Journal of Simulation

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/tjsm20

A simheuristic approach using the NSGA-II to solve

a bi-objective stochastic flexible job shop problem

Camilo Andrés Rodríguez-Espinosa, Eliana María González-Neira & Gabriel

Mauricio Zambrano-Rey

To cite this article: Camilo Andrés Rodríguez-Espinosa, Eliana María González-Neira &
Gabriel Mauricio Zambrano-Rey (2023): A simheuristic approach using the NSGA-II to
solve a bi-objective stochastic flexible job shop problem, Journal of Simulation, DOI:
10.1080/17477778.2023.2231877

To link to this article: https://doi.org/10.1080/17477778.2023.2231877

Published online: 05 Jul 2023.

Submit your article to this journal

View related articles

View Crossmark data

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

https://www.tandfonline.com/action/journalInformation?journalCode=tjsm20
JOURNAL OF SIMULATION
https://doi.org/10.1080/17477778.2023.2231877

RESEARCH ARTICLE

A simheuristic approach using the NSGA-II to solve a bi-objective stochastic

flexible job shop problem
Camilo Andrés Rodríguez-Espinosa, Eliana María González-Neira and Gabriel Mauricio Zambrano-Rey
Departamento de Ingeniería Industrial, Facultad de Ingeniería, Pontificia Universidad Javeriana, Bogotá, Colombia

ABSTRACT ARTICLE HISTORY

This paper addresses a bi-objective problem in flexible job shop scheduling (FJSS) with Received 11 August 2022
stochastic processing times. Following the Just-In-Time philosophy, the first objective is to Accepted 20 June 2023
minimise deterministic Earliness+Tardiness, and the second objective is to minimise the KEYWORDS
Earliness+Tardiness Risk. The second objective function seeks to obtain robust solutions Stochastic Flexible Job Shop
under uncertain environments. The proposed approach is a simheuristic that hybridises the Scheduling; Earliness
non-dominated sorting genetic algorithm (NSGA-II) with Monte Carlo simulation to obtain the +tardiness; Simulation-
Pareto frontier of both objectives. The computational results demonstrate the effectiveness of optimisation; Robustness;
the proposed algorithm under different variability environments. Non-dominated sorting
genetic algorithm (NSGA-II)

1. Introduction
Production scheduling is a decision-making process studied considering stochastic processing times
that involves the allocation of resources to tasks over because, as stated by Pinedo (2016), other sources of
given time periods to achieve one or more specific uncertainty, such as machine breakdowns, can be mod­
goals. Particularly, the job shop scheduling (JSS) pro­ elled as an internal part of the processing times. There
blem is an environment in which each job has are three main approaches for scheduling under uncer­
a particular route of operations that must be followed tainties: reactive, predictive, and predictive-reactive
and each of those can be processed only by a single (Aytug et al., 2005; Z. Li & Ierapetritou, 2008).
machine. The generalisation of this environment is the Reactive scheduling sequences jobs with real-time dis­
flexible job shop scheduling (FJSS), in which each patching algorithms that create partial schedules based
operation can be processed in one or more machines on limited and specific information. Predictive methods
(Pinedo, 2016). Due to its combinatorial nature, the consist in creating a schedule that minimises the effects
JSS and, consequently, the FJSS are classified as NP- of disruptions. In this category, five approaches have
hard (Demir & İ̇̇şleyen, 2013; Y. Yang et al., 2020). The been reported in the literature: stochastic, robust, fuzzy,
FJSS problem exists in many important industries, sensitivity analysis, and parametric programming (Z. Li
such as chemical material processing (Kacem et al., & Ierapetritou, 2008). Predictive-reactive techniques
2002), aeronautical (J. Zhang & Yang, 2016), semicon­ combine the predictive and reactive approaches, but
ductor manufacturing (Tamssaouet et al., 2022), and this process must deal with the right timing to start
steelmaking (J.-Q. Li et al., 2020), among others. the rescheduling action, and also with the action that
Following the just in time (JIT) philosophy, which should be taken. For this work, the predictive approach
traditionally implies not finishing jobs too early, to is chosen because its results are the basis for planning
prevent increments in finished product inventory support activities, which allows the estimation of com­
costs; and not delivering jobs too late, to avoid incur­ pletion times and costs.
ring penalties with customers, the minimisation of the Considering the stochastic environment, robust
sum of Earliness and tardiness is considered as the scheduling is considered to minimise the effects of
main objective function of the study (Gonzalez-Neira the uncertainties in the predictive schedule. For this
et al., 2021; Horng et al., 2012). purpose, the robust measure Earliness+Tardiness Risk
Seeking to have a representation that more closely is used to ensure that the predictive schedule does not
resembles reality, the performance of a schedule can be differ excessively from the realised schedule, maintain­
influenced by stochastic factors, such as machine break­ ing a high-level performance of the objective. In that
downs, the release of jobs or processing times of opera­ way, the obtained schedule uses the limited resources
tions (Horng et al., 2012; Pinedo, 2016; Shen et al., efficiently and allows sufficient flexibility for changes
2017). In this case, the stochastic version of the FJSS is (Al-Hinai & ElMekkawy, 2011).

CONTACT Camilo Andrés Rodríguez-Espinosa [email protected] Departamento de Ingeniería Industrial, Facultad de Ingeniería,
Pontificia Universidad Javeriana, Bogotá, Colombia
© 2023 The Operational Research Society
2 C. A. RODRÍGUEZ-ESPINOSA ET AL.

One way to address the presence of uncertainty in finance-based scheduling (El-Abbasy et al., 2020),
complex optimisation problems is through simheur­ construction (Bagheri-Esfeh et al., 2020), logistic dis­
istics, which are the combination of simulation tech­ tribution network (Wang et al., 2021), and so forth.
niques and metaheuristic algorithms (Juan et al., The rest of the paper is organised as follows.
2015). Simheuristics have been successfully applied Section 2 is dedicated to reviewing studies related to
for solving vehicle routing (González-Martín et al., stochastic FJSS (SFJSS) and/or robust FJSS. In
2016; Juan et al., 2011), and production scheduling Section 3, the problem formulation is presented. In
(González-Neira et al., 2019; Hatami et al., 2018; Section 4, the proposed simheuristic is explained. The
Juan et al., 2014), among others. Simheuristics are results of the simheuristic performance are reported in
extremely flexible, allowing the integration of different Section 5, and then the conclusions and future work
metaheuristic algorithms and simulation techniques are stated in Section 6.
(Juan et al., 2015). According to Chica et al., (2017)
the paradigms of simulation that can be used to build
a simheuristic are Monte Carlo, discrete event, system 2. Literature review
dynamics, and agent-based modelling. Of these, the
Since the aim of this paper is to solve a stochastic and
Monte Carlo simulation requires the least computa­
robust FJSS through a simheuristic, this section is
tional consumption. For that reason, Monte Carlo
divided into three parts. The first subsection reports
paradigm is selected for this study.
some literature related to stochastic and/or robust
Considering that the problem to be solved has two
FJSS problems. The second part presents some works
objectives, a multi-objective optimisation approach
related to the application of simheuristics in optimisa­
has to be considered. Although these types of pro­
tion problems other than scheduling ones. The third
blems address real-life scenarios better, they include
subsection states the elements on which this study
an additional difficulty in terms of handling their
focuses on and its contribution.
objectives, which have to be analysed in terms of the
trade-off between them, called the Pareto optimal set
(Türkyılmaz et al., 2020; Von Lücken et al., 2014).
2.1. Literature in stochastic and/or robust FJSS
Different techniques have been implemented to solve
this problem, such as the weighted sum method, hier­ A search was executed on the SCOPUS and ISI Web of
archical optimisation method, lexicographic ordering, Science databases, using the following terms:
successive Pareto optimisation, methods based on (“Stochastic” OR “Robust”) AND (“Flexible Job Shop
evolutionary algorithms, and methods based on Scheduling”). Around 50 papers were found, and after
swarm intelligence, among others (Türkyılmaz et al., filtering out those related to job shops and merely
2020). For this case, the non-dominated sorting reactive approaches, only 20 papers, dealing with the
genetic algorithm II (NSGA-II) is proposed to obtain flexible job shop problem, are reported hereafter. The
the Pareto frontier of the objective functions using following analysis is presented considering the source
Monte Carlo simulation to evaluate the performance of uncertainty, the objective functions, the method
of the robust measure among different random sce­ used to build the schedule, the robust measure and
narios (Caldeira & Gnanavelbabu, 2021). NSGA-II is the insertion of simulation to account for the stochas­
characterised by its low computational requirements, tic behaviour. Based on our findings, the literature on
elitist approach and parameter-less sharing approach stochastic and/or robust FJSS can be presented in two
to improve solutions in different applications, obtain­ major groups. In the first group of papers (subsection
ing better results than the Pareto-archived evolution­ 2.1.1), it can be seen that the objective function does
ary strategy (PAES) and the strength Pareto not include any robust or stability performance
evolutionary algorithm (SPEA) in terms of the spread metrics. On the contrary, the second group (subsec­
of solutions and convergence near the true Pareto- tion 2.1.2) includes those methods that aim to mini­
optimal front (Deb et al., 2002) and better results mise not only classic production-oriented metrics
than the non-dominated ranking genetic algorithm (e.g., makespan, flowtime or due-date related func­
(NRGA) in terms of the spacing, convergence of the tions) but also robustness and stability functions in
Pareto front and quantity of obtained solutions in order to produce schedules that consider the impact of
a stochastic FJSS problem (Ahmadi et al., 2016). The uncertain events. Both groups are analysed to deter­
NSGA-II simultaneously optimises each objective mine if any of the paradigms of simulation has been
without being dominated by any other solution integrated into the solution approach. Those papers
(Yusoff et al., 2011). The NSGA-II has great perfor­ that have integrated any type of simulation with
mance with a small number of objectives but as this a metaheuristic can be classified as research that has
number increases the performance of this algorithm implemented a simheuristic approach, and this repre­
gets worse (Von Lücken et al., 2014). Some areas sents 45% of the stochastic and/or robust literature
where the NSGA-II has been implemented are: consulted.
 JOURNAL OF SIMULATION 3

2.1.1. Objective functions without robustness/ and falsity memberships, instead of using only one
stability measures membership function. Then, the neutrosophic num­
Within this category, the reported papers are also bers were used to develop an improved teaching-
organised depending on the use of simulation to learning-based optimisation (TLBO) algorithm, using
account for uncertainty (see Table 1). For instance, genetic operators, such as crossover. The performance
Ayyoubzadeh et al. (2021) address the SFJSS under of the proposed approach was studied under small,
two kinds of uncertainty. The first kind, tries to repro­ medium and large uncertainty scenarios, to configure
duce the impact of resource unavailability, the lack of the neutrosophic processing times.
skilled workers and inadequate estimation of opera­ Hybrid simulation-optimisation techniques have
tions, which can be modelled using pre-known prob­ been used for decades, but flourished in the last 20
ability distributions affecting the start and processing years, mainly because of the increase in computa­
times of the jobs. Conversely, certain parameters, such tional power. Particularly, in simulation-based
as machine breakdowns, order cancellation or mod­ optimisation, the optimisation algorithm controls
ification, as well as new order arrivals may not follow the execution by calling the simulation model
well-known probability distributions; hence, a reactive whenever needed (Juan et al., 2015). This is the
scheduling strategy is proposed for this type of disrup­ case of simheuristic approaches, also known as
tion. The authors proposed a non-dominated sorting optimisation via simulation. For example, Gocken
genetic algorithm (NSGA-II) to deal with two conflict­ et al. (2019) proposed an optimisation via simula­
ing objective functions: minimising the tax cost on tion approach for the SFJSS. The optimisation part
surplus energy consumption and minimising the is based on a genetic algorithm (GA), which calls
total cost of jobs lateness based on soft time- upon a discrete-event simulation (DES) model of
windows. The chromosome structure is divided into the system, coded in SIMIO 6.1. Process plans and
two sub-chromosomes: the first sub-chromosome, machine assignments are explored by the GA and
focuses on the stochastic start times, modelled using sent as decision variables to the simulation model,
a uniform distribution. The second sub-chromosome, which returns the objective function value. The
accommodates minor disruptions based on the origi­ arrival rate of parts is exponentially distributed.
nal schedule set by the first sub-chromosome. Also using DES, Gupta and Jain (2022) proposed
Similarly, X. Yang et al. (2016) proposed a NSGA-II a simulation-based optimisation approach for
to tackle stochastic processing times, aiming to mini­ simultaneous preventive maintenance, and random
mise the completion time and total energy consump­ machine breakdown environment with sequence-
tion. The NSGA-II first finds a basic scenario, and dependent setup time (SDST). For the optimisation
then, using parameter sensitivity analysis, the change part, the desirability function approach is used to
of average values of the objective function is observed optimise four different performance measures, i.e.,
based on disturbed processing times, subject to mean flowtime, makespan, mean tardiness, and
a uniform distribution, with different degrees of sever­ total setups. In turn, the simulation model (imple­
ity. The analysis determines the level of machine mented in PROMODEL) is configured to represent
redundancy to absorb such perturbations and its con­ the stochastic behaviour of the system in terms of
sequences in performance loss of the scheduling job arrivals following a Poisson distribution, uni­
scheme. formly distributed processing times, and machine
Likewise, Jin et al. (2021) proposed a population- availability described by two parameters, the mean
based metaheuristic approach to deal with the SFJSS time between failures (MTBF) and the mean time
under stochastic processing times modelled using tri­ to repair (MTTR), which are exponentially distrib­
angular neutrosophic numbers. Neutrosophic sets, uted. As mentioned by Gocken et al. (2019), using
unlike fuzzy sets, quantify the truth, indeterminacy DES enables the representation of complex

Table 1. Works that do not consider robustness/stability measures.

Authors Source of Uncertainty Objective Function Method Use of simulation?
Ayyoubzadeh Start time Processing time Machine Min Tax cost on surplus energy NSGA-II No, random times inserted in
et al. (2021) breakdowns Order cancellation Order consumption Min Total cost of jobs chromosome
modification Order arrivals lateness
X. Yang et al. Processing times Min Completion time Min Total NSGA-II No, parameter sensitivity
(2016) energy consumption analysis based on
uncertainty levels
Jin et al. Processing times Min Makespan (Cmax ) TLBO No, Single-valued
(2021) neutrosophic sets
Gocken et al. Part arrivals Min Total average flowtime of each GA Yes, discrete-event
(2019) part simulation
Gupta and Machine breakdowns Processing times Min Mean flow time Min Makespan Desirability Yes, discrete-event
Jain (2022) (Cmax ) Min Mean tardiness Min function simulation
Total setups approach
4 C. A. RODRÍGUEZ-ESPINOSA ET AL.

environments in which it is hardly to obtain tract­ 2.1.2. Objective functions considering robustness/
able mathematical formulations. However, coupling stability measures
simulation models to optimisation requires much The most significant and widespread objectives
more time to evaluate each scenario, risking high address minimising the completion and flow times,
computational costs. or due-date related functions (see Table 2). Hence,
Another way to simulate sample scenarios and schedules generated to meet such purposes tend to
obtain estimates of objective functions is by cou­ be tight and lack responsiveness to uncertainties.
pling Monte Carlo simulation to an optimisation Therefore, the degradation and stability of the objec­
method. Monte Carlo simulation is used to deter­ tive in the presence of disruptions, can be integrated as
mine which scenario, among the best-found deter­ measures of a schedule’s robustness, among other
ministic ones, shows the best objective function similar metrics (Sajadi et al., 2019). Thus,
considering stochastic parameters (Juan et al., a definition of degradation or regret can be related to
2015). As concerns the optimisation method, heur­ the deviation between the expected value of
istic and metaheuristic approaches are usually a performance measure (e.g., makespan), under
implemented. For example, Dosdoğru et al. (2015) a specific scenario, and its corresponding optimal or
studied the impact of routing flexibility on the pre-established value. Inspired by this robustness
expected makespan, within a FJSS, integrating measure, Hao et al. (2014) proposed a Markov net­
a genetic algorithm with Monte Carlo simulation. work-based estimation of distribution algorithm to
Processing times were randomly generated using solve the SFJSS under uncertain processing times.
a triangular distribution, and the number of opera­ The processing times follow a uniform distribution
tions was uniformly distributed. Focusing on the within defined lower and upper bounds. So, proces­
same objective function, Lin et al. (2022) implemen­ sing time uncertainty can be constructed through the
ted a learning-based grey wolf optimiser to schedule concept of a scenario that corresponds to the assign­
jobs in a semiconductor manufacturing environ­ ment of a reasonable processing time to each opera­
ment. The processing times following a gamma dis­ tion on each machine. Then, a finite set of scenarios is
tribution are also generated with Monte Carlo sampled from a set of random processing times gen­
simulation. On the same lines, Caldeira and erated with the Cartesian product. For each scenario,
Gnanavelbabu (2021) modelled stochastic proces­ the expected regret value is estimated, and several
sing times with a log-normal distribution, to repro­ scenarios are carried out to minimise the maximum
duce non-negative values. The authors applied the expected regret. Alternatively, a variety of processing
Jaya algorithm due to the absence of algorithm- times scenarios can be set in advance to calculate a set
specific parameters that require to be tuned. Thus, of scenarios. Bearing this in mind, Shen et al. (2017)
experiments can focus on the effect of stochastic developed a modified multi-objective evolutionary
processing times without the interference of the algorithm based on decomposition (m-MOEA/D)
optimisation algorithm’s parameters. with two scenario-based robustness measures based
As mentioned before, another source of uncertainty on the relative variation between uncertain makespan
that has attracted interest is machine availability. and workload, and their corresponding deterministic
Rahmati et al. (2018) considered two maintenance values.
scenarios in terms of corrective maintenance (CM), Scenario-based approaches calculate the deviations
caused by machine disruptions between inspection between the efficiency performances obtained from
intervals, and preventive maintenance (PM). The various scenarios, and the performance produced in
event time activity is triggered by a stochastic variable the initial deterministic scenario. Various works have
that denotes shock or the machine degradation level applied this strategy by proposing two-stage
and follows an exponential distribution. The stochas­ approaches. Taking this into account, Nouiri et al.
tic time between two shocks is also exponentially dis­ (2017) proposed a two-stage particle swarm optimisa­
tributed. Moreover, the duration of PM activities, the tion (PSO) to solve the problem considering schedule
recovery level through PM and the duration of CM all stability. In the first stage, the PSO focuses on mini­
follow log-normal distributions, with different para­ mising only the makespan, assuming deterministic
meters. To achieve this, the harmony search optimisa­ parameters, meaning no disruptions. Then,
tion algorithm is coupled with an optimisation the second stage takes the final population of the
module (not explicitly stated as Monte Carlo simula­ PSO as the initial one, to improve a bi-objective func­
tion but it can be assumed) to find robust schedules tion composed of a weighted sum of the makespan and
aiming to minimise the expected makespan (Cmax ). a stability measure. This latter is similar to the regret
This problem shows the power of simulation-based value mentioned before, because its calculated as the
optimisation (i.e., simheuristics) to study various difference between the completion times found in the
sources of uncertainty within a complex manufactur­ first stage and those found in the second stage.
ing setting. Machine breakdowns are inserted into the second
 JOURNAL OF SIMULATION 5

Table 2. Works with objective functions considering robustness/stability measures.

Source of Main Objective
Authors Uncertainty Function Method Robustness measures Use of simulation?
Shen et al. Processing times Min Makespan (Cmax ) m – MOEA/D Relative variety of Makespan and No, random processing times
(2017) Min Max workload inserted into the optimisation
Workload method
Nouiri et al. Machine Min Makespan (Cmax ) 2S – PSO Stability = Min Difference of No, scenario – based approach on
(2017) breakdowns completion times the second stage
Al-Hinai and Machine Min Makespan (Cmax ) 2S – HGA Stability No, scenario – based approach on
ElMekkawy breakdowns the second stage
(2011)
He and Sun Machine Min Makespan (Cmax ) GA plus Right- Robustness Stability No, scenario – based approach on
(2013) breakdowns Shift Route the second stage
Changing
Thi et al. (2022) Machine Min Makespan (Cmax ) GA Three re – Robustness Stability No, scenario – based approach on
breakdowns scheduling the second stage
policies
Buddala and Machine Min Makespan (Cmax ) 2S – TLBO Robustness Stability
Mahapatra breakdowns
(2019)
Ahmadi et al. Machine Min Makespan (Cmax ) NSGA – II NRGA Stability = Min Difference of Yes, incorporated simulation
(2016) breakdowns completion times module
Sajadi et al. Machine Min Realised GA – NSGA–II Stability = Min Difference of Yes, incorporated simulation
(2019) breakdowns Makespan (Cmax ) NRGA completion times module
Xiong et al. Machine Min Makespan (Cmax ) NSGA – II SM1 (Workload, float times) SM2 Yes, Monte Carlo simulation
(2013) breakdowns (distance between central
points)
Mokhtari and Processing times Min Tardy jobs MILP Simulated Machine availability Yes, Monte Carlo simulation
Dadgar Due-dates Annealing (SA)
(2015)
Gnanavelbabu Processing times Min Makespan (Cmax ) MBSA Min Std Cmax Yes, Monte Carlo simulation
et al. (2021) Worker
flexibility
Gomez et al. Processing times TS Max Service level Yes, Monte Carlo simulation
(2021)
Y. Yang et al. Machine Min Makespan (Cmax ) NSGA – II RMc based on the ELM Yes, simulation-based method
(2020) breakdowns
G. Zhang et al. Machine Min Makespan (Cmax ) ICA RMn based on CNN
(2022) breakdowns

stage to evaluate the effect of such disruptions on the is achieved with a genetic algorithm aiming to mini­
predictive schedule performance. It is assumed that mise the makespan. Re-scheduling is done to mini­
information about the uncertainty of machine break­ mise the weighted sum of the relative robustness and
downs is known in advance. Several scenarios are stability measures. The relative robustness measure
created based on which machine breaks down, when is the deviation percentage between the predictive
it occurs and how long the machine is unavailable. The and the realised schedule (after the breakdown), and
first value is found using an empirical relation related the stability measure is the normalised difference
to the machines’ workload, and the second and third between the predicted and realised completion
values are uniformly distributed with different para­ times, only for those operations affected by the
meters. Instead of using PSO, Al-Hinai and breakdown. A follow-up work by D. Sun et al.
ElMekkawy (2011) implemented a two-stage hybrid (2014) included a game theory approach to find the
genetic algorithm (HGA) to deal with random Nash equilibrium between the robustness and stabi­
machine breakdowns. The computational cost of sce­ lity measures. Likewise, Thi et al. (2022) explored
nario-based approaches depends closely on the num­ three different adjustments in the re-scheduling
ber of scenarios, and their efficiency is related to the stage to reduce the makespan, and ensure robustness
type scenario variety and representation of the actual and stability as defined by the aforementioned
behaviour of the system (e.g., long and short duration, works. The pre-scheduled sequence with an uncer­
early or late disruptions, among others). For instance, tain repair time is modified by combined route-
in the latter, 192000 test problems were evaluated in changing and right-shifting; operation insertion on
the second-stage. unaffected machines, and the generation of a new
Following the same two-stage approach, He and schedule that adapts to the breakdown at the lowest
Sun (2013) proposed an idle time insertion strategy cost. Buddala and Mahapatra (2019) also followed
in the pre-schedule (first stage) and a route- the aforementioned strategy but using a two-stage
changing strategy combined with a right-shift policy teaching-learning-based optimisation (2S-TLBO)
and route-changing strategies to keep the reschedul­ and a rescheduling technique called modified
ing (second stage) as stable and robust as possible in affected operations rescheduling (mAOR) was used
the presence of machine disruptions. Pre-scheduling to generate robust schedules in the second stage.
6 C. A. RODRÍGUEZ-ESPINOSA ET AL.

Another way to achieve two-stage strategies is by two surrogate measures: (1) the first uses the machine
implementing a training predictive model in the first workload and float times, considering that the float
stage and a robustness prediction scheduling in time on a machine with a heavier workload is more
the second one, as reported by Y. Yang et al. (2020) important than the one with a lighter workload; and
and G. Zhang et al. (2022). The former aims to mini­ (2) the second uses the average central points between
mise the makespan and a proposed robustness mea­ the float time and breakdown periods. A NSGA-II was
sure called the RMc. The RMc was developed using an implemented to study five different robustness mea­
extensive learning machine (ELM) in two phases: the sures, finding that the second SM performs better for
first phase collects training samples of schedules com­ both small and relatively large cases, and the first one
prising workloads and float times, and the second only on small cases.
phase predicts the value of the RMc. In both phases, Although machine breakdowns can be studied as
a NSGA-II is in charge of producing the schedules. the main source of uncertainty, processing times give
The latter follows the same strategy, although in the us the possibility to capture machine deterioration,
first stage, the training is executed by a convolutional setup times, and worker efficiency, among other sto­
neural network (CNN). Instead of an evolutionary chastic parameters. Focused on the effect of uncertain
approach, an improved imperialist competition algo­ processing times and due dates, Mokhtari and Dadgar
rithm (ICA) is proposed to generate training sche­ (2015) reported a mixed integer linear programming
dules. Another robustness measure was proposed, (MILP) model with the objective of minimising the
called RMn, based on machine breakdown, workload number of tardy jobs. Also, a simulation-optimisation
and the float time of the operation. framework based on a simulated annealing (SA) opti­
The aforementioned two-stage approaches are an miser embedded with local search strategies call up
alternative to simulation in terms of random value a Monte Carlo (MC) simulator in such a way that the
generation for stochastic parameters. However, minimum value of total system availability is satisfied.
Monte Carlo and discrete-event simulation are resour­ In addition to processing times, Gnanavelbabu et al.
ceful tools to generate stochasticity and therefore find (2021) presented a simheuristic approach incorporat­
robust and stable schedules according to known beha­ ing worker flexibility to minimise the expected make­
viours. Ahmadi et al. (2016) addressed the stable sche­ span and standard deviation of the makespan. The
duling of a multi-objective FJSS with random machine optimisation part is realised with a multi-objective
breakdowns. Two evolutionary algorithms, NSGA-II modified backtracking search algorithm (MBSA) fra­
and a non-dominated ranking genetic algorithm mework that integrates the Monte Carlo simulation.
(NRGA), are applied to combine the improvement of To analyse the stochastic variant, four variation levels
the makespan and stability simultaneously. and four probability distributions of the processing
A simulation module, similar to Monte Carlo, is incor­ times were studied. It is interesting to note that the
porated into the optimisation methods, to reproduce log-normal distribution resulted in better quality solu­
machine disruptions following exponential distribu­ tions, while the gamma distribution gave the worst
tions for breakdown occurrences and repair times. results. The Weibull and uniform distributions exhib­
As well as the previous case, Sajadi et al. (2019) also ited competitive performances compared to a log-
applied NSGA-II and NRGA but in the second stage, normal distribution.
and a genetic algorithm was implemented in the first Rather than estimating the classical makespan,
stage to focus on makespan minimisation. Gomez et al. (2021) proposed to maximise the prob­
Surrogate measures (SM) are designed to approx­ ability of the makespan being smaller than
imate the robustness based on some characteristic a predefined value, called the makespan service level.
parameters, such as the float time and free slacks This robustness measure seeks to compensate for the
(Shen et al., 2017). For instance, the average float uncertainty of processing times. The service level of
times, i.e., total slack, are computed as the difference a given schedule, constructed by a tabu search (TS)
between the earliest start time and the latest start time procedure, is evaluated using a Monte Carlo sampling
of an activity. Another slack-based SM is the free slack, procedure. Each operation with a stochastic proces­
defined as the amount of time that an activity can be sing time is sampled, a fixed number of times, accord­
delayed without delaying the start of the immediately ing to a beta distribution to generate a set of scenarios.
following activity. Xiong et al. (2013) argued that
slack-based measures for the robustness of a schedule
2.2. Simheuristics applications in other
focus on the structure of the schedule but neglect the
optimisation problems
uncertainty. Therefore, known information about
machines can be utilised as a baseline to construct Simulation-optimisation and simulation-based opti­
schedules, such as the probability of breakdown occur­ misation offer decision-makers the possibility of deal­
rence and the location of the breakdown for each ing with complexity and uncertainty. In particular, the
machine. Based on those data, the authors proposed use of heuristic and metaheuristic approaches, for the
 JOURNAL OF SIMULATION 7

optimisation part, has gained a lot of interest because arc routing problem (TCARP) with stochastic
most real size problems are NP-hard, difficult to demands. Similarly, Reyes-Rubiano et al. (2019) also
model and require astounding computing resources. dealt with the routing problem of electric vehicles,
Moreover, the lack of precise information and the with the goal of minimising the expected time-based
variability of problem parameters make simulation cost required to complete the freight distribution plan.
models the perfect fit to heuristics/metaheuristics in A simheuristic algorithm that combines Monte Carlo
order to understand and cope with the dynamics of simulation with a multi-start metaheuristic was used
a system. Because of those benefits, simheuristic appli­ to analyse the effect of random travel times and the use
cations are found in various areas, such as transporta­ of energy safety stocks on the routing plans.
tion, logistics, healthcare production, project In portfolio optimisation, the objective is to mini­
management, finance and social systems (Chica mise risk for an expected portfolio return by allocating
et al., 2017). weights to the included assets. As the number of
For instance, Quintero-Araujo et al. (2021) tackled investable assets and the number of periods to con­
the capacitated location routing problem (CLRP), sider grows, the problem becomes NP-hard.
which integrates a facility location problem with Therefore, simheuristic approaches using Monte
a multi-depot vehicle routing problem, aiming to Carlo simulation to deal with stochastic returns mod­
minimise the facility location and routing costs. The elled as random variables, coupled with genetic algo­
authors proposed a biased-randomised iterated local rithms (Nieto et al., 2022), or variable neighbourhood
search metaheuristic combined with Monte Carlo to search (Kizys et al., 2022) have been reported.
simulate stochastic demands. Also focused on location
decisions, Bayliss and Panadero (2023) reported
2.3. Focus of the study and contribution
a learnheuristic algorithm that solves the temporary
medical facility location and population assignment Considering the literature reviewed, this study
problem using a queueing simulation model. In cases focuses on:
related to healthcare quality of service metrics such as • The development of a simheuristic that integrates
people’s travel times, waiting and completion times the NSGA-II with Monte Carlo simulation which
are analysed besides costs. Indeed, the coronavirus has to deal with stochastic processing times.
pandemic came with complex problems for health NSGA-II has obtained efficient results in previous
care systems. Another application of simheuristics studies, and Monte Carlo simulation allows the
was described by Martínez-Reyes et al. (2021) to deal use of any probability distribution to model the
with the distribution of personal protective equipment stochastic parameters.
(PPE) to health centres, clinics, hospitals, and inten­ • Makespan minimisation predominates as the
sive care units (ICUs), considering uncertain main objective of some authors, while the objec­
demands, distribution cost and sustainability of the tive functions related to delivery dates have
health system. The proposed approach consisted of a smaller share. In this study, the minimisation
an iterated local search (ILS) algorithm combined of earliness makes it possible to reduce inventory
with Monte Carlo simulation (MCS). costs and avoid possible product losses due to
Freight transportation is a major problem for com­ spoilage, while the minimisation of tardiness
panies because of the costs, time and environmental guarantees a high level of service for customers.
issues. The inventory routing problem (IRP) combines Only one paper reported in this review consid­
inventory management and delivery route-planning ered a JIT-oriented objective.
decisions. Due to the complexity of this problem, • Processing times are considered stochastic in
several authors have used metaheuristic approaches order to model the influence of exogenous factors
to solve the deterministic case. However, replenish­ on task performance. The most commonly used
ment and delivery depend on stochastic demands. To distributions in the literature for modelling pro­
account for uncertainty, Gruler et al. (2020) proposed cessing times are normal, exponential and uni­
a simheuristic approach that hybridises Monte Carlo form distributions. Specifically, for this research,
simulation within a variable neighbourhood search exponential and log-normal distributions were
(VNS) algorithm to solve the multi-period IRP with considered, because they are non-negative distri­
stochastic customer demands. Usually, IRP problems butions and do not have an upper bound, allow­
aim at minimising inventory and routing costs ing comparisons of the results obtained with both
respecting transportation and warehouse capacities. distributions in the same scenario. On the one
However, time constraints can also be relevant, parti­ hand, Trietsch et al. (2012) have demonstrated
cularly for battery supported vehicles. Keenan et al. statistically that the log-normal distribution is
(2021) propose a simheuristic algorithm that inte­ suitable and plausible to model activity times.
grates Monte Carlo simulation into a strategic oscilla­ Moreover, this distribution has been used to
tion metaheuristic, to deal with the time capacitated model processing times in other scheduling
8 C. A. RODRÍGUEZ-ESPINOSA ET AL.

problems, such as byBaker and Altheimer (2012); optimal or feasible solutions in reasonable computa­
Hatami et al. (2018); Lemos and Ronconi (2015) tional times.
and Castaneda et al. (2022). On the other hand,
Soroush (2010) indicates that, exponential distri­ 3.1.1. Sets
bution is interesting, because it has memory-less I: Set of jobs.
property, leading to useful analysis of stochastic Ji : Set of operations of job i.
scheduling problems. Exponential distribution O: Set of doublets ði; jÞ that correspond to the j-th
has also been analysed in the studies by operation in the set Ji of job i.
Ahmadizar et al. (2010); Jang (2002) and M: Set of machines.
Azadeh et al. (2012). Mij : Set of machines that can process the opera­
• Some of the reviewed studies include tion ði; jÞ.
a robustness measure that quantifies the effect
of the uncertain parameters in the problem 3.1.2. Parameters
simultaneously with the main objective func­ PTijm : Processing time of operation ði; jÞ on
tion, turning the problem into a multi-
machine m. "ði; jÞ 2 O; 2 Mij .
objective one. For this case, the selected
SWi : Lower limit of the due window interval of job
robustness measure analyses, the expected
i. "i 2 I.
increase between the predictive Earliness
EWi : Upper limit of the due window interval of job
+Tardiness and the realised Earliness
i. "i 2 I.
+Tardiness (known as Earliness+Tardiness
L: A very large number.
Risk), which denotes the potential deteriora­
tion of the predictive Earliness+Tardiness and,
consequently, the potential increase of the 3.1.3. Decision variables
related costs (Wu et al., 2020; Xiao et al., Xijm : Binary variable that takes the value of 1 if the
2017). None of the aforementioned studies operation ði; jÞ is performed on machine m, 0 other­
used our proposed metrics. wise. "ði; jÞ 2 O;"m 2 Mij .
Therefore, the contribution of this paper is the study Yhkijm : Binary variable that takes the value of 1 if
of a multi-objective JIT-oriented stochastic and robust operation ði; jÞ is performed immediately after
flexible job shop scheduling problem, through operation ðh; kÞ on machine m, 0 other­
a population-based metaheuristic hybridised with wise.
Monte Carlo simulation. The majority of previous "ði; jÞ 2 O; "ðh; kÞ 2 O [ fð0; 0Þg; "m 2 Mij \ Mhk :
works on FJSS problems under uncertainties have STij : Starting time of the operation ði; jÞ. "ði; jÞ 2 O.
implemented single-solution-based metaheuristics, CTij : Completion time of the operation
and those few that used population-based algorithms ði; jÞ. "ði; jÞ 2 O [ fð0; 0Þg.
have minimised the makespan. In fact, to the best of Ti : Tardiness of job i. "i 2 I.
our knowledge, this is the first work that proposes Ei : Earliness of job i. "i 2 I.
a multi-objective simheuristic to solve a stochastic
and robust FJSS with JIT-oriented measures. 3.1.4. Constrains
Constraint set 1 limits that the operation ði; jÞ has to be
performed in only one machine:
3. Problem formulation X
Xijm ¼ 1 "ði; jÞ 2 O (1)
In this section, a mixed-Integer linear programming
m2Mij
(MILP) model for the deterministic FJSS is presented,
followed by the robust measure for the SFJSS and the Constraint set 2 blinds that the Tardiness has to be
definition of the stochastic processing times used in higher than the difference between the dispatching
this study. time and the upper limit of the due window:
Ti � CTjJi j EWi "i 2 I (2)
3.1. Flexible job shop scheduling
Constraint set 3 ensures that Earliness has to be higher
The model is partially based on the formulation than the difference between the lower limit of the due
reported by Özgüven et al. (2010). The main differ­ window and the dispatching time:
ences are as follows: the objective function, the addi­
Ei � SWi CTjJi j "i 2 I (3)
tional constraints to calculate earliness and tardiness
for each job, and some changes in the relation of Constraint set 4 indicates that the completion time of
binary variables. Another important aspect is that the operation ði; jÞ has to be equal to the starting time
this model uses auxiliary variables to reduce drastically of the operation ði; jÞ plus the processing time in the
the computational costs achieving, in some cases, selected machine:
 JOURNAL OF SIMULATION 9

X �
CTij ¼ STij þ Xijm � PTijm "ði; jÞ 2 O (4) STij � 0 "ði; jÞ 2 O (13)
m2Mij

CTij � 0 "ði; jÞ 2 O [ fð0; 0Þg

Constraint set 5 establishes that the starting time of
the operation ði; jÞ has to be higher or equal than (14)
the completion time of the previous operation of
that job: Ti � 0 "i 2 I (15)

STij � CTiðj 1Þ "ði; jÞ 2 O; j � 2 (5)

Ei � 0 "i 2 I (16)
Constraint set 6 ensures that the starting time of the
operation ði; jÞ has to be higher or equal than the
completion time of the operation ðh; kÞ if this opera­ 3.1.5. Objective function
tion precedes operation ði; jÞ on the same machine. The objective function (Equation 17) seeks to mini­
Sometimes the predecessor can be the operation ð0; 0Þ mise the sum of Earliness and Tardiness ðETÞ of all the
indicating that operation ði; jÞ is the first operation jobs.
executed in machine m: X
X minimise Earliness þ Tardiness ¼ ðEi þ Ti Þ (17)
i2I
STij � CThk L � ð1 Yhkijm Þ "ði; jÞ
m2Mij \Mhk
2 O; "ðh; kÞ 2 O [ fð0; 0Þg (6)
Constraint set 7 limits that only one operation can be 3.2. Stochastic processing times
executed first in each machine: Considering the stochastic environment, two prob­
X
Y00ijm ¼ 1 "m 2 M (7) ability distributions were selected to analyse the effect
ði;jÞ2Ojm2Mij of the stochasticity in the FJSS: log-normal and expo­
nential, which consider only positive values in their
Constraint set 8 guarantees that if the operation domain.
ði; jÞ is performed in machine m then this opera­ �PT is defined as the stochastic processing time,
tion has one predecessor operation in the same and PT as the deterministic processing time which is
machine: used as the expected value for the random variable.
X • Exponential:� Considering the exponential para­
Xijm ¼ Yhkijm "ði; jÞ 2 O; "m �
1
ðh;kÞ2O[fð0;0Þgjm2Mhk
meter λ ¼ PTijm , it is possible to model the
2 Mij processing time of the job i in the j-th operation
on machine m as shown in Equation 18:
(8) � �
1
Constraint set 9 guarantees that if the operation ði; jÞ is �PTijm ,Exp (18)
PTijm
performed in machine m then this operation has at
most one successor operation in the same machine: • Log-normal: If a random variable X is distributed as
X a log-normal distribution, then the LnX is normally
Xijm � Yijhkm "ði; jÞ 2 O; "m
ðh;kÞ2Ojm2Mhk
distributed with mean μ and standard deviation σ.
2 Mij This distribution is used as a model of the influence
on environments with uncontrolled variables
(9) (Liefooghe et al., 2012).
Constraint set 10 does not allow to put the operation Considering that the log-normal parameters μ and
ði; jÞ after the same or other later processes: σ are related to the natural logarithm of the variable, it
X is possible to calculate these parameters by knowing
Yikijm ¼ 0 "ði; jÞ 2 O; "m the expected value and the coefficient of variation
k2Ji jk�j;m2Mik
(CV) of the variable. See Equations 19 and 20:
2 Mij
(10)
� LnðCV 2 þ 1Þ
μijm ¼ Ln PTijm (19)
2
Constraint sets 11 to 16 describe the variables’ nature:
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Xijm 2 f0; 1g "ði; jÞ 2 O; "m 2 Mij LnðCV 2 þ 1Þ
σ ijm ¼ (20)
(11) 2
Thus, it is possible to model the processing time of
Yhkijm 2 f0; 1g "ði; jÞ 2 O; "ðh; kÞ the job i for the j-th operation on machine m as shown
2 O [ fð0; 0Þg"; m 2 Mij \ 2 Mhk (12) in Equation 21:
10 C. A. RODRÍGUEZ-ESPINOSA ET AL.

schedules in uncertain environments to improve deci­

sion-making.

3.3. Robustness measure

As a result of the stochastic behaviour of the processing
times, the completion times of the jobs change, gener­
ating a stochastic behaviour in the objective function.
Figure 2 shows the distribution of the realised Earliness
+Tardiness obtained with a particular scheduling con­
sidering exponential processing times. The blue area
represents the cases in which the value of the objective
function obtains a worse performance than the one
calculated with the expected value of the processing
times, increasing the related costs. Meanwhile, the
white area represents an improvement in this value.
Considering this behaviour of the objective function,
under random processing times, the measure used by
Wu et al. (2020) is adapted for this study.
This Earliness+Tardiness Risk is the average dete­
rioration of the Earliness+Tardiness of the realised
schedule (ET R ) and considers stochastic processing
times, in comparison to the Earliness+Tardiness of the
predictive schedule (ET P ) which uses deterministic pro­
cessing times. Mathematically, the Earliness+Tardiness
Risk (RET ) can be defined as in Equation 22:
Figure 1. Effect of uncertainty on schedules. � ��
RET ¼ E max ET R ET P ; 0 (22)
� � Considering the main objective of minimising
�PTijm ,Log normal μijm ; σ ijm (21)
Earliness+Tardiness, the measure considers as dete­
rioration the cases in which the ET R is highest than the
Figure 1 shows an example of a schedule with three ET P , while in the other cases the deterioration is 0
jobs and four machines, considering the same due considering that this cases benefits the main objective.
window intervals for all jobs. The predictive schedule As a result, the idea of this robustness measure is to
A has an Earliness+Tardiness of 0 time units, making minimise the average deterioration of the schedules
this sequence an optimal one, from the deterministic under random processing times.
point of view. Nevertheless, when the processing times
are considered to be stochastic, the schedule perfor­
mance gets worse obtaining an Earliness+Tardiness of 4. The proposed simheuristic: NSGA-II with
3 time units and 2 time units for schedules B and C, Monte Carlo simulation
respectively. In this case, it can be seen that it is very This section describes the proposed simheuristic.
important to take into account the performance of Bearing in mind the main elements of the NSGA-II
(Deb et al., 2002), the principles of simheuristics (Juan
et al., 2015), and the characteristics of the problem; the
following aspects are hereafter explained: chromo­
some definition, initial population, order and selec­
tion, crossover operator, mutation operator and
Monte Carlo simulation.
Table 3 shows an example of a FJSS problem with
a due window interval that is used for explaining the
proposed NSGA-II in this section.

4.1. Chromosome definition

Based on L. Sun et al. (2019), the coding used for
Figure 2. Distribution of realised Earliness+Tardiness. this problem is a real number from [0, Mij ) for
 JOURNAL OF SIMULATION 11

Table 3. Test problem.

Machines Due window interval
Job Operation M1 M2 M3 Lower limit Upper limit
J1 O1;1 2 4 3 10 14
O1;2 3 - -
O1;3 3 2 3
J2 O2;1 3 - 4 12 15
O2;2 - 4 -
O2;3 3 2 3
J3 O3;1 2 - 1 8 11
O3;2 4 3 4

operations that can be processed (O1;2 : 0:19,

O2;1 : 0:28 and O3;2 : 0:44).
• O2;1 is processed in its first machine, considering
that its fractional part is the greatest among all the
operations that can be processed (O1;2 : 0:19
Figure 3. Representation of the chromosome. and O2;1 : 0:28).
• O2;2 is processed in its first machine, considering
each job operation and each job, Mij being the that its fractional part is the greatest among all the
number of total machines that can process the j- operations that can be processed (O1;2 : 0:19
th operation of job i (Oij ). The integer part of the and O2;2 : 0:22).
number represents the machine selected to process • O2;3 is processed in its second machine, con­
the job. For example, if machines 3, 5, and 7 can sidering that its fractional part is the greatest
process job 2 and the integer part of the gene is 1 among all the operations that can be processed
(considering 0 as the first position), then, machine (O1;2 : 0:19 and O2;3 : 0:43).
5 is selected, because it is the second machine that • O1;2 is processed in its first machine, considering
can process job 2. The fractional part of the num­ that its fractional part is the greatest among all the
ber represents the operations’ priority, which is operations that can be processed
analysed at each step for those operations that (only O1;2 : 0:19).
can be performed. • The last remaining operation O1;3 is processed in
Figure 3 shows an example of the representation its second machine.
of the chromosome. The rows correspond to jobs Figure 4 shows the Gantt chart of the example
and the columns to the number of the operation. which obtains a Earliness+Tardiness of 4 time units.
For instance, the cell located in row J3 column O1 Finally, L. Sun et al. (2019) considers two impor­
corresponds to the operation O3;1 . Therefore, the tant points in the design of the representation of
allocation sequence represented in this example is a solution. The first one is that each representation
as follows: has to be as short as possible, thus reducing the
• O1;1 is processed in its third machine, considering different variables to evaluate. As a second point,
that its fractional part is the greatest among all the each representation has to be a feasible solution to
operations that can be processed (O1;1 : 0:88, guarantee the efficiency of the algorithm, reducing
O2;1 : 0:28 and O3;1 : 0:80). the computational execution time by omitting the
• O3;1 is processed in its first machine, considering unfeasible operations. In this case, with the pro­
that its fractional part is the greatest among all the posed coding, if the integer part is changed the
operations that can be processed (O1;2 : 0:19, machine on which the operation is processed also
O2;1 : 0:28 and O3;1 : 0:80). changes, taking into account that the range only
• O3;2 is processed in its third machine, considering considers the machines that can process the j-th
that its fractional part is the greatest among all the operation of job i. Moreover, if the fractional part

Figure 4. Gantt chart of the chromosome solution of Figure 3 using the instance in Table 3.
12 C. A. RODRÍGUEZ-ESPINOSA ET AL.

is modified, the priority against the other opera­ • AVPRO(average processing time per operation):
tions also varies. Select job i if it has the minimum value of APi for
the current operation j, as in Equation 24:
PJi
4.2. Initial population h¼j PTih�
APi ¼ (24)
jJi j j þ 1
For the initial population of the proposed NSGA-II,
some solutions were generated with the application of where PTih� is the processing times average of all the
three static dispatching rules and three dynamic dis­ possible machines that can perform the h-th task.
patching rules considering the expected values of the • S/RO(slack per remaining operations): Select job i if
processing times. The remaining solutions to reach the it has the minimum value of SRi for the current opera­
number of individuals were randomly generated. tion j, as in Equation 25:
At each iteration, all jobs consider processing their �
max SWi CTiðj 1Þ ; 0
current operation in the machine with the lowest SRi ¼ (25)
availability time. If there are two or more machines jJi j j þ 1
with this characteristic, the machine with the lowest Finally, if there remains a tie between jobs, a job is
processing time is selected. If there is a tie between two selected randomly to schedule it in the selected
or more machines, one of them is chosen randomly. machine and update the current operation of the job
Once all jobs have a candidate machine to process and the availability times. This process has to be done
them, the selection criterion to assign a job is, on the until all the operations have been assigned.
one hand, the job with the lowest availability time
(considering the availability time of the machine
when it finishes the previous operation). On the 4.3. Order and selection
other hand, if there is a tie between two or more
jobs, the following dispatching rules are used to After the initial population is generated, there is
break the tie: a group of N chromosomes that are chosen to be
• EDD (earliest due date): Select job i if it has the recombined. This group will be classified into Pareto
minimum value of EWi frontiers, according to the fast non-dominated sort­
• SPT (shortest processing time): Select job i if it has ing procedure. In this procedure, each solution is
the minimum value of PTijm for the current opera­ compared with every other solution in the popula­
tion j in the selected machine m. tion to find those that are non-dominated. Once
• LPT (longest processing time): Select job i if it has a set of non-dominated solutions is found, these
the maximum value of PTijm for the current solutions are assigned to a first front (F1 ).
Subsequently, the remaining solutions that were not
operation j in the selected machine m.
classified in the F1 front are compared with each
• CR (critical ratio): Select job i if it has the mini­
other once again to select those that are non-
mum value of CRi for the current operation j, as
dominated, to classify them into the F2 front. This
in Equation 23:
! procedure is repeated successively until all the Fn
SWi STij fronts are found, i.e., all the chromosomes are clas­
CRi ¼ max PJi ;0 (23) sified entirely. Later, in each Pareto frontier, the
h¼j PTij�
chromosomes are ordered according to their crowd­
where PTij� is the processing times average of all the ing distance, calculated according to Algorithm 1 for
possible machines that can perform the j-th task. each set I.

Algorithm 1
Crowding distance assignmentðIÞ :
s ¼ jIj ▷ Number of solutions in the Pareto Frontier.
▷ Initialise distance.
for each I, set Distance½i� ¼ 0
▷ Sort I according each objective function.
for each objective function m do ▷ Extreme points are always selected.
I ¼ sortðI; mÞ ▷ For the other points.
Distance½1� ¼ Distance½s� ¼ 1

for i = 2 to ðs 1Þdo
Distance½i� ¼ Distance½i� þ I½i þ 1�:m I½i 1�:m

end for
end for
 JOURNAL OF SIMULATION 13

4.4. Crossover operator

A random number between 0 and 1 is generated
for each position of the chromosome, which
defines how the crossover will be performed to
create an offspring chromosome. When the ran­
dom number is less than the crossover probabil­
ity, the offspring takes the value of the first
parent’s gene in this position. Otherwise, it
takes the value of the second parent’s gene in
this position.
In the example of Figure 5, the yellow forms
represent the parent chromosomes, and the red
form represents the offspring chromosome.
Figure 6 shows the Gantt chart of the example,
which obtains a Earliness+Tardiness of 1 time
Figure 5. Crossover operation. unit.

Furthermore, the crowding distance assigned to the

solutions at the extremes of the frontier is 4.5. Mutation operator
a considerable number (1). Higher crowding dis­
tances are given more weight because the farther Initially, a probability of mutation is defined. For each
apart the chromosomes are, the more diversification resulting chromosome, a random number between 0
occurs during crossover. and 1 is generated to determine whether the mutation
Once all the chromosomes are ordered, a pair of operator should be applied to that chromosome or
chromosomes is selected randomly to apply the cross­ not. If the generated number is less than the defined
over operator until Q offspring is obtained. Next, probability, the mutation is performed. The mutation
a population with size N+Q is obtained. Later, the consists of choosing one job of the chromosome and
chromosomes of the population N+Q are organised reversing the order of the fractional values of the
again by the fast-non-dominated sorting and crowd­ operations. In the example of Figure 7, the obtained
ing distance procedures, and the first N chromosomes offspring inverts the order of the fractional values of
are saved. It is important to emphasise that the pre­ Job 3, from 0.33 and 2.44 to 0.44 and 2.33, respectively.
vious process is repeated until the stopping criterion Figure 8 shows the Gantt chart of the example
is met. which obtains a Earliness+Tardiness of 3 time units.

Figure 6. Gantt chart of the chromosome solution of Figure 5 using the instance in Table 3.

Figure 7. Perform of the mutation.

14 C. A. RODRÍGUEZ-ESPINOSA ET AL.

Figure 8. Gantt chart of the chromosome solution of Figure 7 using the instance in Table 3.

4.6. Monte Carlo simulation study considering the requirement of having the deter­
ministic and stochastic objective functions of each solu­
According to Juan et al. (2015), a simheuristic is an
tion to perform NSGA-II to avoid the risk of losing any
algorithm that integrates any simulation variant into
solution that may be suitable for the Pareto frontier.
a metaheuristic framework to solve complex stochastic
Thus, following the proposed methodology, during
optimisation problems. Thus, each solution generated
the iterative search process the number of replications
through NSGA-II is simulated through a Monte Carlo
of the simulations should be small (fast simulation) to
simulation to evaluate the stochastic objective func­
tion, using the probability distributions defined for the obtain approximate estimates reducing the computa­
stochastic parameters. tional time. Besides, at each generation, only the new
The author indicates that only promising determi­ chromosomes in the population are simulated with
nistic solutions should be simulated in the simheuristic fast simulation because those that are not new (i.e.,
process, in order to reduce computational times. that came from previous generations) were simulated
However, this statement has been modified for this when they appeared for the first time in the

Figure 9. Flow chart diagram of the simheuristic algorithm.

 JOURNAL OF SIMULATION 15

population. Once the NSGA-II procedure meets the chromosomes, number of offspring, number of gen­
stopping criterion, all the chromosomes obtained in erations, crossover probability and mutation probabil­
the last generation have to be simulated again, but, in ity) were selected through an experimental design, and
this case, the number of simulation replicates has to be the rest were selected based on experience and some
extensive (intensive simulation) to obtain high-quality references from the reviewed literature. Table 4 shows
estimates and to reorder all the solutions for the final the values selected for the parameters.
Pareto frontier.
5.2. Benchmark instances
4.7. Simheuristic flow chart Because of the absence of standard benchmark instances
for different variations of the FJSS under stochastic fac­
The proposed simheuristic follows the procedure
tors, some deterministic FJSS instances were selected and
shown in Figure 9, which is based on the methodology
adapted considering the deterministic processing times
proposed by Juan et al. (2015):
as the expected values of the random processing times.
Two sets of benchmarks were evaluated to analyse the
5. Computational experiments performance of the proposed simheuristic:
• Brandimarte (1993), p. 10 instances (Mk01 to
This section presents the computational experiments Mk10) with sizes ranging from 10 jobs and 6 machines
carried out to evaluate the performance of the pro­ to 20 jobs and 15 machines.
posed simheuristic. Since there are not similar works • Hurink et al. (1994), p. 56 instances (abz5 to abz9,
with the same objective functions for comparison, this car1 to car8, la01 to la40, mt06, mt10 and mt20) with
section reports various computational experiments. sizes ranging from 6 jobs and 6 machines to 30 jobs
First, the performance of the simheuristic is evaluated and 10 machines.
by comparing the simulations given by three different Considering that these instances do not have due
deterministic solution approaches to the problem: the windows, their lower limits were generated using the
MILP model, dispatching rules and single-objective method described by Bagheri and Zandieh (2011) as
genetic algorithm. Secondly, the visualisation of shown in Equation 26.
some of the obtained Pareto frontiers is shown. � � � � ��
R R
Thirdly, the behaviour of the objective functions is SWi ,U μi � 1 ; μi � 1 þ (26)
analysed under different coefficients of variation and 2 2
probability distributions. Finally, an experiment is where μi is calculated as shown in Equation 27,
carried out to show the effects of different probability � � X Ji
distributions and coefficients of variation in multi- T�I
μi ¼ 1 þ � PTij� (27)
objective measures of Pareto frontiers. M j¼1
None of the experimental designs performed in this
section fulfilled the homoscedasticity and normality Considering I as the number of jobs, M the number
assumptions of ANOVA. Therefore, the non- of machines, and T and R as parameters related with
parametric test MANOVA (Anderson, 2001) was exe­ the complexity of the due dates, the same values used
cuted in each case, and all conclusions are presented by Bagheri and Zandieh (2011) were used. PTij� is the
with a significance level of 0.05. processing times average of all the possible machines
that can perform the j-th task.
The upper limit of due windows was calculated
5.1. Parameters settings according to Equation 28:
The simheuristic algorithm is coded and executed by EWi ,U ½SWi þ 0:1 � β; SWi þ β� (28)
using NetBeans IDE 8.2 on an Intel(R) Core™ i7
[email protected] GHz with 7.84 GB RAM. The parameter where β is calculated as half of the minimum mean
values chosen for the NSGA-II (number of total processing time per operation, as shown in
Equation 29.
�P �
Ji
Table 4. Simheuristic parametrisation. min j¼1 PT ij�
Parameter Value β¼ (29)
2
Number of chromosomes 1000
Number of offspring 500
Number of generations 750
Crossover probability 30% 5.3. Mixed integer linear programming
Mutation probability 2.5%
Replicates (Fast simulation) 50
Replicates (Intensive simulation) 1000
The proposed MILP model was executed in Jupyter
Maximum time of execution (minutes) 60 notebook with Python 3 and using the PuLP library,
16 C. A. RODRÍGUEZ-ESPINOSA ET AL.

Table 5. Results of MILP model.

Earliness+Tardiness Earliness+Tardiness Risk
Instance State MILP SH Improv PD MILP SH Improv
car7 F 11446 0 1.00% LogN25 1703.89 22.77 98.66%
car7 F 11446 0 1.00% LogN50 5376.54 117.13 97.82%
car7 F 11446 0 1.00% LogN75 10527.93 794.44 92.45%
car7 F 11446 0 1.00% LogN100 16627.19 2448.48 85.27%
car7 F 11446 0 1.00% Exp 12221.26 1817.58 85.13%
la04 F 2973 243 91.83% LogN25 217.35 46.00 78.84%
la04 F 2973 250 91.59% LogN50 664.01 178.18 73.17%
la04 F 2973 237 92.03% LogN75 1289.82 401.03 68.91%
la04 F 2973 223 92.50% LogN100 2051.75 590.26 71.23%
la04 F 2973 243 91.83% Exp 1372.13 548.65 6.01%
Mk01 F 702 0 1.00% LogN25 25.45 0.09 99.66%
Mk01 F 702 1 99.86% LogN50 72.43 2.64 96.35%
Mk01 F 702 0 1.00% LogN75 143.58 13.82 9.37%
Mk01 F 702 0 1.00% LogN100 226.21 26.82 88.14%
Mk01 F 702 1 99.86% Exp 107.94 24.27 77.52%
mt06 Opt 0 0 .00% LogN25 5.98 0.24 95.91%
mt06 Opt 0 0 .00% LogN50 21.69 2.24 89.66%
mt06 Opt 0 0 .00% LogN75 5.24 10.59 78.92%
mt06 Opt 0 0 .00% LogN100 82.97 17.79 78.56%
mt06 Opt 0 0 .00% Exp 79.11 16.75 78.82%

in the computer mentioned in subsection 5.1. All the Table 6. Non-parametric MANOVA for improvement in each
benchmark instances were tested with a time limit of 8 objective function.
Factor ET RET
h. Considering this time constraint, three states can be
Instance 0.0000 0.0000
obtained for each instance”: non-feasible”, which shows Combination 0.0000 0.0000
that the model was not able to obtain a solution satisfy­ Dispatching rule 0.0000 0.0000
Instance*Combination 0.0000 0.0000
ing all constraints in the time frame mentioned above”; Instance*Dispatching rule 0.0000 0.0000
feasible”, which indicates that a solution that satisfies all Combination*Dispatching rule 0.9321 0.0000
constraints was obtained, but it is not guaranteed that Instance*Combination*Dispatching rule 1.0000 0.0000
R2aj 95.69% 97.88%
this best possible value for the objective function; and
”optimal”, which states that a solution with the best
possible value for the objective function was found. optimal or some feasible solutions requires unrea­
The instances that obtained a feasible (F) or sonable computational times in most instances. In
optimal (Opt) solution with the MILP model this case, the use of dispatching rules ensures fea­
were simulated using the log-normal distribution sible solutions that can be obtained in less time. An
(LogN), with coefficients of variation of 25%, 50%, experimental design was carried out to measure the
75%, and 100%, and the exponential distribution performance of the simheuristic in comparison
(Exp), to obtain the Earliness+Tardiness Risk in with simulated solutions of six dispatching rules
each case. The results were compared with the considering the following factors: the combination
solutions obtained with the proposed simheuristic of the probability distribution with the coefficient
(SH) for the same distributions and coefficients of of variation (Log-normal25, Log-normal50, Log-
variation. Table 5 shows the results for all normal75, Log-normal100 and Exponential), dis­
instances that obtained a feasible or optimal solu­ patching rules (EDD, SPT, LPT, CR, AVPRO and
tion with the MILP model for each probability S/RO) and instances as a blocking factor. The
distribution (PD). response variable was the improvement in the
The simheuristic achieves the same Earliness value of each objective function m, calculated as
+Tardiness as the simulated solution given by the the percentage difference between the best value
MILP model for the instance that was optimal, and obtained for that objective function with the sim­
improved by 97.30% the Earliness+Tardiness of the heuristic and the value obtained for the same
instances that obtained a feasible solution. Finally, the objective function with the dispatching rule, as
simheuristic improves by an average of 84.27% the shown in Equation 30.
Earliness+Tardiness Risk in comparison with the results
obtained with the simulation of the solutions given by the Dispatch Rulem Bestm
MILP. Improvementm ¼
Dispatch Rulem
2 Objective Functions
5.4. Dispatching rules (30)

From the previous results, it is concluded that try­ Table 6 shows the p-values and the coefficients of
ing to use the proposed MILP model to obtain the determination obtained for the improvements in each
 JOURNAL OF SIMULATION 17

IMPROVEMENT:EARLINESS+TARDINESS
EDD SPT LPT CR AVPRO S/RO

IMPROVEMENT: EARLINESS + TARDINESS

90.0%
85.0%
80.0%
75.0%
70.0%
65.0%
60.0%
55.0%
50.0%
LOG-NORMAL25 LOG-NORMAL50 LOG-NORMAL75 LOG-NORMAL100 EXPONENTIAL
COMBINATION: PROBABILITY DISTRIBUTION + COEFFICIENT OF VARIATION

Figure 10. Profile chart of Earliness+Tardiness improvement.

IMPROVEMENT: EARLINESS+TARDINESS RISK

EDD SPT LPT CR AVPRO S/RO
IMPROVEMENT: EARLINESS + TARDINESS RISK

70.0%

60.0%

50.0%

40.0%

30.0%

20.0%

10.0%
LOG-NORMAL25 LOG-NORMAL50 LOG-NORMAL75 LOG-NORMAL100 EXPONENTIAL
COMBINATION: PROBABILITY DISTRIBUTION + COEFFICIENT OF VARIATION

Figure 11. Profile chart of Earliness+Tardiness Risk improvement.

COMPARISON: EARLINESS+TARDINESS RISK

Minimum improvement Maximum improvement

100.0%
IMPROVEMENT: EARLINESS+TARDINESS RISK

90.0%
80.0%
70.0%
60.0%
50.0%
40.0%
30.0%
20.0%
10.0%
0.0%
LOG-NORMAL25 LOG-NORMAL50 LOG-NORMAL75 LOG-NORMAL100 EXPONENTIAL
COMBINATION: PROBABILITY DISTRIBUTION + COEFFICIENT OF VARIATION

Figure 12. Profile chart of Earliness+Tardiness Risk improvement against genetic algorithm.
18 C. A. RODRÍGUEZ-ESPINOSA ET AL.

objective function with the non-parametric of variation whereas for Earliness+Tardiness Risk
MANOVA. the best improvements are obtained for higher
According to MANOVA, the first factor (which is coefficients of variation. In the case of dispatching
the combination of the coefficient of variation and the rules, the least improvement of Earliness+Tardiness
probability distribution) and the second factor (which is about 50% against the SR/O dispatching rule and
is the dispatching rule) have both a significant effect in for Earliness+Tardiness Risk is a minimum of 15%
the improvement measure of both objective functions. against the EDD dispatching rule.
However, the interaction of those two factors has
a significant effect only in the improvement measure
5.5. Deterministic approach
of the stochastic objective function.
Average improvements can be observed in Another approach that is more efficient in terms of
Figures 10 and 11 for Earliness+Tardiness and computational time is to solve the deterministic version
Earliness+Tardiness Risk, respectively. It can be of the problem using the single-objective genetic algo­
seen that for the Earliness+Tardiness objective the rithm and then simulating the best solution obtained to
best improvement is reached for lower coefficients calculate the robustness measure. Using the same

Figure 13. Pareto frontiers obtained.

 JOURNAL OF SIMULATION 19

parameters as above, a genetic algorithm is executed, EAR L I N ES S+ TAR D I N ESS R I S K

and, at the end, the best solution is simulated with

EARLINESS + TARDINESS RISK

2000
a total of 1000 replicates considering a combination of
the probability distribution with the coefficient of var­ 1500
iation (Log-normal25, Log-normal50, Log-normal75,
Log-normal100 and Exponential). Then, these results 1000
are compared with the best Earliness+Tardiness Risk
500
obtained by the simheuristic approach, calculating the
improvement as in Equation 31. 0
25 50 75 100
GeneticAlgorithmRET BestRET
ImprovementRET ¼ COEFFICIENT OF VARIATION (%)
GeneticAlgorithmRET
(31) Figure 14. Profile chart of Earliness+Tardiness Risk.

In the main effects plot of Earliness+Tardiness Risk

improvement (Figure 12), it can be observed that the
are presented. These instances consist of 6, 10, 15, 20
solution with the best and the worst Earliness+Tardiness
and 30 jobs, respectively. In these examples, it can be
Risk given by the simheuristic approach improves the
seen that there is a contradictory behaviour of the two
Earliness+Tardiness Risk provided by the metaheuristic
objectives evaluated, and that the minimum value
approach in a range from 69.62% to 88.73% and from
obtained for the Earliness+Tardiness Risk increases
16.03% to 28.65%, respectively, according to the combi­
as the coefficient of variation increases.
nation of the probability distribution and coefficient of
This behaviour can be observed in the Pareto fron­
variation. This shows that the GA is not able to deal with
tiers of the other 61 instances. Nevertheless, it is
the impact of variability of processing times in the
important to perform a statistical analysis of the
Earliness+Tardiness, whereas, at the moment of includ­
results and the behaviour of the Pareto frontier. The
ing an objective function that evaluates the robustness,
experiments carried out for the results obtained are
the simheuristic obtains solutions that minimise the
explained in the following subsections.
impact of uncertainties in the performance of Earliness
+Tardiness.
5.7. Objective functions
It is also essential to analyse the behaviour of objective
5.6. Pareto frontiers functions under different probability distributions and
A total of 1650 Pareto frontiers were obtained. They coefficients of variation. Considering that the log-
consist of five frontiers for each instance, one each for normal distribution was tested with coefficients of
the combination of the probability distribution with variation of 25%, 50%, 75%, and 100% and that the
the coefficients of variation (Log-normal25, Log- exponential distribution has a single coefficient of
normal50, Log-normal75, Log-normal100 and variation of 100%, two analyses were performed: i)
Exponential) of the processing times. comparison of different coefficients of variation
As an example, the five Pareto frontiers for the under a log-normal distribution and ii) comparison
instances mt06 (Figure 13a), car5 (Figure 13b), Mk03 of a log-normal distribution with a coefficient of var­
(Figure 13c), la30 (Figure 13d) and la34 (Figure 13e) iation of 100% and exponential distribution.
Table 7 shows the p-values and coefficients of
determination obtained for each objective function,
Table 7. Non-parametric MANOVA for the objective functions
with different coefficients of variation. considering different coefficients of variation (CV)
Factor ET RET under a log-normal distribution with the non-
Instance 0.0000 0.0000 parametric MANOVA.
CV 0.7514 0.0000 According to the results obtained, the coefficient of
Instance*CV 0.0000 0.0000
R2aj 99.49% 98.68% variation has a significant effect on the stochastic
objective function. According to the main effects
plots and least significant difference test (LSD), with
a significance of 95%, it can be observed that the
Table 8. Non-parametric MANOVA for the objective functions Earliness+Tardines Risk increases as the coefficient
with different probability distributions.
of variation augments (Figure 14).
Factor ET RET
Instance 0.0000 0.0000
Table 8 shows the p-values and the coefficients of
Distribution 0.4466 0.0000 determination obtained with the MANOVA for both
Instance*Distribution 0.7953 0.0000 objective functions under different probability distri­
R2aj 99.56% 98.32%
butions (with CV = 100%).
20 C. A. RODRÍGUEZ-ESPINOSA ET AL.

According to the results obtained, the probability performance criteria proposed by Ahmadi et al. (2016)
distribution (under a coefficient of variation of 100%) were evaluated:
has a significant effect on the value of Earliness • Number of Pareto solutions (NP): the number of
+Tardiness Risk. Specifically, the exponential distribu­ solutions obtained on the first Pareto frontier of
tion presents a better average of Earliness+Tardiness the last generation of the algorithm. High values
Risk (1766.96) than the log-normal distribution of the obtained solutions imply more varied
(1974.254), even though the coefficient of variation is options for the decision-making process.
the same. • Spacing: the standard deviation of the nearest abso­
These tests allow us to highlight the relevance of lute distance from one solution to another (see
performing stochastic analysis (using of simheuristics) Equation 32). This determines the consistency
to deal with uncertainties and, on the other hand, the between the Pareto solutions, which means that
importance of a precise adjustment of the probability the lower this variability between the distances of
distribution used in the method in order to obtain the Pareto frontier solutions is, the higher the quality
good and accurate results. of the algorithm results.
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u
u 1 X NP �2
5.8. Performance measures Spacing ¼ t di d (32)
NP 1 i¼1
Finally, it is important to evaluate the performance of the
NSGA-II algorithm in terms of the quality of the where di is the minimum absolute distance
obtained Pareto frontiers. For this purpose, the between the solution i and another solution, and d

Table 9. Non-parametric MANOVA for performance measures with different coefficients of

variation.
Factor NP Spacing Diversity MMID Run time
Instance .0000 .0000 0.0000 0.0000 0.0000
CV .0000 .0000 0.0000 0.0000 0.3725
Instance*C.V. .0000 .0000 0.0000 0.0000 1.0000
R2aj 75.61% 74.66% 92.16% 51.61% 98.35%

PARETO SOLUTIONS SPACING

59 20
PARETO SOLUTIONS

57
55 15
SPACING

53
10
51
49
5
47
45 0
25 50 75 100 25 50 75 100
COEFFICIENT OF VARIATION (%) COEFFICIENT OF VARIATION (%)

DIVERSITY MMID
10000 0.50
8000 0.48
DIVERSITY

6000 0.46
MMID

4000 0.44
2000 0.42
0 0.40
25 50 75 100 25 50 75 100
COEFFICIENT OF VARIATION (%) COEFFICIENT OF VARIATION (%)

Figure 15. Profile charts.

 JOURNAL OF SIMULATION 21

Table 10. Non-parametric MANOVA for performance measures with different distributions.
Factor NP Spacing Diversity MMID Run time
Instance .0000 .0000 0.0000 0.0000 0.0000
Distribution .0000 .0000 0.0076 0.0008 0.0000
Instance*Distribution .0841 .0520 0.0000 0.0659 0.0000
R2aj 74.27% 7.47% 91.03% 46.57% 99.32%

Table 11. Performance measures by probability distribution.

In the main effects plots of the number of Pareto solu­
Distribution
tions (Figure 15a), spacing (Figure 15b), diversity
Measure Log-normal Exponential
(Figure 15c) and MMID (Figure 15d), and employing
Pareto solutions 46.812 56.458
Spacing 18.810 13.322 the LSD test with a significance of 95%, it can be observed
Diversity 7882.481 7355.057 that the spacing, diversity and MMID increase as the
MMID 0.485 .473
Run time (min) 10.555 7.790 coefficient of variation increases. For the Pareto solutions
measure, the best value is reached with a coefficient of
variation of 50%, followed by a coefficient of variation of
is the average of all the minimum distances of all the 75% and, finally, the lowest value is obtained with coeffi­
solutions. cients of variation of 25% and 100%.
• Diversity: the Euclidean distance between the Table 10 shows the p-values and the coefficients of
most extreme values of the obtained Pareto fron­ determination obtained for each performance mea­
tier (see Equation 33). sure with different probability distributions (log-
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi normal and exponential) under the same coefficient
Diversity ¼ ðMaxET MinET Þ2 of variation (100%) with the non-parametric
þðMaxRET MinRET Þ2 (33) MANOVA.
According to the results obtained, the type of prob­
A higher value of the diversity of the frontier repre­ ability distribution has a significant effect on all the
sents a higher quality in the set of solutions found. performance measures. In Table 11, it is observed that
• Modified mean ideal distance (MMID): the the exponential distribution performs better in all mea­
average distance of the normalised points, con­ sures except diversity compared to the log-normal
sidering the difference of scales in the objective distribution.
functions, to an ideal point ð0; 0Þ. The MMID is
calculated as shown in Equation 34:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 6. Conclusions and future work
1 X
NP
MMID ¼ NETi2 þ NRET 2i (34)
NP i¼1 This research project addresses a bi-objective flexible
job shop problem that seeks to minimise the earliness
where NET and NRET are calculated, as presented
and tardiness in the deterministic environment and to
in Equations 35 and 36, taking into account the mini­
minimise the risk associated with the randomness of
mum and maximum values of each objective function:
processing times in a stochastic environment to obtain
ETi MinET robust solutions. For this purpose, a simheuristic was
NETi ¼ (35)
MaxET MinET developed that hybridises the multi-objective genetic
algorithm NSGA-II and Monte Carlo simulation to
RET i MinRET obtain a Pareto frontier of the two evaluated objective
NRET i ¼ (36)
MaxRET MinRET functions. The performance of the proposed algorithm
This criterion evaluates the efficiency and conver­ was evaluated in two stages. In the first stage, the results
gence power of the algorithm and considering that obtained with the proposed simheuristic were com­
both objectives are to be minimised, lower values pared with the results obtained by other approaches
indicate a higher quality algorithm performance. that provide a particular solution for the problem.
• Run time: the execution time of the algorithm. Those approaches were as follows: simulation of the
Obtaining a good response in a short time repre­ solutions obtained with the MILP model, simulation
sents the high quality of the algorithm to cope of the solutions achieved with the EDD, SPT, LPT,
with emergency situations. CR, AVPRO and S/RO dispatching rules and simula­
Table 9 shows the p-values and the coefficients tion of the solutions obtained with the deterministic
of determination obtained for each performance single-objective version of the genetic algorithm.
measure with different coefficients of variation The second stage presents the analysis of the influence
obtained with the non-parametric MANOVA. of the coefficients of variation and probability distribu­
As shown by the results, the coefficient of variation tions on the evaluated objective functions and the
has a significant effect on all the performance measures. resulting Pareto frontier performance measures.
22 C. A. RODRÍGUEZ-ESPINOSA ET AL.

The comparison of the simheuristic against the ORCID

simulation of the solutions obtained with the MILP
Eliana María González-Neira http://orcid.org/0000-
model shows that the simheuristic performs better 0002-4590-3401
than the simulated MILP model in terms of computa­
tional cost, obtaining better results, though not neces­
sarily the optimum, in less execution time. Against the References
simulation of the solutions achieved with the men­
tioned dispatching rules, the simheuristic improves Ahmadi, E., Zandieh, M., Farrokh, M., & Emami, S. M.
on average by at least 55% the Earliness+Tardiness (2016). A multi objective optimization approach for flex­
ible job shop scheduling problem under random machine
and on average by at least 15% the Earliness breakdown by evolutionary algorithms. Computers and
+Tardiness Risk compared to the evaluated dispatch­ Operations Research, 73, 56–66. https://doi.org/10.1016/j.
ing rules. And against the simulation of the solutions cor.2016.03.009
obtained with the deterministic single-objective ver­ Ahmadizar, F., Ghazanfari, M., & Ghomi, S. M. T. F. (2010).
sion of the genetic algorithm, the simheuristic Group shops scheduling with makespan criterion subject
improves by at least 69% the Earliness+Tardiness to random release dates and processing times. Computers
& Operations Research, 37(1), 152–162. https://doi.org/
Risk. Lastly, in the second stage, an influence of the 10.1016/j.cor.2009.04.002
coefficients of variation on the value of both objective Al-Hinai, N., & ElMekkawy, T. (2011). Robust and stable
functions and most of the performance measures ana­ flexible job shop scheduling with random machine break­
lysed is observed when the probability distribution downs using a hybrid genetic algorithm. International
does not vary. Meanwhile, there is a significant differ­ Journal of Production Economics, 132(2), 279–291.
ence between the exponential and log-normal distri­ https://doi.org/10.1016/j.ijpe.2011.04.020
Anderson, M. J. (2001). A new method for non-parametric
butions under a coefficient of variation of 100% that is multivariate analysis of variance. Austral Ecology, 26(1),
reflected in the stochastic objective function and all 32–46. https://doi.org/10.1111/j.1442-9993.2001.01070.
the performance measures analysed. The above results pp.x
demonstrate the importance of including stochasticity Aytug, H., Lawley, M. A., McKay, K., Mohan, S., & Uzsoy, R.
in the decision-making process, and of making an (2005). Executing production schedules in the face of
accurate distribution fitting in order to obtain consis­ uncertainties: A review and some future directions.
European Journal of Operational Research, 161(1),
tent and appropriate results. 86–110. https://doi.org/10.1016/j.ejor.2003.08.027
For future research, it is recommended to collect Ayyoubzadeh, B., Ebrahimnejad, S., Bashiri, M.,
data from a realistic FJSS, perform probability distri­ Baradaran, V., & Hosseini, S. M. H. (2021). Modelling
bution fitting and evaluate the performance of the and an improved NSGA-II algorithm for sustainable
simheuristic in comparison to dispatching rules with manufacturing systems with energy conservation under
the same fitted probability distributions. In addition to environmental uncertainties: A case study. International
Journal of Sustainable Engineering, 14(3), 255–279.
this, it is suggested to integrate other stochastic factors https://doi.org/10.1080/19397038.2021.1923083
that may occur in a manufacturing environment, such Azadeh, A., Negahban, A., & Moghaddam, M. (2012).
as machine breakdowns, job reprocessing, and job and A hybrid computer simulation-artificial neural network
machine availability time variation, among others. algorithm for optimisation of dispatching rule selection
Another topic that can be studied is the analysis of in stochastic job shop scheduling problems. International
the behaviour of other probability distributions under Journal of Production Research, 50(2), 551–566. https://
doi.org/10.1080/00207543.2010.539281
similar conditions, such as the gamma or Weibull Bagheri-Esfeh, H., Safikhani, H., & Motahar, S. (2020).
distributions. An additional exploration is to evaluate Multi-objective optimization of cooling and heating
this problem with other simheuristic modelling that loads in residential buildings integrated with phase
uses Monte Carlo simulation hybridised with other change materials using the artificial neural network and
metaheuristics, such as tabu search, PSO, and iterated genetic algorithm. Journal of Energy Storage, 32, 101772.
local search, among others. We also suggest hybridis­ https://doi.org/10.1016/j.est.2020.101772
Bagheri, A., & Zandieh, M. (2011). Bi-criteria flexible job-
ing the proposed NSGA-II with other types of simula­ shop scheduling with sequence-dependent setup times—
tion, such as agent-based or discrete-event, and Variable neighborhood search approach. Journal of
making comparisons with the proposed simheuristic. Manufacturing Systems, 30(1), 8–15. https://doi.org/10.
Finally, it is recommended to test the computing bud­ 1016/j.jmsy.2011.02.004
get allocation or the implementation of common ran­ Baker, K. R., & Altheimer, D. (2012). Heuristic solution
dom numbers to make the simheuristic more efficient. methods for the stochastic flow shop problem. European
Journal of Operational Research, 216(1), 172–177. https://
doi.org/10.1016/j.ejor.2011.07.021
Bayliss, C., & Panadero, J. (2023). Simheuristic and learn­
Disclosure statement heuristic algorithms for the temporary-facility location
and queuing problem during population treatment or
No potential conflict of interest was reported by the testing events. Journal of Simulation, 1–20. https://doi.
author(s). org/10.1080/17477778.2023.2166879
 JOURNAL OF SIMULATION 23

Brandimarte, P. (1993). Routing and scheduling in a flexible Gonzalez-Neira, E. M., Montoya-Torres, J. R., &
job shop by tabu search. Annals of Operations Research, Jimenez, J. F. (2021). A multicriteria simheuristic
41(3), 157–183. https://doi.org/10.1007/BF02023073 approach for solving a stochastic permutation flow shop
Buddala, R., & Mahapatra, S. S. (2019). Two-stage scheduling problem. Algorithms, 14(7), 1–22. https://doi.
teaching-learning-based optimization method for flexible org/10.3390/a14070210
job-shop scheduling under machine breakdown. The González-Neira, E. M., Urrego-Torres, A. M., Cruz-
International Journal of Advanced Manufacturing Riveros, A. M., Henao-García, C., Montoya-
Technology, 100(5–8), 1419–1432. https://doi.org/10. Torres, J. R., Molina-Sánchez, L. P., & Jiménez, J. F.
1007/s00170-018-2805-0 (2019). Robust solutions in multi- objective stochastic
Caldeira, R. H., & Gnanavelbabu, A. (2021). A simheuristic permutation flow shop problem. Computers and
approach for the flexible job shop scheduling problem Industrial Engineering, 137, 106026. https://doi.org/10.
with stochastic processing times. Simulation, 97(3), 1016/j.cie.2019.106026
215–236. https://doi.org/10.1177/0037549720968891 Gruler, A., Panadero, J., de Armas, J., Pérez, J. A. M., &
Castaneda, J., Martin, X., Ammouriova, M., Panadero, J., & Juan, A. A. (2020). A variable neighborhood search sim­
Juan, A. (2022). A fuzzy simheuristic for the permutation heuristic for the multiperiod inventory routing problem
flow shop problem under stochastic and fuzzy with stochastic demands. International Transactions in
uncertainty. Mathematics, 10(10), 1760. https://doi.org/ Operational Research, 27(1), 314–335. https://doi.org/10.
10.3390/math10101760 1111/itor.12540
Chica, M., Juan Pérez, A. A., Cordon, O., & Kelton, D. Gupta, S., & Jain, A. (2022). Analysis of integrated preven­
(2017). Why simheuristics? benefits, limitations, and tive maintenance and machine failure in stochastic flex­
best practices when combining metaheuristics with ible job shop scheduling with sequence-dependent setup
simulation. Benefits, Limitations, and Best Practices time. Smart Science, 10(3), 175–197. https://doi.org/10.
When Combining Metaheuristics with Simulation 1080/23080477.2021.1992823
(January 1. https://doi.org/10.2139/ssrn.2919208 Hao, X., Lin, L., Gen, M., & Chien, C.-F. (2014, Janua). An
Deb, K., Pratap, A., Agarwal, S., & Meyarivan, T. (2002). effective markov network based eda for flexible job shop
A fast and elitist multiobjective genetic algorithm: scheduling problems under uncertainty. IEEEVol. 2014.
NSGA-II. IEEE Transactions on Evolutionary Hatami, S., Calvet, L., Fernández-Viagas, V.,
Computation, 6(2), 182–197. https://doi.org/10.1109/ Framiñán, J. M., & Juan, A. A. (2018). A simheuristic
4235.996017
algorithm to set up starting times in the stochastic parallel
Demir, Y., & İ̇̇şleyen, S. K. (2013). Evaluation of mathema­
flowshop problem. Simulation Modelling Practice and
tical models for flexible job-shop scheduling problems.
Theory, 86, 55–71. https://doi.org/10.1016/j.simpat.2018.
Applied Mathematical Modelling, 37(3), 977–988. https://
04.005
doi.org/10.1016/j.apm.2012.03.020
He, W., & Sun, D. (2013). Scheduling flexible job shop
Dosdoğru, A. T., Göçken, M., & Geyik, F. (2015).
problem subject to machine breakdown with route chan­
Integration of genetic algorithm and monte carlo to ana­
ging and right-shift strategies. The International Journal
lyze the effect of routing flexibility. The International
of Advanced Manufacturing Technology, 66(1–4),
Journal of Advanced Manufacturing Technology, 81(5–
501–514. https://doi.org/10.1007/s00170-012-4344-4
8), 1379–1389. https://doi.org/10.1007/s00170-015-7247-
Horng, S.-C., Lin, S.-S., & Yang, F.-Y. (2012). Evolutionary
3
El-Abbasy, M. S., Elazouni, A., & Zayed, T. (2020). Finance- algorithm for stochastic job shop scheduling with ran­
based scheduling multi-objective optimization: dom processing time. Expert Systems with Applications,
Benchmarking of evolutionary algorithms. Automation 39(3), 3603–3610. https://doi.org/10.1016/j.eswa.2011.09.
in Construction, 120, 103392. https://doi.org/10.1016/j. 050
autcon.2020.103392 Hurink, J., Jurisch, B., & Thole, M. (1994). Tabu search for
Flores Gómez, M., Borodin, V., & Dauzère-Pérès, S. (2021). the job-shop scheduling problem with multi-purpose
A Monte Carlo based method to maximize the service machines. OR Spektrum, 15(4), 205–215. https://doi.org/
level on the makespan in the stochastic flexible job-shop 10.1007/BF01719451
scheduling problem. 2021 IEEE 17th International Jang, W. (2002). Dynamic scheduling of stochastic jobs on
Conference on Automation Science and Engineering a single machine. European Journal of Operational
(CASE), Lyon, France (pp. 2072–2077). https://doi.org/ Research, 138(3), 518–530. https://doi.org/10.1016/
10.1109/CASE49439.2021.9551529 S0377-22170100174-6
Gnanavelbabu, A., Caldeira, R. H., & Vaidyanathan, T. Jin, L., Zhang, C., Wen, X., Sun, C., & Fei, X. (2021).
(2021). A simulation-based modified backtracking search A neutrosophic set-based tlbo algorithm for the flexible
algorithm for multi-objective stochastic flexible job shop job-shop scheduling problem with routing flexibility and
scheduling problem with worker flexibility. Applied Soft uncertain processing times. Complex & Intelligent
Computing, 113, 107960. https://doi.org/10.1016/j.asoc. Systems, 7(6), 2833–2853. https://doi.org/10.1007/
2021.107960 s40747-021-00461-3
Gocken, T., Dosdogru, A. T., Boru, A., & Gocken, M. (2019). Juan, A. A., Barrios, B. B., Vallada, E., Riera, D., & Jorba, J.
Integrating process plan and part routing using optimiza­ (2014). A simheuristic algorithm for solving the permu­
tion via simulation approach. International Journal of tation flow shop problem with stochastic processing
Simulation Modelling, 18(2), 254–266. https://doi.org/ times. Simulation Modelling Practice and Theory, 46,
10.2507/IJSIMM18(2)470 101–117. https://doi.org/10.1016/j.simpat.2014.02.005
González-Martín, S., Juan, A., Riera, D., Elizondo, M., & Juan, A. A., Faulin, J., Grasman, S. E., Rabe, M., &
Ramos-González, J. (2016). A simheuristic algorithm for Figueira, G. (2015). A review of simheuristics: Extending
solving the arc routing problem with stochastic demands. metaheuristics to deal with stochastic combinatorial opti­
Journal of Simulation, 12(1), 53–66. https://doi.org/10. mization problems. Operations Research Perspectives, 2,
1057/jos.2016.11 62–72. https://doi.org/10.1016/j.orp.2015.03.001
24 C. A. RODRÍGUEZ-ESPINOSA ET AL.

Juan, A. A., Faulin, J., Grasman, S., Riera, D., Marull, J., & breakdowns. Computers and Industrial Engineering, 112,
Mendez, C. (2011). Using safety stocks and simulation to 595–606. https://doi.org/10.1016/j.cie.2017.03.006
solve the vehicle routing problem with stochastic Özgüven, C., Özbakır, L., & Yavuz, Y. (2010). Mathematical
demands. Transportation Research Part C: Emerging models for job-shop scheduling problems with routing
Technologies, 19(5), 751–765. https://doi.org/10.1016/j. and process plan flexibility. Applied Mathematical
trc.2010.09.007 Modelling, 34(6), 1539–1548. https://doi.org/10.1016/j.
Kacem, I., Hammadi, S., & Borne, P. (2002). Approach by apm.2009.09.002
localization and multiobjective evolutionary optimization Pinedo, M. L. (2016). Scheduling (5th ed.). Springer
for flexible job-shop scheduling problems. IEEE International Publishing. https://doi.org/10.1007/978-
Transactions on Systems, Man, and Cybernetics, Part 3-319-26580-3
C (Applications and Reviews), 32(1), 1–13. https://doi. Quintero-Araujo, C. L., Guimarans, D., & Juan, A. A.
org/10.1109/TSMCC.2002.1009117 (2021). A simheuristic algorithm for the capacitated loca­
Keenan, P., Panadero, J., Juan, A. A., Martí, R., & tion routing problem with stochastic demands. Journal of
McGarraghy, S. (2021). A strategic oscillation simheuris­ Simulation, 15(3), 217–234. https://doi.org/10.1080/
tic for the time capacitated arc routing problem with 17477778.2019.1680262
stochastic demands. Computers & Operations Research, Rahmati, S. H. A., Ahmadi, A., & Govindan, K. (2018).
133, 105377. https://doi.org/10.1016/j.cor.2021.105377 A novel integrated condition- based maintenance and
Kizys, R., Doering, J., Juan, A. A., Polat, O., Calvet, L., & stochastic flexible job shop scheduling problem:
Panadero, J. (2022). A simheuristic algorithm for the Simulation-based optimization approach. Annals of
portfolio optimization problem with random returns Operations Research, 269(1–2), 583–621. https://doi.org/
and noisy covariances. Computers & Operations 10.1007/s10479-017-2594-0
Research, 139, 105631. https://doi.org/10.1016/j.cor. Reyes-Rubiano, L., Ferone, D., Juan, A. A., & Faulin, J.
2021.105631 (2019). A simheuristic for routing electric vehicles with
Lemos, R., & Ronconi, D. (2015). Heuristics for the stochas­ limited driving ranges and stochastic travel times. Sort, 1,
tic single-machine problem with e/t costs. International 3–24. https://doi.org/10.2436/20.8080.02.77
Journal of Production Economics, 168, 131–142. https:// Sajadi, S. M., Alizadeh, A., Zandieh, M., & Tavan, F. (2019).
doi.org/10.1016/j.ijpe.2015.06.014 Robust and stable flexible job shop scheduling with ran­
Li, J.-Q., Deng, J.-W., Li, C.-Y., Han, Y.-Y., Tian, J., dom machine breakdowns: Multi-objectives genetic algo­
Zhang, B., & Wang, C.-G. (2020). An improved Jaya rithm approach. International Journal of Mathematics in
algorithm for solving the flexible job shop scheduling Operational Research, 14(2), 268. https://doi.org/10.1504/
problem with transportation and setup times. Knowledge- IJMOR.2019.097759
Based Systems, 200, 106032. https://doi.org/10.1016/j. Shen, X.-N., Han, Y., & Fu, J.-Z. (2017). Robustness mea­
knosys.2020.106032 sures and robust scheduling for multi-objective stochastic
Liefooghe, A., Basseur, M., Humeau, J., Jourdan, L., & flexible job shop scheduling problems. Soft Computing, 21
Talbi, E. G. (2012). On optimizing a bi-objective flowshop (21), 6531–6554. https://doi.org/10.1007/s00500-016-
scheduling problem in an uncertain environment. 2245-4
Computers and Mathematics with Applications, 64(12), Soroush, H. (2010). Solving a stochastic single machine
3747–3762. https://doi.org/10.1016/j.camwa.2012.02.051 problem with initial idle time and quadratic objective.
Li, Z., & Ierapetritou, M. (2008). Process scheduling under Computers & Operations Research, 37(7), 1328–1347.
uncertainty: Review and challenges. Computers & https://doi.org/10.1016/j.cor.2009.10.002
Chemical Engineering, 32(4–5), 715–727. https://doi.org/ Sun, D., He, W., Zheng, L., & Liao, X. (2014). Scheduling
10.1016/j.compchemeng.2007.03.001 flexible job shop problem subject to machine breakdown
Lin, C., Cao, Z., & Zhou, M. (2022). Learning-based grey with game theory. International Journal of Production
wolf optimizer for stochastic flexible job shop scheduling. Research, 52(13), 3858–3876. https://doi.org/10.1080/
IEEE Transactions on Automation Science and 00207543.2013.784408
Engineering, 19(4), 3659–3671. https://doi.org/10.1109/ Sun, L., Lin, L., Li, H., & Gen, M. (2019). Cooperative
TASE.2021.3129439 co-evolution algorithm with an MRF- based decomposi­
Martínez-Reyes, A., Quintero-Araújo, C. L., & Solano- tion strategy for stochastic flexible job shop scheduling.
Charris, E. L. (2021). Supplying personal protective Mathematics, 7(4), 318. https://doi.org/10.3390/
equipment to intensive care units during the COVID-19 math7040318
outbreak in colombia. a simheuristic approach based on Tamssaouet, K., Dauzère-Pérès, S., Knopp, S., Bitar, A., &
the location-routing problem. Sustainability, 13(14), Yugma, C. (2022). Multiobjective optimization for com­
7822. https://doi.org/10.3390/su13147822 plex flexible job-shop scheduling problems. European
Mokhtari, H., & Dadgar, M. (2015). Scheduling optimiza­ Journal of Operational Research, 296(1), 87–100. https://
tion of a stochastic flexible job-shop system with doi.org/10.1016/j.ejor.2021.03.069
time-varying machine failure rate. Computers and Thi, L. M., Anh, T. T. M., & Hop, N. V. (2022). An improved
Operations Research, 61, 31–45. https://doi.org/10.1016/ hybrid metaheuristics and rule- based approach for flex­
j.cor.2015.02.014 ible job-shop scheduling subject to machine breakdowns.
Nieto, A., Serra, M., Juan, A. A., & Bayliss, C. (2022). A Engineering Optimization, 1–21. https://doi.org/10.1080/
ga-simheuristic for the stochastic and multi-period port­ 0305215X.2022.2098283
folio optimisation problem with liabilities. Journal of Trietsch, D., Mazmanyan, L., Gevorgyan, L., & Baker, K. R.
Simulation, 1–14. https://doi.org/10.1080/17477778. (2012). Modeling activity times by the parkinson distri­
2022.2041990 bution with a lognormal core: Theory and validation.
Nouiri, M., Bekrar, A., Jemai, A., Trentesaux, D., European Journal of Operational Research, 216(2),
Ammari, A. C., & Niar, S. (2017). Two stage particle 386–396. https://doi.org/10.1016/j.ejor.2011.07.054
swarm optimization to solve the flexible job shop predic­ Türkyılmaz, A., Şenvar, Ö., Ünal, Bulkan, S., & Ünal, İ.
tive scheduling problem considering possible machine (2020). A research survey: Heuristic approaches for
 JOURNAL OF SIMULATION 25

solving multi objective flexible job shop problems. Yang, Y., Huang, M., Wang, Z. Y., & Zhu, Q. B. (2020).
Journal of Intelligent Manufacturing, 31(8), 1949–1983. Robust scheduling based on extreme learning machine
https://doi.org/10.1007/s10845-020-01547-4 for bi-objective flexible job-shop problems with
Von Lücken, C., Barán, B., & Brizuela, C. (2014). A survey machine breakdowns. Expert Systems with
on multi-objective evolutionary algorithms for Applications, 158, 113545. https://doi.org/10.1016/j.
many-objective problems. Computational Optimization eswa.2020.113545
and Applications, 58(3), 707–756. https://doi.org/10. Yang, X., Zeng, Z., Wang, R., Sun, X., & Song, H. (2016). Bi-
1007/s10589-014-9644-1 objective flexible job-shop scheduling problem consider­
Wang, Y., Zhang, J., Guan, X., Xu, M., Wang, Z., & ing energy consumption under stochastic processing
Wang, H. (2021). Collaborative multiple centers fresh times. PLOS ONE, 11(12), e0167427. https://doi.org/10.
logistics distribution network optimization with resource 1371/journal.pone.0167427
sharing and temperature control constraints. Expert Yusoff, Y., Ngadiman, M. S., & Zain, A. M. (2011). Overview
Systems with Applications, 165, 113838. https://doi.org/ of NSGA-II for optimizing machining process parameters.
10.1016/j.eswa.2020.113838 Procedia Engineering, 15, 3978–3983. Retrieved from.
Wu, Z., Sun, S., & Yu, S. (2020). Optimizing makespan and https://doi.org/10.1016/j.proeng.2011.08.745
stability risks in job shop scheduling. Computers and Zhang, G., Lu, X., Liu, X., Zhang, L., Wei, S., & Zhang, W.
Operations Research, 122, 104963. https://doi.org/10. (2022). An effective two-stage algorithm based on con­
1016/j.cor.2020.104963 volutional neural network for the bi-objective flexible job
Xiao, S., Sun, S., & Jin, J. J. (2017). Surrogate measures for shop scheduling problem with machine breakdown.
the robust scheduling of stochastic job shop scheduling Expert Systems with Applications, 203, 117460. https://
problems. Energies, 10(4), 543. https://doi.org/10.3390/ doi.org/10.1016/j.eswa.2022.117460
en10040543 Zhang, J., & Yang, J. (2016). Flexible job-shop scheduling
Xiong, J., Xing, L. N., & Chen, Y. W. (2013). Robust sche­ with flexible workdays, preemption, overlapping in
duling for multi-objective flexible job-shop problems operations and satisfaction criteria: An industrial
with random machine breakdowns. International application. International Journal of Production
Journal of Production Economics, 141(1), 112–126. Research, 54(16), 4894–4918. https://doi.org/10.1080/
https://doi.org/10.1016/j.ijpe.2012.04.015 00207543.2015.1134839