methodology that has been used in planning is the: scenario methodology (see for
example [12], [5], and [6]). A scenario is a description/presentation of a possible future state
and the corresponding sequence of events leading to it. The scenario methodology is composed
of three elements: problem structuring, the approach used and the form of interplay among
the actors. In problem structuring some approaches can be used to describe the system and
some approaches to handle with the dynamics of the system. In what concerns the approach
used it is possible to identify two schools: the American school that uses system analysis tools
supported by mathematical and computer models, being primarily computer oriented, and the
French school that uses a less formal approach based on experience, intuition and subjectivity
emphasising dialogue and participation in the form of workshops. The scenario methodology
can be used in both normative and supportive planning processes; this determines the form of
interplay among the actors.
4.1.2 SODA
There is a whole series of approaches for planning and problem solving that can be used in
a supportive way called problem-structuring methods. One of them, and probably one of the
easiest to use, is SODA, the acronym of Strategic Options Development and Analysis (see [9]).
The main tool in this method is cognitive mapping.
A cognitive map is a model of systems of interrelated concepts used by the client to com-
municate how he conceptualises an issue. This model represents the meaning of a concept
by its interrelationship to other concepts in a graphical and manageable way. The facilitator
plays a central role in the elaboration of the cognitive maps.
A SODA project usually develops in two stages: elaboration of individual maps, and
creation of a merged map during a SODA workshop. The facilitator prepares the agenda of
this workshop. The initial purpose of the merged map is to change the minds of each member
of the client group without they feel compromised. The aim is to secure enough agreement
about the nature of the problem that each member is committed to expending energy on
finding a portfolio of actions where that portfolio is the strategy for dealing with the issue.
Cognitive maps follow the problem evolution through time, using appropriate management
SODA can be used both in a process for policy making (strategies) and as a tool for the analysis
of policy issues (planning).
206 P. Borges, L. Sørensen, V. Vidal / Investigação Operacional, 22 (2002) 199-212
5 Multicriteria
How do individuals make personal choices? Which and how many are the factors to take
into consideration and that influences a particular decision? The answers to these questions
will vary, but usually confirm the existence of many factors related to our wishes, values,
possibilities, etc. Those factors can be separated into a few disjoint classes according to their
intrinsic and distinct natures (think about why the horoscopes are built around money, health
and love and not just one of those). Similarly, organisations have to consider many factors
(both internal and external), of different natures, in their decision processes.
For one side, almost all decisions are dependent on multiple factors. For the other, the
human mind can not handle the quantity of information and complexity that sometimes are
generated in the decision process. It might therefore be an advantage to try formalisation by
using methodologies and methods that can support the people involved (decision-makers) in
their tasks.
Choosing among alternatives can be more or less difficult and demand different levels of
analysis and formalism. The best choice might appear clearly and immediately to the decision-
makers, but it might also be very complicated and conflict filled. Two basic approaches can
be identified to formally evaluate alternatives: single-criterion analysis or multiple-criteria
analysis (multicriteria).
In the single-criterion approach alternatives are compared using one single measure or
formula, which is assumed to represent the preference of the organisation. This measure often
corresponds to a single, most important, factor (For example: finding the plant layout that
minimises cost over a number of years or choosing a medical treatment according only to the
survival rate).
The multicriteria approach allows formal consideration of the diversity the several criteria
provided. In this case the wishes of decision-makers can be stated in terms of attributes,
goals or objectives. Attributes correspond to characteristics of the alternatives (Example:
In selecting the vehicles for a taxi fleet a company prefers diesel vehicles). Goals are levels
to be attained by attributes (The fuel consumption should not be excessive). Objectives are
measures to be maximised or minimised (The cars should be as inexpensive as possible). These
three ways of expressing wishes are generally designated as criteria.
To better understand why multicriteria analysis can be an advantage first the single-
criterion philosophy is discussed.
In the single-criterion case finding a good decision or alternative is often left to an analyst
or group that employs the adequate tools (like optimisation) for that purpose. In this approach
there is no doubt whether an alternative is better or worse than another and there will always be
a solution that is “optimal”, if such exists. This makes the choice very easy and unambiguous
but that “optimal” solution is seldom implemented. Why? Because decision-makers are not
committed nor identify themselves to the method and with the results (they might however
P. Borges, L. Sørensen, V. Vidal / Investigação Operacional, 22 (2002) 199-212 207
use the result to help their decision). Two main reasons can be indicated for this failure:
• This process is a simplification of reality and hence has its limitations. The single-
criterion approach only is able to represent the true preferences of the decision-makers
in very uncomplicated cases.
• By delegating the evaluation and choice to an analyst, the decision makers are in fact
withdrawing from the process and give control in crucial decisions to the analyst (For
example, on how to translate social costs of unemployment to monetary units?).
The single-criterion approach is a simplification that makes analysis much less ambiguous
than real life. Multicriteria analysis on the other hand relies on interaction with the decision-
makers to go a step further and bring them into the analysis (they are ultimately the ones who
decide). On top of that the analyst can provide crucial information and insight. This improves
both the quality of the analysis and its acceptance by the decision-makers. For these reasons
we can say that the explicit consideration of the relevant criteria is in practice the only way
to make decisions.
Many MCDA methods deal exclusively with the tasks of evaluation and choice between
alternatives according to criteria. One should however not forget that the generation of al-
ternatives and the choice of criteria are the main ingredients to the evaluation and deserve
therefore a lot of attention. The criteria should cover all the aspects that decision-makers
consider relevant. Agreement should be achieved both on their meaning, inclusion, and on the
way each criterion is constructed and evaluated. Problem structuring methods can be helpful
for this phase. In this discussion, some potential criteria can sometimes be aggregated without
loss of meaning and others might be considered redundant
Some criteria might be conflicting with others in the sense that the available alternatives
show that an improvement in one criteria is accompanied of a deterioration of other. A
characteristic of MCDA is that there will usually not exist an ideal alternative (one “optimal”
alternative, that is the best from all points of view simultaneously), otherwise there is probably
no need for MCDA. This creates a decision conflict. The paradigmatic way to dissolve this
conflict is by generating new alternatives, which due to some structural difference will be ideal
and make the conflict disappear. This can sometimes be achieved through creativity and
radical change. In general however, one has a group of conflicting efficient alternatives. An
alternative is efficient if no other is available that is better from at least one point of view and
not worse in any other.
In the last decades the developments in the area of MCDA produced a large number of methods
and methodologies to support decision making. These methods differ in the way the mecha-
nisms for incorporating decision-maker’s preferences into the decision making process. There
are three basic kinds of strategies for this purpose: “a priori”, “interactive” and “a posteriori”
208 P. Borges, L. Sørensen, V. Vidal / Investigação Operacional, 22 (2002) 199-212
A characteristic of MCDA is that the decision-makers have a lot of control over the pro-
cess. This gives legitimacy to the result and often provides a justification and explanation to
the choice made. Different methods will not necessarily give the exact same result since no
perfect choice method exists. Methods have different assumptions and decision-makers do not
necessarily express preferences exactly in the same way. For these reasons it is advisable to
try more than one single approach and to compare the results. Another possibility is perform-
ing sensitivity analysis to identify critical issues and consolidate the results. Ultimately it is
the decision-makers who decide if they like the result and why. These validation issues are
important for the sake of implementation
The knowledge of the method and the possibility of interacting with a decision support
system allow the decision-makers to reformulate the problem as their knowledge of the situation
improves. In this reformulation, alternatives and criteria can be revised. MCDA as described
here is based on interaction between different individuals and methodological/modelling tools.
This makes MCDA methods natural partners of problem structuring methods (see [3]).
6 Examples
In the following, three examples are given to illustrate how strategy development and planning
take place at different levels, in different organisations, with various purposes, and with differ-
ent support from methodologies and methods. The examples cover: the case of primary schools
in Denmark which are forced by governmental legislation to integrate Information Technology
(IT) in the teaching activities; national energy planning that seeks to balance political goals
with environmentally sound strategies, economic constraints, and technical options; and the
international handling of the global problem of climate change where many nations need to
find ways to effectively implement climate change mitigation measures.
In Denmark, the primary schools currently exhibit tremendous changes. First of all because
the school legislations have been changed giving more and more mandate to the parents of
the students, leaving the teachers and the principal behind to a rather uncertain position.
Secondly, the Government has decided that all primary schools must include IT in all fields
of the subjects and at all levels. Both aspects leave behind questions such as; how can the
schools financially invest in IT (there has not been directed a very large amount of money
for that)? How can the teachers get education so quickly that they can teach the students
about and how to use the technologies? Software programmes are missing for teaching, how
can teachers compensate for that? How can the teaching principles change according to the
usage of IT? What happens to the schools located in towns where the school boards and the
P. Borges, L. Sørensen, V. Vidal / Investigação Operacional, 22 (2002) 199-212 209
municipalities do not think IT should be a high prioritisation? What will be the future roles
of the teachers, the principal, the students, etc.? Planning for the schools includes therefore
large strategic problem areas, which must be dealt with more or less simultaneously.
During several projects where students of the authors have been involved, we have gathered
experience in analysing and using different methodologies for investigating various aspects of
implementing IT in the Danish Primary Schools. Three examples are given here.
Different areas of the planning problems in the schools have to be analysed separately
to make up a whole. The Future Workshop methodology has been applied to analyse one
school class’ expectations to the future use of IT in the schools. The main result was that the
students were rather reluctant to see that IT was a good idea, and they were afraid that the
social dimension of being in the school would vanish.
The SWOT analysis was used in another school to find out what the different actors on this
small village school thought as their personal and the school’s internal strengths, weaknesses,
and external opportunities, and threats in relation to IT. The principal, a teacher, and the
computer manager participated in separately performed interviews. By applying the relatively
simple SWOT analysis as a basis for individual interviews, it became clear that all involved
parties had highly different views of the situation and the future of this small school. The
results of the SWOT analysis were, therefore, used as basis for finding elements for a more
overall view on the school and its situation.
Scenario planning was applied to find out the general future perspectives of the primary
schools of Denmark. Scenarios were used to illustrate future perspectives of which directions
the schools would be heading given different boundaries or decisions. To construct the scenar-
ios, interviews were made using cognitive mapping showing what different individuals would
think of the current and future situation of the schools. These results were combined with
questionnaires and the general reflections of the Government to provide with an overall and
relatively realistic picture of the future schools.
It was the conclusion of all cases that the methodologies applied helped not only the
analysts but also the persons participating in the strategy development and planning to gain
insight into the different expectations of the schools’ actors. Furthermore, the projects provided
a common basis for understanding that the schools’ planning problems were rather complex
and needed some kind of problem structuring approach to be analysed and dealt with properly.
More details on the problem and the cases can be found in [11].
Energy planning has become more and more complex in the last years. Energy resource goals
such as demand and supply balances must be combined with environmentally sound strategies,
economical constraints, technical options, and political priorities and wishes. Analysing a
country’s possibilities for energy strategies therefore becomes a complex task that involves
multiple perspectives, several disciplinary aspects, and numerous constraints and uncertainties
about the future to come.
Here we refer to scenario analysis and scenario planning. Scenario analysis is perhaps the
methodology used most frequently for carrying out analyses and visually showing the effects of
210 P. Borges, L. Sørensen, V. Vidal / Investigação Operacional, 22 (2002) 199-212
the various strategies imposed within the scenarios. Most national energy planning is carried
out in the light of construction of scenarios either using what earlier was referred to as the
principles from the American or French schools of scenario analysis.
However, as the issues of energy planning are presented here, it is clear that the planning
activities include what can be referred to as a strong multicriteria component. Multicriteria
methods can be applied to visibly show what differences in for example priorities mean to the
solutions that are found.
The global, international discussions on climate change issues and mitigation strategies take
place simultaneously while most countries formulate their own national energy and environ-
mental strategies. Many countries, especially the developing countries, meet substantial bar-
riers to taking climate change related measures into consideration when formulating their
national strategies. Priorities and issues are simply too far apart. Furthermore, there exists
no institutional linkage between the often technically and analytical studies that are made in
association with the climate change mitigation country studies, and the relevant national in-
stitutions of the developing countries. The climate change mitigation country studies therefore
often end up having the status of academic exercises that cannot be used to build constructive
and necessary national policies.
The authors have been involved in analysing two aspects in this problem and proposed
ideas for improving the way plans and scenarios related to climate change mitigation are built.
These involve the usage of problem structuring methods and multicriteria decision aid. The
two aspects mentioned here involve the institutional set up for the integration of national and
international strategies, and the methodological construction of climate change mitigation
costing scenarios that are the basis for formulating the international strategies.
In terms of the institutional set-up, the SWOT analysis has been used, in several different
ways, to secure more consistent and integrated strategies where both international regulations
and the developing countries’ planning targets can be integrated. The focus on the institutional
set-up is to secure the co-ordination and communication between the different partners. Using
the SWOT analysis traditionally, one ends up with overall strategies that the single country
can use as basis for more detailed analyses. The SWOT analysis has also been used to assess
and direct the necessary tactical technical options and necessities as basis for the formulation
P. Borges, L. Sørensen, V. Vidal / Investigação Operacional, 22 (2002) 199-212 211
of overall country strategies. Applying the SWOT analysis as the basis for discussing the way
international/national climate change strategies are both linked and constructed, we end up
with a structure that helps building better plans, which have higher probability of successful
implementation.
The analysis of the second aspect, climate change mitigation costing studies, is based
on addressing the individual country emission scenarios. Emission abatement scenarios are
constructed according to feasible and realistic abatement options, whose implementation will
have economic impacts but also other consequences (social, political, environmental) that
cannot unambiguously be aggregated or converted into monetary units. In this case decision-
makers are in the presence of a multicriteria decision situation, as recognised by the analytical
framework proposed for national costing studies. However, and for many reasons, most such
studies do not include a multicriteria evaluation of the proposed options.
The authors have pointed towards the necessity of considering multicriteria analysis within
these analyses. Also experts representing several countries participating in preparatory meet-
ings for the coming new international guidelines have reached the same conclusion. Using
MCDA for this problem is both necessary and inevitable. More details on this study can be
found in [4].
7 Conclusions
This paper provides an overview of the concepts of strategy, planning and multicriteria, their
interrelations and addresses why, how and for what we carry out activities associated. As
a member of any organisation it is relevant and important to know about the reasons and
opportunities of these activities. This paper tries to give that insight.
In the paper, a few soft approaches have been shortly outlined. In [9] more information
can be found on these and more approaches and on their differences in their methodological
support of the strategy development and planning process.
The paper emphasises that Multiple Criteria Decision Aid methods provide a fundamental
basis for organisational problem solving and decision making. Reference [3] gives more insight
into combining MCDA methods with soft OR approaches.
Only three application areas for soft OR and MCDA for strategy development and plan-
ning are mentioned here. Naturally, this way of thinking can be associated with other fields.
Also it shall be mentioned that the combination of the different approaches can be used to
support more complex organisational processes. As an example can be mentioned the support
of restructuring processes for new organisational structures formed as a consequence of the
growing usage and dependence of information technologies.
8 Selected References
[1] Ackoff, R.L., Redesigning the Future – A Systems Approach to Societal Problems. John Wiley &
Sons, New York (1974).
[2] Bogetoft, P. and Pruzan, P., Planning with Multiple Criteria – Investigation, Communication,
212 P. Borges, L. Sørensen, V. Vidal / Investigação Operacional, 22 (2002) 199-212
∗ † ‡
Robinson Hoto Nelson Maculan Marcos Arenales
‡
Fabiano Marques
∗
UEL – CCE – Departamento de Matemática – Campus Universitário – Londrina – PR – Brasil
hoto@uel.br
†
UFRJ – COPPE – Programa de Engenharia de Sistemas e Computação – Rio de Janeiro – RJ –
Brasil
maculan@cos.ufrj.br
‡
USP – ICMC – Departamento de Computação e Estatı́stica – São Carlos – SP – Brasil
{arenales, araxa}@icmc.sc.usp.br
Abstract
It considers the classic knapsack problem. It admits that items must be grouped in
subgroups, and items of a grouping cannot be matched with items of another grouping. The
Compartmented Knapsack Problem (CKP) consists of constructing compartments, whose
dimensions are limited and must be determined, and where each one of them is formed
by items of some grouping. This is a new variation of knapsack that has been studied in
cutting-stock problems. We present in this paper its mathematical formularization and
a resolution method, and as alternative we consider a heuristic procedure to resolve the
problem.
Resumo
Considere o problema clássico da mochila. Admita que os itens devem ser agrupados
em subconjuntos, de modo que itens de um agrupamento não podem ser combinados com
itens de outro. O Problema da Mochila Compartimentada (PMC) consiste em construir
compartimentos, cujas dimensões devem ser limitadas e determinadas, e onde cada um deles
é formado por itens de algum agrupamento. Esta é uma nova variação de mochila que tem
sido estudada em alguns problemas de corte. Neste artigo apresentamos sua formulação
matemática e um método de resolução, e como alternativa propomos um procedimento
heurı́stico para resolver o problema.
1 Introdução
Uma formulação de um caso particular do problema é discutida por Hoto et al. [20] onde
um método branch-and-bound é descrito e, recentemente, Marques [28] e Marques e Arenales
[29] apresentaram alguns procedimentos heurı́sticos para o caso restrito que não será tratado
aqui. O exemplo a seguir ilustra uma Mochila Compartimentada:
Exemplo 1 Considere uma mochila de capacidade igual a 21, que deve ser preenchida
com itens da tabela 1.
Observe que os itens estão agrupados em 2 subconjuntos, pois, os itens 1, 2 e 3 não devem
ser misturados com os itens 4, 5 e 6. Além do mais, considere ainda que as combinações
lineares dos pesos dos itens do agrupamento 1 deverão ter valor entre 5 e 15, já as provenientes
do agrupamento 2 entre 6 e 12. Resumindo, o preenchimento da mochila deverá obedecer as
seguintes condições, onde ai ≥ 0 e inteiro representa o número de itens do tipo i.
3
P 6
P
Condição 1: 5 ≤ pi ai ≤ 15 e 6 ≤ pi ai ≤ 12
i=1 i=4
6
P
Condição 2: pi ai ≤ 21
i=1
3
X
maximizar vj = ui a i
i=1
sujeito a:
3
X
pi a i = w j
i=1
ai ≥ 0 e inteiro , i = 1 , 2 , 3
6
X
maximizar vj = ui a i
i=4
sujeito a:
6
X
pi a i = w j
i=4
ai ≥ 0 e inteiro , i = 4 , 5 , 6
Observe que pode não haver uma combinação linear dos pesos dos itens com valor igual
à capacidade do compartimento, ou seja, uma ou mais mochilas anteriores podem não ter
solução. Neste caso, dizemos que o compartimento associado não é construtivo, no sentido de
que ele não pode ser construı́do, e definimos como nula a sua utilidade, por exemplo:
1) O compartimento 7, que tem capacidade 11 não é construtivo, pois, ele não pode ser
construı́do com os itens do agrupamento 1. Sua utilidade é nula.
2) O compartimento 8, que tem capacidade 12 é construtivo e sua utilidade é 17, pois, pode
3
P
ser construı́do como: pi ai = 5.1 + 9.0 + 7.1 = 12.
i=1
3) O compartimento 10, que tem capacidade 14 é construtivo e sua utilidade é 20, pois,
3
P
pode ser construı́do como: pi ai = 5.1 + 9.1 + 7.0 = 14.
i=1
3
P
Observe ainda que pi ai = 5.0 + 9.0 + 7.2 = 14 é uma outra maneira de construir o
i=1
compartimento 10, porém, adotamos a postura de escolher a combinação linear que produz
o maior valor de utilidade para o compartimento, maiores detalhes serão apresentados na
próxima seção.
existe certamente uma solução ótima em que o compartimento 16 não faz parte. Isso nos
sugere escrever a seguinte definição:
Finalmente uma solução do exemplo 1, indicando uma utilidade total de 37 para a mochila,
é dada por:
Observe que ambas as soluções apresentam o mesmo valor de utilidade (elas produzem os
mesmos itens e em mesmas quantidades), porém, se a cada compartimento for associado um
custo pela sua utilização, a solução alternativa será favorecida e fornecerá um objetivo melhor.
v1=9 v1=9
solução
1 1
v1=9 v12=10
p1=5 p1=5
1 4 4
p1=5 p4=3
v11 =27
solução
1 1 1 alternativa
v12=10
p1=5
4 4
p4=3
os pesos pi inteiros positivos para i ∈ N . Admita que nem todos os itens sejam compatı́veis
entre si, de modo que, uma partição {N1 , . . . , Nk } de N deverá ser considerada. Na mochila
deverão ser construı́dos compartimentos para abrigar itens indexados pelos subconjuntos N s ,
e cujas capacidades são limitadas entre dois valores inteiros, um mı́nimo d min s e um máximo
max
ds , s = 1 , . . . , k.
X
wj = pi ai j , dmin
s ≤ wj ≤ dmax
s , j ∈ Vs (1)
i ∈ Ns
218 R. Hoto et al. / Investigação Operacional, 22 (2002) 213-234
Seja xj o número de vezes que o compartimento j aparece na mochila, cujo custo pela sua
utilização é γj , inteiro e não negativo. Considere ainda yj = 1 se o compartimento de ı́ndice
j foi escolhido para compor a mochila, e yj = 0 no caso contrário. O Problema da Mochila
Compartimentada, doravante designado por PMC, é escrito como:
Modelo do PMC
X X X X
maximizar z = ( ui a i j ) − γ j x j + · · · + ( ui a i j ) − γ j x j (2.1)
j∈V1 i∈N1 j∈Vk i∈Nk
sujeito a: X X X X
pi a i j x j + · · · + pi a i j x j ≤ c (2.2)
j∈V1 i∈N1 j∈Vk i∈Nk
X
dmin
s ≤ pi ai j yj ≤ dmax
s , j ∈ Vs , s = 1 , . . . , k (2.3)
i∈Ns
ai j , xj ≥ 0 e inteiros, yj ∈ {0 , 1}
(2.4)
i ∈ N = N1 ∪ · · · ∪ N k , j ∈ V = V1 ∪ · · · ∪ V k
pi ≤ dmax
s ≤ c , ∀i ∈ Ns , s = 1 , . . . , k (3.1)
ui ui + 1
≥ para i , i + 1 ∈ Ns , s = 1 , . . . , k (3.2)
pi pi + 1
Bobina de Aço
1.o Corte
Bobinas
Intermediárias
2.o Corte
Itens
O método proposto por Carvalho e Rodrigues [40] constrói para cada item i todos os
compartimentos compostos por apenas um tipo de item, e em seguida, resolver a seguinte
mochila:
P P
maximizar u1 a 1 j x j + · · · + un an j xj sujeito a:
j∈V1 j∈Vn
w1 j x j wn j x j
X X
+... + ≤c
j∈V1 j∈Vn
xj ≥ 0 e inteiro, j ∈ V1 ∪ · · · ∪ Vn
No modelo anterior, uj é o custo relativo obtido por meio dos multiplicadores simplex e
xj é o número de vezes que o compartimento aparece na mochila. Note então que o modelo
usado pelos autores é um caso particular do modelo do PMC apresentado aqui. Para que isso
fique ainda mais claro considere o próximo exemplo:
De acordo com a proposta de Carvalho e Rodrigues [40] serão examinados apenas os se-
guintes compartimentos homogêneos:
padrão considerado
pelo gerador de
4 4 4 Carvalho-Rodrigues perda
u41=90 4 4 4
u41=90
padrão NÃO
considerado pelo gerador
3 4 de Carvalho-Rodrigues
utilidade=95
3 4
utilidade=95
Resolvendo a mochila com tais compartimentos o padrão encontrado é composto por duas
unidades do compartimento de largura w4 1 = 3 . 4 = 12, somando uma utilidade de 180
unidades e uma perda de 6 unidades.
Neste mesmo exemplo, entretanto, observamos que esta estratégia não considera padrões
mais expressivos, por exemplo, os dois itens do agrupamento 2 podem ser combinados para
formar um compartimento de largura 15, e que pode ser repetido 2 vezes no preenchimento
da mochila. O padrão compartimentado que acabamos de descrever fornece uma utilidade de
190 unidades e perda nula.
No modelo do PMC, tanto a função objetivo como a restrição fı́sica da mochila são não
lineares. Para resolver o problema, sugerimos decompô-lo em duas etapas: na primeira serão
construı́dos todos os compartimentos dominantes com suas respectivas utilidades, e na segunda
R. Hoto et al. / Investigação Operacional, 22 (2002) 213-234 221
Propriedade 1
Considere o compartimento j ∈ Vs(j) de capacidade wj > dmin s(j) e seu agrupamento Ns(j)
associado. Se existe algum pr para r ∈ Ns(j) , tal que o compartimento de capacidade wj − pr
é construtivo, então o compartimento de capacidade wj é construtivo.
P
Prova Por hipótese temos wj − pr = pi a i ,
i ∈ Ns(j)
P
de onde concluı́mos que wj = pr (ar + 1) + pi a i . ¤
i ∈ Ns(j)
i 6= r
1. Inicialização
1.1 faça W : = ∅;
2.7 se w − pi ∈ C então
fim se 2.7;
2.10 i := i + 1;
2.11 C := C − {0};
¯
2.12 se pmin ≤ dmin
s então faça C : = C − {w ∈ C ¯ w < dmin
s };
2.14 faça W : = W ∪ Ws ;
Observe que uma vez sendo disjuntos os conjuntos Vs , também serão os conjuntos Ws
constituı́dos pelos ı́ndices dos compartimentos construtivos de cada agrupamentos = 1, . . . , k.
Assim, no conjunto W estão todos os ı́ndices dos compartimentos, cujas capacidades possuem
pelo menos uma combinação linear exata dos pesos dos itens indexados por N s . Vejamos agora
as duas etapas da decomposição.
X
maximizar vj = ui a i j (4.1)
i ∈ Ns(j)
sujeito a: X
pi a i j = w j (4.2)
i ∈ Ns(j)
O problema (4.1 – 4.3) foi resolvido pelo algoritmo de Yanasse-Soma [41]. Basicamente o
algoritmo resolve o problema em que wj = dmax s(j) , fornecendo dinamicamente as soluções das
mochilas em que wj = ds(j) , ds(j) + 1 , . . . , dmax
min min
s(j) − 1.
X
maximizar (vj − γj )xj (5.1)
j∈W 0
sujeito a: X
wj x j ≤ c (5.2)
j∈W 0
xj ≥ 0einteiro, j ∈ W 0 (5.3)
R. Hoto et al. / Investigação Operacional, 22 (2002) 213-234 223
O problema (5.1 – 5.3) também pode ser resolvido pelo algoritmo de Yanasse-Soma [41]
usado na resolução do problema (4.1 – 4.3), porém, deve-se efetuar uma pequena modificação
que admite a restrição fı́sica da mochila no formato de desigualdade, e não igualdade.
O custo γj do compartimento j pode ser definido após a sua construção, e ele depende do
compartimento j, sobre tudo dos itens que define o compartimento em questão. Observe que o
problema (4.1 – 4.3) pode apresentar soluções alternativas, assim terı́amos várias possibilidades
para o custo de utilização do compartimento j, e todas elas devem ser consideradas. Afirmamos
que é suficiente escolher apenas aquela que proporciona o menor custo, pois, as demais escolhas
são dominadas pela de menor custo na resolução da mochila (5.1 – 5.3).
1. Inicialização
fim se 2.5
3. Determinação da Compartimentação
fim se 4.2;
Propriedade 2
º
¹
wj 0
w̄j = wj − p (6.1)
p01 1
¹ º
0 w̄j 0
wj = w̄j − p (6.2)
p02 2
¹ º ¹ º
0 wj 0 w̄j
z = 0 u1 + u02 (6.3)
p1 p02
u0
¹ º
0
0
Z =z + wj0 03 (6.4)
p3
p02 − wj0
» 0
p2 − wj0
¹µ » ¼ ¶ 0 ¼ º
u2
Z 1 = z0 + wj0 + p 0
1 − + u01 (6.5)
p01 p02 p01
ZjM T = max{Z 0 , Z 1 } (6.6)
Observe que não poderemos calcular o custo γj pela utilização do compartimento j na mo-
chila, pois, ele depende dos itens que define o compartimento em questão. Assim, adotaremos
a postura de definir o mesmo custo γs para todos os compartimentos indexados em Ws com
itens indexados em Ns , s = 1 , . . . , k.
No passo 3 do COMPEX é feita a escolha dos compartimentos que devem compor a mochila.
O fato de não conhecermos os reais valores das utilidades dos compartimentos nos impede
de determinar os verdadeiros compartimentos dominados, de modo que, o subconjunto W 0
do problema (5.1 – 5.3) pode, no caso de usarmos limitantes, não indexar os verdadeiros
compartimentos dominantes. Assim, resolveremos a mochila (5.1 – 5.3) em W ao invés de
W 0 , quando estivermos usando limitantes para as utilidades dos compartimentos, evitando o
descarte indevido de alguns deles. Vejamos agora o procedimento heurı́stico para resolver o
PMC:
1. Inicialização
fim se 2.4
3. Determinação da Compartimentação
então
fim se 4.4;
senão PARE;
No pior caso o COMPMT com o novo passo 4 atualizará os dados de todos os comparti-
mentos e seu tempo computacional ficará prejudicado, não obstante, ele pode encontrar uma
solução antes de atualizar as utilidades e os custos de utilização de todos os compartimentos
construtivos. Experiências computacionais preliminares não foram favoráveis a esta modi-
ficação, com tempo de execução superior ao do COMPEX, de modo que, abandonamos esta
estratégia de atualizar e refazer a compartimentação.
5 Resultados Computacionais
Na tabela 5 resumimos os resultados que obtivemos para 900 exemplos de mochilas comparti-
mentadas, geradas aleatoriamente. A capacidade de todas as mochilas é 1200, os coeficientes
pi foram gerados no intervalo [3 , dmax
s ], um parâmetro 0 < α < 1 foi gerado aleatoriamente
para definirmos os coeficientes ui = 10 α pi + β, onde β é uma constante. Isso foi feito para
que as utilidades ficassem relacionadas com os pesos. O custo por utilizar um compartimento
foi considerado como nulo.
Os algoritmos foram implementados em Delphi e executados num pentium II, 450 Mhz e
com 160 Mb de RAM.
dmin e dmax : respectivamente, o limite mı́nimo e máximo dos capacidades dos compartimen-
tos;
¯ ¯
calculado pela expressão 100 . ¯(1 − VV al )
alCOM P EX ¯, onde V alCOM P EX é o valor do objetivo
COM P M T ¯
¯
obtido pela solução do COMPEX, eV alCOM P M T o valor do objetivo obtido pela solução do
COMPMT;
O gráfico da figura 4 está fragmentado em três partes de acordo com o número de com-
partimentos: no primeiro fragmento são exemplos com 30 e 200 compartimentos, no segundo
fragmento são exemplos com 150 e 1000 compartimentos e no terceiro fragmento são exemplos
com 900 e 6000 compartimentos.
O próximo gráfico está fragmentado em duas partes de acordo com o número de agrupa-
mentos: 5 para o primeiro fragmento e 20 para o segundo.
Este é o caso em que cada compartimento será utilizado apenas uma vez, ou então não será
utilizado. O PMC 0-1 pode ser resolvido pelo COMPMT ou pelo COMPEX, basta alterar a
mochila do passo 3 para uma mochila 0-1.
R. Hoto et al. / Investigação Operacional, 22 (2002) 213-234 229
1,8
1,7
1,6
1,5
1,4
1,3
1,2
1,1
1 COMPEX
0,9
0,8 COMPMT
0,7
0,6
0,5
0,4
0,3
0,2
0,1
0
1,8
1,7
1,6
1,5
1,4
1,3
1,2
1,1
1 COMPEX
0,9
0,8 COMPMT
0,7
0,6
0,5
0,4
0,3
0,2
0,1
0
X X X X
maximizar z = ( u i a i j ) − γ j yj + · · · + ( u i a i j ) − γ j yj (7.1)
j∈V1 i∈N1 j∈Vk i∈Nk
sujeito a:
X X X X
pi a i j x j + · · · + pi a i j x j ≤ c (7.2)
j∈V1 i∈N1 j∈Vk i∈Nk
X
dmin
s ≤ pi ai j ≤ dmax
s , j ∈ Vs , s = 1 , . . . , k (7.3)
i∈Ns
xj ∈ {0 , 1} , ai j ≥ 0 e inteiros,
(7.4)
i ∈ N = N1 ∪ · · · ∪ N k , j ∈ V = V1 ∪ · · · ∪ V k
Hoto et al. [20] descreve um caso particular do PMC 0-1 apresentado acima, para o qual
o autor restringe a escolha dos compartimentos a apenas um por agrupamento, e cujo modelo
é apresentado a seguir:
X X X X
maximizar z = ( ui a i j ) − γ 1 x j + · · · + ( ui a i j ) − γ k x j (8.1)
j∈V1 i∈N1 j∈Vk i∈Nk
sujeito a:
X X X X
pi a i j x j + · · · + pi a i j x j ≤ c (8.2)
j∈V1 i∈N1 j∈Vk i∈Nk
X
dmin
s ≤ pi ai j xj ≤ dmax
s , j ∈ Vs , s = 1 , . . . , k (8.3)
i∈Ns
X
xj = 1 , s = 1 , . . . , k (8.4)
j∈Vs
xj ∈ {0 , 1} , ai j ≥ 0 e inteiros,
(8.5)
i ∈ N = N1 ∪ · · · ∪ N k , j ∈ V = V1 ∪ · · · ∪ V k
No modelo acima foi introduzida a restrição (8.4) que é justamente a responsável em exigir
apenas um compartimento por agrupamento. Para esse caso particular do PMC 0-1 foi descrito
um método branch-and-bound, cujos detalhes podem ser mais bem estudados em [21].
R. Hoto et al. / Investigação Operacional, 22 (2002) 213-234 231
7 O Caso Restrito
O caso restrito do PMC ocorre quando é preciso limitar o número total de cada tipo de item
a compor a mochila. Vejamos o modelo matemático:
X X X X
maximizar z = ( ui a i j ) − γ j x j + · · · + ( ui a i j ) − γ j x j (9.1)
j∈V1 i∈N1 j∈Vk i∈Nk
sujeito a: X X X X
pi a i j x j + · · · + pi a i j x j ≤ c (9.2)
j∈V1 i∈N1 j∈Vk i∈Nk
X
dmin
s ≤ pi ai j yj ≤ dmax
s , j ∈ Vs , s = 1 , . . . , k (9.3)
i∈Ns
X X
ai j xj + . . . + ai j xj ≤ bi , i ∈ N {limita o total de itens na mochila} (9.4)
j∈V1 j∈Vk
X
ai j ≤ gj , j ∈ Vs , s = 1 , . . . , k {limita o total de itens num compartimento} (9.5)
i∈Ns
X
xj ≤ f , V = V1 ∪ · · · ∪ Vk {limita o total de compartimentos na mochila} (9.6)
j∈V
ai j , xj ≥ 0 e inteiros, yj ∈ {0 , 1}
(9.7)
i ∈ N = N1 ∪ · · · ∪ N k , j ∈ V = V1 ∪ · · · ∪ V k
8 Conclusões
Apresentamos ainda o caso zero-um e o caso restrito que são focos de estudos com outros
autores, sobre tudo o caso restrito, para o qual já foram desenvolvidos alguns procedimentos
que ainda estão sendo aprimorados.
232 R. Hoto et al. / Investigação Operacional, 22 (2002) 213-234
9 Agradecimentos
Ao relator que apontou sugestões e correções valiosas no sentido de tornar clara a compre-
ensão deste texto.
10 Referências
[1] D. A. Babayev, F. Glover and J. Ryan. A new knapsack solution approach by integer equivalent
aggregation and consistency determination. INFORMS Journal on Computing, 9(1), 1997, 43-50.
[2] E. Balas and E. Zemel. An algorithm for large zero-one knapsack problems. Operations Research,
28, 1980, 1130-1154.
[3] R. Bellman. Some applications of the theory of dynamic programming - a review. Operations
Research, 2, 1954, 275-288.
[4] R. Bellman. Dynamic Programming. Princeton University Press, Princeton, NJ, 1957.
[5] R. Bellman and S. E. Dreyfus. Applied Dynamic Programming. Princeton University Press,
Princeton, NJ, 1962.
[6] V. Chvàtal. Linear Programming, W.H.Freeman and Company, New York, 1980.
[7] N. Christofides and C. Whitlock. An algorithm for two dimensional cutting problems. Operations
Research, 25(1), 1977, 30-44.
[8] G. B. Dantzig. Discrete variable extremum problems. Operations Research, 5, 1957, 266-277.
[9] B. L. Dietrich and L. F. Escudore. More coefficient reduction for knapsack-like constraints in 0-1
programs with variable upper bounds. IBM T.J., Watson Research Center, RC 14389, Yorktown
Heights, N. Y., 1989.
[10] D. Fayard and V. Zissimopoulos An approximation algorithm for solving unconstrained two
dimensional knapsack problem. European Journal of Operational Research, 84, 1995, 618-632.
[11] J. S. Ferreira, M. A. Neves and P. F. Castro . A two-phase roll cutting problem. European Jounal
of Operational Research, 44,1990, 185-196.
[12] P. C. Gilmore and R. E. Gomory. A Linear Programming Approach to the Cutting Stock
Problem, Operations Research, 9, 1961, 849-859.
[13] P. C. Gilmore and R. E. Gomory. A Linear Programming Approach to the Cutting Stock
Problem, parth II, Operations Research, 14, 1963, 94-120.
[14] P. C. Gilmore and R. E. Gomory. The theory and computation of knapsack functions. Operations
Research, 14, 1966, 1045-1074.
[15] R. W. Haessler. Solving the two-stage cutting-stock problem. Omega, The International Journal
of Management Science, 7(2), 1979, 145-151.
[16] E. Horowitz and S. Sahni. Computing partitions with applications to the knapsack problem.
Jounal of ACM, 21, 1974, 277-292.
R. Hoto et al. / Investigação Operacional, 22 (2002) 213-234 233
[17] R. Hoto. Otimização no Corte de Peças Unidimensionais com Restrições de Agrupamento. Dis-
sertação de Mestrado, ICMSC-USP, São Carlos, S.P., Brasil, 1996.
[18] R. Hoto e M. Arenales. O problema do corte em bobinas de aço. XX CNMAC, Gramado, R.S.,
Brasil, 1997.
[19] R. Hoto, N. Maculan e M. N. Arenales. O problema do corte em bobinas de aço via geração
de colunas. XXX Simpósio Brasileiro de Pesquisa Operacional, Curitiba, P.R., Brasil, 1998,
267-268.
[20] R. Hoto, M. N. Arenales e N. Maculan. O Problema da Mochila Compartimentada. Relatório
Técnico, Departamento de Matemática, Centro de Ciências Exatas, Universidade Estadual de
Londrina, Londrina, P.R., Brasil, 1999.
[21] R. Hoto. O Problema da Mochila Compartimentada aplicado no Corte de Bobinas de Aço. Tese
de Doutoramento, COPPE-Sistemas/UFRJ, Rio de Janeiro, R.J., Brasil, 2001.
[22] G. P. Ingargiola and J. F. Korsh. A reduction algorithm for zero-one single knapsack problems.
Management Science, 20, 1973, 460-463.
[23] R. E. Johnston and L. R. Khan. Bounds for nested knapsack problems. European Journal of
Operational Research, 81, 1995, 154-165.
[24] G. A. P. Kindervater and J. K. Lenstra. An introduction parallelism in combinatorial optimiza-
tion. Discrete Applied Mathematics, 14, 1986, 135-156.
[25] P. J. Kolesar. A branch-and-bound algorithm for the knapsack problem. Management Science,
13, 1967, 723-735.
[26] E. Yu-Hsien Lin. A bibliographical survey on some well-known non-standard knapsack problems.
INFOR, 36(4), 1998, 274-317.
[27] N. Maculan. Relaxation Lagrangienne: le problème du knapsack 0-1. INFOR, Canadian Journal
of Operational Research and Information Processing, 21, 1983, 315-327.
[28] F. P. Marques. O Problema da Mochila Compartimentada. Dissertação de Mestrado, ICMC-
USP, São Carlos, S.P., Brasil, 2000.
[29] F. P. Marques e M. N. Arenales (2001). “O problema da mochila compartimentada.” Artigo
submetido à Revista Pesquisa Operacional – SOBRAPO.
[30] S. Martello and P. Toth. An upper bound for the zero-one knapsack problem and a branch-and-
bound algorithm. European Journal of Operational Research, 1, 1977a, 169-175.
[31] S. Martello and P. Toth. Branch-and-bound algorithms for the solution of the general unidimen-
sional knapsack problem. In M. Roubens (ed.). Advances in Operations Research, North-Holand,
Amsterdam, 1977b, 295-301.
[32] S. Martello and P. Toth. Knapsack Problems: Algorithms and Computer Implementations. John
Wiley & Sons, Chichester, 1990.
[33] David Pisinger. Algorithms for knapsack problems. Ph.D. Thesis, Department of Computer
Science, University of Copenhagen, Copenhagen, Denmark, 1995.
[34] Hisatoshi Suzuki. A generalized knapsack problem with variable coeficientes. Mathematical
Programming, 15, 1978, 162-176.
[35] N. Y. Soma, H. H. Yanasse and N. Maculan. O problema da mochila. O Problema de Corte e
Empacotamento e Aplicações Industriais, 2.a Oficina Nacional de PCE, XX CNMAC, Gramado,
R.S., 1997, 24-58.
[36] N. Y. Soma, H. H. Yanasse, A. S. I. Zinober and P. J. Harley. A polynomial approximation
scheme for the subset sum problem. Discrete Applied Mathematics, 57, 1995, 243-253.
234 R. Hoto et al. / Investigação Operacional, 22 (2002) 213-234
[37] R. Tarjan. Depth-first search and linear graph algorithms. SIAM Journal on Computing, 1,
1972, 146-160.
[38] J. M. V. Valério de Carvalho. Um problema de corte em duas fases. Tese de Doutoramento,
Universidade do Minho, Portugal, 1991.
[39] J. M. V. Valério de Carvalho and A. J. Guimarães Rodrigues. A computer based interactive
approach to a two-stage cutting-stock problem. INFOR, 32(4), 1994, 243-252.
[40] J. M. V. Valério de Carvalho and A. J. Guimarães Rodrigues. An LP-based approach to a
two-stage cutting-stock problem. European Journal of Operational Research, 84, 1995, 580-589.
[41] H. H. Yanasse, N. Y. Soma. A new enumeration scheme for the knapsack problem. Discrete
Applied Mathematics, 18, 1987, 235-245.
[42] H. H. Yanasse, N. Y. Soma and N. Maculan. An algorithm for determining the k-best solutions
of one-dimensional knapsack problem. Pesquisa Operacional, 20, 1, 2000, 117-134.
J. Valente, R. Alves / Investigação Operacional, 22 (2002) 235-251 235
∗ ∗
Jorge Valente Rui Alves
∗
Faculdade de Economia, Universidade do Porto
{jvalente, ralves}@fep.up.pt
Abstract
Resumo
1 Introdução
data de entrega (due date), verificando-se a existência de um custo sempre que a data de
conclusão Cj de um trabalho não coincida com a respectiva data de entrega, na medida em
que se considera não apenas um custo de atraso por unidade de tempo, como também um
custo de posse do stock, igualmente por unidade de tempo. O objectivo consiste em encontrar
a sequência dos diversos trabalhos que permite minimizar a soma dos custos de posse/atraso
associados ao conjunto de todos os trabalhos.
3) os custos de posse e de atraso podem ser distintos (não apenas entre si, como também
entre os diversos trabalhos), representanto hj e wj , respectivamente, o custo de posse e
o custo de atraso (por unidade de tempo) relativo ao trabalho Jj ;
4) as datas de entrega dos vários trabalhos podem ser distintas, sendo a data de entrega
associada ao trabalho Jj representada por dj ;
6) não é permitida a existência de tempo morto não forçado, pelo que o processador nunca
poderá estar parado desde que exista pelo menos um trabalho disponı́vel para processa-
mento.
O problema P1 é strongly NP-hard na medida em que, e como já foi referido por Ow e
Morton (1989) [7], ele inclui como caso especial a minimização da weighted tardiness, tendo
Lenstra, Rinnooy Kan e Brucker (1977) [5] demonstrado que mesmo este problema mais restrito
é já efectivamente strongly NP-hard. O problema P2 também é strongly NP-hard, dado que
inclui o problema P1 como um caso particular.
inferiores (lower bounds) são calculados por via da relaxação do espaço de estados associ-
ado à formulação de programação dinâmica relativa a este problema. Estes autores referem
parâmetros associados a este problema susceptı́veis de influenciar o desempenho dos algorit-
mos, mas das experiências computacionais por eles efectuadas não é possı́vel retirar conclusões
quanto ao seu impacto concreto, dado o número muito reduzido de instâncias analisadas.
Szwarc (1988) [8] descreve um procedimento de branching que permite igualmente encontrar
uma solução óptima para o problema P1.
O problema P1 é também analisado por Ow e Morton (1989) [7], sendo que estes auto-
res apresentam vários métodos heurı́sticos nos quais o próximo trabalho a ser executado é
seleccionado por via de uma função de prioridade (priority function), tendo realizado testes
computacionais que permitiram concluir que uma destas heurı́sticas superava claramente as
restantes, possibilitando a obtenção de sequências relativamente boas. Estes autores desenvol-
vem ainda um outro algoritmo de aproximação no qual é utilizada uma variante das técnicas
de beam search, tendo sido efectuadas experiências computacionais que demonstraram que
este método permitia obter muito boas soluções mesmo que fossem geradas árvores de pes-
quisa relativamente pequenas. Os testes computacionais efectuados por Ow e Morton foram,
no entanto, realizados sob condições experimentais restritas, na medida em que apenas foi
considerado um subconjunto reduzido das diversas combinações dos parâmetros que podem
influenciar a performance dos algoritmos, muito embora para cada uma dessas combinações
tenha sido analisado um número elevado de instâncias.
estas variantes por via dos procedimentos existentes para o problema de afectação.
Neste artigo iremos apresentar, para os problemas P1 e P2, um método de optimização ba-
seado em programação dinâmica, por um lado, e analisar a performance de algumas heurı́sticas
(entre as quais se inclui a função de prioridade que, de entre todas as analisadas por Ow e
Morton, apresentou a melhor performance), por outro lado. O procedimento de programação
dinâmica foi por nós desenvolvido por via da aplicação ao caso particular do problema early/tardy
de uma formulação geral de problemas de sequenciamento em termos de programação dinâmica
apresentada por Held e Karp (1962) [2]. Assim, temos como objectivo avaliar se o algoritmo de
programação dinâmica constitui uma alternativa válida aos métodos de B&B para a obtenção
de uma solução óptima para o problema P1, na medida em que anteriormente a programação
dinâmica (ou, mais precisamente, uma sua relaxação) apenas tinha sido utilizada para a ob-
tenção de limites inferiores (e não directamente como método optimizante), bem como estudar
a sua viabilidade ao nı́vel do problema P2. O nosso trabalho procura também analisar o de-
sempenho de todo um conjunto de heurı́sticas e estudar o impacto de parâmetros associados
ao problema early/tardy sobre a performance desses métodos para os problemas P1 e P2, bem
como, mais particularmente, averiguar se o bom desempenho da regra EXP-ET (de Ow e
Morton) no contexto do problema P1 se deve apenas ao conjunto particular de combinações
de parâmetros considerado pelos autores ou, pelo contrário, se permanece válido mesmo sob
condições experimentais mais gerais.
Este artigo está organizado da seguinte forma. Na secção 2 iremos apresentar uma descrição
238 J. Valente, R. Alves / Investigação Operacional, 22 (2002) 235-251
dos diversos algoritmos relativos aos problemas P1 e P2. Os métodos utilizados para gerar as
instâncias usadas no teste destes algoritmos serão explicitados na secção 3, sendo os resultados
computacionais descritos na secção 4 e as conclusões finais apresentadas na secção 5.
Nesta secção iremos descrever os algoritmos aplicados aos problemas P1 e P2, quer as heurı́sticas,
quer o procedimento de programação dinâmica. Os algoritmos para estes dois problemas, como
veremos, são similares, sendo que as diferenças se prendem apenas com algumas alterações mo-
tivadas pelo facto de, no problema P2, ser possı́vel que os trabalhos se tornem disponı́veis em
momentos de tempo distintos.
2.1 Problema P1
A heurı́stica Cost Ratio consiste simplesmente em sequenciar os trabalhos por ordem não
decrescente do seu rácio hj/wj , ou seja, do rácio entre o custo de posse e o custo de atraso,
sendo os empates resolvidos de forma arbitrária.
Assim, sempre que o processador se tornar disponı́vel e ainda existirem m trabalhos não
sequenciados, com 1 ≤ m ≤ n, a heurı́stica Cost Ratio seleccionará para execução o trabalho
Ji , 1 ≤ i ≤ m, tal que hi /wi = min1≤l≤m {hl /wl }.
A heurı́stica Greedy não é mais do que, e como o próprio nome indica, uma heurı́stica mı́ope
ou de óptimo local, ou seja, uma heurı́stica que escolhe, em cada passo ou iteração, o trabalho
que se traduzir, nesse momento, num menor acréscimo de custo, sendo eventuais empates
resolvidos de forma arbitrária.
Deste modo, sempre que o processador se tornar disponı́vel num momento de tempo t e
ainda existirem m trabalhos não sequenciados, com 1 ≤ m ≤ n, a heurı́stica Greedy seleccio-
nará para execução o trabalho Ji , 1 ≤ i ≤ m, tal que
n o
= min hl (dl − (t + pl ))+ + wl ((t + pl ) − dl )+
1≤l≤m
procuram reduzir a miopia por este evidenciada. Deste modo, a heurı́stica LookAheadGreedy
calcula a cada iteração, e para cada trabalho, a soma do custo actual relativo a esse trabalho
(o custo a ele associado caso seja sequenciado de imediato) com a variação de custo imedi-
ata provocada nos restantes trabalhos ainda não sequenciados (ou seja, a diferença entre o
custo desses trabalhos caso fossem sequenciados na posição seguinte e o seu custo actual). A
heurı́stica selecciona a cada iteração o trabalho que apresentar o menor valor para esta soma,
sendo eventuais empates resolvidos de forma arbitrária.
Assim, sempre que o processador se tornar disponı́vel num momento de tempo t e ainda
existirem m trabalhos não sequenciados, com 1 ≤ m ≤ n, a heurı́stica LookAheadGreedy
seleccionará para execução o trabalho Ji , 1 ≤ i ≤ m, tal que Custoi = min1≤l≤m {Custol }. A
variável Custol é dada pela expressão
sendo:
A heurı́stica EDD (earliest due date) consiste em sequenciar os trabalhos por ordem não
decrescente da sua data de entrega dj , sendo eventuais empates resolvidos de forma arbitrária.
Deste modo, sempre que o processador se tornar disponı́vel e ainda existirem m trabalhos
não sequenciados, com 1 ≤ m ≤ n, a heurı́stica EDD seleccionará para execução o trabalho
Ji , 1 ≤ i ≤ m, tal que di = min1≤l≤m {dl }.
A regra EDD foi introduzida por Jackson (1955) [3], tendo este autor demonstrado que
este procedimento permitia obter uma solução óptima para o problema no qual o objectivo
consistia na minimização da maior lateness, ou seja, 1||Lmax . Esta regra tem sido utilizada
como heurı́stica ao nı́vel de outros problemas de sequenciamento, nomeadamente no contexto
do problema da total tardiness (1|| wj Tj ), no qual o objectivo consiste em minimizar a soma
P
dos custos de atraso associados aos diversos trabalhos, não sendo considerados custos relativos
a uma conclusão antecipada dos trabalhos. Assim, e de modo análogo ao trabalho realizado
por Ow e Morton (1989) [7], a regra EDD é incluı́da como um exemplo de um método que
ignora os custos associados à earliness, servindo deste modo como referência para a análise de
algoritmos que considerem os custos de posse e de atraso.
A heurı́stica EXP-ET foi desenvolvida por Ow e Morton (1989) [7] e, de entre as várias
regras de prioridade (dispatch priority rules) analisadas por estes autores (excluindo assim os
240 J. Valente, R. Alves / Investigação Operacional, 22 (2002) 235-251
métodos de beam search também por eles considerados), possibilitou a obtenção das melhores
performances. Esta heurı́stica calcula, em cada iteração, uma prioridade associada a cada
trabalho, sendo seleccionado o trabalho que possua a maior prioridade. Assim, sempre que o
processador se tornar disponı́vel num momento de tempo t e ainda existirem m trabalhos não
sequenciados, com 1 ≤ m ≤ n, a heurı́stica EXP-ET seleccionará para execução o trabalho
Ji , 1 ≤ i ≤ m, tal que Pi (si ) = max1≤l≤m {Pl (sl )}, com
Wl , sesl³≤ 0 ´ ³ ´
Wl exp − Hl +Wl (sl /kp) , se0 ≤ sl ≤ Wl
Hl +Wl kp
Hl
Pl (sl ) = ³
(H +W )s
´3 ³
W
´
Hl−2 Wl − l kp l l , se Hl +W l
l
kp ≤ sl ≤ kp
−H , ses ≥ kp
l l
O procedimento de programação dinâmica que iremos agora descrever baseia-se numa for-
mulação geral de problemas de sequenciamento em termos de programação dinâmica apre-
sentada por Held e Karp (1962) [2]. Seja N = {J1 , J2 , ..., Jn } o conjunto dos n trabalhos a
processar e S ⊆ N um subconjunto arbitrário desses mesmos trabalhos. Seja ainda g ∗ (S) o
menor custo total que é possı́vel suportar quando os trabalhos de S estão atribuı́dos às pri-
meiras |S| posições da sequência. O objectivo consiste em sequenciar todos os trabalhos de
N de modo a minimizar o custo total suportado, ou seja, de modo a minimizar g ∗ (N ). O
algoritmo, de acordo com a metodologia da programação dinâmica, divide o problema em n
estágios e actua de forma recursiva, gerando a cada estágio e, 1 ≤ e ≤ n, e a partir do conjunto
N , todos os subconjuntos S com |S| = e (ou seja, todas as combinações dos n trabalhos com
e elementos). Para cada um destes subconjuntos é então encontrado o valor de g ∗ (S) por via
da equação recursiva
½ µX ¶¾
∗ ∗
g (S) = min g (S − {Ji }) + fi pj ,
i∈S j∈S
onde fi (Ci ) representa a função custo associada ao trabalho Ji e sendo inicialmente atribuı́do
o valor de 0 a g ∗ (∅). Assim, para cada estado apenas são considerados os estados do estágio
J. Valente, R. Alves / Investigação Operacional, 22 (2002) 235-251 241
anterior que dele diferem em apenas um trabalho, na medida em que apenas destes estados será
possı́vel aceder ao estado actual, por via do processamento do trabalho na qual esses estados
diferem. Seja Jk , k ∈ S, o trabalho que permite obter gn∗ (S), ou seja, o trabalho ³P que,
´o de
∗
entre todos os trabalhos Ji ∈ S, assegura a minimização de g (S − {Ji }) + fi j∈S pj . O
algoritmo armazena então a informação de que o subconjunto S é precedido pelo subconjunto
(S − {Jk }), o que é equivalente a afirmar que, no subconjunto S, o trabalho Jk é processado na
posição |S|, dado que esta informação se revela necessária para a construção da solução óptima.
De facto, uma vez calculado o valor de g ∗ (N ), é possı́vel determinar a sequência óptima
retrocedendo ao longo dos diversos estágios, transitando sucessivamente de um subconjunto S
para o estado ou subconjunto que o precede.
2.2 Problema P2
2.2.1 As Heurı́sticas
Os métodos heurı́sticos atrás descritos podem ser adaptados ao problema P2 por via da in-
trodução de ligeiras modificações. Na verdade é necessário introduzir algumas alterações ao
nı́vel dos trabalhos que, em cada iteração, são alvo de análise, na medida em que as heurı́sticas
irão agora considerar não todos os trabalhos ainda não sequenciados, como ocorria no problema
P1, mas apenas aqueles que, de entre os trabalhos ainda não processados, já se encontram dis-
ponı́veis. A consideração de datas de disponibilidade distintas para os diversos trabalhos pode,
por outro lado, tornar ainda necessária a introdução involuntária de tempo morto, na medida
em que é possı́vel que num determinado momento de tempo (ou seja, no fim de uma dada
iteração) já tenham sido executados todos os trabalhos que se tornaram disponı́veis até esse
mesmo instante, sendo necessário avançar até à data de disponibilidade seguinte antes de
iniciar a próxima iteração.
O algoritmo gera em cada estágio, mas a partir agora do conjunto N T e não do conjunto N ,
todos os subconjuntos S com |S| = e. Para cada um destes subconjuntos é então encontrado
o valor de g ∗ (S) por via da equação recursiva
n ³ ´o
g ∗ (S) = min g ∗ (S − {Ji }) + fi t(S−{Ji }) + pi ,
i∈S,t(S−{J }) ≥ri
i
sendo que, na inicialização do algoritmo, se define g ∗ (∅) = 0 (assim, para o cálculo de g ∗ (S)
serão agora considerados apenas os estados do estágio anterior que, por um lado, diferem de
S em apenas um trabalho e aos quais, por outro lado, esteja associado um momento de tempo
não inferior à data de disponibilidade desse mesmo trabalho no qual os dois estados diferem).
Seja Jk , k ∈ S, o trabalho que permite obterng ∗ (S), ou seja, o trabalho
³ que, de
´oentre todos os
∗
trabalhos Ji ∈ S, assegura a minimização de g (S − {Ji }) + fi t(S−{Ji }) + pi . O algoritmo
armazena então a informação de que o subconjunto S é precedido pelo subconjunto (S − {J k }),
o que é equivalente a afirmar que, no subconjunto S, o trabalho Jk é processado na posição
|S|, registando igualmente o momento de tempo associado ao subconjunto S, sendo t S =
t(S−{Jk }) +pk . A introdução involuntária de tempo morto pelo algoritmo poderá eventualmente
ocorrer caso num determinado estágio e apenas seja gerado um único nó, na medida em que
isto significa que apenas e trabalhos se encontravam disponı́veis no inı́cio desse estágio. Assim,
será necessário introduzir tempo morto se a data de disponibilidade do próximo trabalho
for superior ao momento de tempo associado ao único estado existente. Nestas condições,
devemos avançar até à data de disponibilidade do próximo trabalho, igualando o momento
de tempo relativo ao único estado existente a essa data de disponibilidade, assegurando deste
modo que em cada estágio e existem pelo menos e trabalhos eventualmente disponı́veis para
processamento.
A complexidade das diversas heurı́sticas encontra-se resumida no Quadro 2.1., sendo que o
procedimento de programação dinâmica possui, para os problemas P1 e P2, tempo de execução
exponencial. De facto, no problema P1 serão sempre gerados 2 n estados, enquanto no problema
P2 o número de estados será igual ou inferior a 2n (a existência de datas de disponibilidade
distintas para os diversos trabalhos contribui para reduzir o número de estados criados). As
heurı́sticas Cost Ratio e EDD consistem, no caso do problema P1, numa simples ordenação dos
trabalhos de acordo com o rácio entre custos e as datas de entrega, respectivamente, o que pode
ser realizado
¡ 2¢
em tempo O (n log n). A complexidade destas heurı́sticas torna-se, no entanto,
igual a O n no contexto do problema P2, dado que não é agora suficiente efectuar uma
ordenação dos trabalhos, na medida em que a existência de datas de disponibilidade distintas
implica a necessidade de, em cada uma das n iterações, proceder à análise do rácio de custos
ou data de entrega de um número máximo de n trabalhos. A complexidade das restantes
heurı́sticas é idêntica para os dois problemas. As heurı́sticas Greedy e EXP-ET calculam,
em cada uma das n iterações realizadas, o custo ou prioridade de um número máximo de
n trabalhos (sendo que este custo ou prioridade pode ser calculado em tempo constante).
A heurı́stica LookAheadGreedy, por sua vez, calcula também, em cada uma das n iterações
realizadas, o custo de um número máximo de n trabalhos. No entanto, o cálculo do custo
associado a cada trabalho implica analisar
¡ 3¢
os restantes trabalhos ainda não executados, daı́
resultando um tempo de execução O n .
J. Valente, R. Alves / Investigação Operacional, 22 (2002) 235-251 243
3 Geração de Instâncias
Ao longo desta secção iremos descrever a metodologia utilizada para gerar as instâncias dos
problemas P1 e P2 que servirão de base à análise da performance dos algoritmos acabados de
descrever. Os valores aleatórios que se revelaram necessários à criação das diversas instâncias
foram gerados recorrendo ao SPSS 8.0 para Windows.
3.1 Problema P1
No que diz respeito ao problema no qual os diversos trabalhos se tornam disponı́veis em si-
multâneo, o método por nós utilizado para gerar as diversas instâncias deste problema coincide
com o adoptado por Abdul-Razaq e Potts (1988) [1].
Para cada trabalho são gerados, a partir de uma distribuição uniforme no intervalo [1, 10],
valores inteiros para o tempo de processamento pj , o custo de posse por unidade de tempo hj
e o custo de atraso por unidade de tempo wj . Para cada trabalho é ainda gerada, a partir
de uma distribuição uniforme no intervalo [T (1 − LF − RDD/2) , T (1 − LF + RDD/2)], um
valor inteiro para a data de entrega dj . O significado associado aos parâmetros T , LF e RDD,
bem como os valores que estes podem assumir, são os seguintes:
– RDD (range of due dates ou dispersão das datas de entrega): este parâmetro deter-
mina a dimensão do intervalo do qual serão retirados os valores das datas de entrega,
influenciando assim a dispersão dessas datas de entrega; RDD ∈ {0.2, 0.4, 0.6, 0.8, 1.0}.
Os valores que o parâmetro LF pode assumir resumem-se a {0.2, 0.4} na medida em que
o problema total holding-tardiness cost é, de certo modo, simétrico, dado que existe um custo
não apenas quando um trabalho é concluı́do após a sua data de entrega, mas também caso um
trabalho seja concluı́do antes da respectiva data de entrega, pelo que a dificuldade inerente
244 J. Valente, R. Alves / Investigação Operacional, 22 (2002) 235-251
3.2 Problema P2
O método utilizado para obter os pj , hj e wj não sofreu quaisquer alterações. Para cada
trabalho é também gerada, a partir de uma distribuição uniforme no intervalo [0, RRD ∗ T ],
um valor inteiro para a data de disponibilidade rj . O parâmetro T representa uma vez mais a
soma dos tempos de processamento de todos os trabalhos pertencentes à instância em análise,
enquanto o significado associado ao parâmetro RRD, bem como os valores que este pode
assumir, são os seguintes:
– RRD (range of release dates ou dispersão das datas de disponibilidade): este parâmetro
determina a dimensão do intervalo do qual serão retirados os valores das datas de dis-
ponibilidade, influenciando assim a dispersão dessas datas de disponibilidade; RRD ∈
{0.2, 0.4, 0.6, 0.8}.
Para cada trabalho será ainda gerado, a partir de uma distribuição uniforme no inter-
valo [T (1 − (1 − RRD)/2 − RDD/2) , T (1 − (1 − RRD)/2 + RDD/2)], um valor inteiro para
a data de entrega dj . Este método de obtenção das datas de entrega não contempla, ao
contrário do que ocorria no procedimento adoptado para o problema P1, diferentes pontos
médios para o intervalo do qual serão retirados os valores das datas de entrega (ou seja, dife-
rentes factores de atraso). De facto, optámos por considerar sempre o mesmo valor médio para
o intervalo que serve de base à obtenção das datas de entrega, sendo que este valor se situa,
para cada instância, precisamente no meio do intervalo compreendido entre o extremo superior
do intervalo correspondente às datas de disponibilidade e a soma dos tempos de processamento
(ou seja,[RRD ∗ T, T ]), pelo que o parâmetro LF foi assim substituı́do por (1 − RRD)/2. O
parâmetro RDD representa uma vez mais a dispersão das datas de entrega, sendo novamente
responsável pela dimensão do intervalo a partir do qual serão geradas as datas de entrega. No
entanto, temos agora que os valores assumidos por este parâmetro para uma dada instância
dependem do valor da dispersão das datas de disponibilidade associado a essa mesma instância,
sendo que:
J. Valente, R. Alves / Investigação Operacional, 22 (2002) 235-251 245
RRD = 0.2 ⇒ RDD ∈ {0.2, 0.4, 0.6, 0.8, 1.0}
RRD = 0.4 ⇒ RDD ∈ {0.2, 0.4, 0.6, 0.8}
RRD = 0.6 ⇒ RDD ∈ {0.2, 0.4, 0.6}
RRD = 0.8 ⇒ RDD ∈ {0.2, 0.4}
De facto, procurámos assegurar que a data de entrega associada a um dado trabalho não
fosse inferior à soma da data de disponibilidade e do tempo de processamento relativos a esse
mesmo trabalho, pelo que se tornou necessário evitar uma grande sobreposição entre o intervalo
do qual são retirados os valores das datas de disponibilidade e o intervalo que serve de base à
obtenção das datas de entrega, de modo a diminuir a probabilidade de serem gerados valores
inválidos para as datas de entrega. Deste modo, o maior valor possı́vel para o parâmetro RDD,
e para um dado valor do RRD, corresponde apenas a uma sobreposição ligeira dos intervalos
correspondentes às datas de disponibilidade e às datas de entrega.
No que diz respeito ao número de trabalhos contidos nas instâncias geradas, foram criadas
instâncias com 5, 10 e 15 trabalhos. O número total de instâncias geradas para cada dimensão
foi igual a 140, dado que foram criados 10 problemas para cada um dos 14 pares possı́veis dos
parâmetros RRD e RDD.
Ao longo desta secção iremos proceder à apresentação e análise dos resultados computacionais,
sendo a performance dos procedimentos heurı́sticos, não apenas no que se refere aos valores
gerados para a função objectivo mas também ao nı́vel dos tempos de execução, comparada
com a associada aos métodos de programação dinâmica.
4.1 Problema P1
No que se refere aos tempos de computação, os métodos heurı́sticos revelam-se, tal como seria
de esperar dada a sua reduzida complexidade temporal, particularmente rápidos, sendo que
para todas as instâncias consideradas a solução gerada por estes procedimentos foi obtida em
menos de um segundo.
No Quadro 4.1. encontramos informação relativa ao tempo de computação exigido pelo pro-
cedimento de programação dinâmica, sendo possı́vel constatar que este procedimento apenas
constitui uma alternativa viável para problemas nos quais se considere um número máximo de
cerca de 10 trabalhos, dado que os tempos de computação necessários para resolver instâncias
com 15 trabalhos se revelam excessivamente elevados. O impacto dos parâmetros LF e RDD
246 J. Valente, R. Alves / Investigação Operacional, 22 (2002) 235-251
VFO = Valor da Função Objectivo da Heurı́stica; VFO* = Valor Óptimo da Função Objectivo
não será analisado ao nı́vel do método de programação dinâmica, na medida em que os pro-
cedimentos efectuados por este algoritmo, e nomeadamente o número de estados gerados, não
são afectados pelo valor destes parâmetros, sendo que tal justifica também a reduzida diferença
entre os tempos de computação mı́nimo e máximo.
amostra utilizada, muito embora não seja de excluir a possibilidade de, para outras amostras,
ser possı́vel obter um comportamento compatı́vel com o que seria de esperar. A análise do
Quadro 4.2. permite ainda constatar que o inferior desempenho da heurı́stica LookAheadGre-
edy face ao procedimento Greedy se deve essencialmente às instâncias com 15 trabalhos, nas
quais se verifica uma diferença considerável entre a performance dos dois métodos. De facto, a
heurı́stica LookAheadGreedy permite a obtenção de melhores resultados para instâncias com
5 trabalhos, sendo a sua performance similar à do procedimento Greedy quando n = 10. A
heurı́stica EXP-ET apresenta sempre, para qualquer valor do número de trabalhos, os melhores
resultados.
No que diz agora respeito ao impacto do parâmetro RDD, e como podemos igualmente
observar no Quadro 4.3., o aumento da dispersão das datas de entrega contribui, de certa
forma, e com a excepção da regra EDD, para uma deterioração do desempenho associado
aos vários algoritmos. De facto, temos que a performance das heurı́sticas (excluı́ndo a EDD)
tende a ser inferior para instâncias nos quais o RDD assume valores mais elevados, muito
embora esta relação não seja perfeitamente monótona, dado que um aumento do parâmetro
RDD origina, por vezes, uma manutenção ou mesmo uma ligeira melhoria da performance.
No que se refere à heurı́stica EDD, verifica-se exactamente o inverso, na medida em que para
este método os melhores resultados são obtidos para instâncias nas quais o valor do RDD é
elevado. A heurı́stica EXP-ET apresenta sempre, para qualquer valor dos parâmetros LF e
RDD, os melhores resultados.
4.2 Problema P2
No que diz respeito aos tempos de computação, temos que os métodos heurı́sticos se revelam
uma vez mais, e tal como seria novamente de esperar dada a sua reduzida complexidade
temporal, particularmente rápidos, sendo que para todas as instâncias consideradas a solução
gerada por estes procedimentos foi novamente obtida em menos de um segundo.
248 J. Valente, R. Alves / Investigação Operacional, 22 (2002) 235-251
No que se refere à influência do parâmetro RRD, e como podemos observar no Quadro 4.6.,
um aumento da dispersão das datas de disponibilidade contribui sempre para uma melhoria da
performance dos diversos métodos heurı́sticos. Deste modo, temos que uma maior dispersão
das datas de disponibilidade associadas aos vários trabalhos permite efectivamente a obtenção
de melhores resultados, na medida em que a existência de datas de disponibilidade distintas
para os diversos trabalhos se traduz numa redução dos trabalhos disponı́veis para execução
em cada iteração, pelo que a possibilidade de as heurı́sticas efectuarem uma escolha errada
será menor. De facto, em várias instâncias, e particularmente ao nı́vel das primeiras iterações,
por vezes existe apenas um único trabalho disponı́vel, pelo que o trabalho seleccionado nestes
casos pelas heurı́sticas concidirá necessariamente com o escolhido por um procedimento que
gere uma solução óptima.
No que diz agora respeito ao impacto do parâmetro RDD, e como podemos igualmente ob-
servar no Quadro 4.6., o aumento da dispersão das datas de entrega contribui, com a excepção
da regra EDD, para uma visı́vel deterioração da performance das diversas heurı́sticas. No que
se refere à regra EDD, o impacto da variação do parâmetro RDD não é claro, na medida em
que não existe uma tendência bem definida ao nı́vel da evolução da performance deste método
em função do valor da dispersão das datas de entrega.
A heurı́stica EXP-ET apresenta sempre, para qualquer valor do número de trabalhos e dos
parâmetros RRD e RDD, os melhores resultados.
250 J. Valente, R. Alves / Investigação Operacional, 22 (2002) 235-251
5 Conclusão
Neste artigo foram consideradas duas variantes (com e sem datas de disponibilidade distintas)
de uma versão geral do problema early/tardy na qual não se admite a introdução de tempo
morto não forçado. O desempenho de um procedimento optimizante baseado em programação
dinâmica e de diversos métodos heurı́sticos foi analisado, não só ao nı́vel dos valores obti-
dos para a função objectivo, como também no que se refere aos tempos de computação. O
procedimento de programação dinâmica revela-se, na prática, uma alternativa viável apenas
para problemas de dimensão relativamente reduzida (cerca de 10 trabalhos). O impacto de
alguns parâmetros associados ao problema early/tardy sobre a performance dos algoritmos foi
igualmente alvo de estudo, tendo sido confirmado sob condições experimentais mais gerais o
relativamente bom desempenho da regra EXP-ET.
6 Referências
[1] Abdul-Razaq, T. S. e C. N. Potts (1988), “Dynamic Programming State-Space Relaxation for
Single-Machine Scheduling”, Journal of the Operational Research Society 39, pp. 141-152.
[3] Jackson, J. R. (1955), “Scheduling a Production Line to Minimize Maximum Tardiness”, Research
Report 43, Management Science Research Project, University of California, Los Angeles.
[4] Lawler, E. L., J. K. Lenstra, A. H. G. Rinnooy Kan e D. B. Shmoys (1993), “Sequencing and
Scheduling: Algorithms and Complexity”, in Logistics of Production and Inventory, S. C. Graves,
A. H. G. Rinnooy Kan e P. H. Zipkin (eds.), Handbooks in Operations Research and Management
Science, pp. 445-522, Amsterdam: North-Holland.
[5] Lenstra, J. K., A. H. G. Rinnooy Kan e P. Brucker (1977), “Complexity of Machine Scheduling
Problems”, Annals of Discrete Mathematics 1, pp. 343-362.
[6] Ow, P. S. (1984), “Heuristic Knowledge and Search for Scheduling”, PhD. Dissertation, Graduate
School of Industrial Administration, Carnegie-Mellon University.
[7] Ow, P. S. e T. Morton (1989), “The Single Machine Early-Tardy Problem”, Management Science
35, pp. 177-191.
J. Valente, R. Alves / Investigação Operacional, 22 (2002) 235-251 251
[8] Szwarc, W. (1988), “Minimizing Absolute Lateness in Single Machine Scheduling with Different
Due Dates” Working Paper, University of Wisconsin, Milwaukee.
[9] Valente, J. (2000), “Estudo de Alguns Problemas de Sequenciamento em Ambiente Determinı́stico”,
Tese de Mestrado em Economia, Faculdade de Economia, Universidade do Porto.
D. Torres / Investigação Operacional, 22 (2002) 253-263 253
‡
Delfim F. M. Torres
‡
R&D Unit Mathematics and Applications, Department of Mathematics, University of Aveiro
delfim@mat.ua.pt
Abstract
We give conditions under which a function F (t, x, u, ψ0 , ψ) satisfies the relation dFdt =
∂F
∂t + ∂F
∂x · ∂H
∂ψ − ∂F ∂H
∂ψ · ∂x along the Pontryagin extremals (x(·), u(·), ψ 0 , ψ(·)) of an optimal
control problem, where H is the corresponding Hamiltonian. The relation generalizes the
well known fact that the equality dH ∂H
dt = ∂t holds along the extremals of the problem, and
that in the autonomous case H ≡ constant. As applications of the new relation, methods
for obtaining conserved quantities along the Pontryagin extremals and for characterizing
problems possessing given constants of the motion are obtained.
Keywords: dynamic optimization, optimal control, Pontryagin extremals, constants of the motion.
1 Introduction
A dynamic optimization continuous problem poses the question of what is the optimal magni-
tude of the choice variables, at each point of time, in a given interval. To tackle such problems,
three major approaches are available: dynamic programming; the calculus of variations; and
the powerful and insightful optimal control. The calculus of variations is a classical subject,
born in 1696 with the brachistochrone problem, whose field of applicability is broadened with
optimal control theory. Dynamic programming is based on the solution of a partial differential
equation, known as the Hamilton-Jacobi-Bellman equation, in order to compute a value func-
tion. Dynamic programming is well designed to deal with optimization problems in discrete
time. All these techniques are well known in the literature of operations research (see e.g.
[3, 4, 31]), systems theory (see e.g. [13]), economics (see e.g. [8, 19] and [22, Capı́tulo 14])
and management sciences (see e.g [12])). Here, we are concerned with the methods and pro-
cedures of optimal control. This approach allows the effective study of many optimization
D. Torres / Investigação Operacional, 22 (2002) 253-263 254
At the core of optimal control theory is the Pontryagin maximum principle – the celebrated
first order necessary optimality condition – whose solutions are called (Pontryagin) extremals
and which are obtained through a function H called Hamiltonian, akin to the Lagrangian
function used in ordinary calculus optimization problems (see e.g. [21, 27])). For autonomous
problems of optimal control, i.e. when the Hamiltonian H does not depend explicitly on time
t, a basic property of the Pontryagin extremals is the remarkable feature that the correspond-
ing Hamiltonian is constant along the extremals (see e.g. [23, 16]). In classical mechanics
this property corresponds to energy conservation (see e.g. [18, 24]), while in the calculus of
variations it corresponds to the second Erdmann necessary optimality condition (see e.g. [9]).
For problems of optimal control that depend upon time t explicitly (non-autonomous prob-
lems), the property amounts to the fact that the total derivative with respect to time of the
corresponding Hamiltonian equals the partial derivative of the Hamiltonian with respect to
time:
dH ∂H
(t, x(t), u(t), ψ0 , ψ(t)) = (t, x(t), u(t), ψ0 , ψ(t)) (1)
dt ∂t
for almost all t (see e.g. [23, 2, 14]). This corresponds to the DuBois-Reymond necessary
condition of the calculus of variations (see e.g. [7]). Recent applications, in many different
contexts of the calculus of variations and optimal control, show the fundamental nature of the
property (1). It has been used in [11, 1, 25] to establish Lipschitzian regularity of minimizers;
in [10] to establish some existence results; and in [29, 30] to prove some generalizations of first
Noether’s theorem. The techniques used in the proof of the relation are also very useful, and
have been applied in contexts far away from dynamic optimization (see e.g. [15]). In this note
we give conditions under which a function F (t, x, u, ψ0 , ψ) satisfies the equality
dF ∂F ∂F ∂H ∂F ∂H
= + · − · , (2)
dt ∂t ∂x ∂ψ ∂ψ ∂x
almost everywhere, along the Pontryagin extremals. For F = H equality (2) reduces to (1).
As a corollary, we obtain a necessary and sufficient condition for F (t, x, u, ψ 0 , ψ) to be a
constant of the motion. From it, one is able to find constants of the motion that depend
on the control and that are not momentum maps, that is, one can find preserved quantities
F (t, x(t), u(t), ψ0 , ψ(t)) along the Pontryagin extremals (x(·), u(·), ψ0 , ψ(·)) of the problem,
which are not of the form ψ(t) · C (x(t)). This is in contrast with the results obtained in [5],
where the conserved quantities are always of the form ψ(t) · C (x(t)). Our condition provides
also a method for the characterization of optimal control problems with given constants of the
motion. All these possibilities are illustrated with examples.
D. Torres / Investigação Operacional, 22 (2002) 253-263 255
2 Preliminaries
Without loss of generality (see e.g. [2]), we will be considering the optimal control problems
in Lagrange form with fixed initial time a and fixed terminal time b (a < b).
called the performance index, among all the solutions of the vector differential equation
and the control u(·) is a r-vector measurable and bounded function satisfying the control
constraint u(t) ∈ Ω,
The set Ω ⊆ Rr is called the control set. In general, the problem may include some boundary
conditions and state constrains, but they are not relevant for the present study: the results
obtained are independent of those restrictions. We assume the functions L : [a, b]×R n ×Ω → R
and ϕ : [a, b]×Rn ×Ω → Rn to be continuous on [a, b]×Rn ×Ω and to have continuous derivatives
with respect to t and x.
We shall now formulate the celebrated Pontryagin maximum principle [23], which is a first-
order necessary optimality condition. The maximum principle provides a generalization of
the classical calculus of variations first-order necessary optimality conditions and can treat
problems in which upper and lower bounds are imposed on the control variables – a possibility
of considerable interest in operations research (see [12]).
Definition 1. A quadruple (x(·), u(·), ψ0 , ψ(·)) satisfying the Hamiltonian system and the
maximality condition is called a (Pontryagin) extremal.
Remark 1. Different terminology for the function H can be found in the literature. The
Hamiltonian H is sometimes called “unmaximized Hamiltonian”, “pseudo-Hamiltonian” or
“Pontryagin function”.
Remark 2. Transversality conditions may also appear in the Pontryagin maximum principle.
These conditions depend on the specific boundary conditions under consideration. Our meth-
ods do not require the use of such transversality conditions and the results obtained are, as
already mentioned, valid for arbitrary boundary conditions.
Remark 3. The maximality condition is a static optimization problem. The method of solving
the optimal control problem (3)–(4) via the maximum principle consists of finding the solutions
of the Hamiltonian system by the elimination of the control with the aid of the maximality
condition. The required optimal solutions are found among these extremals.
The proof of the following theorem can be found, for example, in [23, 2].
Theorem 2. If (x(·), u(·), ψ0 , ψ(·)) is a Pontryagin extremal, then H (t, x(t), u(t), ψ0 , ψ(t)) is
an absolutely continuous function of t and satisfies the equality (1), where on the left-hand side
we have the total derivative with respect to t, and on the right-hand side the partial derivative
of the Hamiltonian with respect to t.
As a particular case of Theorem 2, when the Hamiltonian does not depend explicitly on t,
that is when the optimal control problem is autonomous – functions L and ϕ do not depend
on t – then the value of the Hamiltonian evaluated along an arbitrary Pontryagin extremal
(x(·), u(·), ψ0 , ψ(·)) of the problem turns out to be constant:
First we introduce the concept of an absolutely continuous function in t uniformly with respect
to s.
Definition 2. Let φ(s, t) be a real valued function defined on [a, b]×[a, b]. The function φ(s, t)
is said to be an absolutely continuous function in t uniformly with respect to s if, given ε > 0,
there exists δ > 0, independent of s, such that for every finite collection of disjoint intervals
(aj , bj ) ⊆ [a, b]
X X
(bj − aj ) ≤ δ ⇒ |φ(s, bj ) − φ(s, aj )| ≤ ε (s ∈ [a, b]) .
j j
The proof of the following two propositions can be found in [14, p. 74].
Proposition 3. Let F (t, x, u, ψ0 , ψ), F : [a, b] × Rn × Ω × R− n
0 × R → R, be continuously
differentiable with respect to t, x, ψ for u fixed, and assume that there exists a function G(·) ∈
L1 ([a, b]; R) such that
° °
°∇(t,x,ψ) F (t, x(t), u(s), ψ0 , ψ(t))° ≤ G(t) (s, t ∈ [a, b]) .
Then φ(s, t) = F (t, x(t), u(s), ψ0 , ψ(t)) is absolutely continuous in t uniformly with respect to
s on [a, b].
Proposition 4. Let φ(s, t), φ : [a, b] × [a, b] → R, be an absolutely continuous function in t
uniformly with respect to s satisfying
in a set dense in [a, b]. Then the function φ(t, t) can be uniquely extended to a function m(t)
absolutely continuous on [a, b].
3 Main Result
a.e. in t ∈ [a, b] along any Pontryagin extremal (x(·), u(·), ψ0 , ψ(·)) of the optimal control
problem, then t → F (t, x(t), u(t), ψ0 , ψ(t)) is absolutely continuous and the equality
dF ∂F ∂F ∂H ∂F ∂H
= + · − · (6)
dt ∂t ∂x ∂ψ ∂ψ ∂x
D. Torres / Investigação Operacional, 22 (2002) 253-263 258
Proof. Our proof is an extension of the standard proof of Theorem 2. Let (x(·), u(·), ψ 0 , ψ(·))
be a Pontryagin extremal of the problem. Setting v = u(s) in (5) we obtain that φ(s, t) =
F (t, x(t), u(s), ψ0 , ψ(t)) satisfies
for t in a set of full measure on [a, b]. Proposition 4 then implies that m(t) = φ(t, t) =
F (t, x(t), u(t), ψ0 , ψ(t)) is an absolutely continuous function on [a, b]. It remains to prove that
∂F ∂F ∂H ∂F ∂H
ṁ(t) = (π(t)) + (π(t)) · (π(t)) − (π(t)) · (π(t)) ,
∂t ∂x ∂ψ ∂ψ ∂x
(ii) F (t, x(t), u(t), ψ0 , ψ(t)) = max F (t, x(t), v, ψ0 , ψ(t)) a.e. in a ≤ t ≤ b;
v∈Ω
Solving the Hamiltonian system by the elimination of the control with the aid of the maximality
condition is typically a difficult task. Therefore, it is worthwhile to look for circumstances
which make the solution easier. This is the case when the extremals don’t change the value of
a given function. Indeed, the existence of such a function, called constant of the motion, may
be used for reducing the dimension of the Hamiltonian system (see e.g. [28, Módulo 5]). In
extreme cases, with a sufficiently large number of (independent) constants of the motion, one
can solve the problem completely.
From Theorem 5, one immediately obtains a necessary and sufficient condition for a function
to be a constant of the motion.
Definition 3. A quantity F (t, x, u, ψ0 , ψ) which is constant along every Pontryagin extremal
(x(·), u(·), ψ0 , ψ(·)) of the problem, is called a constant of the motion.
Corollary 7. Under the conditions of Theorem 5, F (t, x, u, ψ0 , ψ) is a constant of the motion
if and only if
∂F ∂F ∂H ∂F ∂H
+ · − · =0 (8)
∂t ∂x ∂ψ ∂ψ ∂x
holds, almost everywhere, along the Pontryagin extremals of the optimal control problem.
Example 1. (n = 4, r = 2, Ω = R2 ) Let us consider the problem
Z b³ ´
(u1 (t))2 + (u2 (t))2 dt −→ min ,
a
x˙1 (t) = x3 (t)
x˙2 (t) = x4 (t)
³ ´
x˙3 (t) = −x1 (t) (x1 (t))2 + (x2 (t))2 + u1 (t)
³ ´
x˙ (t) = −x (t) (x (t))2 + (x (t))2 + u (t) .
4 2 1 2 2
We claim that
F = −ψ1 x2 + ψ2 x1 − ψ3 x4 + ψ4 x3 (9)
is a constant of the motion for the problem. Direct calculations show that
4 4
∂F X ∂F ∂H X ∂F ∂H
+ − = ψ 4 u1 − ψ 3 u2 . (10)
∂t ∂xi ∂ψi ∂ψi ∂xi
i=1 i=1
D. Torres / Investigação Operacional, 22 (2002) 253-263 260
∂H ∂H
From the maximality condition it follows that ∂u 1
= 0 and ∂u 2
= 0, that is, 2ψ0 u1 + ψ3 = 0
and 2ψ0 u2 + ψ4 = 0. Using these last two identities in (10) one concludes from Corollary 7
that (9) is a constant of the motion.
We shall endeavor here to find a method to synthesize optimal control problems with given
constants of the motion. If a function F is fixed a priori, we can regard equality (8) as a partial
differential equation in the unknown Hamiltonian H. Obviously, if this differential equation
admits a solution, then an optimal control problem can be constructed with the constant of
the motion F . We shall illustrate the general idea in special situations.
∂H
Example 2. The Hamiltonian H is a constant of the motion if and only if ∂t = 0. Condition
is trivially satisfied for autonomous problems.
∂H ∂H ∂H
Example 3. Function ψx+Ht is a constant of the motion if and only if H = ∂x x− ∂ψ ψ− ∂t t.
Condition is satisfied, for example, for problems of the form (0 < a < b)
b
L (tx(t), u(t))
Z
dt −→ min ,
a t
ϕ (tx(t), u(t))
ẋ(t) = .
t2
Example 4. We conclude from Corollary 7 that a necessary and sufficient condition for Hψx
to be a constant of the motion is
∂H ∂H ∂H
ψx + ψH − Hx = 0.
∂t ∂ψ ∂x
A simple problem with constant of the motion Hψx is therefore
Z b
L (u(t)) dt −→ min ,
a
ẋ(t) = ϕ (u(t)) x(t) .
Example 5. The following optimization problem is important in the study of cubic polyno-
mials on Riemannian manifolds (see [6, p. 39] and [26]). Here we consider the particular case
when one has 2-dimensional state and n controls:
Z T³ ´
(u1 (t))2 + · · · + (un (t))2 dt −→ min , (11)
0
(
x˙1 (t) = x2 (t) ,
x˙2 (t) = X1 (x1 (t)) u1 (t) + · · · + Xn (x1 (t)) un (t) .
Functions Xi (·), i = 1, . . . , n, are assumed smooth. The Hamiltonian for the problem is
As far as the problem is autonomous, the Hamiltonian is a constant of the motion. We are
interested in finding a new constant of the motion for the problem. We will look for one of the
form
F = k 1 ψ1 x1 + k 2 ψ2 x2 ,
where k1 and k2 are constants. This is a typical constant of the motion, known in the literature
by momentum map (see [5]). First we note that
∂F ∂F ∂F ∂F ∂F
= 0, = k 1 ψ1 , = k 2 ψ2 , = k 1 x1 , = k 2 x2 ,
∂t ∂x1 ∂x2 ∂ψ1 ∂ψ2
and
∂H ¢ ∂H
= ψ2 X10 (x1 )u1 + · · · + Xn0 (x1 )un ,
¡
= ψ1 ,
∂x1 ∂x2
∂H ∂H
= x2 , = X1 (x1 )u1 + · · · + Xn (x1 )un .
∂ψ1 ∂ψ2
Substituting these quantities into (8) we obtain that
The equality is trivially satisfied if k1 = k2 and Xi0 (x1 )x1 = Xi (x1 ), i = 1, . . . , n. We have
just proved the following proposition.
Acknowledgments
The author is in debt to A. V. Sarychev for the many useful advises, comments and suggestions.
The research was supported by the program PRODEP III 5.3/C/200.009/2000.
References
[1] Ambrosio L., Ascenzi O., Buttazzo G. Lipschitz Regularity for Minimizers of Integral Functionals
with Highly Discontinuous Integrands. J. Math. Anal. Appl. 142, 1989, pp. 301–316.
[2] Berkovitz L. D. Optimal Control Theory. Applied Mathematical Sciences 12, Springer-Verlag, New
York, 1974.
[3] Bertsekas D. P. Dynamic Programming and Optimal Control, Vol. I (2nd ed.). Athena Scientific,
Belmont, Massachusetts, 2000.
[4] Bertsekas D. P. Dynamic Programming and Optimal Control, Vol. II. Athena Scientific, Belmont,
Massachusetts, 1995.
D. Torres / Investigação Operacional, 22 (2002) 253-263 262
[5] Blankenstein G., van der Schaft A. Optimal control and implicit Hamiltonian systems. In:
Isidori A., Lamnabhi-Lagarrigue F., Respondek W. (eds). Nonlinear control in the year 2000,
vol. 1 (Paris). Springer, London. 2001, pp. 185–205.
[6] Camarinha M. A Geometria dos Polinómios Cúbicos em Variedades Riemannianas. Ph.D. thesis,
Departamento de Matemática, Universidade de Coimbra, Coimbra, 1996.
[7] Cesari L. Optimization—Theory and Applications. Springer-Verlag, New York, 1983.
[8] Chiang A. C. Elements of Dynamic Optimization. McGraw-Hill Inc, 1992.
[9] Clarke F. H. Optimization and Nonsmooth Analysis. John Wiley & Sons Inc., New York, 1983.
[10] Clarke F. H. An Indirect Method in the Calculus of Variations. Trans. Amer. Math. Soc. 336,
1993, pp. 655–673.
[11] Clarke F. H., Vinter R. B. Regularity Properties of Solutions to the Basic Problem in the Calculus
of Variations. Trans. Amer. Math. Soc. 289, 1985, pp. 73–98.
[12] Connors M. M., Teichroew D. Optimal Control of Dynamic Operations Research Models. Inter-
national Textbook Company, Scranton, Pennsylvania, 1967.
[13] Elgerd O. I. Control Systems Theory. McGraw-Hill Inc, 1967.
[14] Fattorini H. O. Infinite Dimensional Optimization and Control Theory. Encyclopedia of Mathe-
matics and Its Applications 62, Cambridge University Press, Cambridge, 1999.
[15] Freiling G., Jank G., Sarychev A. Non-blow-up Conditions for Riccati-type Matrix Differential
and Difference Equations. Results Math. 37, 2000, pp. 84–103.
[16] Gamkrelidze R. V. Principles of Optimal Control Theory. Mathematical Concepts and Methods
in Science and Engineering 7, Plenum Press, New York, 1978.
[17] Isaacs R. Differential Games – A Mathematical Theory with Applications to Warfare and Pursuit,
Control and Optimization. Dover Publications Inc., Mineola, New York, 1999.
[18] Lauwerier H. A. Calculus of Variations in Mathematical Physics. Mathematical Centre Tracts 14,
Mathematisch Centrum, Amsterdam, 1966.
[19] Léonard D., Van Long N. Optimal Control Theory and Static Optimization in Economics. Cam-
bridge University Press, Cambridge, 1992.
[20] Pereira F. L. Control Design for Autonomous Vehicles: A Dynamic Optimization Perspective.
European Journal of Control 7, 2001, pp. 178–202.
[21] Pinch E. R. Optimal Control and the Calculus of Variations. Oxford University Press, Oxford,
1995.
[22] Pires C. Cálculo para Economistas. McGraw-Hill de Portugal Lda., 2001.
[23] Pontryagin L. S., Boltyanskii V. G., Gamkrelidze R. V., Mischenko E. F. The Mathematical
Theory of Optimal Processes. John Wiley, New York, 1962.
[24] Rund H. The Hamilton–Jacobi Theory in the Calculus of Variations, Its Role in Mathematics and
Physics. D. Van Nostrand Co., Ltd., London–Toronto, Ont.–New York, 1966.
[25] Sarychev A. V., Torres D. F. M. Lipschitzian Regularity of Minimizers for Optimal Control Prob-
lems with Control-Affine Dynamics. Applied Mathematics and Optimization, 41, 2000, pp. 237–
254.
[26] Silva Leite F., Camarinha M., Crouch P. Elastic Curves as Solutions of Riemannian and Sub-
Riemannian Control Problems. Math. Control Signals Systems 13, 2000, pp. 140–155.
[27] Smith D. R. Variational Methods in Optimization. Dover Publications Inc., Mineola, New York,
1998.
D. Torres / Investigação Operacional, 22 (2002) 253-263 263
Os artigos devem ser escritos em Português ou Inglês. A primeira página deve conter a seguinte
informação:
Artigos: autor(es), tı́tulo, nome e número da revista (livro com indicação dos editores), ano,
páginas.
Para garantir boa qualidade de reprodução, sempre que possı́vel as figuras devem estar sob
formatos vectoriais, devendo ser evitados formatos rasterizados (jpg, bmp, gif, etc).
Revista Investigação Operacional
Volume 22 - Número 2 (Dezembro 2002)
ÍNDICE
L. Grilo, I. Themido
Modelação de vendas de produtos de grande consumo: Uma aplicação ao mercado de transfor-
mados de papel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
J. Valente, R. Alves
Estudo de Algumas Variantes do Problema Early/Tardy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
D. Torres
A Remarkable Property of the Dynamic Optimization Extremals . . . . . . . . . . . . . . . . . . . . . . . . . 253
ISSN: 0874-5161
Investigação Operacional
Volume 22
Número 2
Apdio Dezembro 2002
CESUR - Instituto Superior Técnico
Av. Rovisco Pais - 1049 - 001 LISBOA
Telef. 21 840 74 55 - Fax. 21 840 98 84
http://www.apdio.pt