Int. J.

Production Economics 174 (2016) 156167

Contents lists available at ScienceDirect

Int. J. Production Economics

journal homepage: www.elsevier.com/locate/ijpe

Using an agent-based neural-network computational model

to improve product routing in a logistics facility
Till Becker a,b,1, Christoph Illigen c, Bill McKelvey d,n, Michael Hlsmann e,2, Katja Windt f,3
a
Production Systems and Logistic Systems, Faculty of Production Engineering, University of Bremen, Badgasteiner Strae 1, 28359 Bremen, Germany
b
BIBA - Bremer Institut fr Produktion und Logistik GmbH at the University of Bremen, Hochschulring 20, 28359 Bremen, Germany
c
Accenture GmbH, Kaistrae 20, 40221 Dsseldorf, Germany
d
UCLA Anderson School of Management, United States
e
Systems Management, School of Engineering and Science, Jacobs University Bremen, Campus Ring 1, 28759 Bremen, Germany
f
Global Production Logistics, School of Engineering and Science, Jacobs University Bremen, Campus Ring 1, 28759 Bremen, Germany

art ic l e i nf o a b s t r a c t

Article history: This study tests whether a simplied neural-network computational model can make routing decisions
Received 4 April 2014 in a logistics facility more efciently than ve 'intelligent' routing heuristics from the logistics literature.
Accepted 7 January 2016 The experiment uses a real-world simulation scenario based on the Hamburg Harbor Car Terminal, a
Available online 29 January 2016
logistic site faced with managing approximately 46,500 car-routing decisions on a yearly basis. The
Keywords: simulation environment has been built based on a data set provided by the Terminal operator to reect a
Logistics real-world case. The simulation results show that the percent-improvement of the neural-net model's
Ship terminal performance is 48% better than that of the best routing heuristic tested in previous studies. To test the
Routing heuristics applicability of the method with more complex logistic scenarios, we relaxed the sequence constraints
Agent-based computational model
for routing in a subsequent simulations study. If logistic complexity in terms of more freedom in
Neural network model
decision-making is increased, the neural net model's percent-improvement performance of routing
Complexity management
decisions is around three times better than the best-performing heuristic.
& 2016 Elsevier B.V. All rights reserved.

1. Introduction they cannot be mapped to simplistic de-complexied mathematical

Companies aim for short throughput times, high schedule
reliability, and low costs as three central objectives in logistics-
network management. These objectives contribute to the fulll-
ment of customer demands, planning reliability, and nancial
wellbeing (Yaged, 1973; Widrow et al., 1994; Nyhuis and Wiendahl,
2008; Michalewicz et al., 2010). Logistic routing options, however,
have become more and more complex as the number of products,
decision points, and global suppliers and customers have increased
(Warnecke, 1993; Tharumarajah et al., 1996; Choi et al., 2001; Choi
and Hong, 2002; Surana et al., 2005; Wycisk et al., 2008). Many
logistic scenarios, however, are too large and too dynamic so that
n
Corresponding author. Tel.: 1 818 902 5991.
E-mail addresses: tbe@biba.uni-bremen.de (T. Becker),
christoph.illigen@accenture.com (C. Illigen),
mckelvey@anderson.ucla.edu (B. McKelvey),
m.huelsmann@jacobs-university.de (M. Hlsmann),
k.windt@jacobs-university.de (K. Windt).
1 and Wiendahl, 2008; Windt et al., 2010a).
Phone: 49 421 218 50176; fax: 49 421 218 9850176.
2
Phone: 49 421 200 31 90; fax: 49 421 200 3078.
3
Phone: 49 421 200 34 78; fax: 49 421 200 3078.
expressions. Instead, we apply rule-directed analytical meth-
odsdirecting the analytical process toward the best solution (Allen
and Helferich, 1990; Harrington et al., 1992).
From a complexity perspective, the complexity of decision-
making is measured in terms of degrees of freedom, i.e. the
number of possible alternatives or options to choose from (Cramer,
1993; Gell-Mann, 1994, 2002). This calls for the direct manage-
ment of decision alternatives, and especially, the reduction of the
degrees of freedom embedded in a decision to keep it manageable
but without simplifying the decision processes so much such
that they do not reect the actual problem any more. In a logistics
facility, inappropriate product-routing decisions caused by
unmanaged complexity can have a negative reinforcing effect and
cause more drastic delays in succeeding production stages.
Accordingly, fulllment of customer demands becomes increas-
ingly problematic due to the increased risk of failing to achieve
short throughput times and high schedule reliability at lowest
possible cost (Gallager, 1977; Wang and Browning, 1991; Nyhuis
To better manage degrees-of-freedom, we use the simplied
neural network design introduced by LeBaron (2001b) in his

http://dx.doi.org/10.1016/j.ijpe.2016.01.003
0925-5273/& 2016 Elsevier B.V. All rights reserved.
 T. Becker et al. / Int. J. Production Economics 174 (2016) 156167
computational model of a stock market. It has been successfully
applied in McKelvey et al.'s (2009) model for an agent-based
computational market design for smart-parts logistics and
therefore is a promising basis for further developments in dealing
with increased degrees of freedom in logistics systems (Choi et al.,
2001; Surana et al., 2005; Pathak et al., 2007; Wycisk et al., 2008).
Moreover, there is a rising demand for new modeling approaches
since modeling becomes ever more difcult as degrees of freedom,
uncertainty, and time constraints increase (Chambers and Mount-
Campbell, 2002; Iassinovski et al., 2003; Michalewicz et al., 2010).
While some recent progress has been made in using simulations to
improve robustness (Caridi and Sianesi, 2000; Al-Mubarak et al.,
2003; Schikora and Godfrey, 2003; Kchel and Nielnder, 2005),
we see little signicant evidence showing how the use of com-
putational models whether cellular automata, genetic algo-
rithms, or neural network models (NNMs) actually improves
managerial logistics decision-making in complex systems so as to
speed up the ow of goods (i.e. cars in our study) while also
lowering costs (Michalewicz et al. (2010), is an exception). Thus,
the objective of our article is to test whether a computational
model, specically the NNM, can actually improve logistics man-
agement given increasing degrees of freedom.
We perform our test using a simulation model derived from a
real-world logistics scenario because we are interested in the
applicability of our approach by companies in the logistics sector.
We are aware of the limitations when selecting a single, specic
case for testing, such as loss of generality and comparability.
However, we think that the general advantages of NNMs have
been documented elsewhere, and we now want to analyze our
approach in a realistic scenario. By choosing a scenario that has
been used previously for the study of different control approaches
(Windt et al., 2010c), we at least partially overcome the issue of
limited comparability.
In Section 2 we describe our real-world case, the Hamburg
Harbor Car Terminal (hereinafter Terminal), a facility that
receives, stores, reworks, and dispatches approximately 46,500
cars per year. The coordination effort of all activities at the
Terminal represents an illustrative logistic site faced with mana-
ging moderate degrees of freedom. We describe our use of
LeBaron's (2001a, 2001b) NNM in Section 3. This includes
(1) describing his simplied NNM; (2) a baseline test (Test 1) of
how well the simulated car-ow compares to the real-world car-
ow at the Terminal; (3) testing the NNM's management of rout-
ing heuristics (Tests 2 and 3); and (4) testing whether using the
NNM helps better manage car-ows under conditions of increased
complexity (Test 4).
Fig. 1. Carport terminal in Hamburg harbor (Google maps, 2011_07_09).
157
In Section 4 we describe aspects of our method: (1) the Terminal
facilities; (2) the data used for our analyses; (3) the current method
used in the Terminal to manage car ows; (4) the ve routing
heuristics applied by Windt et al. (2010c); (5) our NNM approach;
and nally (6) an alternative design of the simulated database so as
to assess the management performance of the NNM under condi-
tions of increased degrees of freedom. In Section 5 we describe the
results for the baseline test (Test 1), the ve heuristics, the com-
parison between the baseline model, Windt et al.'s (2010c) heur-
istics, and the NNM in terms of throughput times (Tests 2 and 3),
and nally the comparison of all heuristics with increased degrees
of freedom (Test 4). We perform all tests with simulated data, based
on recorded data from the Terminal operator. A conclusion follows.
2. An example management problem: the Hamburg Car
Terminal
Fig. 1 shows an aerial photograph of the Terminal. We use this
Terminal as a scenario for testing whether a NNM offers any
advantage over human decision-making, given a complex situation
having some number of degrees of freedom. The car-ow process at
the Terminal is readily described and offers an example of changing
degrees of freedom. It has a exible production sequence and a
large amount of available real-world data, presents various changes
and exibilities in the car-ow process, and easily identiable
logistics objects (cars) in a spatially and organizationally dened
space (Windt et al., 2010c).
For a better illustration of the processes and the car ow on the
Terminal, Fig. 2 depicts it schematically: On the upper left-hand
side the cars arrive at the Terminal. The majority of the incoming
cars reach the Terminal via vessel. They are unloaded from the ship
and stored in the Incoming Delivery parking area (I). Then, cars are
processed through all relevant treatment stations according to
their requirements before they reach the Outgoing Exit area (O).
Available treatment stations are (1) re-fueling cars (gasoline or
diesel); (2) de-waxing; (3) car-body repair; (4) car wash; (5) paint
shop; and (6) nal inspection. Each car enters the Terminal with a
pre-dened list of orders, which can be divided into treatment and
parking orders, depending on the required treatments (if any) and
the duration of the stay before being sent on. If treatment is
required, between one and ve treatment steps are assigned to
each car. The treatment steps have a specic sequence due to
processing constraints; e.g., re-fueling comes rst; removal of
transport protection (de-waxing) needs to come before washing,
158 T. Becker et al. / Int. J. Production Economics 174 (2016) 156167
Street 4 Car-body Repair
Buildings 5 Car Wash
I Delivery (Incoming) 6 Paint Shop
1 Gasoline Station 7 Final Inspection
2 Diesel Station O Delivery (Outgoing)
3 De-waxing
Parking
I Parking
7 3
1 2
Parking
O
Fig. 2. Schematic representation of the terminal in Hamburg harbor.
which has to come after car-body repair but before painting; and
then nal inspection.
3. Car-routing by using a neural network agent-based com-
putational model
Though NN modeling started in the early 1980s (Hecht-
Nielsen, 1990) and the rst business application was in 1988
(Wong et al., 2000), application of NNMs toward routing
improvement, and even optimization, mostly started in the
1990s (El Ghaziri, 1991; Zhou et al., 1991; Arizono and Ohta,
1992; Ali and Kamoun, 1993; Satake et al., 1994; Wieselthier
et al., 1994; Widrow et al., 1994; Baransel et al., 1995; Wang and
Weissler, 1995; Tsujimura et al., 1997; Feng and Douligeris,
2001; Caputo and Pelagagge, 2008). Our review of the top
production and logistics journals (in March 2014; including
Management Science, Int. J. Operations Management; Int. J. of
Operations & Production Management; Int. J. Production Eco-
nomics; Int. J. Logistics Research and Applications, J. Manu-
facturing Technology Management, however, nds very few
specic instances where a NNM is empirically demonstrated to
show improved management of an actual logistic product-
routing situation such as we describe at the Terminal; and
especially we nd no studies showing that the NNM becomes
demonstrably more effective than human decision making by
logistic experts or traditional routing heuristics as degrees of
freedom (complexity) increase for recent works more broadly
applying agent-based models to logistics, see: Hlsmann et al.,
2011; Caglie et al., 2012; Hu et al., 2012; Ivanov and Sokolov,
2012; Wu and Barnes, 2012; Adhau et al., 2013; Aitken and
Harrison, 2013; Elalouf et al., 2013; He et al., 2013; Vimal and
Vinodh, 2013; Haughton, 2013; He et al., 2014; Marques et al.,
2014; Tangpong et al., 2014).
Given the lack of relevant prior examples, in what follows we
apply a very simple kind of NNM to an actual management
situation the Terminal and compare the NNM's performance
against human decisions and various routing heuristics in the lit-
erature as degrees of freedom are increased.
3.1. LeBaron's neural net computational model
We justify our use of LeBarons (2001a, 2001b, 2003, 2006)
NNM as follows: Our objective is to use the NNM to optimize the
ow of products on a complex, multi degrees of freedom logistics
Parking Parking
Parking
6
5
Parking
4
site; The Terminal processes  46,500 cars per year with each car
potentially having to go through from one to ve different Treat-
ment Stations, which means that there are from one to ve station
choices at each decision point (not counting Final Inspection). The
NNM has to constantly monitor the car-ows at the Terminal and
for the most complex situation pick the best ow-path for each
different car at all decision points. We select LeBaron's NNM
design because he designed it to offer information to decision-
makers in a complex, dynamic, and multi-stage decision envir-
onment. In his original model, the NNM advises traders which
stocks to buy and sell at each period in time in a highly dynamic,
high volume stock market. His NNM is designed to monitor a
process having many degrees of freedom. Similarly, decision-
makers in logistics are faced with large amounts of products and
a dynamically changing situation in a multi-period scenario.
LeBaron's (2001b, p. 443) NNM is separate from the current
agents [traders], like an investment advisor or mutual fund as
he puts it. His NNM monitors the market, updating its six sources
(formulas) of investment information. LeBaron's (2001b, p. 444)
sources of information are predictors that are commonly used in
real markets that are continuously updated: information about
current returns, past returns, price-dividend ratios, and technical
trading rules based on exponential moving averages. Traders may
create up to 250 trading rules by weighting information elements
from the NNM (and/or other traders) differently. We illustrate this
in Fig. 3: It shows (1) the six trading rules on the left; (2) the
hidden layer in the middle comprised of information supplied by
each rule LeBaron uses a very simple, single hidden layer as
opposed to the many hidden layers used in other NNMs; and
(3) on the right-hand side it shows various traders with their
multiple rules (up to 250 each) consisting of various weightings,
n, of the six kinds of information available to stock traders (and
of course traders can learn from other traders as well). LeBaron
uses previous S&P stock-trading data to create the investing
information offered by each investment formula. In his model, a
trader may select from a large range of different weightings of the
basic information set, i.e. information from the NNM or informa-
tion based on the success or failure of other traders.
The parallel between LaBaron's and our application of the NNM
is that it observes a real-world process involving degrees of
freedom. For him, market changes are reduced to six formulas. For
us, car-ow at the Terminal is reduced to ve decision points and
six methods of deciding how to route cars (the Terminal's existing
method and the ve routing heuristics used by Windt et al.
(2010c)).
 T. Becker et al. / Int. J. Production Economics 174 (2016) 156167
Fig. 3. Neural net used by LeBaron (2003): In this Figure we show arrows pointing only from left to right, so as to indicate traders learning from the NNM. In LeBaron's
model, however, traders also learn from each other, which if illustrated in Fig. 1, would mean adding arrows in both up and down directions: i.e. going from trader to trader.
However, the logistics complexity at the Terminal is much simpler than the number of degrees of freedom in a stock market.
3.2. Our baseline test of the NN model
In order to test whether a NNM can help people better manage
car-ows through the various treatment stations at the Terminal,
we rst have to test the simulated data used by the NNM with the
real-world data from the Terminal. We need to show how well our
model replicates key aspects of the real-world logistics system at
the Terminal before we start using the model for experimenting
this is the baseline test (#1). For this article we simulate car ows
at the Terminal. The real-world car-ows are totally determined by
human decisions. This option reects the traditional concept of
routing decisions made by managers.
3.3. Using routing heuristics
In their prior research, Windt et al. (2010c) identify a set of
routing heuristics that appear to offer quicker throughput times
through the several treatment stations than the Terminal's cur-
rently used method. In our rst experiment, we use the NNM to
test which, if any, of these heuristics will show quicker throughput
times than the Terminal's currently used routing decisions.
3.4. NNM's use of the routing heuristics
In the existing logistics process at the Terminal, there are only
weekly updates of the schedule of the cars. In our second
experiment, the NNM continuously observes the various car-ows
and then applies one or more routing heuristics at various decision
points to give each car its best routing. The NNM can offer much
quicker and more frequent routing-decision updates. In the real
world, a NNM could make a routing update for each car instan-
taneously whenever a decision is pending or the situation has
changed. As opposed to this, the frequency of decision updates is
only once a week for human decision makers. In our third
experiment the NNM aims to offer an optimal solution for redu-
cing total car-ow times.
3.5. Increasing the degrees of freedom
For most cars at this Terminal there is really only one degree of
freedom stay parked till moved out of the Terminal or go to
Treatment stations. Therefore, in this fourth experiment we add
degrees of freedom by relaxing the sequence constraints. While
159
this may be unrealistic for most of the cars at this particular
terminal, this experiment applies to other logistic routing situa-
tions where objects could be routed to any one of ve options at
the rst decision point, and then to any one of the four options at
the second decision point, and then to any one of the three options
at the third decision point, and so on to the end. In terms of
decision-making time for making routing decisions similar to a
simplied traveling-salesman problem for each car, the NNM
clearly has an advantage and is a much cheaper decision processor
over the long term than the cost of hiring employees to attempt
making these kinds of decisions for each of the 46,500 cars
being processed. At this Terminal, car-ow decisions need to be
made to route 1800 cars on average per day.
4. Method: car-ow data, heuristics and description of the NN
model
4.1. Structure of the car-ow database
The simulation model in this article builds from a prior simu-
lation developed by Windt et al. (2010c), since it is already
operational and correctly represents the real world in the harbor
context. The original model was validated based on interviews
with experts from the Terminal operator. The schematic ow of
the original simulation is depicted in Fig. 4. To develop their
database, three researchers went to the Terminal and interviewed
experts. They collected all relevant descriptive data about such
things as:
 Number of cars per year;
 Various routings of car ows in and out of parking;
 Various routings cars take through the ve treatment stations;
 Length of queues before each station; and
 Various ways available to speed as many cars through the
Terminal as possible.
The data are drawn from the operator's database, which logs all
operations activities on the Terminal. Each movement and each
treatment of each car is logged, including start and end time for
controlling and billing purposes. Windt et al. then decomposed all
raw data into a material ow scheme, represented by the boxes
shown in Fig. 4. Then, the means and standard deviations are
160 T. Becker et al. / Int. J. Production Economics 174 (2016) 156167
P1
P2
I DP1
P9
I Delivery (Incoming) TPA Treatment Parking Area
Pi Parking area Ti Treatment
Fig. 4. Schematic representation of the model according to Windt et al. (2010b).
calculated for the time each of 46,574 cars took to travel on the
roads, to go through the required treatment queues and stations,
and then through the various subsequent stations, as required.
We ignore the mode of transportation and car-ow rates before
cars arrive at the Terminal in this simulation since we model only
the processes at the Terminal. Fig. 4 illustrates all possible routes
of a car through the Terminal. Routing decisions are made at the
incoming parking area (I), and decision point 1 (DP1) and 2 (DP2).
Cars can be scheduled either for immediate or future delivery, and
they can have either no or one to ve treatment orders. Conse-
quently, we can distinguish between four possible decision sce-
narios (Windt et al., 2010c):
1. Immediate delivery, no treatment: Cars are routed from
incoming delivery (I) via decision point 1 (DP1) to outgoing exit
(O) immediately.
2. Immediate delivery, treatment required: If a car is already
scheduled for delivery upon arrival and a treatment is required,
the car is routed from incoming delivery (I) via decision point 1
(DP1) to the treatment parking area (TPA), through treatments
(T1T5) and then to decision point 2 (DP2). Here a decision has
to be made whether to route the car back to the treatment
parking area (TPA) to wait for space in its next treatment-station
queue, or send it to outgoing exit (O).
3. Future delivery, no treatment: If delivery is not yet scheduled,
the car is routed from incoming delivery (I) via parking areas
(P19) to decision point 1 (DP1) and sent further to the outgoing
exit (O).
4. Future delivery, treatment required: The car is routed from
incoming delivery (I) via parking areas (P19) to decision point 1
(DP1), where it is sent to the treatment parking area (TPA), and
then to its next treatment-station queue until it reaches
decision point 2 (DP2). From here on, each car proceeds as from
TPA if it has further treatments; otherwise it is routed to
outgoing exit (O).
Data describing the treatment-station routing for each of the
46,574 cars requiring some kind of treatment are stored in a
central database. Data for most cars are transferred to the Terminal
before delivery and are thus available at arrival. Missing data are
added manually. Data includes production code, type number,
chassis number (used for identication of each car within the
Terminal), color and ultimate retailer/car dealer. Some treatments
are to x up cars after transportation and others result from cus-
tomization requests from dealers.
The sample of 46,500 cars offers a realistic real-world data-
base for modeling. Although planning of car ow is centrally
organized, with decisions made by employees, the data offer a
platform from which we can use the NNM to experimentally test
for possible benets from adding the routing heuristics and then
Cue T1
Cue T2
TPA DP2 O
Cue T7
O Delivery (Outgoing)
DPi Decision Point
to what extent letting the NNM manage car-ow routing offers
further improvements as we increase the degrees of freedom. The
uncertainty and dynamics at the Terminal caused by the differ-
ent treatment orders that cars may require as well as the length of
queues at the treatment stations result in considerable variance
in car throughput times. Consequently, total times in the different
stations can range from a few minutes to 12 hours (e.g. if a car
has to wait overnight before the treatment at a particular station is
completed). Hence:
Test #1: The data relevant to car processing at each station, upon
which our experiments using the NNM depend, are statistically
equivalent to real-world station-processing data.
4.2. Car-ow routing heuristics
Ideally, each car-routing decision should attempt to minimize
car-ow time CFT (total traveling station-queue treatment
processing times and possible overnight waiting) times. Thus it is
possible that for any given car, the waiting time before entering
into any particular station is constantly changing because of the
routing effects of the queue-waits and station-processing speeds,
say, of the 10 cars ahead of it in the various queues. We illustrate
this in Fig. 5. It shows, for one specic situation, how the many
different car-routing options can affect a car-routing decision at
point DP1 (shown in Figs. 4 and 5). It illustrates one possible
(depth 3)4 set of station queue and treatment sequences and
how they can variously affect the routing decision made at point
DP1. Only three stations are involved in this example, but even so
there are quite a number of alternative route sequencings
depending on waiting time in a station queue, and length of
treatment for any given car since these are always changing.
The original model of Windt et al. (2010c) applies the Term-
inal's Method, which is a human decision process. We exclude a
Random Method from our study because Windt et al. (2010b)
show that random is denitely less effective. We use ve routing
heuristics drawn from the logistics literature: the Minimum
Queue-length Estimation Heuristic, Minimum Cumulative Processing
Heuristic, Simple Rule-based Heuristic, Ant Pheromone Heuristic,
and the Holonic Heuristic. We choose these heuristics based on a
literature search covering the current state of autonomous control
methods applied to logistics (Windt et al., 2010b). For our research,
the foregoing routing heuristics serve as literature-based
4
Depth is the number of successor stations included in calculating the total
driving station queue treatment time used to decide which next station to send
a car to; all of the subsequent treatment-station times have to be calculated before
the next station can be selected. Depth1 includes only the next station and its
queue. Depth 3 means that the times for the two previous stations are added to
the third station's wait times.
 T. Becker et al. / Int. J. Production Economics 174 (2016) 156167 161

start depth = 1 depth = 2 depth = 3 where: F average ow time of a car at station (which is the
average time a car is in a queue and station); Pi completion time of
successor successor treatment of car i; Ri time of entry of car i into queue; Nnumber
2 3 of cars included in the calculation; set of stations included in the
successor analysis; inext car to be processed. The formula is adapted from
1
(Daniel and Guide, (1997), p. 111).
successor successor The Simple Rule-based Heuristic comes from Scholz-Reiter
3 2
et al.'s (2006) Method 2; it compares estimated waiting times
at station-queue points based on the processing time of the pre-
vious 10 cars. Depth 1:
successor successor
1 3
P
N
successor dj wj pj
DP1 j Ns
2 Qi
successor successor
s
3 1 Select station i : mini Q i
where:6 N is the number of cars that have completed treatment
at station i. wj is the time car j has spent waiting in the queue, pj is
successor successor
the processing time of car j, and dj is the driving time required to
1 2
bring car j to the treatment station. The driving time component is
successor
3 not part of the original source, but has been added as a specic
successor successor
component of the car terminal case. The s most recent items are
2 1 taken into consideration (s 10).
The Ant Pheromone Heuristic was adapted to logistics needs by
Armbruster et al. (2006) and Helbing et al. (2006). This heuristic
Fig. 5. Determining successors with depth 3 (Windt et al., 2010a).
starts from how ants gather food. Translated to our Terminal
simulation, this means that cars following this heuristic leave
benchmarks for dening well-accepted logistics performance
indicators, i.e. throughput times (Nyhuis and Wiendahl, 2008). virtual Pheromone marks at each station when they leave. The
Descriptions of the ve different routing heuristics follow, following equations dene the time a car leaves a queue and then
including their dening essence, a brief description of the heur- leaves a treatment station:
istic, and general remarks. Further details about the methods and Z t Z t
in
their characteristics can be found in Windt et al. (2010a). For a Gks t Rks r qpks r dr ; F in
ks t f ks r dr
1 1
review of heuristic applications, see Michalewicz et al. (2010) or
Vidal et al. (2013). where: Fks, Gks is the number of cars that have entered/left queue k
The Terminal's Method [what Windt et al. (2010c) call the at processing stage s; qpks denotes the portion of the queue due to
Standard Method] is an approximation of the method actually the product (car) of type p; Rks is the processing rate at station k on
in
used by human decision makers at the carport Terminal for car stage s; f pks is the inux of product p at station k on stage s;
routing. It is a pre-determined ordering of how cars are sent ttime; the equations are the antiderivatives of ux functions
through the treatment stations, as follows: (1) refueling; (2) de- (Armbruster et al., 2006, p. 109).
waxing; (3) car-body repair; (4) car wash; (5) painting; (6) nal Then we have:
inspection.    1 
The Minimum Queue Length Estimation Heuristic gives top pks t wait in
ks T pks t  F ks Gks t T pks ;
priority to the station with the fewest number of cars in its queue
(Philipp et al., 2007). If all queues have equal length, a random where: pks is a car's total throughput time; wait ks ; T pks is the wait-
ing/processing time in a treatment station.
selection is made. Depth1:
The quantity pks represents the Pheromone concept, i.e. the
X
n
routing choice that offers the minimum queue and treatment
Qi pj r k
j1
processing time option available to a following car in the
sequence. To minimize processing time, then:
Select station i : mini Q i  
where:5 the queue length is Q i is the sum of the processing p;ps; s min pks ; k 1; ; K ; p 1; ; P; s 1; ; S;
times pj of all already waiting elements j plus the remaining where: ps; represents the minimal time available.
processing time of the element k currently being processed, r k . In
For additional Pheromone modeling details, see Armbruster
case the processing times are identical, or the processing times are
et al. (2006, pp. 107109).
not known and an identical average value is assumed, pj 1 8 j and
The Holonic Heuristic comes from Markus et al. (1996), Van Brussel
r k 0. This means that simply the queue lengths are compared
et al. (1998) and Lee and Banerjee (2011). After being processed at
and the shortest queue is chosen.
whatever depth (e.g. depth 2 as portrayed in Fig. 5), a car's routing to
The Minimum Cumulative Processing Heuristic sends the car to
the next station (e.g. depth3) is based on a market-like approach:
the station at the next depth level that shows the lowest QP time,
whichever treatment station can offer the shortest total time (based
where QP[station processing time  (car in stationnumber of
on drivingstation-queue waitingtreatment times) gets the car. If
cars in station queue)]. Depth1:
P several stations at depth3 have the same queue length (delay), the
P  Ri car showing a quicker nal release time is chosen (e.g. it needs fewer
F i i
N
6
The formula is not explicitly stated in the source (Scholz-Reiter et al., 2006)
5
The formula is not explicitly stated in the source (Philipp et al., 2007) and has and has been composed by the authors based on the detailed description in the
been composed by the authors based on the detailed description in the source. source.
162 T. Becker et al. / Int. J. Production Economics 174 (2016) 156167
Fig. 6. Neural net applied in the terminal.
treatments). Depth varies:
Bi di wi pi
Select station i :mini Bi
where: a treatment station is a holon i, Bi is the bid of treat-
ment station i, with di, wi, and pi being driving time, queue waiting
time, and station processing time if the current car decides to
choose station i.
The routing heuristics described above, which are used by
Windt et al. (2010c), show more or less improvement to car-ow
times at the Terminal; they serve as benchmarks from which we
expect that the NNM will show additional improvement for our
sample of 46,500 cars. Hence:
Test #2: One or more of the heuristics will lower car-ow time and
costs at the Terminal.
4.3. Our neural net method
In our use of the NNM, we replace the trading formulas with
the ve car-ow routing heuristics used by Windt et al. (2010c)
illustrated in Fig. 6. It shows (1) the ve routing heuristics relevant
to speeding up the Terminal's car-ow, the average times of the 50
most recent cars and the average expected queue and processing
time of each treatment station on the left; (2) the hidden layer in
the middle is limited to data from each of the ve heuristics; and
(3) on the right-hand side it uses arrows to show that each of the
46,574 cars is given a routing path by the NNM at each decision
point as the cars move through one or more of the Terminal's six
treatment stations.
Our NNM proceeds as follows: Every time a car arrives at DP1
the Terminal has a routing decision to make. To begin, for each car,
it lets each of the ve routing heuristics tell it which is the best
next station to send the car to. It does this by counting how often
each treatment is mentioned by the heuristics. Suppose for
example that the Minimum Queue-length Estimation, Minimum
Cumulative Processing, and the Simple Rule-based Heuristics all
select T1 as the next best station, but the Ant Pheromone and
Holonic Heuristics select T2. Since T1 gets three mentions and T2
gets only two, T1 appears to be the best next station. However, the
NNM also checks for the possibility of a jam at the T1 station
(perhaps because of equipment failure). The NNM checks the
average queue-wait treatment time for the last 50 cars processed
at T1. We focus on the last 50 cars, since after extensive pre-test-
ing, this number turned out as the best performing. This source of
information is also shown in the right hand box in Fig. 6.
If the average is more than 50% higher than the expected mean,
the NNM presumes a jam and routes the car to the next most
optimal station, in this example, T2. The resolution if a jam is
detected by the NNM as follows: Since some cars have to be sent
to T1 because there is no other station they need to be sent to,
eventually they start being processed and the average of the most
recent 50 cars going through T1 starts going down toward the
norm. But the NNM's routing calculations don't stop at the
depth 1 decision (our depth 1 to depth3 example is shown
in Fig. 5). For each of the three station choices shown in Fig. 5, the
NNM uses all ve heuristics to see which of the three depth1
stations gets rst choice, but in doing this it also checks
throughput speeds at depth2 and then for depth 3 and so on
down to depth 5. In short, our NNM considers all possible rout-
ing sequences from depth 1 to depth 5 to make the choice at
depth 1. It continues this multi-depth-sequencing calculation till
the car reaches the nal inspection station.
Since the placement of cars in the various areas labeled Park-
ing in Fig. 4 has no bearing on how fast cars are processed through
the ve treatment stations, we ignore car-storage times in these
parking lots. In terms of Fig. 4, we just focus on processing times
between DP1 and point O. Of the four key logistics performance
indicators described by Nyhuis and Wiendahl (2008), the real-
world data we have available forces us to focus on throughput
times rather than due dates and due-date reliability.7 These
include statistics about how much time each car spent being dri-
ven from temporary waiting in parking lot I to one or another
treatment station, or waiting in a treatment-station parking queue,
and/or being processed in a treatment station. Given this, note that
for stations having treatment times of less than 2 min, the travel
time and possible shortage of drivers becomes a signicant
uncertainty, given that the station-queues only hold 4 cars at
a time.
The number of cars waiting in any given treatment parking queue
plus station-processing speed denes the length of time for any
particular treatment. Given that each car may be tagged for several
treatment stations and their parking queues, a decision has to be
made for each car as to what is the speediest total routing from
decision point DP1 (where station queues and treatments begin),
to the end at point O. This is what the routing heuristics are
supposed to improve. Hence:
Test #3:. The NNM will offer an optimization solution that lowers
both total car-ow time and total cost more effectively than does the
best of any of the routing heuristics (since for any given car it can
7
Other key performance indicators are not considered because of missing
information due to condential restrictions: due dates from dealers are unavailable;
as is reliability of meeting due dates. Machinery utilization is embedded in treat-
ment processing times. Inventory level is also embedded in treatment times; lack of
cars to be processes is surely not a dening aspect.
 T. Becker et al. / Int. J. Production Economics 174 (2016) 156167
always choose the best performing heuristic, or set of best performing
times when it considers more than one depth).
A variation of the scenario: adding degrees of freedom (i.e. complexity)
The Terminal database we have discussed so far comes with
minimum degrees of freedom. For many cars, the most important
scheduling criterion is due date at the dealer, but we do not have
access to this since it is restricted. For many cars the only degree of
freedom is whether they need gas or diesel fuel just after being
driven off the ship. If a car needs car-body repair there are no
other degrees of freedom: the sequence is fuel, de-waxing, car-
body repair, cleaning, painting, and nal inspection. But not all
cars need car-body repair so this adds a degree of freedom. If a car
does not need body repair it may not need painting; this adds
another degree of freedom. Even so, there are not many. With so
few degrees of freedom, there is not much advantage for the NNM
to offer enhanced optimization above what any single routing
heuristic offers. Consequently, we add an articial variation of the
scenario so as to increase the degrees of freedom, since we expect
the NNM to perform even better than the other heuristics as
degrees of freedom are increased.
Given the reality that there are many supply-chains with many
more degrees of freedom, with the possibility of offering a better
demonstration of the power of the NNM to offer an improved
optimization enhancement over any single routing heuristic, we
now switch to a second simulated Terminal database that has no
sequence restrictions built into it. While it does not offer a real-
world test of the Hamburg Terminal, it does offer us a test of
whether a NNM can offer greater optimization improvement as
degrees of freedom increase by taking advantage of what each
individual heuristic offers singly or in combination with others and
by accounting to all ve depths (i.e. successive optimization
decisions as a product goes through multiple treatment stages
with decisions needed at each stage partially illustrated in
Fig. 5). Consequently, the product-routing choices degrees of
freedom are considerably increased in this second database.
The NNM is now solving a semi-traveling-salesman problem all
possible routings from depth1 to depth5 are accounted to.
This sets up our fourth test:
Test 4: As complexity increases in a logistics system, the
performance-optimization improvement offered by the NNM relative
to the several routing heuristics is increased.
5. Simulation results
5.1. Test #1: baseline test
We test T1 by using a two-sample t-test to show the t of the
baseline model to the real-world data by comparing average
throughput times for each treatment station. Since the sample size
exceeds 46,000 cars, we assume normal distributions (Brase and
Brase, 1995; Agresti and Franklin, 2009; Johnson, 2011). If a
Terminal's distribution is skewed, our comparison-test against a
normal distribution would reduce the signicance of our t-test
comparisons. Results are shown in Table 1.
For our baseline test, in Table 1 we combine the following times
for each car for each station: (travel time station-queue wait -
treatment time). For each station, Table 1 shows real-world time
and simulated time. The t-test signicances in the right-hand
column show that there is little meaningful difference between
real-world and simulated times.
We conclude, therefore, that our model serves as an acceptable
basis for further investigation between existing performances at
163
Table 1
Statistical validation of the baseline model.
Treatment station Real world Model p-Value
TT TTR TQ TT TTR TQ P (RW a MO)
Gasoline station 10:50 min 10:50 min p o0.001
Diesel station 12:41 min 12:41 min p o0.001
De-waxing 1:22 min 1:22 min p o0.001
Car-body repair 565:53 min 564:53 min p o0.001
Car wash 6:05 min 6:05 min p o0.001
Paint shop 14:35 min 14:36 min p o0.001
Final inspection 353:33 min 347:46 min p o0.001
TT Treatment time; TTR Traveling Time; TQ Queue time; RW Mean real world;
MO Mean model.
the Terminal and what the several routing heuristics offer by way
of improvement. We thus conrm T1. Moreover, the model now
gives us an appropriate platform for conducting experimental
analyses of the NNM's ability to further enhance optimization in
supply chains featuring more complexity than at the Hamburg
Terminal as it currently exists.
5.2. Test #2: replication of Windt et al.'s application of the routing
heuristics
With respect to the ve routing heuristics, we re-compare their
performance relative to the total throughput time produced by the
Terminal's Method since our car-sample includes only the 46,574
cars that require some kind of treatment, as opposed to the
complete sample of 152,000 that Windt et al. (2010c) study
which includes over 100,000 cars needing no treatment. The
average throughput time from I (incoming) to O (outgoing) in
Fig. 4 for the 46,574 cars is 4462.10 min,8 which is the same as the
average time of the Terminal's Method shown in Table 2; it shows
(1) to what extent each of the ve heuristics and the NNM speeds
up throughput time in minutes relative to the Terminal's Method
(column 2); and (2) relative time reductions by the various heur-
istics in percentage terms (column 3).
The Minimum Cumulative Processing Heuristic cuts throughput
time the most, at  8.71%. The Simple Rule-based Heuristic comes
next. The Ant Pheromone Heuristic is the worst, actually increasing
throughput time by 6.53%, caused by the inertia of Pheromone
build-up and decay. We conrm T2 since one or more routing
heuristics do in fact improve the Terminal's throughput time.
5.3. Test #3: does the NNM offer faster car-routing times than the
heuristics?
The bottom line of Table 2 shows that the NNM offers the most
cost-saving potential. The fact that overnight stays in station
queues are counted as part of average throughput time gives the
NNM the advantage of being able to make routing decisions that
try to minimize overnight stays. In comparison to the Terminal's
Method, the NNM speeds up throughput time by 12.73% (i.e. cuts
568 min per car); this is a 48% improvement over the 8.71%
improvement of the best-performing routing heuristic (i.e. the
Minimum Cumulative Processing Heuristic). This equates to saving
18,375 out of a total of 144,317 car-days at the Terminal. While the
actual costs are not available in the dataset, the NNM is telling us
that the Terminal could increase its revenue considerably by
8
While specic treatments such as de-waxing (takes  2 min) vs. car-body
repair (takes  60 min) vary considerably in treatment time, the 4462.10 min
average includes the hours that cars stay in a treatment station over night or even
over a weekend. Needless to say, optimizing to avoid overnight and over-weekend
delays is crucial for reducing throughput times.
164 T. Becker et al. / Int. J. Production Economics 174 (2016) 156167
Table 2
Heuristics' performance assessment with sequence constraints.
Optimization method Total time mean
TT TTR TQ TN
Terminal's method 74.37 h
Minimum queue-length estimation method 69.74 h
Minimum cumulative processing method 67.89 h
Simple rule-based method 68.18 h
Ant Pheromone method 79.22 h
Holonic method 71.38 h
Neural net 64.90 h
TT Treatment time; TTR Traveling Time; TQ Queue time; TN Overnight time.
Table 3
Heuristics' performance assessment without sequence constraints.
Optimization method Total time mean
TT TTR TQ TN
Terminal's method 73.04 h
Minimum queue-length estimation method 70.96 h
Minimum cumulative processing method 77.08 h
Simple rule-based method 73.45 h
Ant Pheromone method 73.25 h
Holonic method 70.58 h
Neural net 63.75 h
TT Treatment time; TTR Traveling time; TQ Queue time; TN Overnight time.
processing some 18,000 more cars/year than it does at present. T3
is validated.
5.4. Test #4: use of the NNM with increased degrees of freedom
As we noted earlier, much of the car-ow at the Terminal is
strictly sequenced (see Section 2). This sequencing greatly reduces
degrees of freedom, which in turn undermines the advantage of
the NNM over the simpler routing heuristics. Before running the
experiments we did not expect the NNM to offer as much
advantage as is shown in Table 2. Much of this, however, is
because its multi-depth use of the heuristics, which allows it to
take the effects of overnight stays into account.
Given the dominance of sequencing, we conduct a further test of
the NNM's abilities by increasing the degrees of freedom in our
simulated Terminal data. Our Test T4, shown in Table 3, holds that the
NNM's optimization contribution will increase even more than is
shown in Table 2 when degrees of freedom are increased by elim-
inating the sequence restrictions. Results of de-sequencing are shown
in Table 3. Like in Table 2, the averages of Total Times shows the
improvement of throughput times in minutes (column 2) and the gain
or loss in performance by the various heuristics in column 3. The
optimization results by the NNM are shown at the bottom.
We see in Table 4 that when compared to the Terminal's
Method, the NNM speeds up car-ows by 14.27%. This despite that
fact that the Terminal's Method also improves car-ow times
because of the loss of the sequencing constraints. Because of this,
the NNM is challenged with improving optimization even more
than the Terminal's improved performance.
What is most interesting, however, is that while the NNM cuts
car-ow times by 14.27% (12% better than with the sequencing
constraints) the performance of the heuristics changes con-
siderably. The best heuristic in Table 2, Minimum Cumulative Pro-
cessing Heuristic, has its car-ow improvement cut from an 8.71%
improvement in throughput time to a 3.65% worsening. The Simple
Rule-based Heuristic has its performance reduced from an 8.32%
improvement to a 0.56% worsening. In contrast, the Ant Pheromone
Heuristic shows a change from a 6.53% worsening to a 1.50%
improvement in optimization ability, though the actual cut in
throughput time is only 1.5%. The Holonic Heuristic shows a 1%
improvement in optimization. By any measure the NNM shows
even more optimization ability than any of the heuristics when the
 6.22%
 8.71% sequencing constraints are abandoned, that is, when the degrees
 8.32% of freedom (and complexity) are increased. We conclude that T4 is
6.53%
conrmed.
 4.02%
 12.73% We assume that uctuations in the performance of the meth-
ods when changing the experiment setting are caused by the
relatively narrow rule-set of these methods. Some methods per-
form better in more linear, constrained settings, presumably
because of the relatively simple decision rules. Others, such as the
Ant Pheromone Method, need some more time to adapt to a
change in the environment (by re-distributing the Pheromone
levels), but are able to cope with more exible and previously
unknown states of the system (i.e. not pre-dened by rules). NNMs
are learning and adapting methods by design, and are thus more
likely to perform better in many different scenarios.
 4.58%
3.65%
0.56%
 1.50%
 5.10% 6. Summary, discussion and conclusion
 14.27%
6.1. Summary
In this article we begin with a baseline test to show that our
simulated data very well match the empirical database collected
from the Hamburg Harbor Car Terminal (Terminal) and then we
create a neural network model (NNM) patterned after LeBaron's
use of a NNM in his stock-market model. His model watches the
stock market and summarizes information in formulas traders can
use. We let our NNM select among ve routing heuristics so as to
optimize the routing of cars through various treatment stations at
the Terminal. The NNM assigns each car at each decision point to
the routing path offering the fastest throughput time.
The Terminal unloads a total of approximately 152,000 cars off
ships per year. Of these, our simulation focuses only on the 46,574
cars that are routed to one or more treatment stations. There are
ve treatment stations: refueling, de-waxing, car-body repair, car
wash, and painting, and nal inspection. We use ve well-known
routing heuristics from the logistics literature for the NNM to use
to pick the fastest car-ow through the Terminal via each of up to
ve key decision stages and many different path-choice combi-
nations to deal with.
We rst conduct a baseline test to show that our simulated data
show a strong statistically signicant match to the data from the
Terminal. Our rst test, T1, is conrmed using a two-sample t-test,
with p o0.001. Next we test T2, which holds that one or more of
the ve traditional routing heuristics we draw from Windt et al.
(2010c) will improve car-ow through the Terminal. We show that
all but the Ant Pheromone Heuristic quicken throughput time. T3
holds that the NNM will reduce throughput time more than any
single heuristic. T3 is conrmed: the percent of the NNM's
increased car-ow efciency is 2.8 times better than the best
heuristic (from 5.1 to 14.27%). Finally, we eliminate sequencing
constraints in the data so as to increase degrees of freedom. Using
the NNM, we then conrm T4 by showing that the percentage
increase of the NNM's efciency-based treatment-time-saving is
almost three times better than the percent improvement of
the best heuristic, given the conditions of increased degrees of
freedom.
 T. Becker et al. / Int. J. Production Economics 174 (2016) 156167
Table 4
Algorithms vs. neural net performance in sequenced and de-sequenced datasets.
Optimization method Sequenced data
Minimum queue-length estimation method  6.22%
Minimum cumulative processing method  8.71%
Simple rule-based method  8.32%
Ant Pheromone method 6.53%
Holonic method  4.02%
Neural net  12.73%
6.2. Discussion
Besides the successful baseline-test of our model, we have
three additional ndings: (1) the routing heuristics show dramatic
differences, one vs. another; (2) they show dramatic differences
going from sequenced to de-sequenced (degrees of freedom
added) simulation databases; and (3) the NNM is better than any
single heuristic in either scenario, as shown in Table 3. Finally, as
shown in Table 4 some heuristics get better as the degrees of
freedom increase, but others get worse. Note, also, that the NNM
delivers a stable performance and is best with both. It improves
throughput time by 12.73% in the sequenced dataset (e.g. cars go
through stations in a xed order) and by 14.27% in the dataset
with degrees of freedom increased; a 12% improvement above
12.73%.
Second, the Minimum Cumulative Processing Heuristic goes from
best performance in the sequenced dataset to the worst perfor-
mance when degrees of freedom are added. Like the Simple Rule-
based Heuristic, which shows almost as much loss in effective
routing ability, the Minimum Cumulative Processing Heuristic only
focuses on queue length. But, since the latter multiplies queue
length by the treatment-processing time of the station, longer
queue length is magnied dramatically. Since a decision based
only on queue length at one depth (as depicted in Fig. 5) ignores
throughput times in all the routing choices through all the suc-
ceeding stages, it stands to reason that it would falter as degrees of
freedom (i.e. complexity) increase.
Third, the only heuristic that improves its optimizing ability as
the dataset goes from sequenced to de-sequenced is the Holonic
Heuristic the only multi-depth one. It goes from a 4.02% opti-
mization improvement for the sequenced dataset to a 5.10%
improvement when degrees of freedom are increased. This is does
not seem much of an increase, but 5.1% is a 25% improvement over
4.02%; The Holonic heuristic's advantage appears to be based on
the fact that the Holonic heuristic uses information about total
throughput time (i.e. drivingqueue treatment times) at the
next stage stations, rather than past ones (e.g. the Simple Rule
Heuristic focuses on the times of the past 10 cars). The Holonic
Heuristic also has the advantage of taking account of release time,
which is one way of seeing which is the fastest route through the
one or more successive stations (we preferred due-date at the
dealer but these data were not available).
It also appears, fourth, that heuristics focusing on past perfor-
mance are not as optimizing as those looking forward. This seems
logical, and the Ant Pheromone Heuristic is a good example of this.
It improves its optimization score the most when going from the
sequenced dataset to the de-sequenced one, i.e. it goes from 6.53%
worse than the Terminal's Method to 1.50% better after degrees of
freedom are added. Even though it is better than the Terminal's
Method it is not better than the Holonic Heuristic which is 5.10%
better but clearly not as good as the score produced by the NNM,
which shows a 14.27% optimization improvement.
165
Degrees of freedom added Performance change
 4.58% 1.64% decline
3.65% 12.36% best to worst
0.56% 8.88% decline
 1.50% 8.03% improvement
 5.10% 1.08% becomes best
 14.27% Best overall value
6.3. Conclusion
In this article we utilize the Hamburg Harbor Car Terminal to
show that the NNM can manage routing logistics better than
employees and better than any of the specic existing routing
heuristics tested. Given the few minimalist applications of NNMs
for the actual management of product routing to date (based on
our search for relevant research in the top-ranked logistics and
operations research journals), our research takes logistics man-
agement down a fairly untried path. Our ndings follow from
three central questions.
The rst question is: Can logistics routing management be
taken over by an agent-based computational model such as a
NNM? We offer a preliminary positive response to this question in
that we have created simulated data for use by our NNM that
compare very well with the real-world database at the Terminal;
in short our follow-on computational experiments showing that
the NNM has management capabilities have real-world validity.
The second question is: Does the NNM outperform the ve
representative routing heuristics from the logistics literature?
Question 2 is clearly conrmed. Our results show that models such
as the NNM offer product-routing decisions more effectively than
standard optimization heuristics. Further investigations focusing
on truly smart parts (adaptive agents) with abilities such as
interaction, learning, and autonomous behavior are needed in
order to reveal the full potential of smart-parts logistics systems,
and more broadly the management complex adaptive logistics
systems in general (Wycisk et al., 2008; McKelvey et al., 2009).
The third question is: Does adding degrees of freedom in a
logistics system positively affect the performance of the NNM in
relation to the ve routing heuristics? Since our results show that
the NNM outperforms the ve routing heuristics even better as
degrees of freedom are increased, Question 3 is also conrmed.
This indicates that the increased degrees of freedom and dynamics
we have in modern logistics systems call for use of models such as
our NNM. Presumably, then, current routing heuristics should
more and more be replaced with intelligent and learning agent-
based computational models, since logistics systems are them-
selves becoming ever more complex. Additionally, methods such
as our NNM offer useful logistics-management improvements in
scenarios such as the Terminal. To summarize: As modern logistics
systems become more complex, agent-based computational
modeling methods offer promising optimization improvements.
We begin the process of using a NNM to actually manage the
use of routing heuristics for car-routing logistic decision making.
Consequently, there are several targets for further research:
 Will computational models generally perform better in complex
and dynamic scenarios than existing routing heuristics?
 Will our type of NNM also perform better in other real-world
scenarios?
 Will the NNM perform better in real-world scenarios having
demonstrably increased degrees of freedom?
166 T. Becker et al. / Int. J. Production Economics 174 (2016) 156167
 How does the logistics performance of the NNM compare with
the optimal-routing performances of other kinds of agent-based
computational models?
There are also some limitations of our initial investigation. First
of all, our current ndings need to be conrmed by applying our
method to other real-world scenarios especially those consisting
of supply chains extending from Asia to Europe and the US. These
scenarios offer much more sophisticated tests of how well the
NNM can manage routing heuristics. Verifying our preliminary
results across different scenarios will strengthen the results by
basing them on a wider range of real-world applications. Second,
future research should also explicitly focus on other performance
indicators [following Nyhuis and Wiendahl (2008)]; these include
high due-date reliability, high machinery utilization, and low work
in progress. Finally, McKelvey et al. (2009) raise the question as to
whether smart parts can outperform computational models such
as the NNM. Further research is needed that introduces true smart
parts into computational experiments to test whether the NNM
can out-manage smart parts in modern adaptive logistics systems,
as suggested by Carlson and Doyle (2000), Choi et al. (2001),
Surana et al. (2005), Albino et al. (2006), Pathak et al. (2007),
Wycisk et al. (2008), Hlsmann et al. (2009) and McKelvey et al.
(2009).
Acknowledgment
This research was supported by the German Research Foun-
dation (DFG) as part of the Collaborative Research Center 637
Autonomous Cooperating Logistic Processes A Paradigm Shift
and its Limitations. The work of Till Becker has been supported by
the Institutional Strategy of the University of Bremen, funded by
the German Excellence Initiative. The authors thank the Terminal
operator for providing the expert knowledge and the data used for
this study.
References
Adhau, S., Mittal, M.L., Mittal., A., 2013. A multi-agent system for decentralized
multi-project scheduling with resource transfers. Int. J. Prod. Econ. 146 (2),
646661.
Agresti, A., Franklin, C., 2009. Statistics: The Art and Science From Learning Data.
Prentice Hall, Upper Saddle River, NJ.
Albino, V., Carbonara, N., Giannoccaro, I., 2006. Innovation in industrial districts: an
agent-based simulation model. Int. J. Prod. Econ. 104 (1), 3045.
Ali, M.K.M., Kamoun, F., 1993. Neural networks for shortest path computation and
routing in computer networks. IEEE Trans. Neural Netw. 4 (6), 941954.
Aitken, J., Harrison, A., 2013. Supply governance structures for reverse logistics
systems. Int. J. Oper. Prod. Manag. 33 (6), 745764.
Allen, M.K., Helferich, O.K., 1990. Putting Expert Systems to Work in Logistics.
Council of Logistics Management, Oak Brook, IL.
Al-Mubarak, F., Canel, C., Khumawala, B.M., 2003. A simulation study of focused
cellular manufacturing as an alternative batch-processing layout. Int. J. Prod.
Econ. 83 (2), 123138.
Arizono, I., Ohta, H., 1992. Scheduling for minimizing total actual ow time by
neural networks. Int. J. Prod. Res. 30 (3), 503511.
Armbruster, D., de Beer, C., Freitag, M., Jagalski, T., Ringhofer, Ch, 2006. Autonomous
control of production networks using a pheromone approach. Phys. A 363 (1),
104114.
Baransel, C., Dobosiewicz, W., Gburzynski, P., 1995. Routing in multihop packet
switching networks: Gb/s challenge. IEEE Netw. 9 (3), 3861.
Brase, C.H., Brase, C.P., 1995. Understandable Statistics Concepts and Methods. In:
Lexington, D.C. (Ed.), 5th edition Heath and Company, Lexington D.C..
Caglie, A., Kechidi, M., Levy, R., 2012. Complex product and supplier interfaces in
aeronautics. J. Manuf. Technol. Manag. 23 (6), 717732.
Caputo, A.C., Pelagagge, P.M., 2008. Parametric and neural methods for cost esti-
mation of process vessels. Int. J. Prod. Econ. 112 (2), 934954.
Caridi, M., Sianesi, A., 2000. Multi-agent systems in production planning and con-
trol: an application to the scheduling of mixed-model assembly lines. Int. J.
Prod. Econ. 68 (1), 2942.
Carlson, K., Doyle, J., 2000. Highly optimized tolerance: robustness and design in
complex systems. Phys. Rev. Lett. 84 (11), 25292532.
Chambers, M., Mount-Campbell, C.A., 2002. Process optimization via neural net-
work metamodeling. Int. J. Prod. Econ. 79 (2), 93100.
Choi, T.Y., Dooley, K.J., Rungtusanatham, M., 2001. Supply networks and complex
adaptive systems: control versus emergence. J. Oper. Manag. 19 (3), 351366.
Choi, T.Y., Hong, Y., 2002. Unveiling the structure of supply networks: case studies
in Honda, Acura, and Daimler Chrysler. J. Oper. Manag. 20 (5), 469493.
Cramer, F., 1993. Chaos and Order: The Complex Structure of Living Things. VCH,
New York.
Daniel, V., Guide Jr., R., 1997. Scheduling with priority dispatching rules and drum-
buffer-rope in a recoverable manufacturing system. Int. J. Prod. Econ. 53 (1),
101116.
Elalouf, A., Levner, E., Chang, T.C.E., 2013. Routing and dispatching of multiple
mobile agents in integrated enterprises. Int. J. Prod. Econ. 115 (2), 444460.
El Ghaziri, H., 1991. Solving routing problems by a self-organizing map. In: Kohonen, T.,
Mkisara, J., Simula, O., Kangas, J. (Eds.), Articial Neural Networks.. North-Holland,
Amsterdam, pp. 829834.
Feng, G., Douligeris, C., 2001. A neural network method for minimum delay routing
in packet-switched networks. Comput. Commun. 24 (10), 933941.
Gallager, R., 1977. A minimum delay routing algorithm using distributed compu-
tation. IEEE Trans. Commun. 25 (1), 7385.
Gell-Mann, M., 1994. The Quark and the Jaguar. W.H. Freeman, New York.
Gell-Mann, M., 2002. What is complexity?. In: Curzio, A.Q., Fortis, M. (Eds.), Com-
plexity and Industrial Clusters. Physica-Verlag, Heidelberg, Germany, pp. 1324.
Harrington, T.C., Lambert, D.M., Sterling, J.U., 1992. Simulating the nancial impact
of marketing and logistics decisions. Int. J. Phys. Distrib. Logist. Manag. 22 (7),
312.
Haughton, M., 2013. Tackling complexities of cyclic inventory routing under con-
ditions of limited modeling and computing capacity. Int. J. Logist. Res. Appl. 17
(3), 216231.
He, Z., Wang, S., Cheng, T.C.E., 2013. Competition and evolution in multi-product
supply chains: an agent-based retailer model. Int. J. Prod. Econ. 146 (1),
325336.
He, N., Zhang, D.Z., Li, Q., 2014. Agent-based hierarchical production planning and
scheduling in make-to-order manufacturing system. Int. J. Prod. Econ. 149,
117130.
Hecht-Nielsen, R., 1990. Neurocomputing. Addison-Wesley, Reading, MA.
Helbing, D., Seidel, T., Lmmer, S., Peters, K., 2006. Self-organization principles in
supply networks and production systems. In: Chakrabarti, B.K., Chakraborti, A.,
Chatterjee, A. (Eds.), Econophysics and Sociophysics: Trends and Perspectives.
Wiley, Berlin, pp. 535558.
Hu, H., Shi, X., Stahlbock, R., Voss, S., 2012. Computational logistics. In: Proceedings
of the Third International Conference, Shanghai. Springer-Verlag, Heidelberg,
Germany.
Hlsmann, M., Illigen, C., Cordes, P., 2009. Requirements and approaches for a
complexity science-based modelling of international supply networks. In:
Pawar, K.S., Lalwani, C.S. (Eds.), Proceedings of the 14th International Sympo-
sium on Logistics. Loughborough, United Kingdom, pp. 100108.
Hlsmann, M., Scholz-Reiter, B., Windt, K., 2011. Autonomous Cooperation and
Control in Logistics: Contributions and Limitations Theoretical and Practical
Perspectives. Springer, Berlin.
Iassinovski, S., Artiba, A., Bachelet, V., Riane, F., 2003. Integration of simulation and
optimization for solving complex decision making problems. Int. J. Prod. Econ.
85 (1), 310.
Ivanov, D., Sokolov, B., 2012. The inter-disciplinary modeling of supply chains in the
context of collaborative multi-structural cyber-physical networks. J. Manuf.
Technol. Manag. 23 (8), 976997.
Johnson, R., 2011. Statistics: Principles and Methods. Wiley, Hoboken, New York.
Kchel, P., Nielnder, U., 2005. Simulation-based optimisation of multi-echelon
inventory systems. Int. J. Prod. Econ. 93 (1), 505513.
LeBaron B., 2001a. A builders guide to agent-based nancial markets. Brandeis
University Working Paper.
LeBaron, B., 2001b. Empirical regularities from interacting long-and short-memory
investors in an agent-based stock market. IEEE Trans. Evolut. Comput. 5 (5),
442455.
LeBaron, B., 2003. Calibrating an agent-based nancial market. Brandeis University
Working Paper.
LeBaron, B., 2006. Agent-based nancial markets: matching stylized facts with
style. In: Colander, D. (Ed.), Post Walrasian Macroeconomics: Beyond the
Dynamic Stochastic General Equilibrium Model. Cambridge University Press,
New York, pp. 221235.
Lee, H., Banerjee, A., 2011. Executable design and control framework using recongur-
able manufacturing holons and its evolution algorithm. Int. J. Prod. Res. 49 (5),
14051423.
Markus, A., Kis Vancza, T., Monostori, L., 1996. A market approach to holonic
manufacturing. CIRP Ann. Manuf. Technol. 45 (1), 433436.
Marques, A., Lacerda, D.P., Riehs Camargo, L.F., Teixeira, R., 2014. Exploring the
relationship between marketing and operations: neural network analysis of
marketing decision impacts on delivery performance. Int. J. Prod. Econ. (only
online at present)
McKelvey, B., Wycisk, C., Hlsmann, M., 2009. Designing an electronic auction
market for complex smart parts logistics: options based on LeBarons com-
putational stock market. Int. J. Prod. Econ. 120 (2), 476494.
Michalewicz, Z., Schmidt, M., Michalewicz, M., Chiriac, C., 2010. Adaptive Business
Intelligence. Springer-Verlag, Berlin.
 T. Becker et al. / Int. J. Production Economics 174 (2016) 156167
Nyhuis, P., Wiendahl, H.P., 2008. Fundamentals of Production Logistics: Tools and
Applications Theory. Springer, Berlin.
Pathak, S.D., Day, J., Nair, A., Sawaya, W.J., Kristal, M., 2007. Complexity and adap-
tivity in supply networks: Building supply network theory using a complex
adaptive systems perspective. Decis. Sci. J. 38 (4), 547580.
Philipp, T., de Beer, C., Windt, K., Scholz-Reiter, B., 2007. Evaluation of autonomous
logistic processes Analysis of the inuence of structural complexity. In:
Hlsmann, M., Windt, K. (Eds.), Understanding Autonomous Cooperation and
Control in Logistics The Impact on Management, Information and Commu-
nication and Material Flow. Springer, Berlin, pp. 303324.
Satake, T., Morikawa, K., Nakamura, N., 1994. Neural network approach for mini-
mizing the makespan of the general job-shop. Int. J. Prod. Econ. 33 (13), 6774.
Schikora, P.F., Godfrey, M.R., 2003. Efcacy of end-user neural network and data
mining software for predicting complex system performance. Int. J. Prod. Econ.
84, 231253.
Scholz-Reiter, B., Freitag, M., De Beer, C., Jagalski, T., 2006. Proceedings of the 5th
CIRP International Seminar on Computation in Manufacturing Engineering
(CIRP ICME6). University of Naples, Naples, Italy, pp. 317320.
Surana, A., Kumara, S., Greaves, M., Raghavan, U.N., 2005. Supply-chain networks: a
complex adaptive systems perspective. Int. J. Prod. Res. 43 (20), 42354265.
Tangpong, C., Hung, K.-T., Li, J., 2014. Agent-system co-development in supply chain
research: propositions and demonstrative ndings. J. Oper. Manag. (Online)
Tharumarajah, A., Wells, A., Nemes, L., 1996. Comparison of the bionic, fractal and
holonic manufacturing system concepts. Int. J. Comput. Integr. Manuf. 9 (1),
217226.
Tsujimura, Y., Gen, M., Ishizaki, S., 1997. Optimal routing in multiple I/O data net-
work using neural network and perturbed energy function. Comput. Eng. 33
(34), 477480.
Van Brussel, H., Wyns, J., Valckenaers, P., Bongaerts, L., Peeters, P., 1998. Reference
architecture for holonic manufacturing systems: PROSA. Comput. Ind. 37 (3),
255274.
Vidal, T., Crainic, T.G., Gendreau, M., Prins, C., 2013. Heuristics for multi-attribute
vehicle routing problems: a survey and synthesis. Eur. J. Oper. Res. 231 (1),
121.
Vimal, K.E.K., Vinodh, S., 2013. Application of articial neural network for fuzzy
logic based leanness assessment. J. Manuf. Technol. Manag. 24 (2), 274292.
167
Wang, C.J., Weissler, P.N., 1995. The use of articial neural networks for optimal
message routing. IEEE Netw. 9 (2), 1624.
Wang, Z., Browning, D.W., 1991. An optimal distributed routing algorithm. IEEE
Trans. Commun. 39 (9), 13791387.
Warnecke, H., 1993. The Fractal Company: A Revolution in Corporate Culture.
Springer, Berlin.
Widrow, B., Rumelhart, D.E., Lehr, M.A., 1994. Neural networks: applications in
industry, business and science. Commun. ACM 37 (3), 93105.
Wieselthier, J.E., Barnhart, C.M., Ephremides, A., 1994. A neural network approach
to routing without interference in multihop radio networks. IEEE Trans. Com-
mun. 42 (1), 166177.
Windt, K., Becker, T., Asenov, D., Arbabzadah, F., 2010a. A generic implementation
approach of autonomous control methods in production logistics. In: Pro-
ceedings of the 8th IEEE International Conference on Control & Automation.
Xiamen, China, pp. 629633.
Windt, K., Becker, T., Jeken, O., Gelessus, A., 2010b. A classication pattern for
autonomous control methods in logistics. CRC, University of Bremen, Bremen,
Germany (Working paper).
Windt, K., Becker, T., Kolev, I., 2010c. A comparison of the logistics performance of
autonomous control methods in production logistics. In: Sihn, W., Kuhlang, P.,
(Eds.), Sustainable Production and Logistics in Global Networks. Proceedings of
the 43rd CIRP International Conference on Manufacturing Systems. Neuer
Wissenschaftlicher Verlag, Vienna, Austria, pp. 576583.
Wong, B.K., Bodnovich, R.A., Selvi, Y., 2000. A bibliography of neural network
business applications research: 19941998. Comput. Oper. Res. 27 (1112),
10451076.
Wu, C., Barnes, D., 2012. A dynamic feedback model for partner selection in agile
supply chains. Int. J. Oper. Prod. Manag. 32 (1), 79103.
Wycisk, C., McKelvey, B., Hlsmann, M., 2008. Smart parts logistics systems as
complex adaptive systems. Int. J. Phys. Distrib. Logist. Manag. 38 (2), 108125.
Yaged Jr., B., 1973. Minimum cost routing for dynamic network models. Networks 3
(3), 193224.
Zhou, D.N., Cherkassky, V., Baldwin, T.R., Olson, D.E., 1991. A neural network
approach to job-shop scheduling. IEEE Trans. Neural Netw. 2 (1), 175179.