Você está na página 1de 13

A Fuzzy Set Approach to Activity Scheduling for Product Development Author(s): J. R.

Wang Source: The Journal of the Operational Research Society, Vol. 50, No. 12 (Dec., 1999), pp. 12171228 Published by: Palgrave Macmillan Journals on behalf of the Operational Research Society Stable URL: http://www.jstor.org/stable/3010631 . Accessed: 21/02/2014 10:33
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Palgrave Macmillan Journals and Operational Research Society are collaborating with JSTOR to digitize, preserve and extend access to The Journal of the Operational Research Society.

http://www.jstor.org

This content downloaded from 200.69.103.254 on Fri, 21 Feb 2014 10:33:06 AM All use subject to JSTOR Terms and Conditions

Journal of the Operational Research Society (1999) 50, 1217-1228

Society Ltd. All rights reser-ved.0160-5682/99 $15.00 ( 1999 Oper-ational Resear-ch k/jors http://www.stockton-press.co.u

A fuzzy set approachto activity scheduling for productdevelopment


JR Wang ROC Tcaiwatn, Feng Chia Univer-sity, Industriesneed to effectively manage their productdevelopmentprocesses to reduce the productdevelopmenttime and cost. Due to incomplete design informationat the early stage of productdevelopment,the durationof each activity is difficult to estimate accurately.The objective of this research is to develop a methodology to schedule product developmentprojectshaving imprecisetemporalinformation.The researchproblemis formulatedas a frizzy constraint the satisfactiondegrees of satisfactionproblemand a new method based on possibility theoiy is proposedto determnine fuzzy temporalconstraints.Based on the proposed method, a fuzzy scheduling procedureis developed to constructa schedulewith the least possibility of being late and to maximizethe satisfactiondegreesof all fuizzytemporalconstraints. Moreover,the computationalefficiency of the proposed approachis also discussed. The proposed methodology can produce more satisfactoryschedules in an uncertainproductdevelopmentenvironment. Keywords: project scheduling;fuzzy set theory;productdevelopment

Introduction In recent years, new products and changes in existing products are occurring at an increasing rate, causing the life cycle of a product to decrease. To maintain market shares, industriesneed to effectively manage their product development processes in order to reduce the product development time and bring their products to market as early as possible. A productdevelopmentprocess is considered as a series of several phases: product specification, conceptual design, detail design, prototype, and testing.' Each phase of a productdevelopmentprocess consists of a set of interrelated activities. A modem-day project for productdevelopment,for example to develop an electronic product, usually consists of thousands of activities. It is difficult for an individualproject managerto keep track of all activities. Unlike the manufacturingprocess, it is also difficult to predict accuratelythe durationof an activity at the project initialization stage, due to the high level of design imprecision and engineering changes frequently occurring during the product development.2'3 The imprecise temporal information furthermakes project management tasks more difficult. Poor management of product development projects may result in a critical delay of productsgoing to market and cause great sales losses. The concept of concurrent engineering (CE)4 has emerged to reduce the lead time of product development
Dr. Juite (Ray,) Wang, Associ(te Professor, Dept. of Corresponldenice: Road, SeatEngineering, Feng Chia University, 100 WVenhwa Indiustrial Tailvan407, ROC. wen, Taichiuing,
E-mail: rdwang Wfcu.edu.tw

by coordinating various life cycle considerations concurrently. A product development project based on the CE concept is perfonmedby teams with members from different disciplines. An R&D organizationusually has several projects under development simulataneously. Therefore, activities for various projects have to compete for limited available resources in order not to delay projects. The complex resource dependencies among activities within a productdevelopmentprojectand between differentprojects furthercomplicate the managementtasks. The objective of this researchis to develop a methodology to schedule product development projects having imprecise temporal infoirnation. To effectively allocate scarce resources to activities in an uncertain product development project, it is necessaiy to incorporate the into the problem solving imprecise temporal informnation process. Fuzzy set theory5is used to representthe imprecise temporal informationand our research problem is formulated as a fuzzy constraint satisfaction problem.6 A new method based on possibility theory7is proposed to handle the fuzzy temporal constraints.Since the durationof each activity is imprecise, the actual project finish time is difficult to predict. A fuzzy scheduling procedure based on beam search8is developed to determinea schedule with the least possibility of being late. The start time of each activity is assigned with the maximum satisfactiondegrees of all fuzzy temporalconstraints. In the next section, the literaturerelated to this research is presented. The activity scheduling problem for product development is then modelled with the fuzzy set theory. A new method is introduced to measure the satisfaction

This content downloaded from 200.69.103.254 on Fri, 21 Feb 2014 10:33:06 AM All use subject to JSTOR Terms and Conditions

12 Vol. 50,No, Research Society oftheOperational 1218 Journal

degree of a fuzzy temporalconstraint,and a fuzzy scheduling procedureis developed to producethe most satisfactory schedules. An illustrativeexample and some computational results are also presented. Finally, we provide a summary of this researcheffort. Literature review The project evaluation and review technique (PERT) and critical path method (CPM) have long been established in industryas tools for planningand managingprojects.9One of their shortfallsis their rigid analysis and specificationof a project structure.Many studies have been conducted on extending the classical concept of PERT to deal with more complex problems. GeneralActivity Network (GAN)9 is a stochastic project network with a more flexible network structure in the form of stochastic event nodes. Each activity (on arc) is assigned with an execution probability in addition to its stochastic duration. Several researchers have developed scheduling approaches on GAN with a 10 11A GAN model represents limited numberof resources. the stochastic temporalinformationwith probabilitydistributions. Therefore, the computationrequiredfor temporal analysis or resource allocation decisions is based on probability theory. The objective is to minimize the expected projectmakespan,expected project lateness, etc. However, scheduling of activities on GAN with limited resources is computationally too expensive and theoretically too complex." Moreover, a product development project is usually unique in nature and is often described as being 'open-ended' or 'ill-defined.' Therefore, it is difficult to collect enough data to obtain the distributionof the random variable for each activity duration.'2 The activity networkwith GeneralizedPrecedenceRelations (GPR)13generalizes the classical PERT with strict precedence relations between activities. Four types of precedence relations that are present in many practical situationsare introducedto provide more general representation. Several researchers have extended the resourceconstrainedproject scheduling problem with GPR to mini14 However, all temporalparamize the project makespan. meters in an activity network with GPR are precise numbers. Therefore, the satisfaction of a temporal constraintis either true or false. In contrastto the scheduling models with GAN or GPR, this researchrepresentsthe imprecise temporalinformation with fuzzy set theory. The computation required for temporal analysis or resource allocation decisions are based on fuzzy arithmetic and fuzzy ranking methods5 that are less complex than probabilitytheory. In addition, as the temporalparametersare representedwith fuzzy sets, the satisfactionof a fuzzy temporal constraintis not about truthor falsehood, but is a degree within [0, 1].6 ect may be The schedulingof a productdevelopmentproj scheduling project viewed as a resource-constrained

problem, which has been investigated extensively in the literature.'5'16 Since the resource constrainedproject scheduling problem is NP-complete, optimizationapproaches'6 are not suitablefor solving practical-sizedproblems.Therefore, heuristic approaches'7-'9 have been developed to construct a schedule effective for practical use. With the advent of concurrent engineering, Belhe and Kusiak20 considered the flow of information in the design process as a pull system like a just-in-time manufacturingsystem and developed a dynamic schedulingheuristic to minimize weighted lateness. Little research has been performed on extending the resource-constrained project scheduling problem with fuzzy set theory. The CPM and PERT networks have been extended with fuzzy durationfor each activity.12 21 However, the resource capacity constraints are ignored. Hapke and Slowinski22 applied twelve dispatching rules with fuzzy sets to generate a set of schedules and the best schedule with the minimum fuzzy makespan is selected. However, it is known that the dispatching rules only consider local aspects of the solution space. The computational efficiency of their approach is not investigated. In addition, their approach only determines the fuzzy start time for each activity. The assignmentof a crisp starttime for each activity to execute a project is not discussed. Modeling the scheduling problem with fuzzy set theory
Representation of imprecise temporal information with fuzzy numbers

In many practical situations, the temporalparametersof a product development project, such as the duration of particularactivities or the project deadline, often cannot be given precisely at the project initialization stage due to incomplete design information.Unless a product development project being scheduled is quite similar to previous projects, previous experience is of limited relevance. may not be suitable Therefore, statistical approaches9-11 in this situation. However, one may estimate the activity durationfrom the activity that performsa similar function in previous projects. Information about what duration is more plausible than another is often available and can be predicted by experienced project managers. Therefore, fuzzy set theory may provide an alternativeand convenient frameworkfor handling this issue. 12,23 be the membership Let t be an activity durationand [tL(x) that the activity t the possibility function of to estimate to a time scale. x for all belonging durationt is precisely x, known with xo, then is precisely If the activity duration = la). If the 0 = (Figure for x all ,tt(xo) 1, and :Axo, [lt(x) t two is of between the value is and known duration extreme time points xo and xl, then pft(x)= 1 for all x E [xo,xl]; otherwise Mt-(x)= 0 (Figure Ib). If the duration is fuzzily known, the more or less possible values of t are

This content downloaded from 200.69.103.254 on Fri, 21 Feb 2014 10:33:06 AM All use subject to JSTOR Terms and Conditions

1219 scheduling toactivity setapproach Wang-Fuzzy

restrictedby fut(x)that takes the values between [0, 1]. The values belonging to the core of t (i.e., {xlt-(x) = 1} are consideredas the most possible values for t and the values = O} are the least outside the support of t (i.e., {xl,ut(x) possible ones. The values outside the core and inside the supportof t are in between (Figure Ic). For computational efficiency, the trapezoidalfuzzy number t = (a, b, c, d) is used to representthe imprecise activity duration.A duration between b and c is the most possible durationand a durationless than a and larger than d is the least possible one (Figure Id). The ready time and deadline of a project are often determined by the preference of project managers. For example, a project manager may expect that a project should be completed between d, and d2, but no later than d3, because it may delay the product entering the market. Fuzzy sets can also be used to representthe preferencesof project managers regarding different values of project deadline. Let d be a project deadline and [1a(x) be the membershipfunction to representthe preferenceof assigning x to the project deadline, for all x belonging to a time scale. The values belonging to the core of d are considered as the most preferred values and the values outside the support of d are the least preferable ones. The values outside the core and inside the supportof d are in between. For example, the preferredproject deadline can be represented as a trapezoidal fuzzy number d = (dl, dl, d2, d3) (Figure 2). Similarly, the preferredready time of a project can also be represented as a trapezoidal fuzzy number b = (bl, b2, b3, b3) (Figure 2).

Modeling the activity schedulingproblemfor product development The activity scheduling problem with imprecise temporal parametersis describedas follows. A productdevelopment project P has a ready time b and a deadline d between which all its activities ai(1 < i <?n) have to be performed. There may exist somneprecedence relationships between activities. The preceding activity produces information required by the succeeding activity. Instead of material flow through the manufacturingprocess, the type of flow between activities in a product development project is mostly information. Besides, in order to be successfully executed, each activity ai has specific duration ti and its execution requires the exclusive use of a number of resources (e.g., engineers, laboratory, etc.) defined by a vector Ni = (nil, ni2, .. ., niq), whose elements determine resources for the project is defined by a vector n where each element indicates the R = (ml, M2,... mq), availability of resources rl, r2,. . ., rq. This research considers the researchproblem as a fuzzy constraint satisfaction problem.6 A constraint satisfaction is defined by means of a set V = {vI, v2, ... ., I problem24 of variables,a domainDi specifying for each variablevi'the values to which it may be assigned, and a set C = {c1,C2, ... I CJ of constraints on variables, where Ci: D1 x ... x DP - {true,false}, for i = 1, . . ., w. The problem is to find an assignment z E D1 x ... x DP, such thatV1< < 1ci(z) = true. As the domain of each variable becomes fuzzy, the satisfactionof a constraintis a matterof degree within [0, 1]. The fuzzy constraint satisfaction formulation of our researchproblem is defined as follows: Fuzzy variable: sti: the starttime of activity ai(1 ? i ? n) Fuzzy temporal constraints:
i

the usage of resources rl, r2, . . . ,r,. The availability of

/k~~~~~~~~p t 8ti iA ~ ~~
o 0

'2

xo (a)

Time 0

Xo XI (b)
. .
-

- Time

' . O......................

For all activities,the starttime sti * Ready time constraints: of activityai should be greaterthan or equal to its earliest starttime esti: d
>

bc

Time ?

b
(d)

Tinie

sti > esti.(1

The earliest starttime of activity ai is defined:

(c)

of fuzzy activity duration. Figure 1 Interpretation

t
esti -<

1 tjG , Max{jsti, estj b,

if ai is precededby aJ if ai has no predecessors (2)

Ready time

Deadline

b1 b2 b3 d, d2 d3 Time readytimeanddeadline of theproject Figure 2 Representation

* Deadline constraints:For all activities, the finish time of activity ai should be less than or equal to its latest finish time ift1: lfti. sti (D ii A< (3)

withfuzzynumbers.

This content downloaded from 200.69.103.254 on Fri, 21 Feb 2014 10:33:06 AM All use subject to JSTOR Terms and Conditions

Society Vol. 50,No. 12 1220 Journal oftheOperational Research

The latest finish time of activity ai is defined:


J

Min{ifti, Ifij E t1},

if ai precededsa1 if ai has no successors (4)

id,

then the deadline constraintis partiallysatisfied;otherwise, the constraintis unsatisfied. Due to the actual finish time not really being known in advance, the satisfaction of the deadline constraintis a degree within [0, 1]. Measure of the satisfaction degree of a fuzzy constraint Comparisonindex based on possibility theory

* Precedenceconstraints If activity ai precedes a1 in the partialorder,then: stj ? tj < stj. (5)

It means that the finish time of activity ai should be less than or equal to the starttime of activity aj. * Resource capacity constraints For any time t, let J = {ailsti < t < sti D ti}. Then for all rq}, such that: time t, and all resourcesrk E {r,, r2, .
Eaic-J nik,<- Mk, (6)

Dubois and Prade27 developed a set of four rankingindices within the framework of possibility theory.7 The four indices, based on possibility measure fl and necessity measureN, can be used for comparingtwo fuzzy numbers. Definition: Possibility measure Given possibility distribution F, the possibility of realizing fuzzy event A is: min(,ui(x), ,uA(x)) flp(A) = sup N (7)

where mkis the total availableresourceunits for resourcerk. The objective is to determinea starttime for each activity such that not only all the fuzzy temporal constraints are satisfied, but also the project finish time should meet the project deadline. For details of fuzzy arithmetic (e.g., D and E), please refer to Klir and Yuan.5 As the temporal parametersare fuzzy, the constraints become flexible and their satisfactions depending on the choice of starttimes are degrees within the range [0, 1]. In addition,it may not be possible to find a schedule that fully satisfies all the constraintswith a degree of one. Thus, it is necessary to develop a method that is able to determinean acceptable schedule that 'partially' satisfies all the constraintsin an uncertainproduct development environment. This concept is similar to the fuzzy linear programming25 and goal programming.26 For example, assume that a fuzzy start time sti, fuzzy durationti, and fuzzy latest finish time Ifti for activity ai are (4, 6, 6, 6), (4, 5, 6, 7), and (10, 10, 10, 12), respectively. The fuzzy finish time fti is calculated as (8, 11, 12, 13) which is partiallyoverlappedwith the latest finish time if ti (Figure 3). The possible crisp activity finish time is within the range [8, 13]. If the activity finish time is less than 10, then the deadline constraint 'sti D ti ? If ti' is fully satisfied. However, if the activity finish time is within [10, 12],

Definition: Necessity measure Given possibility distributionF, the necessity of realizing fuzzy event A is: min(1 - II(x), ,i2(x)) Np(A) = inf xA (8)

Given a fuzzy number A, one can derive four sets of numbershaving A as a fuzzy bound: (1) The set of numberspossibly greaterthan or equal to A is [A, +oo) with membershipfunction (Figure 4a):
IflA,oc+Y) = Sup HA(X)
X ,,

(9)

(2) The set of numbers necessarily greater than A is ]A, +oo) with membershipfunction (Figure 4a):
I]A +?) (Y)

inf (1 - i

(x))

( 0)

(3) The set of numberspossibly smaller than or equal to A is (-oo, A] with membershipfunction (Figure 4b):
'(_c,A](Y)

= sup [A (X) x v

(11)

(4) The set of numbers necessarily smaller than A is (-oo, A[ with membershipfunction (Figure 4b): =inf (I - [, (x)) yu(_O A vy) x I
ii Alx

(12)

14
I

4A [A,?+c),

.0Latest
finish tim .-Finish time

O_
0

1.0

x,y

0 Figure 3

10 11 12

13

Time

(a)

(b)

of the satisfactionof a fuzzy deadline Examnple constraint.

Figure4 Fuzzysets:(a) [A,+oo) and]A,+oo), (b) (-oo, A] and(-oo, A[.

This content downloaded from 200.69.103.254 on Fri, 21 Feb 2014 10:33:06 AM All use subject to JSTOR Terms and Conditions

setapproach toactivity scheduling 1221 JWang-Fuzzy

Given two fuzzy numbers M and N, the four indices are defined and used to assess the possible relationships between fuzzy numbersM and N: PG(M, N) =fl([N,
u;( > ,,

pi" (u)

inf I:-,jk(v)

I.0
. i )NSG(m,n)

+oo)) (13)
0.0 J1

sup minC'(,uM), MM(v))

PSG(in, W) 0 1 2 3~ 4 5 6 7 8 9 10 u, v

PSG(M, N) =f1lM(]N, +oo)) = sup inf min(y,u(u), 1 -,u(v)) NG(M, N)


= =

n) of two indices PSG(im-, Figure 6 Geometricalinterpretation

(14)

andNSG(im, n). Satisfactiondegree of a fuzzy constraint

N&([N, +oo)) inf sup max(1 - [L,(u), ,uk(v))


It v;vI, < It

(15)

NSG(M, N) = Ni(]N, +oo)) = 1 - sup min(,Y(u),


it '< v

According to the four indices defined above, the satisfaction degree of a fuzzy temporal constraint can be determined. Definition a and b. The satisfacGiven two fuzzy temporalparameters tion degree of the fuzzy temporal constraint 'a > b' is defined as the weighted sum of PG(a, b) and NG(a, b):
g(a > b) = , x PG(a,b) + (1-,f) x NG(a, b), (17)

Pk(V))

(16)

The four indices take values in [0, 1]. PG(M, N) (respectively, PSG(M, N)) implies that the grade of possibility of proposition 'M is greaterthan or equal to N' (respectively, 'M is strictly greater than N'). NG(M, N) (respectively, NSG(M, N)) implies that the grade of necessity of proposition 'M is greaterthan or equal to N' (respectively, 'M is strictly greaterthan N'). The four ranking indices can be interpretedgeometriConsidertwo fuzzy numbers cally for betterunderstanding. as shown in Figures 5 and h=(3,6,6,9) m=(2,4,4,7) and 6. PG(mi,h) is used to compare the left leg (the worst part) of n with the right leg (the best part) of mi.The index will be high, when the left leg of h is smallerthan or equal to the right leg of uin.NG(mi,h) compares the left legs of both miand h. The index will be larger,if the left leg of h is smaller than the left leg of mi. One can interpret that PG(mi,h) is the upper bound of the constraint 'M > N' to be satisfiedandNG(mi, h) is its lower bound.PSG(mi, h) and NSG(m,h) can be interpretedsimilarly. According to Eqs. (13)(16), four indices between miand h are calculated as follows: PG(mi,h) = 0.67 NG(m, h) = 0.20 PSG(i, h) = 0.17 NSG(mii, h) = 0.00

where ,Bis the preferenceratio, 0 < ,B< 1. PG(a, b) and NG(a, b) use the right leg (optimistic) and left leg (pessimistic) of a to comparewith the left leg of b. We may interpret PG(a, b) as the possibility of 'a > b' and NG(a, b) as the necessity of 'a > b.' Thereforethe satisfaction degree that 'a is greaterthan or equal to b' is between NG(a, b) and PG(a, b). A parameter,B, called preference ratio, is defined to incorporate the attitude of a project manager towards pessimism versus optimism. If the attitude of a project manager is optimistic, then ,B should be greaterthan 0.5. On the other hand, /Bshould be less than 0.5, if his/her attitudeis toward pessimism. According to Eq. (17), the following rule is used to identify the relationshipbetween two fuzzy temporalparameters a and b: If

g(a b) > g(b )a)


b) < g(b a)

then aib
then b >, a

else if g(a

else a = b Similarly, the satisfaction degree of the fuzzy temporal constraint'a > b' can also be defined. Definition Given two fuzzy temporalparameters a and b. The satisfaction degree of the fuzzy temporal constraint "a > b" is defined as the weighted sum of PSG(a, b) and NSG(a, b): i g(a > b) x PSG(a, (-r b) + x NSG(a, (18) b),

'UAL
1.0

U'j (U) .......

su

p-r

(v,.

0.67
0.2

,/

PG(m,n) ~~NGQMi,i)

0 1 2 3 4 5 6 7 8 9 10 u, v of twoindices Figure5 Geometrical interpretation PG(mi, h) and NG(inm, n)

where ,B is the preference ratio, 0 < ,B< 1.

This content downloaded from 200.69.103.254 on Fri, 21 Feb 2014 10:33:06 AM All use subject to JSTOR Terms and Conditions

1222 Journal oftheOperational Vol. 12 Research Society 50,No.

For example, consider the same example of two fuzzy numbersimand h shown in Figures5 and 6. The satisfaction degrees of fuzzy temporalconstraints'im> n' and 'im> n' are calculated as follows (assume /B= 0.5): g(m ) n) = 0.5 x 0.67 + 0.5 x 0.2 = 0.435 g(m > h) = 0.5 x 0.17 + 0.5 x 0.0 = 0.085

Schedule risk Due to the imprecise activity duration,the actual project finish time for a productdevelopmentproject is difficult to predict. Therefore, this paper intends to find a schedule with the least possibility of being late, corresponding to the fuzzy project deadline defined by project managers. A schedule risk is defined to determine the possibility of a project schedule to be late in an uncertainproduct development environment.
Definition (Schedule risk)

activities for delay to ensure the feasibility with respect to temporal and resource capacity constraints at any time point during the schedule construction. The fuzzy scheduling procedureis described as follows. Fuzzy SchedulingProcedure: Denote: E: set of eligible activities with all their predecessors completed. S: set of activities that have been scheduled. V: set of events to characterize the completion of activities. P: set of activities in process. R: a vector to indicate resource availability. D: the minimumnumberof delaying activities to resolve resource conflicts. DS: delaying set that contains alternative groups D of delaying activities. ino,, the currenttime. AR(E): aggregated resource requirementsfor activities in E. Step 1: (Initialization) Set E <- 0,P <-0, V ,S b. <-and-t0an n , Step 2: (Construction of the temporal window for each activity) Calculateearliest starttime esti and latest finish time Ift for activity ai, 1 < i < n. Step 3: (Constructionof the set of eligible activities) Set E <- {ala Eactivities with their predecessors completed. If E = q and V = q, then stop. Step 4: (Test the resourceconflict) If AR(E) > R, then go to step 5; else go to step 6. Step 5: (Conflictresolution) set DS = {D E El the (1) Find delaying < for all which contains alternik mk rk}, EaiEE-D native groups of activities that can be delayed to resolve resource conflicts. (2) Select a group D E DS with the minimal total tardinessf(D) for delay, where: f (D) =
E
a,ED

Given a scheduleS, the risk of S to be late is defined as the satisfaction degree of the fuzzy constraint 'project finish timeft is greaterthan project deadline d:

Risk(S)= gt

> d)
x NSG(ft, d), (19)

= /Bx PSG(ft, d) + (1 -,)

where /3is the preferenceratio, 0 < ,B< 1. A schedule is late if its project finish time passes the project deadline. When both project finish time and deadline are fuzzy, the schedule being late is not either true or false, but is a degree within [0, 1]. Therefore, a schedule risk is defined in terms of the satisfaction degree of fuzzy
temporal constraint 'ft > d.'

Fuzzy scheduling procedure

The resource-constrainedproject scheduling problem is NP-hard and it seems unpromisingto use any exact algorithms to solve practice-size problems.17In this paper, a fuzzy schedulingprocedurewith beam search8is developed to producea schedule (fuzzy starttime assignmentfor each activity) with the least possibility of being late in reasonable computing time. Then, the crisp start time of each activity is determinedbased on possibility theory to maximize the satisfaction degrees of all fuzzy temporal constraints. The proposed fuzzy scheduling procedure generates a schedule by proceeding forward chronologically. At any time point, the eligible set of activities that can be scheduled is determined according to their precedence constraints. Due to the limited resource capacities, these activities may contend for the same resources. A minimal total tardiness heuristic is used to choose the conflicting

Max{t E)ltj, 0.

(20)

(3) Find the earliest finishing activity ai from the set Y =PUE-D. Set E < E-D. Add new arcs (ai, aj) to the activity network, where aj E D. Then, update the earliest start times of aj and all its successors. (4) Go to step 6. Step 6: (Resource allocation) For activity ai E E, do 10 0 tii, S <- S U {a}l, P <(1) sti *<- fti1o ti PU{ai}, R -R-MN;

This content downloaded from 200.69.103.254 on Fri, 21 Feb 2014 10:33:06 AM All use subject to JSTOR Terms and Conditions

1223 toactivity scheduling setapproach JWang-Fuzzy

(2) Update the earliest starttimes of affected succeeding activities; (3) V <- VU {(ai,fti)}. Step 7: (Advance to next time point) (1) Remove (a, t) with the smallest t from V; - Max{t,0,t,, t}; (2) t10},<N (3) SetR <- R+Ni,P<- P-{a}; (4) Go to Step 3. Step 1 initializes the procedureand sets currenttime t,,O,,, to the project ready time b. Step 2 calculates the temporal window for each activity that is determinedby the forward propagationof project ready time and backwardpropagation of project deadline according to the precedence relations between activities. Step 3 constructs the set E of eligible activities whose immediate predecessors Step 4 examines the rehave been completed at t source capacity constraints.If there is a resource conflict, then go to step 5 to identify the group D of delaying activities with the minimal total tardinessand next remove D from E for conflict resolution. In step 6, the activities in E are scheduled and the completion events corresponding to these activities are inserted into V. Step 7 upwith the completion time of the earliest dates t0n,W finishing activity and next updatesthe resource availability as well as the set of activities in process. Then, go to step 3. The same process continues until E and V become empty. At every scheduling decision time point, if the set of eligible activities is in conflict, then go to step 5 to determine the delaying activities with the minimal total tardinessto resolve conflicts. Step 5 requiresthe generation of all the alternativegroups of delaying activities and the worst case time complexity is of the order of + JEllog2 JEl),where JElis the numberof activities OC(EI2 in E. In the worst case, the number of times that step 5 is repeated is at most equal to the number of activities in a given problem. Therefore, the overall worst case time complexity of the fuzzy scheduling procedure is O(nlEl2+ njEl log2 JEl),where n is the numberof activities for a given problem. To obtain better solution quality, the fuzzy scheduling procedureis extended with beam search.8The beam search is based on the concept of the enumerative branch and bound methods that have been widely used to obtain optimal solutions for the scheduling problem. However, it is usually time-consuming,because the nodes at any level of the searchtree are all expandedand the size of the search tree grows exponentially. On the contrary, beam search selects only the most promising nodes at each level for furtherexpansion. The remaining nodes at that level are of nodes retained is discarded permanently. The numnber The evaluation is search. the w of width called the beam function the valuation on f, which is used to made based predict the 'goodness' of a node. The advantage of beam

search is that the search tree grows constantly,because the constant numberof nodes is chosen for expansion. This paper uses the total tardinessof delaying activities (Eq. (20)) as the valuation function for selecting the promising nodes for expansion. The computationalexperiment of fuzzy beam search procedure will be dis-cussed below. Assignmentof crisp start time The fuzzy scheduling procedure constructs a sequence of activities and assigns a fuzzy start time to each activity. Next, the crisp start time for each activity should be determined for executing a project. The objective is to maximize the satisfaction degrees of all fuzzy temporal constraints. According to Eqs. (1) and (3), the fuzzy starttime of an activity should be greaterthan or equal to its fuzzy earliest starttime and less than or equal to its fuzzy latest starttime; i.e., esti ? sti < 6sti. (21)

The fuzzy latest starttime lst{of activity ai is determinedby its fuzzy latest finish time 1ftiminus its durationti: lstj = Ifti E)ti (22)

In the fuzzy scheduling procedure discussed above, the assignmentof a fuzzy starttime for each activity is always greaterthan or equal to its fuzzy earliest starttime. Therefore, the assignment of a crisp start time si for activity ai only needs to consider selecting a value within its fuzzy start time and fuzzy latest start time with a membership grade as high as possible. Thus, si should be within the interval determinedas follows: 0 where a = PG(1sti,sti). (23) n Sti, Si E (-oo, 1Rti] of In other words, si is definedby the intersectionof oc-cuts can One to a is equal PG(lsti, sti). (-oc, Isti] and sti, where interpret oc as the satisfaction degree of the start time assignment for an activity. For example, considerthe case in Figure 7 that the fuzzy starttime Ati(= (a, b, c, d)) is less than or equal to the fuzzy latest starttime lsti(= (el, el, e2, e3)). The range of si with the maximum satisfaction degree is determinedas follows: Y.= PG(lsti, sti) = 1.0 Si E [b, c] n (-oo, e2j = [b, c] (24) (25)

It is said that si can be given within the interval [b, c] with the satisfaction degree equal to one. One can set si to b, if (s)he prefers to startthe activity as early as possible. Consider the second case in Figure 7 that sti is greater than or equal to 5ti(= (e', e', e, e'3)). To find the maximum satisfaction degree for both sti and lst, si is equal to the x-axis of the intersectionpoint between the right leg of

This content downloaded from 200.69.103.254 on Fri, 21 Feb 2014 10:33:06 AM All use subject to JSTOR Terms and Conditions

1224 Journal oftheOperational 12 Research Vol. Society 50,No.

(e ,

e',

e, e)

(a,bcd)

(e ,e ,e,,e3)

Table1 Activityinformation corresponding to theprecedence graph in Figure8


ActivityNo. Duration (10, 12, 12, 14) (3, 4, 4, 7) (9, 10, 10, 14) ResourcesRequired (1, 1, 0) (1, 2, 2) (1, 0, 1)

est'

/2

st,

1s

s bc

Time

Figure7 Assignment of crispstart tome si for activity ai.


lstl and the left leg of st1. Accordingly, s1 is determinedas

follows:
c' = PG(lst, sti) = o S1=(-o,S*]n[S*,d] =s

(26)
(27)

a, a2 a3 a4 a5 a6 a7 a8 a9 a1o a,, a12

(9, 10, 10, 14)


(13, 15, 15, 20) (23, 25, 25, 30) (6, 7, 7, 9) (5, 5, 5, 9) (4, 6, 6, 11) (18, 20, 20, 25) (12, 15, 16, 20) (25, 31, 42, 47)

(2, 1, 3)
(3, 1, 0) (1, 2, 3)

(3, 2, 3)
(1, 1, (2, 3, (1, 0, (0, 2, (2, 3, 1) 1) 1) 1) 1)

In summary,this paperdevelops a unified frameworkbased on possibility theory to determinethe satisfaction degrees of temporal constraintsand to assign a crisp start time to each activity with the maximum satisfactiondegrees of all fuzzy temporalconstraints. Illustrated example In this section, an example is used to illustratethe approach developed. A product development project consists of twelve activities, which is representedwith the precedence graph shown in Figure 8. Three types of resources rl, r2, and r3 are required to perform the project and the availability of resources is R = (3, 3, 3). Assume that the ready time and deadline of the project are (0, 3, 3, 3) and (110, 120, 120, 130) respectively. The corresponding activity informationis listed in Table 1. The beam width is set to one and the preferenceratio ,Bis set to 0.5 for the average case in this example. The iterationsperformedby the developed fuzzy scheduling procedure are demonstrated in Table 2. During iterations 1 and 2, no resource conflicts are observed. In iteration 3, activity a4 needs three units of resource r3 which have been used by activities a2 and a3. Therefore, activity a4 has to be delayed. After a2 and a3 are completed, the identified eligible set E in iteration 4 includes a4, a5, and a6. A resource conflict is identified, because the resource availability (R = (3, 3, 3)) does not meet the

aggregated

resource

requirements

(AR(E) = (6, 4, 6)).

Thus, step 5 is used to determine the delaying activities. Table 3 shows the delaying set DS, which contains the alternative groups of delaying activities to resolve the resource conflicts. We can identify that DI listed in Table 3 has the minimal total tardiness (0, 0, 0, 28). Therefore, activities a4 and a5, are selected for delay and then a6 can be processed. The similarprocess iteratesuntil all activities are completed. A frizzy schedule is generatedand listed in Table 4. The activities are scheduled in the following order: (a,,a2, a3,a6, a5, a4, a9, a8, a7, alo, a,I,a12). The obtained
fuzzy project makespan is ft = (122, 144, 155, 200) which

Table2 Theiterations performed by the fuzzyscheduling


procedure,as R = (3, 3, 3) Iteration Eligible Conflicting Delayed Activities in Activities no. set resources activities process completed

1
2 3 4 5

6
7 8
9

a1 a2, a3 a4 a4, a5, a6 a4, a5, a9 a4,a8,a9 a7,a8,a9 a7, a8


a7, a10

None None r3 rl, r2, r3 rl, r2, r3 rI,r2,r3 rl, r2, r3

ri
r,

None None a4 a4, a5 a4, a9 a8,a9 a7,a8 a7


a1o

a1 a2, a3 a2 a6 a5 a4 a9 a8
a7

a1 a2 a3 a6 a5 a4 a9 a8
a7

10

11
12

a1o,all

None

None

None
a12

None
None

None
None

a1o,all a1o
a12

a,1 a1o
a12

Table3 The identified 4 set in iteration delaying


Delaying set (DS) Groupof delaying activities {a4,a5} {a4,a6} "a "a6) Total tardiness (0, 0, 0, 28) (0, 0, 0, 49) (0, 0, 0, 55)

9e

DI
D2

Figure8 Example of precedence graph withtwelveactivities.

D3-

This content downloaded from 200.69.103.254 on Fri, 21 Feb 2014 10:33:06 AM All use subject to JSTOR Terms and Conditions

J Wang-Fuzzy setapproach toactivity scheduling 1225

Table 4 The schedule generatedfor the precedencegraph in Figure 8, as R = (3, 3, 3) Activity (ai, i = 1 - 12) a1 a2 a3 a4 a5 a6 a7 a8 a9 alo all
a12

Fuzzy start time (sti) (0,3,3,3) (10,15,15,17) (10,15,15,17) (55,65,65,81) (42,50,50,61) (19,25,25,31) (73,86,86,114) (68,81,81,105) (64,75,75,95) (79,93,93,123) (79,93,93,123) (97,113,113,153)

Fuzzy latest start time (Isti) (-1, 29, 41, 65) (13,41,53,75) (0,30,42,65) (20,45,57,78) (14,40,52,74) (-7, 27, 38, 60) (34,55,67,87) (24,53,64,82) (23,52,63,83) (33,58,69,87) (43,62,74,93) (63,78,89,105) (122,144,155,200)

Fuzzyfinish time (fti) (10,15,15,17) (13,19,19,24) (19,25,25,31) (64,75,75,95) (55,65,65,81) (42,50,50,61) (79,93,93,123) (73,86,86,114) (68,81,81,105) (97,113,113,153) (91,108,109,143) (122,144,155,200) Projectrisk

Start time (Si) 3 15 15 63 50 25 79 74 71 83 85 101 0.875

Sat. degree 1.00 1.00 1.00 0.74 1.00 1.00 0.42 0.45 0.61 0.25 0.42 0.25

Fuzzy Makespan

overlaps' with the planned project due date d = (110, 120, 120, 130). According to Eq. (19), the project risk is calculated: Risk(P) = 0.5 x PSG(ft, d) + (1 - 0.5) x NSG(ft,d) = 0.5 x 1.00 + 0.5 x 0.75=0.875 The last two columns in Table 4 display the crisp starttime and the correspondingsatisfactiondegree for each activity, accordingto Eq. (23). For example, the crisp starttime for activity a4 is calculated (see Figure 9): a = PG(lst4, st4) = 0.74.

R = (4, 3, 4) to the project. Otherwise, (s)he may increase the resource availability again. In summary,the developed methodology allows projectmanagersto evaluate decisions for resourceallocation in an uncertainproductdevelopment environment. Computational experiments Primary computationalexperiments are conducted to test the proposedfuzzy schedulingprocedureon a set of benchmark problems to evaluate its efficiency. The set of problems contains three problem types including 30 activities, 60 activities and 90 activities. Each problem type contains 86 problems. The benchmark problems are randomly selected from the project scheduling problem library(PSLIB).8 Due to the fact that all the benchmarkproblems created from PSLIB are deterministic, we randomly fuzzify its data. For each crisp duration ti, four numbers that are used to define the fuzzified activity duration
ti = (ai, bi, ci, di) are randomly generated from the interval (T x ti, ti + (1 - y) x ti) and ranked in ascending order. In Table 5 The iterationsperformedby the fuzzy scheduling algorithm, as R = (4, 3, 3) Iteration Eligible Conflicting Delayed Activities in Activities no. set resources activities process completed

n 0?74st4= 62.42 S4 = 074(-oo, 1st4]

63.

In orderto reduce the risk of the projectbeing late, project managers may consider including more resources in the project. It is observed from Table 2 that resource r1 is in conflict from iterations 4 to 9. Therefore assigning an additional one unit of r1 may obtain the most benefit. Table 5 shows iterations performed when the resource availabilityis changedto R = (4, 3, 3). Table 6 summarizes the obtained schedule and the fuzzy project makespan is reduced to (109, 129, 140, 180) with the risk level 0.65. One can observe that resource r3 becomes the bottleneck resource,because it is in conflict from iterations3 to 9. If r3 is also increased one unit, then the shortermakespan (92, 113, 125, 155) with risk level 0.438 can be obtained(Table 7). If one accepts this risk level, then (s)he may assign

1 2 3
4
5

a, a2, a3 a4
a4, a5, a6 a4, a8 a4,a8,a9

None

None

None
r3
r1, r2, 3 r2, r3 r1,r2,r3 t-2, 3 I12,
r3 r3

None
a4
a4 a4, a8 a8,a9

a, a., a3 a3
a5, a6
a6 a4

a, a2 a3
a5 a6 a4

1st4 (20,45,57,78)
1.0
A....

t4

=(55,65,65,81)

0.74

6 7
8
9

a7, a8,a9
a7,a8 a7, al0

a7,a8
a7 alo

a9
a8
a7

a9
a8 a7

62.42

Time

of crispstart Figure9 Assignment timeS4for activity a4.

10 11 12

a1o,a, I None
a12

None None None

None None None

alo, a, I a1o
a12

all

alo
a12

This content downloaded from 200.69.103.254 on Fri, 21 Feb 2014 10:33:06 AM All use subject to JSTOR Terms and Conditions

1226 Journal oftheOperational Research Society Vol. 50,No.12

Table 6 The generatedschedule, as R = (4, 3, 3) Activity (ai, i 1 - 12) a, a2 a3 a4


a5

Fuzzy start time (sti) (0,3,3,3) (10,15,15,17) (10,15,15,17) (42,50,50,61) (19,25,25,31) (19,25,25,31) (60,71,71,94) (55,66,66,85) (51,60,60,75) (66,78,78,103) (66,78,78,103) (84,98,98,133)

Fuzzy latest start time (lsti) (-1,29,41,65) (13,41,53,75) (0,30,42,65) (20,45,57,78) (14,40,52,74) (-7,27,38,60) (34,55,67,87) (24,53,64,82) (23,52,63,83) (33,58,69,87) (43,62,74,93) (63,78,89,105) (109,129,140,180)

Fuzzyfinish time (fti) (10,15,15,17) (13,19,19,24) (19,25,25,31) (51,60,60,75) (32,40,40,51) (42,50,50,61) (66,78,78,103) (60,71,71,94) (55,66,66,85) (84,98,98,133) (78,93,94,123) (109,129,140,180) Projectrisk

Start time (si) 3 15 15 50 25 25 70 66 60 75 77 94 0.650

Sat. degree 1.00 1.00 1.00 1.00 1.00 1.00 0.87 0.93 1.00 0.70 0.87 0.70

a6 a7 a8 a9 alo all a 12 Fuzzy Makespan

Table 7 The generatedschedule, as R = (4, 3, 4) Activity (ai, i 1 '12) a, a2 a3 a4


a5

Fuzzy start time (sti) (0,3,3,3) (10,15,15,17) (10,15,15,17) (13,19,19,24) (22,29,29,38) (22,29,29,38) (49,60,60,78) (35,44,44,58) (45,54,54,68) (49,60,60,78) (55,67,67,87) (67,82,83,108)

Fuzzy latest start time (lsti) (-1,29,41,65) (13,41,53,75) (0,30,42,65) (20,45,57,78) (14,40,52,74) (-7,27,38,60) (34,55,67,87) (24,53,64,82) (23,52,63,83) (33,58,69,87) (43,62,74,93) (63,78,89,105) (92,113,125,155)

Fuzzyfinish time (fti) (10,15,15,17) (13,19,19,24) (19,25,25,31) (22,29,29,38) (35,44,44,58) (45,54,54,68) (55,67,67,87) (40,49,49,67) (49,60,60,78) (67,80,80,108) (67,82,83,107) (92,113,125,155) Projectrisk

Start time (si) 3.0 15.0 15.0 19.0 29.0 29.0 60.0 44.0 54.0 60.0 67.0 87.0 0.438

Sat. degree 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00

a6 a7 a8 a9 alo all
a12

Fuzzy Makespan

the currentexperiment, y is set to 0.8. The project ready time and deadline are fuzzified in the similar way. The optimal solutions to these problems were found using the fuzzy version of the A* algorithm29 (see Appendix). Table 8 shows the computationalresults for the three problemsets regardingdifferentvalues of beam width. The average deviations from the optimum solutions are almost less than 0.1, except for the problem set III for beam width = 1. The solution quality can be improved by increasing the beam width (Figure 10). Moreover, the

average CPU time is increased linearly, as the beam width increases (Figure 11). Conclusion The contributionof this research was to develop a fuzzy activity scheduling model to handle uncertain temporal information in product development. A new method was developed to determine the satisfaction degrees of fuzzy temporal constraints. Based on the proposed method, a

Table 8 Computational results regardingdifferentbeam widths Beam width 1 Problem set I II III No. of Activities 30 60 90 Avg dev 0.084 0.065 0.116 Std dev 0.111 0.106 0.125 Avg dev 0.052 0.031 0.088 3 Std dev 0.095 0.052 0.116 Avg dev 0.043 0.028 0.076 5 Std dev 0.083 0.066 0.100

This content downloaded from 200.69.103.254 on Fri, 21 Feb 2014 10:33:06 AM All use subject to JSTOR Terms and Conditions

1227 toactivity scheduling setapproach JWang-Fuzzy


0.140 0.120 0.100 . 0.080
16

tional performance,the following pruningrules29are used: a node n weakly dominates n', if (1) n and n' have the same scheduled set of activities. (2) Let E be the set of eligible activities. Both n and n' have the same eligible set of activities. If the earliest start times of all activities,in E for node n are less than or equal to the correspondingactivities in n'. Node n' can be pruned,if n weakly dominates n'. The fuzzy version of A* algorithm is presented as follows. A* algorithm Step 1. Set OPEN *- OPEN U {no}, where OPEN is the set storing unexpandednodes and no is the initial node for a given problem. Step 2. If OPEN 0 0, then go to step 3; else exit with failure and no solution exists. Step 3. Remove from OPEN a node n for which f(n) is minimum. Set CLOSE <- CLOSEU {n}, where CLOSEis the set storing expandednodes. Step 4. For the node n, performthe Resource Allocation Procedure.If a schedule is found, then the schedule is an optimal solution and exit successful. Otherwise,go to Step 5. Step 5. Find the delaying set DS = {D c El EajEE-D nik < 1Mk for all rk}. Step 6. For every D c DS, generatesuccessor n' of node n with the new added arcs (ai, aj) to the corresponding activitynetwork.Then, updatethe earlieststart times of activity aj as well as all its successors, where ai is the earliest finishing activity in Y and aj c D. Step 7. For every successor n' of n: (1) calculatef(n'), wheref(n') = schedule risk of
n'.

Ag
0,076

> 0.060
o0.040 X 0.020 * 0.028

3 1 3

0.000 5 Beam width

for solutions fromoptimal deviations Figure10 Plotof average sets. threeproblem


1.0

0.90:5 0.808
0.7

a- 0.6 ' 0.5


' 0.4 g 0.3 ? 0.2 0.10 o.0 1 2 3 4 5 6
/

V.4#0 325

9 Bearnwidtlh

beamwidthsfor different CPUtimeregarding Figure11 Average set I. problem fuzzy scheduling procedure was developed not only to obtain a schedule with the least possibility of being late, but also to maximize the satisfaction degrees of all fuzzy temporal constraints. Finally, computational studies were provided to examine the solution quality. It is concluded that the proposed methodology can produce more satisfactory schedules in an uncertainproductdevelopment environment.

Appendix: The fuzzy version of A* algorithm for activity scheduling A* algorithm, which is a kind of the best-first search algorithm,is describedin most introductoryartificialintelligence texts.30This paper adapts the A* scheduling algorithm developed by Bell and Park29to obtain optimum solutionsfor examiningthe solution quality of the proposed fuzzy schedulingprocedure.At each step of the A* search process, the most promising node is chosen for further expansion. If the selected node has no resource conflict, then an optimal schedule is found and the search process terminatessuccessfully. Otherwise,we expand the selected node by delaying some activities to generate the node's successors for resolving resource conflicts. Newly generated nodes are then added to the set OPEN which have not been expanded.Again the most promising node is selected andthe process continues.Note thatthe evaluationfunction of a node is determinedby the schedule risk according to Eq. (19). In addition, in order to obtain better computa-

(2) if n' is neitheron OPEN nor on CLOSE,add it to OPEN; (3) if n' already resides on OPEN or CLOSE, comparen' with n" c OPEN or CLOSE.If n' is dominated by n", then discard n'. If n' dominates n", then substitute n" with n'. If the matching node n' c CLOSE,then move it back to OPEN. Step 8. Go to step 2. Resource Allocation Procedure Step 1. (Finding the eligible activities) Find the set E of activities whose predecessors have been completed. Step 2. (Test of resourceconflict) (1) If EaeE Enik > mk for some rk, then stop and returnthe status 'conflict'.

This content downloaded from 200.69.103.254 on Fri, 21 Feb 2014 10:33:06 AM All use subject to JSTOR Terms and Conditions

Vol. 50,No.12 Society Research oftheOperational 1228 Journal

(2) If all activities have been completed, then stop and return the status 'complete', else go to Step 3. Step 3. (Resource allocation) For activity ai E E, do
t7,1, S<- S U {ai}, sti fti ti, t10ov,ED P <-PU {ai}, R R R-Nj; (2) Update the earliest start times of affected succeeding activities; (3) V <- V U {(ai,fiti)}.

(1)

*-

Step 4.

(Advance to next time point) (1) Remove (a, t) with the smallest t from V; (2) ti11TV+Maxfit,O,,,,t}; (3) SetR<- R+Ni,P< -P-{a}; (4) Go to Step 3.

For the proof of the optimality of A* algorithm, please refer to Bell and Park29 as well as Russell and Norvig30 for details.
researchis partiallysupportedby grantNo. NSC Acknowledgenments-This 88-2213-E-035-021 from the National Science Council of the Republic of China.

References 1 Ulrich KT and Eppinger SD (1995). Product Design and Development.McGrawHill: New York. 2 Williams T, Eden C, AckermannF, and Tait A, (1995). The effects of design changes and delays on projectcosts. J Opl Res Soc 46: 809-818. 3 GiachettiRE, YoungRE, RoggatzA, EversheimW,andPerrone G, (1997). A methodology for the reductionof imprecisionin the engineeringprocess. Eur J Oper Res 100: 277-292. 4 Prasad B, (1996). Concurrent Engineering Fundamentals: IntegratedProduct and Process Organization.Prentice Hall: London. 5 Klir GJ, and Yuan B, (1995). Fuzzy Sets and Fuzzy Logic: Theoryand Applications.PrenticeHall: New Jersey. 6 Dubois D, FargierH, and Prade H, (1994). Propagationand satisfactionof flexibleconstraints. In:YagerRR, andZadehLA, (eds). Fuzzy Sets, Nelural Networks and Soft Computing. Kluwer:Dordrecht. 7 Dubois D, and Prade H, (1988). Possibility Theory: An Approachto Computerized Processing of Uncertainty. Plenum Press:New York. 8 Ow PS, and Morton TE, (1988). Filtered beam search in scheduling.Int JProd Res 26: 35-62. 9 ElmaghrabySE, (1977). Activity Networks:Project Planning and Control by Network Models. John Wiley & Sons: New York. 10 NeumannK, (1984). Recent developmentsin stochasticactivity networks.INFOR 22: 219-248.

11 Foulds L, and Neumann K, (1989). Temporalanalysis, cost minimization,and schedulingof projectswith stochasticevolution structure. Asia Pac J Oper Res 6: 167-191. 12 Lootsma FA, (1989). Stochastic and fuzzy PERT.Eur J Oper Res 43: 174-183. 13 ElmaghrabySE, and KamburowskiJ, (1992). The analysis of activity networks under generalized precedence relations (GPRs). Mgt Sci 38: 1245-1263. 14 De Reyck B, and HerroelenW, (1998). A branch-and-bound procedure for the resource-constrainedproject scheduling problem with generalized precedence relations. Eur J Oper Res 111: 152-174. 15 OzdamarL, and Ulusoy G, (1995). A survey on the resourceconstrained projectschedulingproblem.IIE Trans27: 574-586. 16 Herroelen W, De Reyck B and Demeulemeester E, (1998). Resource-constrained project scheduling: a survey of recent developments.ComputOps Res 25: 279-302. 17 Davis EW,and Patterson JH, (1975). A comparisonof heuristic and optimumsolutionsin resource-constrained projectscheduling. Mgt Sci 21: 944-955. R and Tamarit 18 Alvarez-Valdes JM, (1989). Heuristicalgorithms for resource-constrainedproject scheduling: a review and empirical analysis. In: Slowinski R, and Weglarz J, (eds). Advances in Project Scheduling. Elsevier: Amsterdam, 113134. 19 Lee JK, and Kim YD, (1996). Search heuristics for resource constrainedprojectscheduling.J Opl Res Soc 47: 678-689. 20 Belhe U, and Kusiak A, (1995). Scheduling design activities IEEE Ton RoboticAutom12: 15with a pull system approach. 21. 21 Nasution SH, (1994). Fuzzy criticalpath method.IEEE T Syst Man Cyb 24: 48-57. 22 HapkeM, and SlowinskiR, (1996). Fuzzy priorityheuristicsfor projectscheduling.Fuzzy Set Syst 83: 291-299. 23 Dubois D, and Prade H, (1989). Processing fuzzy temporal knowledge. IEEE T Syst Man Cyb 19: 729-744. 24 Dechter R, and Pearl J, (1988). Network-basedheuristics for constraintsatisfactionproblems.Art Intell 34: 1-38. 25 Zimmermann HJ, and Hintz GW, (1989). A Method to Control Flexible ManufacturingSystems. Eur-J Oper Res 41: 321334. and Extensions.Lexing26 Ignizio JP,(1976). Goal Programming ton Books: Massachusetts. 27 Dubois D, and PradeH, (1983). Rankingfuzzy numbersin the setting of possibility theory.InformSci 30: 183-224. 28 Kolisch R, and SprecherA, (1997). PSLIB - a projectschedulEur J Oper Res 96: 205-216. ing problemlibrary. 29 Bell CE, and Park K, (1990). Solving resource-constrained project scheduling problems by A* search. Nav Res Log 37: 61-84. 30 Russell S, and Norvig P, (1995). Artificial Intelligence: A London. Modem Approach.Prentice-Hall:

Received July 1998, accepted June 1999 after one revision

This content downloaded from 200.69.103.254 on Fri, 21 Feb 2014 10:33:06 AM All use subject to JSTOR Terms and Conditions

Você também pode gostar