International Journal of Production Research

ISSN: 0020-7543 (Print) 1366-588X (Online) Journal homepage: http://www.tandfonline.com/loi/tprs20

Extracting priority rules for dynamic multi-

objective flexible job shop scheduling problems
using gene expression programming

Gurkan Ozturk, Ozan Bahadir & Aydin Teymourifar

To cite this article: Gurkan Ozturk, Ozan Bahadir & Aydin Teymourifar (2018): Extracting
priority rules for dynamic multi-objective flexible job shop scheduling problems using
gene expression programming, International Journal of Production Research, DOI:
10.1080/00207543.2018.1543964

To link to this article: https://doi.org/10.1080/00207543.2018.1543964

Published online: 09 Nov 2018.

Submit your article to this journal

Article views: 54

View Crossmark data

Full Terms & Conditions of access and use can be found at

http://www.tandfonline.com/action/journalInformation?journalCode=tprs20
International Journal of Production Research, 2018
https://doi.org/10.1080/00207543.2018.1543964

Extracting priority rules for dynamic multi-objective flexible job shop scheduling problems
using gene expression programming
a,b∗
Gurkan Ozturk , Ozan Bahadira,b and Aydin Teymourifara,b
a Department of Industrial Engineering, Faculty of Engineering, Eskisehir Technical University, 26555, Eskisehir, Turkey;
b Computational Intelligence and Optimization Laboratory (CIOL), Eskisehir Technical University, 26555, Eskisehir, Turkey

(Received 11 January 2018; accepted 22 October 2018)

In this paper, two new approaches are proposed for extracting composite priority rules for scheduling problems. The sug-
gested approaches use simulation and gene expression programming and are able to evolve specific priority rules for all
dynamic scheduling problems in accordance with their features. The methods are based on the idea that both the proper
design of the function and terminal sets and the structure of the gene expression programming approach significantly affect
the results. In the first proposed approach, modified and operational features of the scheduling environment are added to
the terminal set, and a multigenic system is used, whereas in the second approach, priority rules are used as automatically
defined functions, which are combined with the cellular system for gene expression programming. A comparison shows that
the second approach generates better results than the first; however, all of the extracted rules yield better results than the
rules from the literature, especially for the defined multi-objective function consisting of makespan, mean lateness and mean
flow time. The presented methods and the generated priority rules are robust and can be applied to all real and large-scale
dynamic scheduling problems.
Keywords: dynamic job shop scheduling; priority rules; simulation; gene expression programming; multi-objective
optimisation

1. Introduction
Although several types of scheduling models exist in the literature, many assumptions are made for the simplification of
them. For example, while processing and setup times are usually assumed to be fixed, they can instead be considered as the
functions of several parameters (Yin et al. 2013; Ji et al. 2015). In many of the studied models, buffers are assumed to be
infinite; however, different buffer constraints can also be considered (Liu and Kozan 2016). Scheduling models can also be
more accurately designed by considering basic processes such as the maintenance schedule of the resources and the work
plan of the staff (Teymourifar and Ozturk 2018).
Despite all of the simplifying assumptions, because of the combinatorial structure of scheduling problems, all of them are
difficult to solve (Rudek 2013; Keshavarz and Salmasi 2014; Behjat and Salmasi 2017). For example, the number of possible
sequences of n jobs in one machine queue with deterministic information is n!. For m machines, the number of all possible
sequences is n!m , which is very large, even for n = m = 10. Considering different constraints and objective functions makes
the sequencing problem even more difficult. However many exact methods and heuristics have been proposed for such
problems, they can also be solved using priority rules (PRs) with polynomial complexity (Subramaniam et al. 2000a, 2000b;
Allahverdi and Aydilek 2010; Soroush 2010). PRs are robust and useful methods that are applicable to all scheduling
problems, and they generally achieve fair results. PRs use one or more features of jobs or shop floors to define job priorities in
order to minimise the inventory and/or tardiness costs or other characteristics (Holthaus and Rajendran 1997). Subramaniam
et al. have noted that in real manufacturing systems, PRs are the most preferred approach for solving scheduling problems.
However, there are no universal rules that yield the best results in all models. The development of robust PRs that achieve
good results in all dynamic scheduling environments is an essential matter for production management systems, which are
generally modelled as dynamic problems (Subramaniam et al. 2000a). Thus, the use of simulations is an inevitable part of
the modelling and performance evaluation of complex systems.

*Corresponding author. Email: gurkan.o@anadolu.edu.tr; gurkanozturk@gmail.com

© 2018 Informa UK Limited, trading as Taylor & Francis Group

2 G. Ozturk et al.

Many studies in the literature concern new methods of PR extraction and evaluation that take advantage of techniques
such as simulation, artificial intelligence and heuristics (Teymourifar and Ozturk 2018; Teymourifar et al. 2018). Real-
world scheduling problems are mostly multi-objective in nature, with objective functions comprising multiple different
functions such as job makespan, earliness, tardiness and lateness; therefore, the extracted PRs should be compatible with
multi-objective problems (Talbi 2009). It is also necessary to choose one or more suitable measures in accordance with the
objectives based on the stated goals (Teymourifar and Ozturk 2018).
In this article, two new approaches are proposed to extract PRs for dynamic multi-objective scheduling problems using
simulation and gene expression programming (GEP), which is an evolutionary method that creates computational models
for complex systems. Unlike previous ones, in this study, we analyze the effect of the elements of the terminal set and also
the structure of GEP on the results. In the first approach, as the most of previous studies, the multigenic structures is used to
rule extraction. Different from the previous studies, in addition to the job and shop floor features, modified and operational
features are included in the terminal set. We demonstrate that using modified and operation-based features in the terminal
set is a more effective approach. In the second approach, a cellular GEP system is proposed, in which simple priority rules
(SPRs) and composite dispatching rules (CDRs) are used as automatically defined functions (ADFs) in the terminal set, and
the result is a combination of rules. In previous studies, GEP was applied to combine features, but we show that combining
PRs, which is done using the cellular structures is a more efficient approach. In fact, each PR represents a feature and their
combination can represent the characteristics of the scheduling environment better. This approach will also prepare a base
for using multi-cellular structures in future studies.
In both approaches, elements of the function set serve as the basis for linear and non-linear combinations. The rules
extracted using these methods are compared with the classical rules and also with PRs designed via genetic programming
(GP) from a successful study in the literature. The results show that all rules obtained in the present study exhibit signifi-
cantly better performances than the rules considered for comparison. It is also found that the second proposed approach is
more efficient than the first. The suggested method generates robust PRs that can be applied to extract specialised PRs for
particular scheduling environments, with some modifications.
The remainder of this article is organised as follows. First, the problem is defined, and previous studies are summarised;
then, the proposed models and their implementation steps are described. In the following sections, results and comparisons
are given, and several designs of experiments (DOEs) are applied to determine the significance levels of the results. The
conclusion and future works are discussed in the last section.

2. Problem definition
In any scheduling problem, a set of jobs (J = {Ji |1 ≤ i ≤ n}), which consist of operations (Oij ), must be processed on a
subset of machines (M = {Mk |1 ≤ k ≤ m}). In a JSSP, each job has a fixed route; therefore, this type of problem can be
considered as a job sequencing problem on the machines. In a flexible JSSP (FJSSP), each job can proceed along more than
one route; therefore, such a problem can be modelled as a combination of assignment and sequencing sub-problems. The
simulated scheduling environment considered in this paper is a deterministic dynamic FJSSP (DFJSSP), in which the jobs
have different release times (∀i ∈ Ji : ri ≥ 0) and all information about the machines and jobs is known in advance. Other
assumptions about the scheduling model made in this article are the standard assumptions about DFJSSPs (Holthaus and
Rajendran 1997; Gholami and Zandieh 2009; Adibi, Zandieh, and Amiri 2010): (i) according to the resource constraint, each
machine can process only one operation at a time; (ii) according to the sequence constraint, the processing of any operation
of a job can start only after the completion of the previous operation of that job; (iii) job preemption is not permitted; (iv) the
resources are unlimited, except for the machines; (v) there are no interruptions or breakdowns during the scheduled period;
(vi) the jobs are independent of each other; (vii) the processing times include the assembly and setup times; and (viii) all
machines are idle at time 0.
The terminology and notations used in this paper are presented in Table 1.
There are many studies in the literature about PR extraction for dynamic scheduling problems, including those of
Cheng (1985), Chang (1994), Baker (1984), Cheng and Jiang (1998), Teymourifar and Ozturk (2018) and Teymourifar
et al. (2018).
In this study, the model from Tay and Ho’s study (Tay and Ho 2008) is simulated, in which the multi-objective function
n
consists of the makespan
n (C max = Max{C i |i = 1, 2, . . . , n}), the mean lateness ( L̄ = ( i=1 |Ci − di , 0|)/n) and the mean
flow time (F̄l = ( i=1 (Ci − ri ))/n), as expressed in Equation 1.

Cmax L̄ F̄l
F= + + (1)
3 3 3
 International Journal of Production Research 3

Table 1. Notations.
t Current time
N Number of available jobs
M Number of machines
nj Total number of operations of job j
pj Total processing time of job j (total work content of job j)
pij ≈ pjm Processing time of job j on machine m
pi+1j Total work content of jobs in the same queue as the next operation of job j
p̄m Average operation processing time on machine m
Rj Release time of job j
rjm The earliest time at which operation (j, m) can start (operational release time)
Rej Remaining processing time of job j
Dj Due date of job j
Sj Slack time of job j (= Dj − Rej − t)
nrj Number of remaining operations of job j
Cj Completion time of job j
Lj Lateness of job j (= |Cj − Dj —)
Cmax Makespan (= Max{Ci | i = 1, 2, . . . , n})
n
|Ci − di , 0|
L̄ Mean lateness (= i=1 )
n n
(C i − r i )
F̄l Mean flow time (= i=1 )
n

3. Previous works
JSSPs and FJSSPs have been studied extensively in the literature due to their wide applicability and computational difficulty.
However, real-world scheduling problems in manufacturing systems are more complex and have numerous variables (Park
et al. 2015; Manupati et al. 2016; Qu, Wang, and Shivani 2016; Teymourifar and Ozturk 2018). Production environments
are generally dynamic and stochastic because of different skills; shortages or breakdowns in resources; variable processing,
maintenance and repair times; and the events related to the arrival of new jobs, which make problems more difficult (Sharma
and Jain 2015; Karunakaran et al. 2017; Xiong et al. 2017). Several approaches have been proposed for solving these
complex problems. For example, Allahverdi and Mittenthal (1994) proposed a method of transforming the problem of
scheduling n jobs on m parallel machines with random breakdowns and processing times into a deterministic, unrelated
parallel machine scheduling model with modified processing times, wherein the number of breakdowns is modelled as
a generalised Poisson process. However, PRs are a more practical method of solving complex scheduling problems in
polynomial time, and they generally yield good results (Durasević and Jakobović 2018b, 2018a). PRs are applied in many
fields of production and logistics, especially in real-time, online and highly complex manufacturing scheduling systems
(Pickardt et al. 2013; Branke, Hildebrandt, and Scholz-Reiter 2015; Heger et al. 2016). For example, Manupati et al. (2016)
reported that PRs are one of the most frequently used approaches for scheduling in semiconductor manufacturing. PRs
sequence jobs according to one or more feature of the jobs, operations, and scheduling environment. Based on the parameters
used, PRs are categorised into four classes: (i) SPRs, which use one parameter in the model, such as the release time,
processing time or due date; (ii) CDRs, which use more than one parameter; (iii) weighted priority indexes; and (iv) heuristic
scheduling rules (Panwalkar and Iskander 1977; Tay and Ho 2008). Generally, CDRs demonstrate better performance in
multi-objective scheduling problems (Tay and Ho 2008). In the literature, several SPRs and CDRs have been presented,
including (i) first in, first out (FIFO) and (ii) arrival time (AT), which are both effective for minimising the maximum
value and variance of the flow time; (iii) earliest due date (EDD), which is a simple rule that achieves good performance
in minimising the variance of the tardiness, especially in single-machine scheduling problems; (iv) slack per remaining
operation (SOPN), which mostly is used for evaluating rules with respect to tardiness-related objectives; (v) cost over time
(COVERT), which is efficient for tardiness-related objectives; (vi) modified due date (MDD), which is a simple and effective
rule for minimising tardiness; and (vii) shortest process time (SPT), which decreases the mean flow time and the percentage
of tardy jobs and is also very effective for mean flow time and mean tardiness minimisation under highly loaded shop floor
conditions.
There are several approaches to the use of PRs. Along with using the classic rules in the literature, it is also common to
design a specific rule of the scheduling environment. The multiple PR approach, in which different rules are applied to each
machine or work centre, has also been studied in the literature (Zahmani et al. 2015). Also, different approaches can be used
to develop different types of PRs, such as dynamic rules and fixed rules (Tay and Ho 2008; Kapanoglu and Alikalfa 2011).
4 G. Ozturk et al.

SO, there are many studies in the literature about the extraction and application of PRs for scheduling problems (Baker 1974;
Holthaus and Rajendran 1997; Rajendran and Holthaus 1999; Jayamohan and Rajendran 2000, 2004; Kanet and Li 2004;
Ingimundardottir and Runarsson 2016). Blackstone, Phillips, and Hogg (1982) reviewed the research on the application
of PRs in manufacturing. Rajendran and Holthaus (1999) studied PRs for dynamic JSSPs (DJSSPs) and dynamic FSSPs
(DFSSPs). They proposed three PRs and compared them with 13 other PRs. In their study, the evaluation was performed
in accordance with the objectives of minimising the mean flow time, the maximum flow time, the variance of the flow
time, the proportion of tardy jobs, the mean tardiness, the maximum tardiness and the variance of the tardiness. Jayamohan
and Rajendran (2000) designed several PRs for various objective functions, such as the mean, maximum and variance of
the flow time and tardiness, and compared them with the classical rules for dynamic scheduling problems. Allahverdi and
Tatari (1996) used simulations to compare rules for finding the best makespan in stochastic machine dominance problems
in flow shop scheduling environments. Anderson and Nyirenda (1990) presented two PRs for minimising tardiness in a
job shop. These rules are simple to implement and do not require any parameter estimations during their implementation.
Chiang and Fu (2007) designed a PR for JSSPs with due-date-based objectives and compared it with 18 PRs from the
literature with respect to the tardy rate, mean tardiness and maximum tardiness. Rochette and Sadowski (1976) compared the
performance of PRs in JSSP-based experimental designs. Grabot and Geneste (1994) used fuzzy logic to build aggregated
rules and thus achieved a compromise among the satisfaction of several criteria. Sels, Gheysen, and Vanhoucke (2012)
designed and compared various PRs and their robustness in JSSPs under different objective functions. Doh et al. (2013)
suggested a practical PR for FJSSPs that employed a combination of operation/machine selection and job sequencing rules in
accordance with objective functions for the makespan, total flow time, mean tardiness, number of tardy jobs, and maximum
tardiness. Jeong and Kim (1998) presented a real-time dynamic scheduling model based on simulation and PRs for flexible
manufacturing systems. Pierreval and Mebarki (1997) presented a scheduling method based on the dynamic selection of
PRs. Lu, Huang, and Yang (2011) integrated order review/release (ORR) and PRs in a simulated JSSP for optimising due-
date- and flow-time-related performance measures. Land, Stevenson, and Thürer (2014) integrated load-based order release
heuristics and a well-established production control concept for job shops with PRs for efficient scheduling and the reduction
of shop floor workloads.
PRs are usually designed by experts in a time-consuming trial-and-error process. However, in recent years, evolutionary
methods and hyper-heuristics have been widely used as practical approaches to automating the design of heuristics for pro-
duction scheduling problems (Hildebrandt, Heger, and Scholz-Reiter 2010; Branke, Hildebrandt, and Scholz-Reiter 2015;
Branke et al. 2016; Nguyen et al. 2018, 2019). The dynamic approaches like GP and GEP are able to discover sophisticated
policies to deal with complex and dynamic production environments (Nguyen, Mei, and Zhang 2017; Hein et al. 2018). Tay
and Ho (2008) used a GP approach to extract PRs for multi-objective DFJSSPs. A GP algorithm is an evolutionary algorithm
(EA) in which the individuals, which represent computer programmes, are composed of terminals and functions. In their
study, the terminal set and the function set were chosen to be T = {R, D, p, ct, Re, n, P̄} and F = {+, −, ×, /, ADF(x1 , x2 )},
respectively. ADFs were introduced by Koza (1992) as a way of reusing codes in GP and improving the efficiency
of the evolutionary computation (Koza 1992; Ferreira 2006a, 2006b). The ADF used in the terminal set designed by
Tay and Ho (2008) was ADF(x1 , x2 ) = x1 ∗ x2 , and the PRs extracted after 5 runs of the GP algorithm on the train-
ing set were (i) R + 2p + 2P + n, (ii) R + p + 2P, (iii) 7P + 11p + 12(n + R), (iv) (R + D) + 2(R + P) + p − n and (v)
(R + D) + (P + p) − 2(R/n). These PRs are robust and show good performance for all multi-objective scheduling prob-
lems (Tay and Ho 2008). Nie et al. used GEP to construct PRs for solving dynamic single-machine scheduling problems
(Nie et al. 2010). In a more generalised model, Nie et al. proposed a heuristic for reactive scheduling in a job shop where
jobs arrive over time. The method decomposes the problem into a number of sub-problems, which are the dynamic single
machine scheduling problem with job release dates. The jobs are scheduled using PRs, which are constructed by a frame-
work consisting of GEP and simulation. They demonstrated that the performance of the rules extracted with this approach
is significantly better than some prominent rules from the literature (Nie et al. 2013b). In a similar study, Nie et al. used
GEP to construct reactive scheduling policies for DFJSSP. To evaluate the performance of the extracted policies under dif-
ferent conditions, three factors, such as the shop utilisation, due date tightness, problem flexibility, were considered in the
simulation model. The results showed that GEP can evolve effective policies under a big range of processing conditions and
performance measures (Nie et al. 2013a).
Another novel research method that has many applications in this domain is based on hyper-heuristics, which can be
explained as ‘heuristics for choosing heuristics’ to automate the process of selecting or combining simpler heuristics in
order to solve complex computational search problems (Ochoa et al. 2009). Park et al. (2016) used a new co-evolutionary
GP-based hyper-heuristic (GP-HH) approach that evolves an ensemble of PRs to solve a DJSSP. In a similar study,
Park et al. (2018) developed some measures to analyze the behaviour of the rule ensembles. Branke, Hildebrandt, and
Scholz-Reiter (2015) discussed several approaches for representing PRs in hyper-heuristic searches and reported that a suit-
able choice of representation is of key importance for a successful EA because the representation determines the search
 International Journal of Production Research 5

neighbourhood and the complexity of the rules. They empirically investigated the suitability of the derandomized evo-
lution strategy with covariance matrix adaptation (CMA-ES) for the neural network and linear representations and GP
for tree representation in a dynamic stochastic job shop scenario. Park et al. (2015) proposed a new approach in which
an ensemble of rules is evolved using GP and cooperative co-evolution, called Ensemble Genetic Programming for Job
Shop Scheduling (EGP-JSS), which generally produces more robust rules than single-rule GP. Pickardt et al. (2013) used
a hyper-heuristic combined with a GP algorithm to evolve a composite rule from basic characteristics of jobs via an EA
that searches for a good assignment of rules to work centres. Hunt, Johnston, and Zhang (2014) used a GP-HH method
to evolve PRs for static and dynamic two-machine job shop environments. For the static model, the proposed method
could represent and find optimal PRs. The results showed that the performance of the methods depended on the testing
instances.
Other aspects of the GEP and GP like feature extraction have also used in scheduling. Zhang et al. used GEP to extract
five attributes with the most influence on the total energy consumption for solving an energy-efficient flexible job shop
scheduling problem (Zhang et al. 2017). Also, there is growing research community for solving other dynamic problems
owing to their several applications in the prediction of energy, temperature, CO2 emissions etc. by GP and GEP (Bagatur
and Onen 2018; Hein et al. 2018; Hong, Jeong, and Koo 2018; Samadianfard et al. 2018).
Other methods have also been applied for the extraction of PRs. For instance, Qu, Wang, and Shivani (2016) applied a
reinforcement learning approach that considered the real-time production environment to generate optimal policies under
various manufacturing process conditions. This method was found to achieve better performance than most common PRs
under different objectives and constraint conditions.

4. GEP algorithm for extracting PRs

In the DFJSSP simulated in this study, jobs arrive at the shop at different times, and since the model starts at a time that
is equal to the minimum release time for the jobs, when all machines are idle, the first operations of the released jobs are
assigned to the machines in a random order and begin to process. Then, the event-based simulation moves to the next event
in the system, which is the completion of an operation (at which time the corresponding machine begins to idle) or the
release of an operation. After the system update at the next time point, if an operation is released, it is then assigned to a
machine according to the least waiting time assignment heuristic (Tay and Ho 2008). When a machine is idle and there are
operations waiting in its queue, one operation is selected according to a PR and starts processing.
In this study, we propose a new GEP-based approach for extracting new PRs as combinations of features and PRs
from the literature. GEP was proposed by Candida Ferreira Ferreira (2006b); it is similar to GP, but it has a linear struc-
ture (Zhou et al. 2003). The structural organisation of GEP genes is more understandable in terms of the open reading
frames (ORFs). Because the gene coding sequence in an ORF begins with a codon, continues with the amino acid codons
and ends at a termination codon, the start point in GEP is the first gene, but the finish point is not the last gene. The
three parts of a GEP gene are the head (h), tail (t) and domain (dc), which contain the functions, elements and numer-
ical values used in the computation. The head contains only functions, whereas the tail can include both elements and
functions. The length of the tail is calculated from the head size as t = h(nmax − 1) + 1, where nmax is the maximum num-
ber of elements that the function uses. The position of the domain is after the tail, and its size is equal to the tail size
(Ferreira 2006a).
GEP modelling can be based on different structures, including (i) structures without ADFs, for unigenic and multigenic
systems, and (ii) structures with ADFs, for unicellular and multi-cellular systems. In the multigenic systems, there is more
than one gene of a specific size, where sub-expression trees (sub-ETs) are linked with each other. The homeotic genes that
have a normal gene structure specify the interactions between the sub-ETs. In the cellular systems, ADFs are used instead
of Sub-ETs, which are reusable components Koza (1994). If there is only a single homeotic gene, the system is unicellular,
whereas if the chromosome contains more than one homeotic gene, the structure is multicellular (Ferreira 2006a, 2006b). In
the next two subsections, new approaches are proposed for PR extraction: one method is based on multigenic systems with
static linking, and the other is based on cellular systems. In both methods, the special design of the function and terminal
sets enhances the performance.

4.1. The first approach, based on the multigenic system

It is well known that in dynamic scheduling problems, the job and shop floor parameters affect the results. For the design of
efficient PRs, it is necessary to extract effective features of the problem. Moreover, the proper combination of the parameters
and the assignment of suitable coefficients to them are also important. When GEP is used for PR extraction, the parameters
are provided through a suitable selection of terminal and function sets. If the basic features, such as pij , Pj , Dj and Rj , are
6 G. Ozturk et al.

Figure 1. Genotype - phenotype of a PR encoded as an expression tree.

Figure 2. A CDR is extracted using a three-genic chromosome, whose sub-ETs are linked based on the homeotic genes. a. Chromosomes
of Sub-ETs, b. Sub-ETs, c. The linkage between the Sub-ETs based on the homeotic genes and the resulting rule. In the expression tree,
SE stands for sub-ET.

used in the terminal set, then composite PRs will be extracted as combinations of these features. In Figure 1, an example of
the genotype and phenotype of a CDR is shown.
The main components of the multigenic structure considered in GEP are the chromosomes and the ETs. When a genome
contains more than one gene, each of them is independently translated as a sub-ET (Ferreira 2006a, 2006b). In a multigenic
system, multiple CDRs are expressed as different sub-ETs. The sub-ETs of multigenic systems are linked to constitute more
complex CDRs, as presented in Figure 2.
The terminal set elements used in the first approach are presented in Table 2. It is necessary to consider both linear
and non-linear combinations of the PRs to achieve better results. The function set {+, −, /, ∗} provides suitable types of
combinations of the members of the terminal set. Even ‘square’ and ‘square root’ operators can be provided.

4.2. The second approach, based on the cellular system

In a multigenic system, sub-ETs can preserve composite features of the scheduling environment. These features are generally
more effective than the basic features. In cellular systems, linked ADFs are used to represent features more efficiently than
sub-ETs. In the second approach, SPRs and CDRs are used as ADFs in the terminal set for GEP, and consequently, the
performance of the extracted rules is improved.
As seen in Figure 3, in a cellular system, the ADFs, which can represent either SPRs as EDD or CDRs such as the
critical ratio and slack time, are linked with homeotic genes.
In the second approach, the operator set {+, −, ∗, /} is used as the function set for GEP, and the terminal set is presented
in Table 2 (Lin, Goodman, and PunchIII 1997). As seen, in the first approach, the members of the terminal set are the
features of the scheduling environment and/or the jobs, whereas in the second approach they are objective functions.
 International Journal of Production Research 7

Table 2. GEP terminal sets.

First Approach
Feature name Notation
Arrival time Rj
Due date Dj
Processing time pij
Remaining time Rej
Total work content Pj
Operation due date ODDij
Critical operation due date CRODDij
Second Approach
Rule name Notation
First In, First Out (FIFO) Zj = rij
Least Arrival Time (AT) Zj = Rj
Shortest Processing Time (SPT) Zj = pjm
Least Remaining Work Content (LRWRK) Zj = Rej
Least Total Work Content (LTWRK) Zj = Pj
Least Number of Operations (LnOps) Zj = nj
Least Remaining Number of Operations (LRnOps) Zj = nrj
Earliest Due Date (EDD) Zj = Dj
Earliest Modified Operation Due Date (MOD) Zj = max(Rj + (Dj − Rj )Rej /Pj , t + pjm
Minimum Slack Time (SLK) Zj = Sj
Earliest Operation Due Date (ODD) Zj = Rj + (Dj − Rj )Rej /Pj
Slack per Operation (SOP) Zj =[1 − (Sj )/Nj )]/pjm
S /N , if Sj ≥ 0
Slack per Remaining Operation (SOPN) Zj = j j
Sj Nj , otherwise.
Critical Ratio (CR) Zj = [1 − (Dj − t)/Rej )]/pjm
+
Cost Over Time (COVERT) Zj = [1 − (Sj )+ /2Rej )] /pjm

Figure 3. (a) PRs as ADFs used in the terminal set, (b) a cell is constructed based on the homeotic genes and the result is a combination
of rules.
8 G. Ozturk et al.

4.3. Algorithm

Algorithm 1: GEP algorithm for rule extraction

1 Input: The parameters for problem generation (problem size, flexibility rate and due date assignment tightness factor).
2 Input: The parameters for GEP (PopSize, mutation rate, recombination rate, elitism, gene length, head length, number of homeotic genes).
3 Output: The extracted PR set

4 Create the initial population, P0 = {pj |j = 1, . . . , PopSize}; (each individual pj corresponds to a PR)
5 for t = 0; t < maxiter ; t = t + 1 do
6 Generate the problem set for Pt , St = {sit |i = 1, . . . , ns}, based on the parameters (problem size, flexibility and
tightness factor);
7 foreach pj ∈ Pt do
8 foreach sit ∈ St do
9 Run the simulation on problem si to obtain the objective values for pj (makespan, Cmaxij , mean lateness
L̄ij , mean flow time, F̄ij ) ;
10 end
  
11 Calculate the fitness of pj as follows: fj = 13 nsi=1 Cmaxij + L̄ij + F̄ij ;
12 end
13 Generate the next population Pt+1 by applying the operators (selection, recombination, mutation, elitism) ;
14 end
15 Find the best extracted rule;

The steps of PR extraction are summarised in Algorithm 1. First, an initial population is generated. Each individual in
the population is an ET that corresponds to a PR. During each iteration of the GEP algorithm, a problem set St , which con-
tains ns problems, is created according to three parameters: problem size, flexibility rate and due date assignment tightness
factor. The simulation runs and calculates the makespan (Cmax), mean lateness (L̄) and mean flow time (F̄) values for each
individual. The fitness value of each rule is then calculated as the sum of the achieved objective function values for all
problems in the set. Then, the evolutionary operators are applied to the current population to generate the next one. These
steps continue until the maximum number of iterations (maxiter ) is reached, and at the end of the algorithm, a set of robust
PRs is extracted.

4.4. Numerical experiments

The benchmarks used in this study were taken from Tay and Ho’s study. In these benchmarks, the generated problem
sizes (J × M ) are as follows: 10 × 5, 20 × 5, 50 × 5, 20 × 10, 50 × 10, 100 × 10, 50 × 15, 100 × 15 and 200 × 15. The
processing times of the operations are generated according to the distribution Pijk = U[(M )/2, (M ) × 2], where the standard
deviation of the processing times is equal to 5 (dev(Pijk , Pijl ) = 5). The release times of the jobs are generated according to
the following distribution (Tay and Ho 2008):

U[0, 40], if n Jobs ≥ 50;
rj = (2)
U[0, 20], otherwise.
i
The due dates of the jobs are assigned according to Baker’s model, as follows: di = ri + c × nj=1 Pij (Baker 1984). In
scheduling problems, a flexibility of f % means that at most this percentage of all machines can process a given operation.
The benchmarks are created with flexibility rates of 100%, 50% and 20% and with tight (1.2), moderate (1.5), loose (2) and
mixed due date tightness factors. In the mixed set of due date tightness factors, 34% of the entries are created with the tight
factor, 33% with the moderate factor and 33% with the loose factor. Using the various possible combinations of the problem
size, flexibility rate and due date factor parameters, a total of 108 benchmarks are generated (Tay and Ho 2008).
In this study, the proposed PR extraction steps were performed using these benchmarks. The algorithm solved each of
these 108 data sets using each PR obtained in each iteration. After the objective function for each benchmark was calculated
according to Equation 1, the fitness value of the rule was calculated as the sum over all benchmarks. Since the parameters
used are very tight, for the comparison of the extracted rules, additional benchmarks were generated after rule extraction
 International Journal of Production Research 9

Table 3. Parameters and operators of the GEP algorithm.

Parameter Value
Population size (PopSize) 20
Maximum number of iterations (maxiter ) 100
Head length 5, 10
Gene length 16, 31
Number of homeotic genes/cells 1
Mutation rate (for homeotic genes) 0.3%
One-point recombination rate 90%
Elitism 10%
Selection type Tournament selection

using looser parameters similar to those encountered in many real-world problems. For example, values of 2.5 and 3 were
added for the due date tightness factor.
In addition, distributions of 5 × U[(M )/2, (M ) × 2] and U[(M )/8, (M ) × 8] were used for processing time generation.
The operators used in the GEP algorithm are listed in Table 3.

5. Analysis of experimental results

The algorithms were implemented in Python 2.7.10 on a system with an Intel(R) Xenon(R) 2.93 GHz dual processor and
8.00 GB of RAM. The convergence of the designed algorithm for the benchmarks with 50% flexibility is illustrated in
Figure 4, which shows the stability of the algorithm. Similar patterns are also valid for the other flexibility rates.
Well defined operators of GEP can prevent early convergence, so more than one rule is usually extracted. In this case,
all of these rules are evaluated.
The PRs extracted with the first approach are presented in Equations 3 and 7.

ODD + Re (3)

(ODD − CRODD3 ) ∗ Re (4)

Figure 4. The convergence of the designed algorithm for benchmarks with 50% flexibility.
10 G. Ozturk et al.

In both of these rules, the first element of the homeotic genes is not a function, so they appear as the unigenic structures.
The PRs extracted with the second approach are presented in Equations 5 and 6.

LTWRK/(3 + LnOps − LRnOps) (5)

(EDD + ((LRnOps + TWORK)/(LRWRK) − TWORK) ∗ LnOps)) ∗ LRnOps (6)

(ODD − CRODD3 ) ∗ Re (7)

Example 1 Let suppose that there are three jobs with the properties given as follows waiting to process on Machine 1.

ODD Re TWRK nOps RnOps

J1 57.70 32.55 32.55 5 5
J2 30.73 19.03 28.81 5 3
J3 32.20 17.31 26.30 5 4
The chromosomes of rules defined in Equations (3) and (5) and the corresponding PRs are given in below.

Chromosome → Priority Rule

[+ | ODD | Re |· · · ] → ODD + Re
[/ |ADF1 | + | 3 | -| ADF2| ADF3 | · · · ] → LTWRK/(3 + LnOps − LRnOps)

In order to show the working principles of the PRs, the priorities of jobs are calculated according to the rules. In this way,
J3 and J2 are selected using the first and second rules, respectively.

ODD + Re LTWRK / (3+ LnOps + LRnOps)

J1 57.70 + 32.55 = 80.25 32.55 / (3 + 5 - 5) = 10.85
J2 30.73 + 19.03 = 49.76 28.81 / (3 + 5 - 3) = 5.76
J3 32.20 + 17.3 1= 49.51 26.30 / (3 + 5 - 4) = 6.58

To compare these rules with the best known PRs from the literature, the objective function values were calculated
in the same way as in the rule extraction stage. In addition, each simulation was repeated 30 times to account for
the randomness of the problems. The average results of 30 simulations for the 20%, 50% and 100% flexibility rates
are summarised in Table 5. The indexes used to identify the rules in these tables are given in Table 4. The first
4 rules are classic PRs, and SPT + LRWRK + SLK (Teymourifar and Ozturk 2018), AT + SPT + (2*LTWRK) and
(7*LTWRK) + (11*SPT) + 12*(LnOps + AT) (Tay and Ho 2008) are CDRs from the literature.
As seen in Table 5, for the 20% flexibility rate, the rules extracted with the second approach yield the best results by far
with respect to the F and F̄l objective functions, with rule 10 being the best. For the L̄ objective function, rule 11 shows
the best performance, but all extracted rules produce good results. Although FIFO yields the best result in terms of Cmax , its
results are the worst for the other objective functions.

Table 4. The indexes used to identify the rules in the result tables.
Rule index Rule description
1 FIFO
2 SPT
3 EDD
4 ODD
5 SPT + LRWRK + SLK
6 AT + SPT + (2*LTWRK)
7 (7*LTWRK) + (11*SPT) + 12*(LnOps + AT)
8 ODD + Re
9 (ODD-CRODD3 )*Re
10 LTWRK/(3 + LnOps-LRnOps)
11 [EDD + [(LRnOps + LTWRK)/(LRWRK-LTWRK)]*LnOps]*LRnOps
 International Journal of Production Research 11

Table 5. Results of the PRs for the different flexibility rates.

Rule index Cmax L̄ F̄l F Average CPU time
Flexibility Rate: 20 %
1 1014.43 685.29 878.19 859.31 6.31
2 1050.49 474.90 603.94 709.78 5.06
3 1082.09 388.91 388.29 619.76 4.37
4 1090.14 383.87 348.57 607.52 4.27
5 1087.30 380.55 371.54 613.13 4.86
6 1079.72 383.32 382.63 615.23 5.45
7 1074.63 392.71 389.98 619.1 6.20
8 1093.63 379.92 361.94 611.83 4.70
9 1089.99 381.57 357.32 609.63 5.75
10 1091.45 382.06 330.94 601.48 5.89
11 1089.27 388.35 334.31 603.98 7.05
Flexibility Rate: 50 %
1 955.59 640.66 835.07 810.44 6.47
2 985.54 436.33 568.42 663.43 5.09
3 942.29 318.81 309.44 523.51 4.36
4 945.61 316.86 274.94 512.47 4.39
5 945.79 313.47 299.01 519.42 4.93
6 937.86 317.04 313.42 522.77 5.56
7 931.15 325.53 324.44 527.04 6.29
8 947.88 314.38 287.79 516.68 4.72
9 943.22 314.03 283.72 513.65 5.73
10 947.32 316.85 258.08 507.42 5.83
11 942.16 317.31 260.14 506.54 7.07
Flexibility Rate: 100 %
1 889.09 583.53 778.45 750.36 6.67
2 915.64 392.95 528.26 612.29 5.29
3 809.30 256.04 256.99 440.78 4.45
4 813.57 255.43 228.12 432.37 4.45
5 812.06 251.45 250.88 438.13 5.06
6 805.41 256.72 265.23 442.45 5.61
7 801.31 265.14 276.08 447.51 6.36
8 813.25 254.72 238.86 435.61 4.86
9 812.69 253.12 235.23 433.68 5.84
10 814.27 257.15 214.29 428.57 5.97
11 813.62 256.15 215.26 428.34 7.22

For the 50% flexibility rate, the extracted rules yield the best results with respect to the F and F̄l objective functions.
While rule 5 achieves the best result in terms of the L̄ objective function, the extracted rules, especially rules 10 and 11,
produce similar results.
As seen in Table 5 and Figure 5, the results for the 100% flexibility rate are similar to those for the 20% and 50%
flexibility rates, and the extracted rules clearly achieve the best results in terms of the mean flow time and the multi-
objective function. For the L̄ objective function, although rule 5 shows the best performance, the rules extracted with the
first approach yield similar results. Meanwhile, rule 7 achieves the best performance with respect to the Cmax criterion,
whereas all of the extracted rules yield fair results.
To demonstrate the effectiveness of the extracted rules, a comparison is also presented based on additional benchmarks
generated according to different parameters. The results obtained for benchmarks generated with due date tightness factors
of 1.2, 1.5, 2 and 2.5 are shown in Table 6. For these benchmarks, the other parameters are the same as those used for the
previous benchmarks.
As seen in Table 6, the results are similar for benchmarks generated with due date tightness factors of 1.5, 2, 2.5 and 3.
For processing times generated according to distributions of 5 × U[(M )/2, (M ) × 2] and U[(M )/8, (M ) × 8], the
results are as shown in Table 7, respectively.
As seen in Tables 6 and 7, the extracted rules show significantly better performances for the instances created with looser
parameters. This feature can be an advantage for their application to real-world problems.
12 G. Ozturk et al.

Figure 5. Comparison among the results of the extracted rules for the 100% flexibility rate according to the multi-objective function (In
this figure, rules 1 and 2, which are FIFO and SPT, are excluded because according to the multiobjective function their results are worse
than others. The other rules are indexed in the same order in which they are listed in the result tables).

Table 6. Comparison based on benchmarks generated with different due date

tightness factors.
Rule index Cmax L̄ F̄l F Average CPU time
tightness factors: 1.2, 1.5, 2 and 2.5
1 893.07 559.49 780.47 744.34 17.90
2 920.06 374.07 530.44 608.19 14.24
3 944.96 303.91 317.61 522.16 12.37
4 950.53 303.2 283.81 512.51 12.24
5 946.41 299.02 309.02 518.15 14.02
6 941.42 305.27 319.1 521.93 15.53
7 938.94 314.52 331.61 528.36 17.63
8 951.47 302.02 284.07 512.52 13.43
9 947.09 302.13 281.79 510.34 16.30
10 953.72 306.02 266.82 508.85 16.55
11 946.8 304.68 268.78 506.75 19.80
tightness factors: 1.5, 2, 2.5 and 3
1 891.1 520.91 780.13 730.71 19.47
2 919.11 346.17 531.35 598.88 15.43
3 942.1 283.59 321.27 515.65 13.40
4 946.86 283.96 284.3 505.04 13.26
5 947.35 280.79 310.24 512.79 14.92
6 937.46 284.97 319.57 514 16.82
7 934.87 293.56 331.23 519.89 18.54
8 946.13 283.57 284.62 504.77 14.46
9 952.23 284.03 281.85 506.04 17.81
10 950.44 288.07 266.59 501.7 18.12
11 945.98 285.7 268.73 500.13 21.85

The average CPU times for solving all of 108 problems have been written in the 5th column of Tables 5–7. Although
the rule extraction process lasts for some hours, the solution times of each problem for all rules are less than 1 second on
average. Whereas, more time is needed to solve these problems by optimisation methods.
To confirm that the extracted rules are better than the others, analysis of variance (ANOVA) and Tukey tests were
performed. In the two-way ANOVA test, whose results are shown in Table 8, the fitness of the compared rules and the
flexibility of the benchmarks were chosen as the factors. The null hypothesis (H0 ) was that there are no differences between
the factors or their interaction effects, whereas the alternative hypothesis (H1 ) was that at least one level of these factors is
different.
Since P − value < 0.05, H0 is rejected with more than 95% confidence. The R2 value of this test is 99,96%, meaning that
it represents the model properly. To identify which of the rules is different, it was necessary to also perform a Tukey test,
 International Journal of Production Research 13

Table 7. Comparison based on processing times generated according to the

distribution 5 × U[(M )/2, (M ) × 2] and U[(M )/8, (M ) × 8].
Rule index Cmax L̄ F̄l F Average CPU time
pij ≈ 5 × U[(M )/2, (M ) × 2]

1 4451.11 2978.93 3984.98 3805.01 21.05

2 4599.07 1994.07 2660.87 3084.67 16.54
3 4700.6 1604.96 1620.29 2641.95 14.02
4 4723.66 1589.95 1428.56 2580.72 13.71
5 4720.46 1582.42 1523.44 2608.77 15.51
6 4687.63 1614.15 1657.67 2653.15 17.77
7 4648.04 1675.89 1906.81 2743.58 20.76
8 4721.24 1588.01 1422.25 2577.17 15.07
9 4722.26 1589.29 1425.34 2578.96 18.53
10 4740.44 1619.48 1332.61 2564.18 18.96
11 4728.89 1597.68 1329.05 2551.87 23.06
pij ≈ U[(M )/8, (M ) × 8]

1 2524.02 1558.62 2204.94 2095.86 17.90

2 2584.26 1008.13 1480.49 1690.96 14.50
3 2721.88 842.62 999.77 1521.42 12.33
4 2738.75 841.88 920.19 1500.27 12.23
5 2742.89 829.08 960.72 1510.9 14.21
6 2708.48 829.53 976.88 1504.97 16.52
7 2665.55 847.81 1080.56 1531.31 18.40
8 2744.95 846.33 929.3 1506.86 13.73
9 2740.87 842.63 921.96 1501.82 17.00
10 2742.76 867.18 843.73 1484.56 17.31
11 2738.71 859.35 840.4 1479.48 21.07

Table 8. Two-way ANOVA test for the fitness of the

rules versus the flexibility of the benchmarks.
Source DF SS MS F P
DR 7 18483.08 2640.44 282.91 0
Error 72 671.98 9.33
Total 79 19155.06

Table 9. Results of the Tukey test for the

benchmarks used for rule extraction.
Rule index N Mean Grouping
7 10 531.217 A
3 10 528.018 B
5 10 526.817 B
4 10 517.455 C
8 10 516.724 C-D
9 10 515.873 D
11 10 512.952 E
10 10 512.489 E

whose results are presented in Table 9. FIFO and SPT were not included in this test because they do not yield good results
with respect to the multi-objective function.
In Figure 6(a), the results of the Tukey test is visulalized for the data given in Table 9. Figure 6(b,c) are presented for
benchmarks generated with different parameters in order to show the robustness of the extracted PRs. It is obvious that in
the all cases PRs generated according to the second approach are definitely better than the others.
These results statistically prove that the extracted rules, particularly those generated with the second approach, yield
different and better results than the other rules.
14 G. Ozturk et al.

Figure 6. The results of the Tukey test for (a) the benchmarks used for rule extraction, (b) the ones whose processing times are generated
according U[M /8, M × 8] distribution and (c) the ones whose processing times are generated according 5 × U[M /2, M × 2] distribution.
In this figure, the new rules are compared with the rules 6 and 7, which derived using GP and also the classic rules 3 and 4.

6. Conclusion and future work

PRs are practical scheduling methods for real dynamic environments because they have reasonably low computational com-
plexity and generate fair results. Many robust SPRs and CDRs have been reported in the literature; however, it can be more
useful to extract specific rules for particular scheduling environments. There are two important approaches to rule extrac-
tion. In the first approach, after the detection of efficient features of the scheduling environment, the detected features are
combined to construct composite rules. Generally, job characteristics are used as the parameters of PRs; however, simulation
results show that operation-based parameters are more sensitive and that using combinations of job- and operation-based
parameters improves the results. As a more efficient method, the second approach to CDR extraction involves the linear
or non-linear combination of several PRs for the construction of composite rules. The results show that such PRs exhibit
better performance because PRs based on modified features such as MOD, CR, SLK and CRODD represent the environ-
ment better than those based on basic features such as R, p, P and DD. In this paper, these two approaches are applied in
combination with GEP as an evolutionary learning method for the extraction of new PRs. A multigenic structure is used for
the GEP algorithm in the first approach, whereas a cellular structure is applied in the second approach. In the first approach,
sub-ETs preserve the respective composite features of the scheduling environment, whereas in the second approach, the
evolution process is applied to PRs in the form of ADFs that represent modified features of the jobs, operations and the
environments, and consequently, the performance of the extracted PRs is improved. The repeated simulation of dynamic
scheduling problems for the extraction of robust rules is a time-consuming process; therefore, the design of the terminal set
and the parameters is very important to improve the quality of the search. Extracting and using more efficient features in
these sets enhances the results; however, applying unnecessary features decreases efficiency. For example, it has also been
observed that LRnOps, ODD and CRODD, as modified features, are generally efficient, and their inclusion in the terminal
set has been shown to be useful. For the function set, in addition to the ‘ + ’ and ‘ − ’ operators, using operators such as ‘*’
and ‘/’ to achieve non-linear combinations of the PRs leads to better results. In this study, based on these considerations, 4
rules were extracted, which can be considered as combinations of PRs or as rules based on composite features of the jobs
and the shop floor. The extracted rules are robust and can be applied to all real problems as the scheduling of semiconductor
production. They demonstrate significantly better performance than all of the other PRs considered for comparison in terms
of both the multi-objective function and the single objective functions applied in this study.
In future work, we plan to propose fuzzy PRs for more dynamic scheduling problems with stochastic processing times,
machine breakdowns, repair times and unavailability.

Acknowledgments
The authors would like to thank the editor and three anonymous referees whose criticisms significantly improved the manuscript.

Disclosure statement
No potential conflict of interest was reported by the authors.

Funding
This work was supported by the Anadolu University Scientific Research Projects under grant number 1703F080.
 International Journal of Production Research 15

ORCID
Gurkan Ozturk http://orcid.org/0000-0002-9480-176X

References
Adibi, M., M. Zandieh, and M. Amiri. 2010. “Multi-Objective Scheduling of Dynamic Job Shop using Variable Neighborhood Search.”
Expert Systems with Applications 37 (1): 282–287.
Allahverdi, A., and H. Aydilek. 2010. “Heuristics for the Two-Machine Flowshop Scheduling Problem to Minimise Makespan with
Bounded Processing Times.” International Journal of Production Research 48 (21): 6367–6385.
Allahverdi, A., and J. Mittenthal. 1994. “Scheduling on m Parallel Machines Subject to Random Breakdowns to Minimize Expected
Mean Flow Time.” Naval Research Logistics (NRL) 41 (5): 677–682.
Allahverdi, A., and M. F. Tatari. 1996. “Simulation of Different Rules in Stochastic Flowshops.” Computers & Industrial Engineering 31
(1): 209–212.
Anderson, E., and J. Nyirenda. 1990. “Two New Rules to Minimize Tardiness in a Job Shop.” The International Journal of Production
Research 28 (12): 2277–2292.
Bagatur, T., and F. Onen. 2018. “Development of Predictive Model for Flood Routing using Genetic Expression Programming.” Journal
of Flood Risk Management 11: S444–S454.
Baker, K. R. 1974. Introduction to Sequencing and Scheduling. Hoboken, NJ: John Wiley & Sons.
Baker, K. R. 1984. “Sequencing Rules and Due-Date Assignments in a Job Shop.” Management Science30 (9): 1093–1104.
Behjat, S., and N. Salmasi. 2017. “Total Completion Time Minimisation of No-Wait Flowshop Group Scheduling Problem with Sequence
Dependent Setup Times.” European Journal of Industrial Engineering 11 (1): 22–48.
Blackstone, J. H., D. T. Phillips, and G. L. Hogg. 1982. “A State-of-the-Art Survey of Dispatching Rules for Manufacturing Job Shop
Operations.” The International Journal of Production Research 20 (1): 27–45.
Branke, J., T. Hildebrandt, and B. Scholz-Reiter. 2015. “Hyper-Heuristic Evolution of Dispatching Rules: A Comparison of Rule
Representations.” Evolutionary Computation 23 (2): 249–277.
Branke, J., S. Nguyen, C. W. Pickardt, and M. Zhang. 2016. “Automated Design of Production Scheduling Heuristics: A Review.” IEEE
Transactions on Evolutionary Computation 20 (1): 110–124.
Chang, F.-C. R. 1994. “A Study of Factors Affecting Due-Date Predictability in a Simulated Dynamic Job Shop.” Journal of
Manufacturing Systems 13 (6): 393–400.
Cheng, T. 1985. “Analysis of Job Flow-Time in a Job Shop.” Journal of the Operational Research Society, 36 (3): 225–230.
Cheng, T., and J. Jiang. 1998. “Job Shop Scheduling for Missed Due-Date Performance.” Computers & Industrial Engineering 34 (2):
297–307.
Chiang, T.-C., and L.-C. Fu. 2007. “Using Dispatching Rules for Job Shop Scheduling with Due Date-Based Objectives.” International
Journal of Production Research 45 (14): 3245–3262.
Doh, H.-H., J.-M. Yu, J.-S. Kim, D.-H. Lee, and S.-H. Nam. 2013. “A Priority Scheduling Approach for Flexible Job Shops with Multiple
Process Plans.” International Journal of Production Research 51 (12): 3748–3764.
Durasević, M., and D. Jakobović. 2018a. “Comparison of Ensemble Learning Methods for Creating Ensembles of Dispatching Rules for
the Unrelated Machines Environment.” Genetic Programming and Evolvable Machines 19 (1–2): 53–92.
Durasević, M., and D. Jakobović. 2018b. “Evolving Dispatching Rules for Optimising Many-Objective Criteria in the Unrelated Machines
Environment.” Genetic Programming and Evolvable Machines 19 (1–2): 9–51.
Ferreira, C. 2006a. “Automatically Defined Functions in Gene Expression Programming.” In Genetic Systems Programming, 21–56.
Springer.
Ferreira, C. 2006b. Gene Expression Programming: Mathematical Modeling by an Artificial Intelligence. Berlin: Springer.
Gholami, M., and M. Zandieh. 2009. “Integrating Simulation and Genetic Algorithm to Schedule a Dynamic Flexible Job Shop.” Journal
of Intelligent Manufacturing 20 (4): 481–498.
Grabot, B., and L. Geneste. 1994. “Dispatching Rules in Scheduling Dispatching Rules in Scheduling: A Fuzzy Approach.” The
International Journal of Production Research 32 (4): 903–915.
Heger, J., J. Branke, T. Hildebrandt, and B. Scholz-Reiter. 2016. “Dynamic Adjustment of Dispatching Rule Parameters in Flow Shops
with Sequence-Dependent Set-Up Times.” International Journal of Production Research 54 (22): 6812–6824.
Hein, F., C. Almeder, G. Figueira, and B. Almada-Lobo. 2018. “Designing New Heuristics for the Capacitated Lot Sizing Problem by
Genetic Programming.” Computers & Operations Research 96: 1–14.
Hildebrandt, T., J. Heger, and B. Scholz-Reiter. 2010. “Towards Improved Dispatching Rules for Complex Shop Floor Scenarios: A
Genetic Programming Approach.” In Proceedings of the 12th Annual Conference on Genetic and Evolutionary Computation,
257–264. ACM.
Holthaus, O., and C. Rajendran. 1997. “Efficient Dispatching Rules for Scheduling in a Job Shop.” International Journal of Production
Economics 48 (1): 87–105.
Hong, T., K. Jeong, and C. Koo. 2018. “An Optimized Gene Expression Programming Model for Forecasting the National co2 Emissions
in 2030 using the Metaheuristic Algorithms.” Applied Energy228: 808–820.
16 G. Ozturk et al.

Hunt, R., M. Johnston, and M. Zhang. 2014. “Evolving Machine-Specific Dispatching Rules for a Two-Machine Job Shop using Genetic
Programming.” In: IEEE Congress on Evolutionary Computation (CEC), 2014, 618–625. IEEE.
Ingimundardottir, H., and T. P. Runarsson. 2016. “Evolutionary Learning of Linear Composite Dispatching Rules for Scheduling.” In
Computational Intelligence, 49–62. Springer.
Jayamohan, M., and C. Rajendran. 2000. “New Dispatching Rules for Shop Scheduling: A Step Forward.” International Journal of
Production Research 38 (3): 563–586.
Jayamohan, M., and C. Rajendran. 2004. “Development and Analysis of Cost-Based Dispatching Rules for Job Shop Scheduling.”
European Journal of Operational Research 157 (2): 307–321.
Jeong, K.-C., and Y.-D. Kim. 1998. “A Real-Time Scheduling Mechanism for a Flexible Manufacturing System: Using Simulation and
Dispatching Rules.” International Journal of Production Research 36 (9): 2609–2626.
Ji, M., D. Yao, J. Ge, and T. Cheng. 2015. “Single-Machine Slack Due-Window Assignment and Scheduling with Past-Sequence-
Dependent Delivery Times and Controllable Job Processing Times.” European Journal of Industrial Engineering 9 (6):
794–818.
Kanet, J. J., and X. Li. 2004. “A Weighted Modified Due Date Rule for Sequencing to Minimize Weighted Tardiness.” Journal of
Scheduling 7 (4): 261–276.
Kapanoglu, M., and M. Alikalfa. 2011. “Learning If–Then Priority Rules for Dynamic Job Shops using Genetic Algorithms.” Robotics
and Computer-Integrated Manufacturing 27 (1): 47–55.
Karunakaran, D., Y. Mei, G. Chen, and M. Zhang. 2017. “Dynamic Job Shop Scheduling under Uncertainty using Genetic Program-
ming.” In Intelligent and Evolutionary Systems: The 20th Asia Pacific Symposium, IES 2016, Canberra, Australia, November
2016, Proceedings. Springer, 195–210.
Keshavarz, T., and N. Salmasi. 2014. “Efficient Upper and Lower Bounding Methods for Flowshop Sequence-Dependent Group
Scheduling Problems.” European Journal of Industrial Engineering 8 (3): 366–387.
Koza, J. R. 1992. Genetic Programming: On the Programming of Computers by Means of Natural Selection. Vol. 1. Cambridge: MIT
press.
Koza, J. R. 1994. Genetic Programming II: Automatic Discovery of Reusable Subprograms. Cambridge: MIT Press.
Land, M., M. Stevenson, and M. Thürer. 2014. “Integrating Load-Based Order Release and Priority Dispatching.” International Journal
of Production Research 52 (4): 1059–1073.
Lin, S.-C., E. D. Goodman, and W. F. Punch III. 1997. “A Genetic Algorithm Approach to Dynamic Job Shop Scheduling Problem.” In
ICGA, 481–488.
Liu, S. Q., and E. Kozan. 2016. “Parallel-Identical-Machine Job-Shop Scheduling with Different Stage-Dependent Buffering Require-
ments.” Computers & Operations Research 74: 31–41.
Lu, H., G. Q. Huang, and H. Yang. 2011. “Integrating Order Review/Release and Dispatching Rules for Assembly Job Shop Scheduling
using a Simulation Approach.” International Journal of Production Research 49 (3): 647–669.
Manupati, V., A. Revanth, K. Srikanth, A. Maheedhar, and M. S. Reddy. 2016. “Real-Time Rule-Based Scheduling System for Integrated
Delivery in a Semiconductor Manufacturing using Evolutionary Algorithm-Based Simulation Approach.” In Artificial Intelligence
and Evolutionary Computations in Engineering Systems, 981–989. Springer.
Nguyen, S., Y. Mei, B. Xue, and M. Zhang. 2018. “A Hybrid Genetic Programming Algorithm for Automated Design of Dispatching
Rules.” Evolutionary Computation 1–31. doi:10.1162/evco_a_00230.
Nguyen, S., Y. Mei, and M. Zhang. 2017. “Genetic Programming for Production Scheduling: A Survey with a Unified Framework.”
Complex & Intelligent Systems 3 (1): 41–66.
Nguyen, S., M. Zhang, M. Johnston, and K. C. Tan. 2019. “Genetic Programming for Job Shop Scheduling.” In Evolutionary and Swarm
Intelligence Algorithms, 143–167. Springer.
Nie, L., L. Gao, P. Li, and X. Li. 2013a. “A GEP-Based Reactive Scheduling Policies Constructing Approach for Dynamic Flexible Job
Shop Scheduling Problem with Job Release Dates.” Journal of Intelligent Manufacturing 24 (4): 763–774.
Nie, L., L. Gao, P. Li, and X. Shao. 2013b. “Reactive Scheduling in a Job Shop where Jobs Arrive over Time.” Computers & Industrial
Engineering 66 (2): 389–405.
Nie, L., X. Shao, L. Gao, and W. Li. 2010. “Evolving Scheduling Rules with Gene Expression Programming for Dynamic Single-Machine
Scheduling Problems.” The International Journal of Advanced Manufacturing Technology 50 (5–8): 729–747.
Ochoa, G., J. A. Vázquez-Rodríguez, S. Petrovic, and E. Burke. 2009. “Dispatching Rules for Production Scheduling: A Hyper-Heuristic
Landscape Analysis.” In IEEE Congress on Evolutionary Computation, 2009. CEC’09, 1873–1880. IEEE.
Panwalkar, S. S., and W. Iskander. 1977. “A Survey of Scheduling Rules.” Operations Research 25 (1): 45–61.
Park, J., Y. Mei, G. Chen, and M. Zhang. 2016. “Niching Genetic Programming Based Hyper-Heuristic Approach to Dynamic Job
Shop Scheduling: An Investigation into Distance Metrics.” In Proceedings of the 2016 on Genetic and Evolutionary Computation
Conference Companion, 109–110. ACM.
Park, J., Y. Mei, S. Nguyen, G. Chen, and M. Zhang. 2018. “An Investigation of Ensemble Combination Schemes for Genetic
Programming Based Hyper-Heuristic Approaches to Dynamic Job Shop Scheduling.” Applied Soft Computing 63: 72–86.
Park, J., S. Nguyen, M. Zhang, and M. Johnston. 2015. “Evolving Ensembles of Dispatching Rules using Genetic Programming for Job
Shop Scheduling.” In European Conference on Genetic Programming, 92–104. Springer.
Pickardt, C. W., T. Hildebrandt, J. Branke, J. Heger, and B. Scholz-Reiter. 2013. “Evolutionary Generation of Dispatching Rule Sets for
Complex Dynamic Scheduling Problems.” International Journal of Production Economics 145 (1): 67–77.
 International Journal of Production Research 17

Pierreval, H., and N. Mebarki. 1997. “Dynamic Scheduling Selection of Dispatching Rules for Manufacturing System.” International
Journal of Production Research 35 (6): 1575–1591.
Qu, S., J. Wang, and G. Shivani. 2016. “Learning Adaptive Dispatching Rules for a Manufacturing Process System by using Reinforcement
Learning Approach.” In 2016 IEEE 21st International Conference on Emerging Technologies and Factory Automation (ETFA),
1–8. IEEE.
Rajendran, C., and O. Holthaus. 1999. “A Comparative Study of Dispatching Rules in Dynamic Flowshops and Jobshops.” European
Journal of Operational Research 116 (1): 156–170.
Rochette, R., and R. P. Sadowski. 1976. “A Statistical Comparison of the Performance of Simple Dispatching Rules for a Particular Set
of Job Shops.” The International Journal of Production Research14 (1): 63–75.
Rudek, R. 2013. “Minimising Maximum Lateness in a Single Machine Scheduling Problem with Processing Time-Based Aging Effects.”
European Journal of Industrial Engineering 7 (2): 206–223.
Samadianfard, S., E. Asadi, S. Jarhan, H. Kazemi, S. Kheshtgar, O. Kisi, S. Sajjadi, and A. A. Manaf. 2018. “Wavelet Neural Networks
and Gene Expression Programming Models to Predict Short-Term Soil Temperature at Different Depths.” Soil and Tillage Research
175: 37–50.
Sels, V., N. Gheysen, and M. Vanhoucke. 2012. “A Comparison of Priority Rules for the Job Shop Scheduling Problem under Different
Flow Time-and Tardiness-Related Objective Functions.” International Journal of Production Research 50 (15): 4255–4270.
Sharma, P., and A. Jain. 2015. “Performance Analysis of Dispatching Rules in a Stochastic Dynamic Job Shop Manufacturing System with
Sequence-Dependent Setup Times: Simulation Approach.” CIRP Journal of Manufacturing Science and Technology 10: 110–119.
Soroush, H. 2010. “Single-Machine Scheduling with Inserted Idle Time to Minimise a Weighted Quadratic Function of Job Lateness.”
European Journal of Industrial Engineering 4 (2): 131–166.
Subramaniam, V., G. Lee, G. Hong, Y. Wong, and T. Ramesh. 2000a. “Dynamic Selection of Dispatching Rules for Job Shop Scheduling.”
Production Planning & Control 11 (1): 73–81.
Subramaniam, V., T. Ramesh, G. Lee, Y. Wong, and G. Hong. 2000b. “Job Shop Scheduling with Dynamic Fuzzy Selection of Dispatching
Rules.” The International Journal of Advanced Manufacturing Technology 16 (10): 759–764.
Talbi, E.-G. 2009. Metaheuristics: From Design to Implementation. Vol. 74. Hoboken, NJ: John Wiley & Sons.
Tay, J. C., and N. B. Ho. 2008. “Evolving Dispatching Rules using Genetic Programming for Solving Multi-Objective Flexible Job-Shop
Problems.” Computers & Industrial Engineering 54 (3): 453–473.
Teymourifar, A., and G. Ozturk. 2018. “New Dispatching Rules and Due Date Assignment Models for Job Shop Scheduling Problems.”
International Journal of Manufacturing Research 13 (4): 302–329.
Teymourifar, A., G. Ozturk, Z. K. Ozturk, and O. Bahadir. 2018. “Extracting New Dispatching Rules for Multi-Objective Dynamic
Flexible Job Shop Scheduling with Limited Buffer Spaces.” Cognitive Computation 0 (0): 1–11. doi:10.1007/s12559-018-9595-4.
Xiong, H., H. Fan, G. Jiang, and G. Li. 2017. “A Simulation-Based Study of Dispatching Rules in a Dynamic Job Shop Scheduling
Problem with Batch Release and Extended Technical Precedence Constraints.” European Journal of Operational Research 257
(1): 13–24.
Yin, Y., C.-C. Wu, W.-H. Wu, and J.-C. Chen. 2013. “Single-Machine Group Scheduling with a General Learning Effect.” European
Journal of Industrial Engineering 7 (3): 350–369.
Zahmani, M. H., B. Atmani, A. Bekrar, and N. Aissani. 2015. “Multiple Priority Dispatching Rules for the Job Shop Scheduling Problem.”
In 2015 3rd International Conference on Control, Engineering & Information Technology (CEIT), 1–6. IEEE.
Zhang, L., Q. Tang, Z. Wu, and F. Wang. 2017. “Mathematical Modeling and Evolutionary Generation of Rule Sets for Energy-Efficient
Flexible Job Shops.” Energy 138: 210–227.
Zhou, C., W. Xiao, T. M. Tirpak, and P. C. Nelson. 2003. “Evolving Accurate and Compact Classification Rules with Gene Expression
Programming.” IEEE Transactions on Evolutionary Computation 7 (6): 519–531.