Journal of Manufacturing Systems 46 (2018) 264–271

Contents lists available at ScienceDirect

Journal of Manufacturing Systems

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

Review

A multi-attribute, rank-dependent utility model for selecting

dispatching rules
Isaac Pergher ∗ , Adiel Teixeira de Almeida
Universidade Federal de Pernambuco, CDSID – Center for Decision Systems and Information Development, Av. Acadêmico Hélio Ramos, s/n – Cidade
Universitária, Recife, PE, CEP 50.740-530, Brazil

a r t i c l e i n f o a b s t r a c t

Article history: The use of dispatching rules is one of the most popular solutions used when handling dynamic scheduling
Received 26 July 2017 problems. However, how best to select a suitable dispatching rule or rules depends on the stochas-
Received in revised form tic behavior of the performance of the system with respect to multiple conflicting attributes. Previous
14 December 2017
studies based on multi-objective optimization approaches have not explored the subjective informa-
Accepted 12 January 2018
tion presented by the decision-maker in relation to the multiple impacts that arise from the decisions
made using the dispatching rule. This study takes into consideration a rank-dependent utility approach
Keywords:
combined with a multi-attribute utility theory to identify the best dispatching rule for dynamic job
Dispatching rules
Simulation
shop environments. An additive, non-expected, multi-attribute utility function is modeled to represent
Rank-dependent utility the decision-maker’s preferences and attitude to risk over the multiple stochastic consequences associ-
Multi-attribute utility theory ated with each simulated heuristic. A numerical simulation with realistic data from a furniture hardware
Maut industry is performed to demonstrate the usefulness of the proposed decision approach, and the practical
Rdu issues that emerge are discussed.
© 2018 Published by Elsevier Ltd on behalf of The Society of Manufacturing Engineers.

Contents

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
2. Multi-attribute rank dependent utility model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
3. Numerical application and results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
4. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
5. Concluding remarks and future research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270

1. Introduction

Dispatching rules are an effective solution for dynamic schedul-

ing problems. A large number of these heuristics has been rule by taking into consideration the uncertainties of the system
developed over the past decades for dealing with various types and the trade-off assessment among conflicting attributes is often
of manufacturing environments. However, depending on the dis- a difficult task.
patching rule implemented in the system, the results achieved can In the case of dynamic environments in which the schedule
be satisfactory for one performance attribute (or objective) but should be updated as jobs arrive and are completed, recent studies
poor for another. As the aim of scheduling decisions is to maxi- have focused on developing new dispatching rules or on compar-
mize the performance of multiple attributes, selecting a suitable ing performances under various experimental conditions. In some
studies, the best results are presented for each possible combina-
tion of various factors and performance objectives. Xiong et al. [1]
∗ Corresponding author. investigated the influences of the due-date tightness and percent-
E-mail addresses: eng.isaac@hotmail.com (I. Pergher), almeida@cdsid.org.br age of jobs. These have extended technical precedence constraints
(A.T. de Almeida). on the scheduling of mold and die manufacturing systems and fit-

https://doi.org/10.1016/j.jmsy.2018.01.007
0278-6125/© 2018 Published by Elsevier Ltd on behalf of The Society of Manufacturing Engineers.
 I. Pergher, A.T. de Almeida / Journal of Manufacturing Systems 46 (2018) 264–271
ting these in with batch releases. Using an analysis of variance, the
performance of 20 dispatching rules (including four new rules that
this paper proposes) have been evaluated with regard to the total
tardiness and percentage of tardy jobs. Jia et al. [2] addressed the
problem of scheduling lots and configuring machines in semicon-
ductor assembly and test systems. They did this by using the Greedy
Randomized Adaptive Search Procedure (GRASP) integrated with a
discrete event simulation. Various heuristics have been assessed as
to four objectives: weighted sum of key device shortages; weighted
throughput; number of machines used; and makespan. In addition,
they proposed three new rules based on the GRASP logic for setting
up machines and assigning lots to them. Sharma and Jain [3] simu-
lated nine dispatching rules while considering the effects of various
shop utilization levels and sequence-dependent setup times. The
performance of the system is measured in terms of the makespan,
mean flow time, maximum flow time, mean tardiness, maximum
tardiness, number of tardy jobs, total setups and mean setup time.
Hübl et al. [4] studied the impacts of 12 dispatching heuristics on
the average production lead time in view of the covariance level
between the processing time and production lead time.
Various methodologies have been used to address the problem
of selecting dispatching rules from a multidimensional perspective.
El Bouri and Amin [5] developed a multi-criteria decision approach
that integrates ordered weighted averaging (OWA) with data envel-
opment analysis (DEA). The OWA operator is used to assess the
decision-maker’s (DM’s) optimism level with regard to the simula-
tion results of 20 dispatching rules. This is followed by using a DEA
model for aggregating the OWA scores to identify the most efficient
solution for each combination of scenario and due-date tightness
level. The performance criteria for their study were: makespan;
mean flow time; machine utilization; mean tardiness; mean sum
of earliness and tardiness; percentage of jobs completed on time;
and maximum job lateness. Parthanadee and Buddhakulsomsiri [6]
presented an approach for identifying the most appropriate set of
dispatching rules for scheduling the canned fruit manufacturing
process. Using a simulation, the impacts of nine scheduling rules
on the flow time, number of tardy jobs, and average tardiness are
explored. The performance data from the simulation experiment
are compared by means of Tukey’s multiple comparisons test and
the multi-criteria TOPSIS method.
Fuzzy logic, in particular, has been applied for coping with fac-
tors such as the ambiguity and uncertainties that inherently exist in
multi-objective scheduling problems. Asadi [7] focused on single-
machine scheduling problems with the objectives of minimizing
the makespan, total completion time, and total weighted comple-
tion time. Two models based on the likelihood profile approach
and fuzzy mixed integer nonlinear programming were proposed
to incorporate the fact that workers learn quickly on the job how
to exercise their duties and that this reduces processing times.
Abd et al. [8] proposed a Taguchi method integrated with fuzzy
logic for the multi-objective scheduling problem in robotic flexible
assembly cells. This approach identifies the best solution for four
control scheduling factors, namely, the sequencing rule, dispatch-
ing rule, cell utilization, and due-date tightness. These are then
used to optimize the performance of the cell with respect to the
makespan; total tardiness; and the number of tardy jobs. Huang and
Süer [9] developed a fuzzy dispatching-rule-based genetic algo-
rithm (GA) to find the best combination of rules that provides the
highest overall fuzzy satisfaction level. This fuzzy model consid-
ers the consequences with respect to the makespan, average flow
time, maximal tardiness, and total tardiness. To determine the best
schedule, a two-level fuzzy evaluation approach is performed. First,
an individual fuzzy membership function is established to meet
each performance objective. Next, the fuzzy operator aggregates
these functions to obtain the overall fuzzy satisfaction level.
265
The combinatorial nature of several scheduling problems allows
the use of population-based metaheuristics such as genetic pro-
gramming (GP) and GA. Cheng and Huang [10] combined GA with
a distributed release time control mechanism to minimize the total
earliness and tardiness of job completion in unrelated, parallel
machine scheduling problems. In addition, they developed a mixed
integer linear programming model so as to evaluate the perfor-
mance of the proposed algorithm. Ðurasević et al. [11] investigated
the use of GP for scheduling parallel, unrelated machines using
arbitrary performance criteria. The performance of the proposed
heuristic was compared with other GP methods for evolving prior-
ity functions in relation to weighted tardiness, weighted number of
tardy jobs, flow time, and makespan. Nguyen et al. [12] presented
four new multi-objective GP-based hyperheuristic methods for
evolving scheduling rules in job-shop environments. In their study,
the Pareto-front analysis is applied to assess the trade-offs among
makespan, normalized total weighted tardiness, and mean abso-
lute percentage error. Quin et al. [13] developed a multi-objective
method for a semiconductor wafer fabrication system that com-
bines GP with a simulation to generate composite dispatching rules.
They compared the performance of the proposed method against
traditional dispatching rules while taking into consideration nine
simulated scenarios. The best results are identified for each of
the following measures: average delivery time; wafer cassette
throughput; mean processing cycle time; and due-date satisfaction
rate. In a similar production environment, Chang et al. [14] pro-
posed a GA-based simulation optimization model to dynamically
select appropriate rules for three scheduling levels: lot dispatch-
ing, batch dispatching, and automated guided vehicle dispatching.
The response surface methodology was applied in order to opti-
mize the GA parameters. Zhang et al. [15] explored incorporating
GA into tabu search so as to improve the efficiency of a schedule
and to maintain its stability in job shop environments where thee
are random job arrivals and machine breakdowns. The schedule
efficiency factor is evaluated based on the makespan, while the
starting time deviations for all the jobs between the new schedule
and the original schedule represent the schedule stability factor.
The use of artificial intelligence and machine learning methods
usually provides better results as compared with traditional dis-
patching heuristics. However, the large computational effort and
time required to deal with various shop-floor conditions can make
such methods impractical for real-time scheduling problems. Thus,
traditional rules seem appealing owing to the low computational
effort they require and their ease of implementation.
The above brief literature review indicates a lack of studies
that take into consideration the DM’s preferences and his/her
attitude toward risk owing to uncertainties inherent to the mul-
tidimensional performance of the job shop production system.
Moreover, current models identify the best performing heuristics
without assessing the trade-offs among probabilistic consequences
associated with multiple attributes. As a result, the solution rec-
ommended may be inadequate, when the DM is willing to forego
the advantages of a high level of performance in one attribute to
compensate for the poor performance of another attribute. Hence,
an approach that explores the DM’s trade-offs and preference
judgments in relation to the uncertainties of multiple conflicting
attributes is required to address such problems.
In this paper, we present a multi-attribute rank-dependent util-
ity model in order to identify the most appropriate dispatching
rule(s) for dynamic job shop environments. This non-expected
utility model uses a discrete event simulation to assess the perfor-
mance of various heuristics in relation to the number of tardy jobs,
mean flow time, and mean tardiness. Given these results, the rank-
dependent utility (RDU) [16] approach is applied to model the DM’s
behavior under risk. The RDU extends Expected Utility Theory (EUT)
by incorporating the DM’s subjective perception of probabilities,
266 I. Pergher, A.T. de Almeida / Journal of Manufacturing Systems 46 (2018) 264–271

which are modeled as nonlinear decision weights on each prospec-

tive outcome. These decision weights are obtained by transforming 1. Typify the DM and other actors
the probabilities of ranked outcomes using a weighting function,
which is very useful for representing the DM’s behavior toward
risk. The RDU value of the prospects is then aggregated into an addi-
tive non-expected multi-attribute function, in accordance with the 2. Structure the set of alternatives
independence condition required by Multi-Attribute Utility The-
ory (MAUT) [17]. For multi-criteria decision problems involving
probabilistic consequences, MAUT allows compensations among
decision attributes based on the DM’s preferences and behavior. 3. Simulate the alternatives
Therefore. MAUT incorporates the subjectivity and stochastic char-
acteristics inherent to this scheduling problem. MAUT has been
applied to support managerial decisions related to the newsven-
dor problem with partial backlogging [18] and the planning- and 4. Estimate the set of discrete prospects
control-parameter definition problem for constant work in process
(CONWIP) production systems [19].
The remainder of this article is organized as follows. In Section
2, a multi-attribute non-expected utility model for the dispatching
5. Elicit unidimensional utility and
rule selection problem is presented. The case study and experimen-
weighting functions
tal results are presented in Section 3 and discussed in Section 4.
Conclusions and future research directions are presented in Section
5.
6. Calculate RDU
2. Multi-attribute rank dependent utility model
 
2.1 Let A = a1 , a2 , . . .ai , . . ., an be a discrete set with n poten-
7. Elicit multi-attribute utility functions
tial options of dispatching rules available to the DM. Each ai is
associated with probabilistic occurrences that represent the per-
formance of the system with respect to the number of tardy
jobs (N), mean tardiness (T̄ ), and mean flow time (FT ). In order
8. Conduct sensitivity analysis
to incorporate these  uncertainties into the decision model, the 
prospect qi,j ∈ Q = xij1 , pij1 ; xij2 , pij2 ; . . .; xijb , pijb ; . . .; xijm , pijm ;
is estimated for each combination of ai and attribute j, while taking
into consideration m (b = 1, 2, . . ., m) possible outcomes xijm that
9. Final solution
are ranked from best to worst and their corresponding probability
pijm . The proposed model integrates the RDU and MAUT approaches
in order to identify the option with the highest non-expected multi- Fig. 1. Structure of the proposed model.
attribute utility. This model assumes that the DM’s preferences
between prospects are nonlinear in the probability dimension and
the multi-attribute value of ai is decomposed into an additive form. To address the uncertainties of qi,j using RDU, the unidimen-
The structure of the decision approach in the context of selecting sional utility uj and probability weighting functionswj should be
a dispatching rule(s) is shown in Fig. 1. It is based on a framework elicited. In this study, two distinct procedures are applied to elicit
presented by de Almeida et al. [20] that has been widely applied these functions. First, we apply the trade-off method (TO-method)
for building multi-criteria decision making (MCDM/A) models in proposed by Wakker and Deneffe [22] to assess the DM’s utility for
the context of risk analysis, reliability and maintenance. each of the attributes. This method was selected because it does not
In the first step, the actors involved in this decision-making pro- use questions based on changes in probabilities in order to obtain
cess should be identified. The DM plays the central role as he/she the preference information, and it provides valid utilities under the
provides preference information regarding multiple consequences RDU concept. Moreover, the TO-method was also the starting point
in such a way that the associated objectives of the company can of non-parametric elicitation models presented in previous studies
be achieved. Other relevant actors for this process are a decision such as Munier [23], Krawczyk [24], Van de Kuilen and Wakker [25],
analyst and experts. The decision analyst provides methodological Abdellaoui [26], and Bleichrodt and Pinto [27]. The elicited utilities
support for applying the proposed model, while experts provide the are then used as the inputs of the Midweight method [25] for assess-
decision analyst with technical information about the production ing wj functions. The advantage of using this method is that there
system. is no prior assumption about the form of wj (Kuilen and Wakker
The subsequent step involves determining the dispatching [25]), which would unnecessarily require more information from
rules that will be included in the set of alternatives. There- the DM. The probability weighting function increases strictly as wj :
after, a simulation experiment [21] should be developed to obtain [0,1] → [0,1] and satisfies wj (0) = 0 and wj (1) = 1.
 
the performance values of each ai with respect to the decision The RDU value of a prospect qi,j , denoted by VRDU i,j qi,j ,
attributes considered in this scheduling problem. Given the sim- depends on both the uj function and decision weight b of each
 
ulation results, one can build a set of prospects Q to represent the outcome xijb ranked from best to worst. VRDU qi,j can be obtained
behavior of probabilistic occurrences associated with FT , T̄ , and N. by solving (1).
For this purpose, qi,j is determined from the histogram hij ∈ H of the
simulation outputs produced in R replications, where the midpoint   m
of m class intervals and the relative frequency of hij are denoted by VRDUi,j qi,j = b u (xb ) (1)
b=1
xijm and pijm respectively.
 I. Pergher, A.T. de Almeida / Journal of Manufacturing Systems 46 (2018) 264–271 267

Table 1
Summary of processing times and standard lot sizes.

klt Lot size Processing time (h)

S1 S2 S3 S4 S5 S6 S7

k11 1000 Lognormal (4.7; 1.1) Gamma (1.8; 0.4)

k12 1500 Lognormal (5.5; 1.5) Gamma (1.7; 0.4)
k13 3000 Lognormal (6.1; 1.9) Gamma (2.4; 0.8)
k21 200 Lognormal (2.8; 0.9) Lognormal (1.5; 0.5) Lognormal (3.3; 0.98) Gamma (2.0; 0.5)
k22 250 Lognormal (4.2; 1.3) Lognormal (1.9; 0.5) Lognormal (5.0; 1.0) Gamma (2.1; 0.5)
k23 350 Lognormal (3.5; 1.1) Lognormal (2.4; 0.5) Lognormal (7.5; 1.8) Gamma (1.1; 0.4)
k31 6000 Normal (3.0; 1.2) Normal (0.8; 0.3) Lognormal (2.9; 0.9) Gamma (2.2; 0.5)
k32 16000 Normal (3.3; 1.3) Normal (1.7; 0.8) Lognormal (3.0; 1.5) Gamma (2.7; 0.9)
k33 6500 Normal (2.8; 0.9) Normal (0.9; 0.5) Lognormal (1.9; 0.7) Gamma (1.3; 0.4)
k41 25000 Normal (2.0; 0.9) Lognormal (4.2; 1.7) Gamma (3.5; 1.1)
k42 25000 Normal (2.5; 1.0) Lognormal (2.8; 0.8) Gamma (3.1; 1.1)
k43 25000 Normal (3.1; 1.2) Lognormal (3.3; 1.3) Gamma (1.2; 0.9)
k51 200 Lognormal (1.8; 0.5) Lognormal (3.1; 1.2) Gamma (0.9; 0.2)
k52 200 Lognormal (2.8; 0.8) Lognormal (2.1; 0.9) Gamma (1.9; 0.8)
k53 200 Lognormal (4.2; 1.8) Lognormal (3.1; 1.3) Gamma (2.1; 0.9)

(mean value, standard deviation).

whereb = wj (pb + pb+1 + . . . + pm ) − wj (pb+1 + . . . + pm ) , b = Table 2

1, 2, . . ., m − 1, and m = wj (pm ). When wj (p) = p for all p ∈ (0, 1),
 
VRDUi,j qi,j becomes identical to the EUT.
 
After assessing VRDUi,j qi,j , the next step is to obtain an overall
utility function that aggregates the RDU value of attributes based on
the DM’s preferences for trade-off judgments. For this type of pref-
erence modeling, MAUT was chosen because it presents a very solid
and consistent axiomatic framework for decisions involving multi-
ple probabilistic consequences. The model proposed assumes that
the DM’s preference structure is consistent with the additive inde-
pendence condition explained in reference [17], which is checked
during the elicitation protocol with MAUT. This consistency has
been often found in practice as supported by other practical appli-
cations of MAUT [18], [28] and [29].
For a DM with an additive independence condition, the best
compromise corresponds to the alternative ai that maximizes the
value of the multi-attribute non-expected utility function given in
(2). The parameters FT , T̄ and N are the scaling constants elicited
via comparison of the pay-off lotteries in a probabilistic context
[17,20]. The following property holds: FT + T̄ + N = 1.
MRDU(ai ) = [FT¯ VRDU i,FT¯ (qi,FT¯ ) + T̄ VRDU i,T̄ (qi,T̄ ) + N VRDU i,N (qi,N )]
The last step involves examining the sensitivity of the overall
utility results with respect to variations in the scaling constants. For
this purpose, we use Kendall’s ␶ nonparametric statistic test [30] to
assess the significance of the differences between the rankings of
alternatives obtained by using (2) and the rankings generated by
the Monte Carlo simulation. This step also includes a final report
that specifies the main assumptions of the modeling process and
its impact on the proposed solution.
3. Numerical application and results
In this section, a job shop system that manufactures furni-
ture hardware kits under the make-to-order policy was modeled.
The production environment comprises an order entry unit and
seven workstations (see Fig. 2). The order entry unit deter-
mines the delivery date for confirmed customer orders. A
workstationSh (h = 1, . . ., 7) consists of a set of equipment that
is set up to perform the production operations. The system has
fifteen types of customer orders that are divided into five main
product families (F1 , F2 , . . ., Fl , . . ., and F 5 ) with three kit-types
klt (t = 1, 2, and3) in each Fl .
As shown in Fig. 2, the production routing of F1 begins at S3 with
the material mixing and injection molding processes. Workstation
7 (S7 ) is an assembly unit where the parts from S3 are packed in kits
Summary of workstations characteristics.
Sh Work shift Capacity
(Hours/day) Increment Number of orders
1 16 Not adjusted 2
2 1
3 3
4 2
5 1
6 3
7 3
with other parts produced by others. After a quality inspection, the
finished kits are transferred for shipping and then dispatched to
customers. The kits belonging to F2 enter the system at S1 —which
is a machining unit that is used to cut and debur raw materials
from the warehouse. Then the semi-finished parts are transferred
to S2 where they are welded together to form an insert unit. In S3 ,
this insert is placed on core pins in the mold cavity and the melted
thermoplastic material is injected under pressure around it. The
parts are then transferred to S7 for the final assembly.
(2) The production of F3 starts with the metal stamping process at
S4 . At this stage, progressive dies are used to produce a finished
part without performing any secondary operations. Semi-finished
stamped parts from S4 are transferred to S5 , S6 , and S7 for the
threading process, galvanic surface treatment, and final assembly,
respectively. The job routing of F4 and F5 through the system begins
at S4 and S1 respectively, and both families require operations at S6
and S7 .
The statistical parameters of processing times and standard lot
sizes (in units of number of finished kits) are summarized in Table 1.
In relation to the shop floor capacity, each Sh can deal with a
maximum number of customer orders simultaneously as given in
Table 2. The system runs two 8.5 h shifts per day and shuts down
for 30 min at the end of a shift while its capacity remains constant.
In the first step (Fig. 1), the production manager and the oper-
ational analyst of each workstation were characterized as the DM
and specialists. Next, a simulation study was developed using Micro
Saint Sharp software to examine the performance of this produc-
tion system under various dispatching rules. The development of
the simulation model is based on the following assumptions:
• Raw materials are readily available.
• Within the composition of a complete kit, each klt requires only
a single type of part produced at Sh .
• Kits are produced in a standard lot size.
268 I. Pergher, A.T. de Almeida / Journal of Manufacturing Systems 46 (2018) 264–271
Fig. 2. Job shop system under study.
• The transfer time between two workstations is not taken into
consideration.
• Set-up times are sequence independent and included in the pro-
cessing times.
• No stable bottleneck workstation can be identified.
• Each customer order comprises only one kit type klt .
• Customers do not accept partial deliveries.
Order inter-arrival times at the order entry stage are random
and follow an exponential distribution with a mean of 0.753 h and
the type of customer order is determined by a uniform distribu-
tion [1,15]. This arrival–processing time combination results in an
overall workstation utilization of 0.85.
The klt due date is assigned in accordance with the total work
content method [31] presented in (3).
 
DDklt = ATklt + ˛TPTklt
where DDklt is the due date of klt ; TPTklt represents the total process-
ing time; ATklt , the arrival time; and ␣, the due-date tightness factor.
In this application, ␣ was set to fulfil approximately 28.6% of late
orders when the first-in-first-out rule (FIFO) is used to prioritize
orders that are awaiting processing at Sh .
Table 3 presents the dispatching rules considered in the simula-
tion experiment and their priority index Iklt ,Sh calculated for every
order klt that awaits processing at Sh . The set of rules tested includes
those that can be easily implemented in the system under study. In
Table 3, PT, T, and RT denote the process time, current time in the
system, and sum of the remaining processing times, respectively.
With respect to the cost over time (COVERT) and apparent tardi-
ness cost (ATC) rules, the values of K = 1, 1.3, 1.5, and 1.7 are used
in the simulation for both rules.
Each rule is replicated 30 times and the system is observed until
2000 orders have been completed. The warm-up period is set to
48 h. After simulating the rules listed in Table 3, the set of prospects
Q can be established by applying the procedure described in step
4. In step 5, the unidimensional utility functions uj are assessed in
accordance with the TO-method [22]. The elicitation of a standard
sequence of positive outcomes is achieved by defining reference
outcomes y < Y and a starting outcome x0 . The analyst then asks
the DM to express an outcome x1 , in which he/she is indifferent
between (x1 , p; y, 1 − p) and (x0 , p; Y, 1 − p) for a given constant
value of probability p. After x1 has been defined, the analyst can
elicit the value of x2 that causes the DM to be indifferent between
the lotteries (x2 , p; y, 1 − p) and (x1 , p; Y, 1 − p). The indifferences
established by the TO-method imply the equality u (x2 ) − u (x1 ) =
u (x1 ) − u (x0 ). We can continue to determine indifferences between
lotteries until xo is equal to the value for which u (x∗ ) = 1. In this
application, p was set as 0.5, and 16 indifference values were estab-
lished for each attribute.
When assessing the flow time attribute, the starting parameters
were y = 9,Y = 9.5, andx0 = 10. Using a curve-fitting procedure over
the utility points that had been elicited, a monotonically decreasing
exponential utility function, given by (4), was adjusted. This utility
function translates the DM’s conservative attitude in relation to
the risks that are associated with the lotteries evaluated within the
consequence space csFT = [10; 50] h.
uFT = 3.19249e−(0.130448xijm ) (4)
(3)
For the number of tardy jobs, a monotonically decreasing expo-
nential utility function (5) represents the DM’s risk-averse attitude
for lotteries with y = 70,Y = 85,x0 = 100, and csN = [100; 1200]
orders.
uN = 1.25483e−(0.00219468xijm ) (5)
The utilities associated with the tardiness attribute shows a
decreasing risk aversion attitude has been modeled as shown in (6),
for which smaller results of tardiness are significantly more desir-
able than higher outcomes. This exponential function was obtained
by applying y = 0, Y = 0.4, x0 = 0.8, and csT̄ = [0.9; 90] hours.
uT̄ = 0.990507e−(0.0420056xijm ) (6)
Using the midweight method, we assess the probability cor-
responding to five weights wj−1 (0.125),wj−1 (0.25), wj−1 (0.5),
wj−1 (0.75), and wj−1 (0.875). Starting with wj−1 (0.5), the DM
is asked to specify the value of p ∈ [0, 1] such that he/she
is indifferent between x1 ∼ (x2 , p; x0 , 1 − p), where x1 is the
utility midpoint of x2 and x0 . Having obtained wj−1 (0.5) = p,
we elicit wj−1 (0.25) = p for p ∈ [0, p] based on the lotter-
ies (x1 , p; x0 , 1 − p) ∼ (x2 , p ; x0 , 1 − p ). In this application, the
sequence of weights elicited were wj−1 (0.25), wj−1 (0.125),
wj−1 (0.75), and wj−1 (0.875). In order to translate the DM’s
 I. Pergher, A.T. de Almeida / Journal of Manufacturing Systems 46 (2018) 264–271 269

Table 3
Dispatching rules tested.

Rules Priority index Iklt ,Sh References

FIFO Iklt ,Sh = min(AT ) [32–35]

EDD Iklt ,Sh = min(DD)
STP Iklt ,Sh = min(PT )
MSLACK Iklt ,Sh = min(DD − T − RT )
LWKR Iklt ,Sh = min(RT )
MINSLK Iklt ,Sh = min(DD − PT ) [6]
MDD Iklt ,Sh = min (max(DD, T + PT )) [36]
−RT +PT ))
ATC Iklt ,Sh 1
= max( PT e(−( max(0,(DD−T
KRT
))) [37]
−RT ))
COVERT Iklt ,Sh 1
= max( PT max(0, (1 − ( max(0,(DD−T
KRT
)))) [38]

FIFO: first-in- first-out; EDD: earliest due date; STP: shortest processing time; MSLACK: minimum slack time; LWKR: Least work remaining; MINSLK: minimum slack; MDD:
modified due date; ATC: apparent tardiness cost; and COVERT: cost over time.

Table 4 The results obtained on applying the MRDU function in (2) so as

Results obtained using the multi-attribute rank-dependent utility model.
to aggregate the RDU value of each qi,j are summarized in Table 4.
 
ai Rules VRDU i,j qi,j MRDU (ai ) Rank As observed in the last column of Table 4, COVERT with K = 1.7
offers the most desirable trade-off among the multiple attributes
¯
FT N T̄ and the highest non-expected utility among the set of alterna-
1 FIFO 0.2151 0.4039 0.5452 0.432 15◦ tives examined, MRDU (a15 ) = 0.558. The proposed solution for this
2 EDD 0.2630 0.4847 0.6295 0.509 6◦ scheduling problem was found by considering the DM’s preferences
3 SPT 0.2861 0.6883 0.3690 0.517 4◦
and risk attitude regarding the stochastic behavior of the perfor-
4 LWKR 0.2690 0.6981 0.2675 0.482 11◦
5 MINSLK 0.2285 0.4537 0.5292 0.452 13◦ mance values associated with flow time, number of tardy jobs, and
6 MSLACK 0.2349 0.4494 0.5697 0.466 12◦ tardiness.
7 MDD 0.2510 0.4980 0.5993 0.503 9◦ In step 8, a sensitivity analysis was conducted to verify the
8 ATC K = 1 0.2308 0.3966 0.5814 0.443 14◦ robustness of the final solution in relation to variations in the scal-
9 ATC K = 1.3 0.2464 0.5233 0.5919 0.512 5◦
10 ATC K = 1.5 0.2575 0.4897 0.6174 0.507 7◦
ing constants. Using MATLAB software version R2015a, a total of
11 ATC K = 1.7 0.2957 0.4771 0.6573 0.520 3◦ 1,000.000 independent ranks of alternatives were randomly gener-
12 COVERT K = 1 0.2627 0.4976 0.5981 0.504 8◦ ated using a uniform distribution with parameters [0.75j ; 1.25j ],
13 COVERT K = 1.3 0.2378 0.5425 0.5297 0.498 10◦ where j is the scaling value previously assessed. Kendall’s ␶ test
14 COVERT K = 1.5 0.2712 0.5817 0.5653 0.535 2◦
[30] was then implemented for testing the null hypothesis: no cor-
15 COVERT K = 1.7 0.2822 0.6451 0.5379 0.558 1◦
relation exists between the original ranking presented in Table 4
and each ranking randomly obtained at this stage.
For 86.42% of the rankings simulated, the ␶ statistic exceeded
subjective perception regarding probabilities assessed in a non-
0.800. The remaining 13.58% of the results yielded a ␶ between 0.71
parametric manner, we adjust Prelec’s two-parameter weighting
and 0.799. The null hypothesis was rejected for the p-value ≤0.0001
function [39] given by (7). Prelec’s weighting function was chosen
in both cases, which indicates that a15 remains invariant within the
because it satisfies a property called subproportionality and it is
range (±25%) that is simulated for FT , T̄ , and N. This result reflects
well-suited for very small and very large probabilities as explained
the robustness of the solution recommended for the problem of
by Wakker [40].
dispatching rule selection in relation to the DM’s preference among
˛ the set of alternatives.
w (p) = e−ˇ(− ln p) (7)

For the tardiness attribute, the increasing convex function with

4. Discussion
␣ = 0.69 and ˇ = 1.55 represents a DM’s risk aversion attitude for
lotteries involving small probabilities. With respect to the num-
In this study, the most appropriate dispatching rule for the sys-
ber of tardy jobs, a combination of pessimism and over-weighted
tem under study was found by taking into consideration the DM’s
small probabilities were modeled by means of the inverse s-shape
trade-off judgments regarding probabilistic occurrences associated
curve (␣ = 0.38 and ˇ = 0.93). In a manner similar to the tardiness
with multiple attributes. When compared witha14 , MRDU (a14 ) =
attribute, a convex function was adjusted (␣ = 0.73 and ˇ = 1.93)
0.535; the compromise solution enables a better non-expected util-
to translate this DM’s risk attitude for probabilistic consequences
ity on the FT and N attributes, whereas a14 reduces the average
associated with the flow time attribute.
time of tardy orders. In accordance with the preference information
Once the uj and wj functions have been assessed, one can cal-
modeled in (2), this DM foregoes the advantage in the tardiness
culate the RDU value for each prospect qi,j ∈ Q to reflect the DM’s
attribute yielded by a14 to obtain the advantage of reducing the
  in the probability dimension. Table 4 sum-
nonlinear perception
average flow time and number of tardy jobs. As seen in Table 4,
marizes VRDUi,j qi,j which was obtained by using (1). In step
a11 —which is ranked in the third position—produces the best result
7, the scaling constants j were elicited in accordance with the
for the T̄ attribute. In addition, the average time that a customer
multi-attribute elicitation protocols of Keeney and Raiffa [17]. This
order spends in the system is lower than that for the alternatives
process consists of obtaining the indifference value between lotter- a15 and a14 (VRDU = 0.2957 in contrast with VRDU = 0.2822
ies with three-dimensional probabilistic consequences N, T̄ , FT . 11,FT 15,FT

At different stages of the elicitation, these lotteries were also used
to assess the utility independence and conditional utility func-
tions prior to the assessment of the scaling constants. As a result,
the additive independence condition was validated, and the val-
ues of the scaling constants were assessed as follows: N = 0.498,
T̄ =0.371, and FT = 0.131.
and VRDU = 0.2712). However, these performance values are
14,FT
not compensated for by the DM, as a15 provides a better perfor-
mance with respect to the number of tardy jobs.
When the performance evaluation is based on a single dimen-
sion of this scheduling problem,a11 ,a4 , and a11 provide the best
performance for FT , N, and T̄ respectively. However, there are
270 I. Pergher, A.T. de Almeida / Journal of Manufacturing Systems 46 (2018) 264–271
some disadvantages to the performance of the system caused by
the selection of these alternatives. For instance, a11 reduces the
non-expected utility of the N attribute by 26%, whereas a4 has the
negative impact of increasing the tardiness dimension by 50%. The
use of MAUT allows the translation of the DM’s trade-off judg-
ments among multiple attributes to a substitution rate [17], which
is represented by j . Thus, the values of j elicited through ques-
tions involving probabilistic choices cannot be simply interpreted
as the relative degree of importance of the attributes. Although
Kendall’s ␶ test demonstrated significant effects for a15 within the
[0.75j ; 1.25j ] range, a careful evaluation of the MAUT assump-
tions, especially those required by the additive form, is needed.
This is important for finding a multi-attribute utility function that
is representative of the DM’s preference structure.
Assuming that the DM’s subjective perception of probability is
one of the factors that affects the selection of dispatching rules,
the use of RDU allows a consideration of the deviations from the
expected utility explained by Allais’ paradox [41]. In RDU, the util-
ity function uj and the weighting function wj are the subjective
parameters that characterize the DM’s probability-dependent risk
attitude. wj can be usefully incorporated into the evaluation of each
prospect, as human behavior puts too much weight on small and
very large probabilities. Based on the wj functions elicited, note
that the DM’s preferences are dependent on the value of the prob-
ability of obtaining an outcome xijm . When this decision problem
is modeled using EUT where wj (p) = p for all p ∈ (0, 1), the rank-
ing of the alternatives obtained has a moderate correlation (␶ = 0.7,
and p-value ≤0.0001) with that given in Table 4. Although a15 and
a14 remained in the top position, a9 is ranked in the third position
(instead of a11 ), and the best unidimensional results are obtained
with a3 ,a4 , and a2 for FT , N and T̄ respectively.
Although there is a multiple objective nature in the problem of
selecting dispatching rule(s) for a dynamic job shop system, few
papers that involve applying compensatory MCDM/A methods can
be found. Several studies focus on the Pareto-front analysis, Fuzzy
Set Theory, and weighted approaches so as to deal with various con-
flicting performance objectives. However, these techniques do not
take into consideration the DM’s preferences and his/her nonlinear
perception in the probability dimension. For a DM with an additive
independence condition [17], the proposed model enables trade-
offs amongst attributes to be analyzed in order to recommend a
dispatching rule(s) that reflects the target performance that the DM
intends to meet. In addition, it is economically feasible to evaluate
a large group of alternatives, as the DM preferences are elicited in
the consequence space of each attribute.
In another study involving the trade-off of probabilistic conse-
quences, Pergher and de Almeida [19] integrated a discrete event
simulation with MAUT so as to identify appropriate settings for
the planning parameters in production systems controlled by CON-
WIP [42]. Owing to the characteristics of the problem, where the
work-in-process and finished-goods inventory do not often cause
rare occurrences, the uncertainties associated with the planning
parameters have been modeled in accordance with the EUT. How-
ever, that approach could not be applied in the current study. So,
this study evaluates the deviations from the linearity of the EUT
assuming that the dispatching rules may cause rare probabilistic
occurrences in the performance of FT , T̄ , and N. Therefore, using
RDU is more appropriate for this particular decision context.
In this non-expected utility model, it is assumed that Prelec’s
function [39] is able to translate the DM’s risk attitude into the
probability dimension. Furthermore, the DM’s preference struc-
ture is consistent with MAUT requirements related to using an
additive form to aggregate all criteria in a single analytic function.
To keep the experimental setting reasonable, we did not consider
the preferences of multiple decision-makers. As to the simulation
experiment, the due date of customer orders was determined by
the total work content method [31], with ␣ being set to achieve
approximately 28.6% of late orders under FIFO. In addition, a pure
job-shop system with an independent setup time and a constant
production lot size was simulated to demonstrate the usefulness
of the proposed model. Hence, future research could explore other
types of job-shop production systems, for example, it could con-
sider dependent setup times.
5. Concluding remarks and future research
In this paper, we present a non-expected multi-attribute util-
ity model in order to identify the most suitable dispatching rule(s)
for a dynamic job shop system. A numerical application using real-
istic data from a furniture hardware industry was undertaken to
demonstrate the applicability of the proposed model. Fifteen dis-
patching rules were evaluated under three performance attributes:
number of tardy jobs, mean flow time, and tardiness. The alter-
native a15 (COVERT with a parameter K = 1.7) is the compromise
solution recommended for the problem of dispatching rule selec-
tion. According to the DM’s preferences which was modeled within
the theoretical framework of RDU and MAUT, a15 provides the most
satisfactory compensations across multiple attributes and the high-
est non-expected utility among the set of alternatives. The use of
the RDU approach coupled with MAUT allows the DM’s trade-off
judgments to be incorporated among multiple attributes and the
attitude towards risk caused by the uncertainties of the perfor-
mance of the system in a non-expected multi-attribute function.
Once implemented, the performance of any set of dispatching
rules can be assessed while considering the various order inter-
arrival times, due date assignment method, lot size, product mix,
and processing time parameters. In addition, this model can be
adapted to deal with other types of production environments and
performance attributes. However, the additive independence prop-
erties required by the MRDU function in (2) should be validated.
Acknowledgements
This work had partial support from CNPq (the Brazilian Research
Council). The CNPq was not involved in the research for nor in
writing this paper.
The authors are most grateful to the Editor and the anonymous
reviewers for considering the potential of a previous version of this
paper and for encouraging its improvement, by offering an insight-
ful critique and valuable suggestions.
References
[1] Xiong H, Fan H, Jiang G, Li G. A simulation-based study of dispatching rules in
a dynamic job shop scheduling problem with batch release and extended
technical precedence constraints. Eur J Oper Res 2017;257(1):13–24.
[2] Jia S, Bard JF, Chacon R, Stuber J. Improving performance of dispatch rules for
daily scheduling of assembly and test operations. Comput Ind Eng
2016;90:86–106.
[3] Sharma P, Jain A. Performance analysis of dispatching rules in a stochastic
dynamic job shop manufacturing system with sequence-dependent setup
times: simulation approach. CIRP J Manuf Sci Technol 2015;10:110–9.
[4] Hübl A, Jodlbauer H, Altendorfer K. Influence of dispatching rules on average
production lead time for multi-stage production systems. Int J Prod Econ
2013;144(2):479–84.
[5] El Bouri A, Amin GR. A combined OWA–DEA method for dispatching rule
selection. Comput Ind Eng 2015;88:470–8.
[6] Parthanadee P, Buddhakulsomsiri J. Simulation modeling and analysis for
production scheduling using real-time dispatching rules: a case study in
canned fruit industry. Comput Electron Agric 2010;70(1):245–55.
[7] Asadi H. Apply fuzzy learning effect with fuzzy processing times for single
machine scheduling problems. J Manuf Syst 2017;42:244–61.
[8] Abd K, Abhary K, Marian R. Multi-objective optimisation of dynamic
scheduling in robotic flexible assembly cells via fuzzy-based Taguchi
approach. Comput Ind Eng 2016;99:250–9.
 I. Pergher, A.T. de Almeida / Journal of Manufacturing Systems 46 (2018) 264–271
[9] Huang J, Süer GA. A dispatching rule-based genetic algorithm for
multi-objective job shop scheduling using fuzzy satisfaction levels. Comput
Ind Eng 2015;86:29–42.
[10] Cheng C, Huang L. Minimizing total earliness and tardiness through unrelated
parallel machine scheduling using distributed release time control. J Manuf
Syst 2017;42:1–10.
[11] Ðurasević M, Jakobović D, Knežević K. Adaptive scheduling on unrelated
machines with genetic programming. Appl Soft Comput 2016;48:419–30.
[12] Nguyen S, Zhang M, Johnston M, Tan KC. Automatic design of scheduling
policies for dynamic multi-objective job shop scheduling via cooperative
Co-evolution genetic programming. IEEE Trans Evol Comput
2014;18(2):193–208.
[13] Qin W, Zhang J, Sun Y. Multiple-objective scheduling for interbay AMHS by
using genetic-programming-based composite dispatching rules generator.
Comput Ind 2013;64(6):694–707.
[14] Chang X, Dong M, Yang D. Multi-objective real-time dispatching for
integrated delivery in a Fab using GA based simulation optimization. J Manuf
Syst 2013;32(4):741–51.
[15] Zhang L, Gao L, Li X. A hybrid genetic algorithm and tabu search for a
multi-objective dynamic job shop scheduling problem. Int J Prod Res
2013;51(12):3516–31.
[16] Quiggin J. A theory of anticipated utility. J Econ Behav Organ
1982;3(4):323–43.
[17] Keeney RL, Raiffa H. Decision Making with multiples objectives, Preferences,
and Value Tradeoffs. New York: John & Wiley & Sonns; 1976.
[18] Brito AJM, de Almeida AT. Modeling a multi-attribute utility newsvendor with
partial backlogging. Eur J Oper Res 2012;220:820–30.
[19] Pergher I, de Almeida AT. A multi-attribute decision model for setting
production planning parameters. J Manuf Syst 2017;42:224–32.
[20] de Almeida AT, Cavalcante CAV, Alencar MH, Ferreira RJP, de Almeida-Filho
AT, Garcez TV. Multicriteria and multi-objective models for risk. Reliability
and Maintenance Decision Analysis. International Series in Operations
Research & Management Science, 231. New York: Springer; 2015, 416p.
[21] Law AM. Simulation Modeling and Analysis. 5th ed. WCB/McGraw-Hill; 2015.
[22] Wakker P, Deneffe D. Eliciting von Neumann-Morgenstern utilities when
probabilities are distorted or unknown. Manag Sci 1996;42:1131–50.
[23] Munier BR. On bespoke decision-Aid under risk: the engineering behind
preference elicitation. IMA J Manag Math 2016:1–21,
http://dx.doi.org/10.1093/imaman/dpw018.
[24] Krawczyk M. Probability weighting in different domains: the role of affect
fungibility, and stakes. J Econ Psychol 2015;51:1–15.
271
[25] van de Kuilen G, Wakker PP. The midweight method to measure attitudes
toward risk and ambiguity. Manag Sci 2011;57(3):582–98.
[26] Abdellaoui M. Parameter-free elicitation of utility and probability weighting
functions. Manag Sci 2000;46:1497–512.
[27] Bleichrodt H, Pinto JL. A parameter-free elicitation of the probability
weighting function in medical decision analysis. Manag Sci 2000;46:1485–96.
[28] Medeiros CP, Alencar MH, de Almeida AT. Multidimensional risk evaluation of
natural gas pipelines based on a multicriteria decision model using
visualization tools and statistical tests for global sensitivity analysis. Reliab
Eng Syst Saf 2017;165:268–76.
[29] de Almeida AT, Alencar MH, Garcez TV, Ferreira RJP. A systematic literature
review of multicriteria and multi-objective models applied in risk
management. IMA J Manag Math 2016;28:153–84.
[30] Kendall MG. Rank Correlation Methods. 4th ed. London: Charles Griffin; 1970.
[31] Baker KR, Bertrand JWM. A comparison of due-date selection rules. AIIE Trans
1981;13(2):123–31.
[32] Pinedo M. Scheduling −Theory, Algorithms, and Systems. 5th ed. New York:
Springer; 2016.
[33] Bauer A, Browne J, Bowden R, Duggan J. Shop Floor Control Systems: From
design to implementation. London: Chapman & Hall; 1994.
[34] Blackstone JH, Philips DT, Hogg GL. A state-of-the-art survey of dispatching
rules for manufacturing job shop operations. Int. J. Prod Res
1982;20(1):27–45.
[35] Kutz M. Mechanical engineers’ handbook. Manufacturing and Management,
vol 3. New Jersey: John Wiley & Sons, Ins; 2015.
[36] Baker KR, Bertrand JWM. A dynamic priority rule for scheduling against
due-dates. J Oper Manag 1982;3(1):37–42.
[37] Vepsalainen A, Morton T. Priority rules for job shops with weighted tardiness
costs. Manag Sci 1987;33:1035–47.
[38] Carroll DC. Heuristic Sequencing of Single and Multiple Component Jobs Ph.D.
Thesis. Cambridge, MA: Sloan School of Management, M.I.T; 1965.
[39] Prelec D. The probability weighting function. Econometrica
1998;66(3):497–527.
[40] Wakker PP. Prospect theory for risk and ambiguity. UK: Cambridge:
University Press; 2010.
[41] Allais M. Le comportement de l’Homme rationnel devant le risque: critique
des postulats et axiomes de l’Ecole Américaine. Econometrica
1953;21:503–46.
[42] Hopp WJ, Spearman ML. Factory Physics. 3rd ed. Long Grove: Waveland Press;
2011.