Accepted Manuscript

Minimizing flowtime in a flowshop scheduling problem with a biased

random-key genetic algorithm

Carlos E. Andrade, Thuener Silva, Luciana S. Pessoa

PII: S0957-4174(19)30160-5
DOI: https://doi.org/10.1016/j.eswa.2019.03.007
Reference: ESWA 12533

To appear in: Expert Systems With Applications

Received date: 26 October 2018

Revised date: 3 March 2019
Accepted date: 4 March 2019

Please cite this article as: Carlos E. Andrade, Thuener Silva, Luciana S. Pessoa, Minimizing flowtime
in a flowshop scheduling problem with a biased random-key genetic algorithm, Expert Systems With
Applications (2019), doi: https://doi.org/10.1016/j.eswa.2019.03.007

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service
to our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and
all legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT

Highlights

• Proposition of a new feature for the Biased Random-Key Genetic Algo-

rithm

• Identification of lower and upper bounds, as well as some optimal values

for classical instances

T
• Developed genetic algorithm are strong contenders for large scale problems

IP
CR
US
AN
M
ED
PT
CE
AC

1
 ACCEPTED MANUSCRIPT

Minimizing flowtime in a flowshop scheduling problem

with a biased random-key genetic algorithm

Carlos E. Andradea , Thuener Silvab , Luciana S. Pessoab,∗

T
a AT&T Labs Research,

200 South Laurel Avenue, Middletown, NJ 07748 USA

b Department of Industrial Engineering, PUC-Rio,

IP
Rua Marquês de São Vicente, 225, Gávea - 22453-900 Rio de Janeiro, RJ, Brazil.

CR
Abstract
In this paper, we advance the state of the art for solving the Permutation Flow-

US
shop Scheduling Problem with total flowtime minimization. For this purpose,
we propose a Biased Random-Key Genetic Algorithm (BRKGA) introducing
on it a new feature called shaking. With the shaking, instead to full reset the
population to escape from local optima, the shaking procedure perturbs all in-
AN
dividuals from the elite set and resets the remaining population. We compare
results for the standard and the shaking BRKGA with results from the Iterated
Greedy Search, the Iterated Local Search, and a commercial mixed integer pro-
gramming solver, in 120 traditional instances. For all algorithms, we use warm
M

start solutions produced by the state-of-the-art Beam-Search procedure. Com-

putational experiments show the efficiency of proposed BRKGA, in addition to
identify lower and upper bounds, as well as some optimal values, among the
solutions.
ED

Keywords: Flowshop Scheduling, Biased Random-Key Genetic Algorithm,

Metaheuristics.
PT

1. Introduction

Scheduling problems refer to the assignment of jobs to machines – productive

resources – as well as the definition of the processing start and end times for each
CE

job in each machine, in order to optimize one or more criteria. This class of prob-
lems is among the most studied in the operational research field and presents
applications in several areas (Pinedo, 2016). Fuchigami & Rangel (2018) re-
view articles published in the last 25 years and identify case studies in various
AC

productive sectors. According to the authors, scheduling problems found appli-

cations in a variety of fields, where a sequence of tasks must be executed. They

∗ Corresponding
author. Tel: +55 (21) 3527-2180
Email addresses: [email protected] (Carlos E. Andrade),
[email protected] (Thuener Silva), [email protected] (Luciana S. Pessoa)

Preprint submitted to Expert Systems with Applications March 8, 2019

 ACCEPTED MANUSCRIPT

include the primary, processing and assembling industries, as much as in the

telecommunication and transportation industries, education, government, and
among others.
One of the most important scheduling problems is the Permutation Flow-
shop Scheduling Problem (PFSP). This problem considers a set of n jobs to
be scheduled on a set of m machines. Each job j has a processing time pj` on

T
machine `, and must be scheduled on each machine in sequence and, therefore,
can be executed only one machine at time. Given a schedule S, we denote by

IP
Cj` the completion time of job j on machine `, and by Cj the completion time of
the job j on the last machine. This work deals with the PFSP whose objective
function value is given by the summation of Cj overall jobs, also known as total

CR
flow time.
Due to its NP-hard nature (Garey et al., 1976), few exact methods have
been developed for solving it to optimality. Among them are the branch-and-
bound methods of Chung et al. (2002); Della Croce et al. (2002); Madhushini

US
et al. (2009), and van de Velde (1991). On the other hand, there are many
heuristic approaches in the literature. Extensive comparative studies on heuris-
tic methods were developed by Framinan et al. (2005) and Pan & Ruiz (2013).
Regarding constructive algorithms, Pan & Ruiz (2013) present a compre-
AN
hensive set of methods. This survey described the LR constructive method
proposed by Liu & Reeves (2001) with O(n3 m) running time. Such method was
improved by Fernandez-Viagas & Framinan (2015), leading to the O(n2 m) FF
algorithm. Afterward, Abedinnia et al. (2016) and Rossi et al. (2017) described
new constructive methods which, despite some potential results on the quality
M

of the solutions, have a higher complexity than FF thus making them inefficient
for larger instances. Recently, Fernandez-Viagas & Framinan (2017) proposed
a Beam-Search Constructive Heuristic (BSCH) based on the FF method by
ED

Fernandez-Viagas & Framinan (2015).

Other methods applied to PFSP are the metaheuristics that constitute a
powerful tool to explore the search space by combining simple constructive and
local search methods with a (pseudo) random mechanism to escape from local
PT

optima (Martí et al., 2017). Pan & Ruiz (2012) analyse comparatively the best
performing metaheuristic methods developed so far including genetic algorithms
(Tseng & Lin, 2009; Zhang et al., 2009; Tseng & Lin, 2010; Xu et al., 2011),
estimation of distribution algorithms (Jarboui et al., 2009), discrete differential
CE

evolution algorithms (Pan et al., 2008; Tasgetiren et al., 2011), iterated greedy
algorithms (Dubois-Lacoste et al., 2011; Pan et al., 2008), iterated local searches
(Dong et al., 2009), among others. Additionally, Pan & Ruiz (2012) included
in the analysis new iterated greedy and iterated local search methods as well as
AC

their population-based variants. A comparative evaluation showed the superi-

ority of the new heuristics concerning the existing methods, not to mention the
simple iterated search algorithms performed better than the variant based on a
population scheme.
To the best of our knowledge, the state-of-the-art in heuristic methods for
the permutation flowshop problem with total flowtime criterion was obtained
by Fernandez-Viagas & Framinan (2017) when applying the beam-search con-

3
 ACCEPTED MANUSCRIPT

structive heuristic in the framework of an Iterated Local Search (ILS) proposed

by Dong et al. (2013). Computational experiments revealed the efficiency of the
BSCH as a constructive heuristic when applied in isolation as well as the benefit
of using it to generate initial solutions for a metaheuristic method.
In this context, we observe that significant progress has been achieved by ap-
plying hybrid methods based on local search and constructive algorithms. How-

T
ever, population-based methods seem to remain underexplored. Therefore, this
paper aims to fill this gap by presenting a Biased Random-Key Genetic Algo-

IP
rithm (BRKGA) for minimizing total flowtime in a permutation flowshop prob-
lem. The proposed method incorporates the beam-search constructive heuristic
(Fernandez-Viagas & Framinan, 2017) as well as a new feature called shaking to

CR
the standard BRKGA (Gonçalves & Resende, 2011) to evolve the population of
solutions. Another contribution of this paper refers to the establishment of pri-
mal and dual bounds by running the related Integer Linear Program (ILP) model
on a commercial solver, as much as the proof of optimality for some instances.

US
Computational study on 120 classical benchmark instances from literature was
carried out to demonstrate the effectiveness of the proposed method.
The remainder of this paper is organized as follows. Section 2 brings the for-
mal definition and the mathematical model of the problem. Section 3 describes
AN
biased random-key genetic algorithms as well as its customization for tackling
the flowshop problem with total flowtime minimization criterion. Addition-
ally, an Iterated Greedy Search and an Iterated Local Search are presented,
respectivelly, in Sections 4 and 5. In order to evaluate our proposals, extensive
computational experiments were carried out. Detailed results and analysis are
M

shown is Section 6. Concluding remarks are in Section 7.

2. Problem definition and formulation

ED

Permutation flowshop scheduling problems require that the jobs be schedule

in sequence, regardless the time they are schedule. Therefore, any job’s per-
mutation generates a valid schedule. In the permutation flowshop scheduling
PT

problems (PFSP) with total flowtime minimization criterion, a set of n jobs

J = {1, 2, . . . , n} have to be scheduled, following the same order, on a set of m
machines M = {1, 2, . . . , m}. Each job j ∈ J has a processing time pj` > 0 on
machine ` ∈ M . Job processing interruption is not allowed and all parameters
CE

are integer numbers. Given a schedule S, we denote by Cj` the completion time
of job j on machine `. We denote by Cj the completion time of the job j in the
last machine. According to the three-field notationP of Graham et al. (1979), the
n
problem under study can be defined as Fm |prmu| j=1 Cj .
AC

Mixed integer programming model (1a)-(1g) is based on the positional-

variables formulation of Wagner (1959), and described by Stafford (1988) and
Della Croce et al. (2011). In that model, let Xij be the binary decision variable
such that Xij = 1 if job i is the j th job of the sequence and Xij = 0 otherwise.
pos
Let Cj` ≥ 0 denote the completion time of the job on the j th position on
pos
machine ` ∈ M . Assume that C0` = 0 for all ` ∈ M .

4
 ACCEPTED MANUSCRIPT

n
X pos
min Cjm (1a)
j=1
Xn
s.t. Xij = 1 ∀j ∈ {1, . . . , n}, (1b)

T
i=1
Xn

IP
Xij = 1 ∀i ∈ {1, . . . , n}, (1c)
j=1
n
X ∀j ∈ {1, . . . , n}

CR
pos pos
Cj` ≥ Cj−1,` + pi` Xij (1d)
∀` ∈ {1, . . . , m},
i=1
Xn
pos pos ∀` ∈ {1, . . . , m − 1}
Cj,`+1 ≥ Cj` + pi,`+1 Xij (1e)
∀j ∈ {1, . . . , n},

US
i=1
pos ∀j ∈ {1, . . . , n}
Cj` ≥0 (1f)
∀` ∈ {1, . . . , m},
Xij ∈ {0, 1} ∀i, j ∈ {1, . . . , n}. (1g)
AN
Objective Function (1a) aims to minimize the sum of the completion times
of all jobs. Constraint (1b) and (1c) ensure that each job is attributed to
one position and vice-versa. Constraint (1d) prevents the overlapping of a job
M

and its predecessor on a machine. Constraint (1e) ensures that a job cannot
be processed on a machine before its completion on the preceding machine.
Constraints (1f) and (1g) give the definition domain of the variables.
ED

3. Biased Random-Key Genetic Algorithm

The Biased Random-Key Genetic Algorithm (BRKGA) has been applied to

several optimization problems, as reviewed in Prasetyo et al. (2015), including
PT

those in the telecommunications context (Resende, 2012), hub location (Lopes

et al., 2016; Pessoa et al., 2017), facility location (Biajoli et al., 2019), and vehicle
routing (Grasas et al., 2014; Prasetyo et al., 2018). When it comes to schedul-
CE

ing problems, recent works tackled a bi-objective batching machine scheduling

problem (Cabo et al., 2018), a single batch processing machine scheduling (Li
& Zhang, 2018), project scheduling problems (Almeida et al., 2018), and a
flowshop problem with stepwise objective function with payoff maximization
criterion (Pessoa & Andrade, 2018). BRKGA has been used sucessfully also in
AC

large-scale real problems such as network planning (Andrade et al., 2015a) and
wireless updates on connected cars (Andrade et al., 2017).
BRKGA is a population-based metaheuristic where each chromosome en-
codes a solution through a vector of random numbers. The biased mating of
two parents is responsible for inducing the offspring inheritance from the best
parent characteristics (Gonçalves & Resende, 2011). The initial population is

5
 ACCEPTED MANUSCRIPT

created with p chromosomes, each one composed of n genes. A gene is a ran-

dom real number (random-key) uniformly drawn from the interval [0, 1]. A
chromosome is mapped into a feasible solution by a decoder function which also
computes the cost of such a solution. In the genetic algorithm context, the cost
of a solution is referred to as the fitness of the individual which is related to its
capacity to survive through generations.

T
As illustrated in Figure 1, after all individuals have their fitness value de-
termined, the population is divided into two sets: the elite set, containing the

IP
fittest individuals, and the non-elite set, with the rest of the population. Af-
terward, the population evolves to a new generation in a three steps process
while keeping its size. First, in the reproduction step, the elite set, containing

CR
pe individuals is entirely copied to the new population. Then, the worst pm
non-elite individuals are replaced by mutants generated randomly in the same
way that the initial population. The third step, called crossover or mating, is
accomplished with the production of p − pe − pm offsprings. The generation of

US
an offspring requires the random choice of a parent from the elite set, and an-
other parent from the non-elite set. Besides, a parameterized uniform crossover
scheme applies a probability ρ > 0.5 of inheriting each gene value from the elite
parent.
AN
Once the reproduction, mutation and crossover steps are finished, each so-
lution is given to the decoder so that the fitness of each individual is calculated
starting a new evaluation-selection-evolution cycle. The population evolves gen-
eration after generation until a stopping criterion is reached. In other words,
the algorithm runs until a maximum processing time or a number of iterations
M

is met. Following, we customize a biased random-key genetic algorithm for the

flowshop problem with total flowtime minimization criterion.
ED

best Reproduction
fitness ELITE pe

Elite parent
PT

Crossover X

p
NON-
CE

ELITE Non-elite parent

AC

worst pm
fitness

Mutants

Figure 1: BRKGA scheme for population evolution.

6
 ACCEPTED MANUSCRIPT

3.1. Representation and decoder

A chromosome that encodes a solution for a flowshop problem is composed
of n genes, i.e., a random-key for each job, as shown in Figure 2. The decoder
sorts the vector of random-keys so that the sequence of jobs related to the sorted
genes corresponds to a feasible solution for the problem.

T
Jobs 1 2 3 4 5 6 7 8 9 10

IP
Unsorted genes 0.00 0.71 0.74 0.87 0.14 0.16 0.65 0.75 0.38 0.81

Genes

CR
sorting

Sorted genes 0.00 0.14 0.16 0.38 0.65 0.71 0.74 0.75 0.81 0.87

Jobs sequence 1 5 6

US 9 7 2

Figure 2: A decoder for scheduling problems.

3 8 10 4
AN
3.2. Warm-start
The standard implementation of a BRKGA creates the initial population by
randomly generating all individuals. However, Reeves (1995), Zhang et al. (2009),
Andrade et al. (2013; 2015b), Lopes et al. (2016), and Pessoa & Andrade (2018)
M

have shown the benefits of providing good solutions as chromossomes in the ini-
tial population. In this work, we used solutions from the Beam Search Construc-
tive heuristic (BSCH – Fernandez-Viagas & Framinan, 2017) as warm-starters
ED

to BRKGA.
BSCH builds, level by level, a search tree with a partial solution in each node
while solutions are constructed element by element. BSCH is parameterized so
that the number of nodes explored at each level can vary from only one node to
x nodes. In the case where BSCH explores only one node at each iteration, this
PT

method becomes equivalent to the simplest FF constructive method (Fernandez-

Viagas & Framinan, 2015). FF builds a job sequence from scratch appending
jobs at the end of a partial solution. At each iteration, all jobs that are not in the
CE

partial schedule are (re)evaluated according to an index function that considers

the completion time and the idle time originated by the placement of such a
job at the end of the partial solution. BSCH behaves in a similar way as FF,
but when exploring more than one node, BSCH also computes an estimate of
AC

which nodes are the most promising to be kept as exploring branches. A forecast
index is used to compare different nodes by computing (additionally to the idle
time and the completion time of scheduled jobs) an estimate of the contribution
of unscheduled jobs to the completion time and, consequently, to the objective
function. For more details, refer to Fernandez-Viagas & Framinan (2017).

7
 ACCEPTED MANUSCRIPT

3.3. Local Search

The hybridization of evolutionary algorithms with local search was proposed
by Moscato (1989) with the goal of improving the fitness of an individual by
refining its characteristics. Indeed, local searches modify some components of
an initial solution to take it to a local minimum in its neighborhood.
For flowshop problems, two methods are commonly used: interchange and

T
insert moves (den Besten & Stützle, 2001). When the interchange move is
applied, a neighbor solution is reached by swapping two jobs from their original

IP
positions. On the other hand, the insert method explores the neighborhood of
a solution by moving a job to another position.
By following an iterative improvement strategy, once an initial solution is

CR
determined, the local search method is applied to it, and an improved neighbor
is chosen to replace it in the next iteration. Otherwise, if no improved solution is
found, the initial solution is returned as locally optimum (Hoos & Stützle, 2004).
Two or more local search methods can be combined in order to better explore the

US
solution space through movements over different neighborhoods. A procedure
that orchestrates the calls is the Variable Neighborhood Search (VND – Hansen
& Mladenović, 2003), which determine the order for applying the local search
methods.
AN
The interchange and insert neighborhood sizes are, respectively, n(n − 1)/2
and (n − 1)2 . Since exploring insert neighborhood is the most time-consuming
method, we allow only one iteration on it, i.e., we only explore the first-degree
neighbors and return the best found. On the other hand, since interchange
M

neighborhood is smaller than the insert neighborhood, we can afford to search

for the local minimum.
Therefore, our local search is VND-based, where in each iteration, we search
in the insert neighborhood first, and then, we explore the interchange neighbor-
ED

hood searching for local optima. The main VND loop stops when no improved
solution is found, or maximum BRKGA time is reached.
Applying the local search for each individual of the population usually does
not help to reach reasonable solutions, since it is time-expensive and impairs
PT

BRKGA to evolve the number of generations needed to good performance.

Therefore, we call the local search only on the best individual of the popu-
lation on each τ iterations, and the best individual value is different from the
best solution found so far.
CE

3.4. Shaking
In some situations, the diversity of the population can be compromised,
AC

i.e., when there is no difference between the fitness values of the whole elite
population or when there are too many iterations without improvement. In such
situations, it is usual to restart the search by resetting the whole population.
However, this multi-start strategy can be too drastic because it destroys the
convergence structure of the population.
In other words, the full resetting destroys useful parts of solutions (block of
genes) built over multiple generations of chromosome crossovers. These well-

8
 ACCEPTED MANUSCRIPT

Algorithm 1: Shaking.
Input : Population P where Pe is the elite set and Pne is the non-elite
set; shaking intensity ψ.
Output: Changed population.
1 foreach ind ∈ Pe do

T
2 Let S be the solution induced by ind ;
3 for k ← 1 to ψ do

IP
4 Swap two neighboring jobs hi, i + 1i in S, where i is uniformly
chosen from [1, n − 1];
5 Swap a pair jobs hi, ji in S, at the ith 6= j th uniformly chosen

CR
random positions;
6 Encode ind according to S;
7 Replace the members of Pne with random samples;
return changed population P .
8

US
conserved parts of the genome usually appear in good solutions. To avoid such
AN
a drastic move, in this work we introduce a new BRKGA feature called shaking.
Instead of fully resetting the whole population, the shaking method applies
random modifications to the individuals in the elite set, and resets the non-elite
individuals. Therefore, we guarantee diversity on the non-elite set and preserve
the convergence structure on the shaken elite set. Algorithm 1 describes the
M

shaking method, which is controlled by a shaking intensity parameter ψ.

The main loop (lines 1–6) iterates over all individuals in the elite set. Each
elite individual is decoded such that a solution S is extracted from it (line 2).
ED

The inner loop (lines 3–5) applies the perturbations ψ times. In the first per-
turbation, the algorithm swaps the positions of a randomly chosen job i and
its neighbor (line 4). The second perturbation swaps positions of two randomly
chosen jobs (line 5). The algorithm then re-encodes the solution as a chromo-
some in the elite set (line 6). To finish, the algorithm replaces all non-elite
PT

chromosomes by random individuals (line 7), and returns the new population
(line 8). Note that the shaking is done in the problem solution space rather
than the BRKGA representation space. Therefore, the algorithm decodes the
CE

chromosome, applies the perturbations, and then re-encode the shaken solution
to the BRKGA space. Note that this can be done without re-encoding the
solution by tracking the order of the jobs in the chromosome.
Although we can do the shaking anytime during the evolutionary process,
AC

we chose to apply it in the following situations:

• There were many iterations (R) without improvements on the best solution
of the current population. In this situation, the number of perturbations
is fixed proportion of the number of jobs, i.e., ψ = λn where n is the
number of jobs, and λ is a fixed value in the interval [0, 1];

9
 ACCEPTED MANUSCRIPT

• When the diversity of the population was compromised, i.e, there was no
difference between the fitness values of the whole elite population. In this
case, ψ = λn where λ is uniformly drawn from the interval [0.05, 0.20];
• There was also a special case when there was no improvement on the
overall best solution for too many iterations (R? ). For this case, the

T
perturbation is similar to the previous one but more aggressive, selecting
λ from the interval [0.5, 1.0];

IP
In some cases, even the shaking cannot help BRKGA escape from local optima.
In such situations, we use the traditional full population restart keeping the

CR
warm-start solution though. We call the full restart when the algorithm reaches
R?? iterations without improvement in the best solution.
Algorithm 2 depicts the overall framework. The algorithm starts with a pop-
ulation of random individuals and a warm-start solution from other methods
and initialize some local variables (lines 2–5). In the main loop (lines 6–31),

US
the algorithm first evolves the population according to BRKGA rules (omit-
ted for shortness; refer to Gonçalves & Resende (2011) for complete evolution
procedure). On every L iterations, the local search is executed (line 9).
In the next blocks, the algorithm calls the shaking procedure with different
AN
intensities according to the population status. If the current solution has not
been improved in the last R iterations, the algorithm performs a random shaking
(line 13). At this point, if the population is too homogeneous, a weak shaking is
performed (line 16). Next, the algorithm checks for updates in the best solution
M

and performs a strong shaking if the best solution has not improved for the last
R? iterations (line 21). However, as described earlier, the shaking may be not
enough. In this case, the algorithm performs a full-population reset if it reaches
R?? iterations without improvements in the best solution (line 26).
ED

4. Iterated Greedy Search

In order to conduct a comparative study with the proposed BRKGA, we also

PT

implemented an Iterated Greedy Search (IGS) method. Ruiz & Stützle (2007)
described the first application of IGS for the Permutation Flowshop Scheduling
Problem (PFSP) with makespan criterion. The superiority of the method was
CE

shown by Pan et al. (2008) when compared to the Discrete Differential Evolution
method of Pan et al. (2007) for problems with the makespan and the total
flowtime minimization criteria. IGS was also used successfully for the sequence-
dependent setup times flowshop problem with makespan and weighted tardiness
AC

objectives (Ruiz & Stützle, 2008), the unrelated parallel machine scheduling
(Fanjul-Peyro & Ruiz, 2010), and the blocking PFSP with makespan criterion
(Tasgetiren et al., 2017).
IGS builds a chain of solutions by iteratively generating new starting solu-
tions. At each iteration, the method partially destroys a starting solution and,
next, reconstructs the solution to the feasibility. Then, a local search method
is applied to the incumbent solution to find an improving neighbor. After each

10
 ACCEPTED MANUSCRIPT

Algorithm 2: General BRKGA framework with Shake.

Input : Limits of the number of iterations without improvement R, R? ,
R?? ; number of iterations to perform a local search L; number
of jobs n; stop criteria.
Output: Best solution.

T
1 Initialize population P with random individuals and the warm-start
solution;

IP
2 bestsol ← warm-start solution;
3 lastsol ← warm-start solution;
4 bestsoliter ← 1;

CR
5 iter ← 1;
6 while a stopping criterion is not reach do
7 Evolve P for one generation;
8 Let cursol be the value of the best solution in P ;
9
10

11
12
if iter = 0 mod L then

if cursol = lastsol then

stableit ← stableit + 1;
US
Do a local search in the best individual of P ;
AN
13 if stableit ≥ R then
14 λ ← rand([0.0, 1.0]);
15 shake(P, λn);
16 if elite individuals in P have the same fitness then
M

17 λ ← rand([0.05, 0.20]);
18 shake(P, λn);
19 else
ED

20 stableit ← 0;
21 if cursol ≥ bestsol then
22 bestsoliter ← bestsoliter + 1;
23 if bestsoliter = R? then
PT

24 λ ← rand([0.50, 1.00]);
25 shake(P, λn);
26 if bestsoliter = R?? then
Replace population P random with individuals and the
CE

27
warm-start solution;
28 else
29 bestsoliter ← 0;
AC

30 bestsol ← cursol ;
31 lastsol ← cursol ;
32 return the best solution bestsol .

11
 ACCEPTED MANUSCRIPT

destruction-reconstruction-local search cycle, an acceptance criterion evaluates

such solution to decide whether it will replace the initial solution to the next
iteration. This loop is repeated until a stopping criterion is met (Stützle &
Ruiz, 2017).
When customizing an IGS heuristic to solve a problem, five procedures must
be developed: (1) a deterministic or random method to construct an initial

T
solution; (2) a destruction procedure to remove some components of a solution;
(3) a reconstruction procedure that takes the partial solution incumbent from

IP
step (2) and generate a complete feasible solution; (4) a local search to explore
the neighborhood of the solution reached by the reconstruction step; and (5)
an acceptance criterion which balances between the highest diversity, when any

CR
new solution is accepted, or the strongest intensification, when only improving
solutions replaces the initial one to the next (2)–(4) loop iteration. Note that the
strength of the destruction, given by the number of affected components, must
be calibrated to reflect a trade-off between diversification and intensification.

US
The IGS heuristic developed for the problem tackled in this work takes as ini-
tial solution the scheduling obtained by the beam-search constructive heuristic
(Fernandez-Viagas & Framinan, 2017), described in Section 3.2. The destruction
procedure removes a certain number of randomly chosen jobs from a candidate
AN
solution (given a destruction parameter), generating a partial solution. Then,
the removed jobs are considered for being appended at the end of the partial
solution, observing the evaluation function proposed by Fernandez-Viagas &
Framinan (2015) for the FF constructive method, as mentioned in Section 3.2.
The local search phase applies the VND method (Hansen & Mladenović, 2003).
M

In this way, insert neighborhood is explored alternately with interchange neigh-

borhood, both using the best improvement search strategy. The incumbent
solution is accepted only if its cost is better than the best solution found so
ED

far.

5. Iterated Local Search

PT

Iterated Local Search (ILS) has been applied for diverse flowshop problems
(Stützle, 1998; Ribas et al., 2013; Juan et al., 2014; Xu et al., 2017). When
it comes to total flowtime minimization criterion, the reader is refered to the
works of Dong et al. (2009; 2011a;b). However their results were overcame in the
CE

recent paper of Fernandez-Viagas & Framinan (2017) which shows the benefits
of incorporating the beam-search procedure into the ILS method.
Analogously to the IGS metaheuristic, presented in Section 4, the ILS gener-
ates an initial local optimum solution, which is the starting point to explore the
AC

solution space. Then, a loop coordinates an iterative process of perturbation-

improvement-evaluation steps until a stopping criterion is reached (Lourenço
et al., 2019).
As proposed by Fernandez-Viagas & Framinan (2017), ILS applies the beam-
search procedure to construct the first solution. This is the same strategy ap-
plied in the BRKGA and IGS methods implemented in this work.

12
 ACCEPTED MANUSCRIPT

Within the cycle of perturbation-improvement-evaluation, the perturbation

step is performed by swapping two randomly chosen jobs. The number of swaps
determines how much characteristics of the original solution will be kept. The
greater this number, the greater the diversification. Therefore, this parameter
must be carefully calibrated. Improvement step is accomplished by applying
the Variable Neighborhood Descent (VND) strategy for local search, as for the

T
IGS (Section 4), exploring alternately insert and interchange neighborhoods
while enhancing solutions are found. Finally, the solution obtained by the im-

IP
provement procedure is compared to the current best solution. The lowest cost
solution is kept as the best solution found up to this point.

CR
6. Experimental results

6.1. Instances and computational environment

US
To test the proposed methodologies, we use 120 instances suggested by Tail-
lard (1991). We warm-started all studied algorithms using the solutions from
the beam-search procedure proposed by Fernandez-Viagas & Framinan (2017),
which are the state-of-the-art constructive heuristic for PFSP. We used the
AN
BSCH(n) version which stretchs the search over “n” jobs. Since in this paper, all
studied algorithms use solutions from BSCH(n) as incumbents, we disregard the
BSCH(n) running time.
The computational experiments were performed in a cluster of identical ma-
chines with a Intel Xeon E5530 CPU at 2.40GHz and 64 GB of RAM running
M

CentOS Linux. Heuristics proposed in this paper were implemented in Julia lan-
guage. For BSCH(n), we used the original version provided by Fernandez-Viagas
& Framinan (2017), implemented in C#. Model (1a)–(1g) was solved with IBM
ILOG CPLEX 12.8 solver. To have a fair comparison among the algorithms, all
ED

implementations use a single thread. We performed experiments with stopping

criterion of 5, 15, 30, and 60 minutes with 30 independent experiments for each
instance for these algorithms.
We run CPLEX with two different setups. In the first, CPLEX uses a single
PT

thread and stops either when it finds an optimal integer solution, or it reaches
the maximum time as defined for the heuristics (CPLEX). This setup is meant
to achieve direct comparison with the other methods proposed in this paper.
CE

In the second setup, we run CPLEX for 24 wall-clock hours using 12 threads
(CPLEX-24h). Such configuration looks to find optimal solutions or, at least,
compute the best possible lower bounds. Since this configuration uses far more
time and computer power per run than the other configurations, we use the
AC

results only for reporting.

6.2. Parameter settings

We performed a full-factorial design of experiments for ILS and IGS, because
the number of control parameters is small. We varied the number of swap moves
(for ILS) and the intensity of the destruction (for IGS) between 2 and 20. The
experiments were carried out on 24 instances, each two chosen from an instance

13
 ACCEPTED MANUSCRIPT

group. The experiments suggested that seven swaps moves for ILS and seven
jobs in the destruction phase for IGS, enable both algorithms to achieve the best
results.
Since BRKGA algorithms require a large number of parameters, we tuned
such settings employing the iterated F-race and the irace package (López-
Ibáñez et al., 2016). irace suggested: population size of p = n × 9, where n

T
is the number of jobs; elite size pe = 0.30 × p; number of mutants introduced
at each generation pm = 0.22 × p; probability of inheriting each allele from

IP
elite parent ρ = 0.55. For variations with local search, the suggested number
of local search iterations is 10. For variants without shaking, irace suggested
a full population reset after 1,000 iterations without improvements in the best

CR
solution. For variants where we apply shaking, irace suggest R = 1,000 and
λ = 10. We made R? and R?? proportional to R, and irace suggested R? =
5 × R and R?? = 10 × R. For details in the parameters and configurations, refer
to Appendix A.

6.3. Result analysis

US
We performed the comparison among the proposed algorithms regarding
solution quality and computational effort. For solution quality, we compute the
AN
classical Relative Percentage Deviation (RPD) and associated averages. Let I
be a set of instances. Let A be the set of algorithms, and assume that set RA
enumerates the independent runs for algorithm A ∈ A (as defined in Section 6.1,
30 runs for the heuristics and one run for CPLEX).1 We defined Cir A
as the total
M

flowtime obtained by algorithm A in instance i on run r, and Cibest as the

best total flowtime found across all algorithms for instance i. The Relative
Percentage Deviation (RPD) from the best solution of instance i is defined as
ED

A
Cir − Cibest
RPD A
ir = , ∀A ∈ A, i ∈ I, and r ∈ RA . (2)
Cibest

The general Average Relative Percentage Deviation (ARPD) for algorithm A

PT

is defined as
100 X X
ARPD A = RPD Air . (3)
|I| · |RA |
i∈I r∈RA
CE

Note that we can obtain different averages by subsetting set I, for example, per
instance or groups of instances.
To measure the computational effort, we use the Average Relative Percent-
age Computational Time (ARPT) proposed on Fernandez-Viagas & Frami-
AC

nan (2015) for deterministic heuristics. However, we have slightly adapted that
definition to accommodate randomized algorithms with multiple runs. First,
instead of account the total running time, we use the time to best solution

1 Note that this is necessary because randomized algorithms generate a distribution of so-

lutions (and convergence times) from runs with different (starting) random number generator
states, and one must account for all these results for a proper statistical analysis.

14
 ACCEPTED MANUSCRIPT

found in a given run, which we define as Tir A

. In other words, Tir
A
is the time
that algorithm A needs to converge on run r (find the best solution in that
run) for instance i. The reason for this approach is that all experiments are
carried out until the maximum time allowed, even though the algorithm may
converge earlier to the best solution found in that run. We also use the results
of all independent runs for all algorithms, as it is done for ARPD. With these

T
considerations, ARPT is defined for each algorithm A as
X X T A − ACT

IP
1 ir i
ARPT A = 1 + (4)
|I| · |RA | ACT i
i∈I r∈RA

CR
where X X A
1 Tir
ACT i = , ∀i ∈ I. (5)
|A| |RA |
A∈A r∈RA
ACT i is the average of Tir A
US
for instance i considering all algorithms. Note
that in Equation (4) starts adding one to the average which guarantees that
ARPT is always a positive number, allowing different transformations (such
AN
as logarithmic ones) when necessary. Therefore, while ARPD is the average
deviation from the value of the best solutions, ARPT is the average deviation
from the average computational time.
In terms of notations, for standard BRKGA variations (with full population
reset), we used the suffix STD. For variations using shaking, the suffix SK was
M

used. Similarly, variations using local search and variants not using local search,
have the prefixes LS and NLS, respectively. For other algorithms, we used their
usual acronyms.
ED

Figure 3 depicts the Pareto frontier among the algorithms regarding their
ARPD and ARPT. Both metrics were computed considering either the best-
known solution in the literature or the best solution found by the algorithms
for each instance. The stop criterion was 60 minutes. Note that as small the
PT

ARPD is, better is the quality of the solutions produced by the algorithm; and
as small the ARPT is, faster the algorithm is to produce the solutions. Hence,
the best algorithms according to ARPD and ARPT should be positioned on
the lower-left of the frontier. First, note that the standard versions of BRKGA
CE

presented large ARPD when compared to the other algorithms. For the re-
maining heuristics, one can observe a downwards tendency in the deviation and
an increase on the ARPT. ILS starts this tendency with the best ARPT but
the worst ARPD. IGS presented a good ARPD improvement concerning ILS
AC

with tiny ARPT increment. BRKGA-SK-NLS follows the tendency, improving the
ARPD at computational time cost. BRKGA-SK-LS show a considerable quality
improvement setting apart from all other algorithms. However, its ARPT is the
largest among the heuristics. CPLEX presented the second best ARDP although
it uses much more time than the heuristics, as expected. Table 1 shows the
ARPD and ARDT for each group of instances. Note that the last line brings
the general average results depicted in the Pareto frontier.

15
 ACCEPTED MANUSCRIPT

0.30 BRKGA−STD−NLS

Ave. Relative Percentage Deviation (ARPD)

BRKGA−STD−LS

0.25
ILS

T
0.20 IGS

IP
BRKGA−SK−NLS

0.15 CPLEX

CR
0.10 BRKGA−SK−LS

0 5 10 15

US
Ave. Relative Percentage Time (ARPT)

Figure 3: Pareto frontier using the average relative percentage deviations and times.
AN
To better assess the differences between the algorithms, we performed exper-
iments using different stopping times. However, the parameters used for these
experiments have the same values of those tuned for 60-minute runs. Figure 4
depicts the boxplots for the RPD and Table 2 brings the RPD average and
standard deviation of each algorithm and scenario. It is worth mentioning that,
M

to compute the RPD, we use the best solution found in each scenario not the
overall best-known solution for each instance. The rationale is to assess the
behavior of the algorithms for unknown instances for which the bounds are also
ED

unknown. Note that BRKGA-SK-LS presents the smallest deviations consistently,

although CPLEX appears to present the best results for the 5 minutes scenario.
Indeed, CPLEX shows the second best results for each other scenario followed
by BRKGA-SK-NLS. In the sequence, IGS depicts better results than ILS. The
PT

Table 1: Average relative percentage deviations (ARPD) and times (ARPT) per instance
group. The last line shows the general ARPD and ARPT for all instances and runs.
CE

Instance BRKGA-SK-LS BRKGA-SK-NLS BRKGA-STD-LS BRKGA-STD-NLS CPLEX IGS ILS

Group
ARPD ARPT ARPD ARPT ARPD ARPT ARPD ARPT ARPD ARPT ARPD ARPT ARPD ARPT

20 × 5 0.00 0.06 0.00 0.13 0.02 1.51 0.02 1.63 0.00 78.57 0.00 0.01 0.00 0.01
20 × 10 0.00 0.04 0.00 0.20 0.09 2.04 0.10 2.18 0.00 45.80 0.00 0.01 0.00 0.01
20 × 20 0.00 0.03 0.00 0.14 0.09 2.62 0.12 2.27 0.00 27.94 0.00 0.01 0.00 0.01
50 × 5 0.11 1.16 0.24 1.62 0.46 0.92 0.46 0.97 0.14 4.44 0.37 0.66 0.39 0.54
AC

50 × 10 0.28 1.31 0.52 1.60 0.86 0.24 0.86 0.29 0.58 3.78 0.61 1.47 0.77 0.99
50 × 20 0.30 1.19 0.49 1.37 0.97 0.68 0.99 0.61 0.43 3.39 0.49 1.14 0.71 0.92
100 × 5 0.06 2.92 0.17 1.13 0.19 0.45 0.20 0.49 0.07 24.35 0.17 0.08 0.19 0.08
100 × 10 0.12 2.83 0.22 1.38 0.26 0.57 0.27 0.66 0.15 9.07 0.25 0.20 0.28 0.01
100 × 20 0.21 2.43 0.35 1.75 0.44 0.58 0.45 0.73 0.37 4.30 0.39 0.30 0.47 0.03
200 × 10 0.02 2.43 0.04 1.03 0.04 0.95 0.05 1.32 0.02 7.10 0.04 0.00 0.04 0.00
200 × 20 0.03 2.60 0.05 0.81 0.05 0.90 0.06 1.37 0.03 9.97 0.05 0.01 0.05 0.01
500 × 20 0.01 2.67 0.01 1.17 0.01 0.76 0.01 0.94 0.01 4.85 0.01 0.12 0.01 0.19
General 0.10 1.64 0.17 1.03 0.29 1.02 0.30 1.12 0.15 18.63 0.20 0.33 0.24 0.23

16
 ACCEPTED MANUSCRIPT

5 min 15 min
0.5

0.4
Relative Percentage Deviation (RPD)

0.3

0.2

T
0.1

IP
0.0

30 min 60 min
0.5

CR
0.4

0.3

0.2

0.1

0.0
US
S

LS

LS

LS

EX

LS

LS
S

EX

S
−L

IG

IL

−L

−L

IG

IL
N

−N

−N

−N
PL

PL
SK

SK

D
−
SK

ST

SK

ST

D
C

C
A−

A−
ST

ST
A−

A−
A−

A−
AN
KG

KG
A−

A−
KG

KG
KG

KG
KG

KG
BR

BR
BR

BR
BR

BR
BR

BR

Algorithms

Figure 4: Distribution of RPDs for each algorithm considering different stopping criteria.
M

standard BRKGA variations presented more variability of results, reaching the

worst averages.
Since the results from each algorithm are very close to each other, we applied
ED

pairwise tests using Wilcoxon rank sum test with Hommel’s p-value adjusting.
Considering a confidence interval of 95%, Table 3 shows the p-values for each
pair of algorithms and stopping times. This table presents some interesting re-
sults. The first thing to note is the results presented by CPLEX. In the 5-min sce-
PT

nario, CPLEX produced significantly better solutions than the other algorithms,
but BRKGA-SK-LS. Indeed, we cannot confirm the statistical difference between
CE

Table 2: The average RDP (ARDP) and standard deviation for each algorithm given a max-
imum running time. The best results are in bold.

Algorithm 5 min. 15 min. 30 min. 60 min.

AC

BRKGA-SK-LS 0.09 ± 0.15 0.09 ± 0.14 0.08 ± 0.13 0.10 ± 0.15

BRKGA-SK-NLS 0.16 ± 0.23 0.16 ± 0.23 0.16 ± 0.23 0.17 ± 0.24
BRKGA-STD-LS 0.26 ± 0.33 0.27 ± 0.35 0.27 ± 0.35 0.29 ± 0.38
BRKGA-STD-NLS 0.26 ± 0.33 0.27 ± 0.35 0.27 ± 0.36 0.30 ± 0.39
CPLEX 0.07 ± 0.13 0.10 ± 0.18 0.12 ± 0.20 0.15 ± 0.23
IGS 0.17 ± 0.23 0.18 ± 0.24 0.18 ± 0.24 0.20 ± 0.25
ILS 0.19 ± 0.28 0.21 ± 0.30 0.22 ± 0.30 0.24 ± 0.32

17
 ACCEPTED MANUSCRIPT

Table 3: The p-values from pairwise Wilcoxon rank sum tests adjusted by Hommel’s proce-
dure. The tests use the RPD distributions of each algorithm, considering different stopping
times, at 95% interval of confidence. The p-values in bold indicate situations where we cannot
confirm the statistical difference between the algorithms. A minus sign (−) indicates that
Algorithm A found solutions that are significant better than solutions found by Algorithm B.
A plus sign (+) indicates the opposite case.

T
Pair Stopping time
Algorithm A / Algorithm B 5 min. 15 min. 30 min. 60 min.

IP
BRKGA-SK-LS / BRKGA-SK-NLS 0.000 − 0.000 − 0.000 − 0.000 −
BRKGA-SK-LS / BRKGA-STD-LS 0.000 − 0.000 − 0.000 − 0.000 −
BRKGA-SK-LS / BRKGA-STD-NLS 0.000 0.000 0.000 0.000

CR
− − − −
BRKGA-SK-LS / CPLEX 0.410 0.560 0.021 − 0.001 −
BRKGA-SK-LS / IGS 0.000 − 0.000 − 0.000 − 0.000 −
BRKGA-SK-LS / ILS 0.000 − 0.000 − 0.000 − 0.000 −
BRKGA-SK-NLS / BRKGA-STD-LS 0.000 − 0.000 − 0.000 − 0.000 −
BRKGA-SK-NLS / BRKGA-STD-NLS
BRKGA-SK-NLS / CPLEX
BRKGA-SK-NLS / IGS
BRKGA-SK-NLS / ILS
US
0.000
0.000
1.000
1.000
−
+
0.000
0.102
0.356
0.001
−

−
0.000
0.541
0.039
0.000
−

−
−
0.000
1.000
0.021
0.000
−

−
−
AN
BRKGA-STD-LS / BRKGA-STD-NLS 1.000 1.000 1.000 1.000
BRKGA-STD-LS / CPLEX 0.000 + 0.000 + 0.000 + 0.000 +
BRKGA-STD-LS / IGS 0.000 + 0.000 + 0.000 + 0.000 +
BRKGA-STD-LS / ILS 0.000 + 0.000 + 0.000 + 0.000 +
BRKGA-STD-NLS / CPLEX 0.000 + 0.000 + 0.000 + 0.000 +
M

BRKGA-STD-NLS / IGS 0.000 + 0.000 + 0.000 + 0.000 +

BRKGA-STD-NLS / ILS 0.000 + 0.000 + 0.000 + 0.000 +
CPLEX / IGS 0.000 − 0.020 − 0.110 0.760
CPLEX / ILS 0.000 − 0.007 − 0.039 − 0.457
ED

IGS / ILS 1.000 0.541 0.410 0.526

CPLEX and BRKGA-SK-LS for 5-min and 15-min scenarios (p-value > 0.05). How-
ever, as we allow more running time, CPLEX results diminish returns, presenting
PT

significant worse results than BRKGA-SK-LS, but significantly better results than
BRKGA standard versions, ILS, and IGS. We cannot confirm the difference be-
tween CPLEX and BRKGA-SK-NLS for 15, 30, and 60-min scenarios (p-values >
CE

0.10).
These results show that CPLEX quickly found good solutions from BSCH(n)
incumbent, but it got stuck in local minima. Allowing more running time enables
the heuristics scape from these local minima and improve the results in the long
AC

run. Despite CPLEX results, BRKGA-SK-LS produced significantly better solutions

than the other algorithms in all scenarios. The solutions of BRKGA-SK-NLS for
30 and 60-min scenarios are statistically better than those of ILS and IGS.
Also, note that we cannot confirm the difference between ILS and IGS for any
scenario (p-values > 0.41). BRKGA-STD-LS and BRKGA-STD-NLS presented the
worse results among the algorithms.

18
 ACCEPTED MANUSCRIPT

Figure 5 shows performance profiles for all algorithms where X axis shows
the time needed to reach a target solution value while the Y axis shows the
cumulative probability to reach a target solution value for the given time in the
X axis. In Figure 5a, the target values are the median solution values for each
instance. For Figure 5b, the values of the best-known solutions for each instance
are used. Note that, in both average and best solution cases, IGS (purple line

T
with inverted triangles) presented a higher probability of finding a good solution
with little running time. However, both BRKGA-SK-LS (red line with circles) and

IP
BRKGA-SK-NLS (golden line with upside triangles) found better solutions in the
long run in both average and best solution cases. BRKGA-STD-LS (dark green
line with plus signal) and BRKGA-STD-NLS (light green line with crosses) have

CR
better probability than CPLEX (blue line with diamonds) in the average case, but
they fall short for the best solutions target. ILS presented a good cumulative
probability in the first minutes of optimization but then falls short for long runs.
It is interesting to note CPLEX behavior. Contrasting the relatively small ARPD

US
depicted in Figure 4 and Table 2 with the cumulative probability distribution
shown in Figure 5b, one can argue that CPLEX finds solutions close to the best-
known solutions, but it fails to reach them.
Note that IGS and ILS spent most of their time in the local searches. Consid-
AN
ering only the best results for each algorithm, IGS iterated 23,578±63,566 times
on average, while ILS iterated 8,914±41,221. These numbers indicate that both
IGS and ILS used their diversification mechanism only a few times, leading to a
premature convergence. While this behavior helps to find good and best solu-
tions for small instances, it makes IGS and ILS to fall short in solving large and
M

more interesting instances. BRKGA variants were able to evolve much more
iterations when compare to the IGS. BRKGA-SK-LS and BRKGA-SK-NLS iterates,
on average, 192,016 ± 365,730 and 362,696 ± 636,657, respectively. The standard
ED
0.4

BRKGA−SK−LS CPLEX BRKGA−SK−LS CPLEX

BRKGA−SK−NLS IGS BRKGA−SK−NLS IGS
1.0

BRKGA−STD−LS ILS BRKGA−STD−LS ILS

BRKGA−STD−NLS BRKGA−STD−NLS
Cumulative probability

Cumulative probability
PT

0.3
0.8
0.6

0.2
CE 0.4

0.1
0.2
0.0

0.0
AC

0 600 1200 1800 2400 3000 3600 0 600 1200 1800 2400 3000 3600

Seconds to reach the median solution Seconds to reach the best solution

(a) (b)

Figure 5: Running time distributions to median and best-known solutions found. The identi-
fication marks correspond to 2% of the points plotted for each algorithm.

19
 ACCEPTED MANUSCRIPT

variants went even further: BRKGA-STD-LS with 658,133 ± 1,959,411 iterations,

and BRKGA-STD-NLS with 729, 059 ± 2,052,841 iterations, on average. Therefore,
although BRKGA variations used more time to converge, they potentially ex-
plored a much larger portion of the solution landscape, which, consequently, led
to better solutions.
Table 4 summaries the results regarding the number of good solutions found,

T
and it is partitioned into two sections. The first group of columns (2–6) shows
the performance considering 20 instances (17% of) for which solution was proven

IP
optimal (instances TA1 to TA20). There, column “# OPT” represents the num-
ber of instances for which the algorithm found an optimal solution; column
“% Opt” shows a percentage of the number of optimal solutions found; and

CR
column “% Run” shows a percentage of the number of runs on which the al-
gorithm found an optimal solution. The two columns under label “Prop. diff.”
show, respectively, the average of the proportional difference between the opti-
mal solution value and the achieved value (%), and its corresponding standard

US
deviation (σ). Note that both BRKGA-SK versions, IGS and ILS were able to find
an optimal solution in all runs, which is really remarkable for a heuristic. The
STD versions missed few optimal solutions, but the proportional deviations were
minimal.
AN
The second segment of the table shows the results for instances with unknown
optimum solutions. The description is similar to the first segment, but the
results are compared to the best solution found by the algorithms. One can note
that BRKGA-SK-LS found the best solution in almost all instances, with a clear
advantage margin from the other approaches. However, note that although the
M

other algorithms did not reach so many best solutions, the average proportional
difference is less than 0.34%. BRKGA-SK-LS found a good solution in 29% of the
runs. However, for the majority of the instances, it did find the best solution for
ED

at least one run. The similar behavior is also observed for the other approaches.
We improved the best-known upper bounds (Fernandez-Viagas & Frami-
nan, 2017) for 36 of 120 instances (30%), mostly concentrated in the large and
difficult ones. The average gap for the non-optimal instances was 6.54 ± 3.77%
PT

(average ± standard deviation) with minimum of 1.43, and maximum of 13.59.

Table 4: Algorithm performance on instances with known and unknown optimum solutions.
CE

Known Optima (20 instances) Unknown Optima (100 instances)

Algorithm Optima Prop. diff. Best Prop. diff.
# Opt % Opt % Run % σ # Best % Best % Run % σ
AC

BRKGA-SK-LS 20 100.00 100.00 – – 96 96.00 18.32 0.14 0.16

BRKGA-SK-NLS 20 100.00 100.00 – – 12 12.00 11.03 0.23 0.25
BRKGA-STD-LS 20 100.00 64.00 0.15 0.14 13 13.00 6.23 0.36 0.40
BRKGA-STD-NLS 18 90.00 64.83 0.17 0.14 11 11.00 5.77 0.37 0.40
CPLEX 20 100.00 100.00 – – 15 15.00 15.00 0.21 0.25
IGS 20 100.00 100.00 – – 13 13.00 11.10 0.27 0.25
ILS 20 100.00 100.00 – – 12 12.00 11.12 0.30 0.31

20
 ACCEPTED MANUSCRIPT

Depending on the application, such numbers represent an excellent compro-

mise between solution cost and optimization time. It is interesting to notice
that for 25 of those 36 instances (69%), at least one of the BRKGA variants
obtained results better than IGS. For the other 11 instances (30%), both algo-
rithms obtained the same results. Concerning the solution values, CPLEX-60min
and CPLEX-24h obtained the same solution values, indicating that the extra

T
time and computer power did not help in improving solutions. This fact may
also indicate that the current solutions are very close to the optimal. However,

IP
CPLEX-24h did improve the lower bounds in all cases. On average, CPLEX-60min
presented an integrality gap of 15.53 ± 28.50%, when CPLEX-24h show an aver-
age gap of 5.24 ± 3.93%. The detailed results for all algorithms, per instance, is

CR
described on Appendix B.
Table 5 shows the average percentage difference between the incumbent so-
lution value given by BSCH(n) and the value of the final solution. The average
was computed over all 30 independent runs for each algorithm (one run per in-

US
stance for CPLEX) using 60 minutes, and they are categorized by instance group
shown in the first column. We highlighted the best results in bold. The last two
lines show the average across all instances (“Average”), and the percentage of
the total runs where the algorithms could not improve the incumbent solution
AN
(“No Impr. %”). In general, the algorithms can improve the initial solution
from BSCH(n). The maximum improvement was found by BRKGA-SK-LS and is
about 3%, followed by IGS with 2.64%. In a few cases, the algorithms got stuck
with the initial solutions. In BRKGA-SK variations, such situation happens in
less than 10% of the runs, when for standard variations the no-improving rate
M

is about 23%. IGS and ILS could not improve BSCH(n) solutions in 14% and
15.30% of the cases, respectively.
ED

Table 5: Average percentage improvement concerning the warm-start solution from BSCH(n).
The last line shows the percentage of runs where no improved solution was found.

Instance BRKGA
CPLEX IGS ILS
Group SK-LS SK-NLS STD-LS STD-NLS
PT

20 × 5 -0.75 -0.75 -0.73 -0.73 -0.75 -0.75 -0.75

20 × 10 -0.68 -0.68 -0.59 -0.59 -0.68 -0.68 -0.68
20 × 20 -0.69 -0.69 -0.60 -0.57 -0.69 -0.69 -0.69
50 × 5 -0.56 -0.44 -0.22 -0.22 -0.54 -0.31 -0.29
50 10 -0.62 -0.38 -0.05 -0.05 -0.33 -0.29 -0.14
CE

×
50 × 20 -1.05 -0.86 -0.39 -0.37 -0.92 -0.85 -0.64
100 × 5 -0.19 -0.09 -0.06 -0.05 -0.18 -0.09 -0.07
100 × 10 -0.21 -0.10 -0.06 -0.06 -0.17 -0.08 -0.04
100 × 20 -0.41 -0.26 -0.17 -0.17 -0.24 -0.22 -0.15
200 × 10 -0.06 -0.04 -0.03 -0.03 -0.05 -0.03 -0.03
200 20 -0.07 -0.05 -0.05 -0.04 -0.07 -0.05 -0.05
AC

×
500 × 20 -0.01 -0.01 -0.01 -0.01 -0.01 -0.01 -0.01
Average -0.44±0.53 -0.36±0.50 -0.25±0.43 -0.24±0.42 -0.39±0.51 -0.34±0.50 -0.30±0.48
No Impr.% 5.39 9.17 23.70 23.80 0.19 14.00 15.30

21
 ACCEPTED MANUSCRIPT

7. Conclusions

In this paper, we addressed the Permutation Flowshop Scheduling Prob-

lem (PFSP) where one wants to schedule jobs aiming to minimize the total
flow time. PFSP is a fundamental problem with applications in several indus-
tries and public activities, and almost all other fields where a sequence of tasks

T
must be executed. We presented four Biased Random-Key Genetic Algorithms
(BRKGA), and an Iterated Greedy Search (IGS) procedure to solve the PFSP.

IP
We compared these approaches with the Iterated Local Search procedure (ILS),
considered to be the state-of-the-art algorithm for the PFSP.
We also applied a commercial mixed integer programming solver proving the

CR
optimality for a subset of instances, and providing the lower bounds for the oth-
ers. Another important contribution of this paper is the introduction of a new
technique for BRKGAs called shaking, where instead of reseting the population
entirely, we “shake” the elite population and reset all non-elite individuals.

US
It is challenging to find a scalable solution for PFSP. All BRKGA variations
were able to explore the solution landscape very well and found a good deal of
best solutions. Variants using the shaking strategy overcame the other algo-
rithms, presenting strong results. Therefore, instead of reseting the population
AN
entirely as the standard BRKGA does, the population shaking strategy was able
to introduce diversification changing the elite set without losing the optimiza-
tion track completely. ILS and IGS presented good results for small instances,
but they fell short in the large ones, mainly because of the little exploration
these strategies used. Therefore, BRKGA algorithms pose as viable methods
M

and strong contenders in solving permutation flowshop problems at large scale.

For future works, we would like to apply the BRKGA with shaking strategy to
other scheduling problems, to assess its efficacy for those scenarios.
ED

Acknowledgement

We thank the anonymous reviewers for providing comments that improved

PT

this paper, Victor Fernandez-Viagas and Jose M. Framinan for had kindly pro-
viding the original code of the Beam-Search Constructive Heuristic, and Rakesh
K. Sinha for comments in the reviewed version of this manusscript. This study
was financed in part by PUC–Rio, Escola Nacional de Seguros and the Coor-
CE

denação de Aperfeiçoamento de Pessoal de Nível Superior – Brasil (CAPES) –

Finance Code 001.
AC

References

Abedinnia, H., Glock, C. H., & Brill, A. (2016). New simple constructive
heuristic algorithms for minimizing total flow-time in the permutation flow-
shop scheduling problem. Computers & Operations Research, 74 , 165–174.
doi:10.1016/j.cor.2016.04.007.

22
 ACCEPTED MANUSCRIPT

Almeida, B. F., Correia, I., & Saldanha-da Gama, F. (2018). A biased random-
key genetic algorithm for the project scheduling problem with flexible re-
sources. TOP , 26 , 283–308. doi:10.1007/s11750-018-0472-9.
Andrade, C. E., Byers, S. D., Gopalakrishnan, V., Halepovic, E., Majmundar,
M., Poole, D. J., Tran, L. K., & Volinsky, C. T. (2017). Managing mas-

T
sive firmware-over-the-air updates for connected cars in cellular networks. In
Proceedings of the 2nd ACM International Workshop on Smart, Autonomous,

IP
and Connected Vehicular Systems and Services CarSys ’17 (pp. 65–72). ACM.
doi:10.1145/3131944.3131953.
Andrade, C. E., Miyazawa, F. K., & Resende, M. G. C. (2013). Evolution-

CR
ary algorithm for the k-interconnected multi-depot multi-traveling salesmen
problem. In Proceedings of the 15th Annual Conference on Genetic and Evolu-
tionary Computation GECCO’13 (pp. 463–470). New York, NY, USA: ACM.
doi:10.1145/2463372.2463434.

US
Andrade, C. E., Resende, M. G. C., Zhang, W., Sinha, R. K., Reichmann, K. C.,
Doverspike, R. D., & Miyazawa, F. K. (2015a). A biased random-key genetic
algorithm for wireless backhaul network design. Applied Soft Computing, 33 ,
AN
150–169. doi:10.1016/j.asoc.2015.04.016.
Andrade, C. E., Toso, R. F., Resende, M. G. C., & Miyazawa, F. K. (2015b).
Biased random-key genetic algorithms for the winner determination problem
in combinatorial auctions. Evolutionary Computation, 23 , 279–307. doi:10.
M

1162/EVCO_a_00138.
den Besten, M., & Stützle, T. (2001). Neighborhoods revisited: An experimen-
tal investigation into the effectiveness of variable neighborhood descent for
ED

scheduling. In MIC’2001–4th Metaheuristics International Conference (pp.

545–549).
Biajoli, F. L., Chaves, A. A., & Lorena, L. A. N. (2019). A biased random-
key genetic algorithm for the two-stage capacitated facility location problem.
PT

Expert Systems with Applications, 115 , 418–426. doi:10.1016/j.eswa.2018.

08.024.
Birattari, M., Yuan, Z., Balaprakash, P., & Stützle, T. (2010). F-Race and
CE

iterated F-Race: an overview. In Experimental methods for the analysis of

optimization algorithms (pp. 311–336). Springer Berlin Heidelberg. doi:10.
1007/978-3-642-02538-9_13.
Cabo, M., González-Velarde, J. L., Possani, E., & Ríos Solís, Y. A. (2018).
AC

Bi-objective scheduling on a restricted batching machine. Computers and

Operations Research, 100 , 201–210. doi:10.1016/j.cor.2018.07.004.
Chung, C.-S., Flynn, J., & Kirca, O. (2002). A branch and bound algorithm to
minimize the total flow time for m-machine permutation flowshop problems.
International Journal of Production Economics, 79 , 185–196. doi:10.1016/
S0925-5273(02)00234-7.

23
 ACCEPTED MANUSCRIPT

Della Croce, F., Ghirardi, M., & Tadei, R. (2002). An improved branch-and-
bound algorithm for the two machine total completion time flow shop prob-
lem. European Journal of Operational Research, 139 , 293–301. doi:10.1016/
S0377-2217(01)00374-5.
Della Croce, F., Gupta, J., & Tadei, R. (2011). A matheuristic approach for

T
the total completion time two-machines permutation flow shop problem. In
EvoCOP (pp. 38–47).

IP
Dong, X., Chen, P., & Huang, H. (2011a). An improved iterated local search
algorithm for the permutation flowshop problem with total flowtime volume

CR
122 LNEE of Lecture Notes in Electrical Engineering. URL: www.scopus.com.
Dong, X., Chen, P., Huang, H., & Nowak, M. (2013). A multi-restart iterated
local search algorithm for the permutation flow shop problem minimizing total
flow time. Comput. Oper. Res., 40 , 627–632. URL: http://dx.doi.org/10.

US
1016/j.cor.2012.08.021. doi:10.1016/j.cor.2012.08.021.
Dong, X., Huang, H., & Chen, P. (2009). An iterated local search algorithm for
the permutation flowshop problem with total flowtime criterion. Computers
AN
& Operations Research, 36 , 1664–1669. doi:10.1016/j.cor.2008.04.001.
Dong, X., Huang, H., & Chen, P. (2011b). Study on iterated local search algo-
rithm for permutation flowshop problem with total flowtime objective volume
225 CCIS of Communications in Computer and Information Science. URL:
M

www.scopus.com.
Dubois-Lacoste, J., López-Ibáñez, M., & Stützle, T. (2011). A hybrid TP+PLS
algorithm for bi-objective flow-shop scheduling problems. Computers & Op-
ED

erations Research, 38 , 1219–1236. doi:10.1016/j.cor.2010.10.008.

Fanjul-Peyro, L., & Ruiz, R. (2010). Iterated greedy local search methods
for unrelated parallel machine scheduling. European Journal of Operational
Research, 207 , 55–69. doi:10.1016/j.ejor.2010.03.030.
PT

Fernandez-Viagas, V., & Framinan, J. M. (2015). A new set of high-performing

heuristics to minimise flowtime in permutation flowshops. Computers & Op-
erations Research, 53 , 68–80. doi:10.1016/j.cor.2014.08.004.
CE

Fernandez-Viagas, V., & Framinan, J. M. (2017). A beam-search-based con-

structive heuristic for the PFSP to minimise total flowtime. Computers &
Operations Research, 81 , 167–177. doi:10.1016/j.cor.2016.12.020.
AC

Framinan, J. M., Leisten, R., & Ruiz-Usano, R. (2005). Comparison of heuristics

for flowtime minimisation in permutation flowshops. Computers & Operations
Research, 32 , 1237–1254. doi:10.1016/j.cor.2003.11.002.
Fuchigami, H. Y., & Rangel, S. (2018). A survey of case studies in production
scheduling: Analysis and perspectives. Journal of Computational Science,
25 , 425–436. doi:10.1016/j.jocs.2017.06.004.

24
 ACCEPTED MANUSCRIPT

Garey, M. R., Johnson, D. S., & Sethi, R. (1976). The complexity of flowshop
and jobshop scheduling. Mathematics of operations research, 1 , 117–129.
Gonçalves, J. F., & Resende, M. G. C. (2011). Biased random-key genetic
algorithms for combinatorial optimization. Journal of Heuristics, 17 , 487–
525. URL: 10.1007/s10732-010-9143-1. doi:10.1007/s10732-010-9143-

T
1.
Graham, R. L., Lawler, E. L., Lenstra, J. K., & Kan, A. R. (1979). Optimization

IP
and approximation in deterministic sequencing and scheduling: a survey. In
Annals of discrete mathematics (pp. 287–326). Elsevier volume 5.

CR
Grasas, A., Ramalhinho, H., Pessoa, L. S., Resende, M. G., Caballé, I., & Barba,
N. (2014). On the improvement of blood sample collection at clinical laborato-
ries. BMC Health Services Research, 14 , 12. doi:10.1186/1472-6963-14-12.

US
Hansen, P., & Mladenović, N. (2003). Variable neighborhood search. In
F. Glover, & G. Kochenberger (Eds.), Handbook of Metaheuristics (pp. 145–
184). Kluwer Academic Publishers.
Hoos, H. H., & Stützle, T. (2004). Stochastic Local Search: Foundations &
AN
Applications. Elsevier / Morgan Kaufmann.
Jarboui, B., Eddaly, M., & Siarry, P. (2009). An estimation of distribu-
tion algorithm for minimizing the total flowtime in permutation flowshop
scheduling problems. Computers & Operations Research, 36 , 2638–2646.
M

doi:10.1016/j.cor.2008.11.004.
Juan, A. A., Lourenço, H. R., Mateo, M., Luo, R., & Castella, Q. (2014).
ED

Using iterated local search for solving the flow-shop problem: Parallelization,
parametrization, and randomization issues. International Transactions in
Operational Research, 21 , 103–126.
Li, X., & Zhang, K. (2018). Single batch processing machine scheduling with
PT

two-dimensional bin packing constraints. International Journal of Production

Economics, 196 , 113–121. doi:10.1016/j.ijpe.2017.11.015.
Liu, J., & Reeves,P C. R. (2001). Constructive and composite heuristic solu-
CE

tions to the P// Ci scheduling problem. European Journal of Operational

Research, 132 , 439–452. doi:10.1016/S0377-2217(00)00137-5.
Lopes, M. C., Andrade, C. E., Queiroz, T. A., Resende, M. G. C., & Miyazawa,
AC

F. K. (2016). Heuristics for a hub location-routing problem. Networks, 68 ,

54–90. doi:10.1002/net.21685.

López-Ibáñez, M., Dubois-Lacoste, J., Cáceres, L. P., Birattari, M., & Stützle,
T. (2016). The irace package: Iterated racing for automatic algorithm con-
figuration. Operations Research Perspectives, 3 , 43–58. doi:10.1016/j.orp.
2016.09.002.

25
 ACCEPTED MANUSCRIPT

Lourenço, H. R., Martin, O. C., & Stützle, T. (2019). Iterated local search:
Framework and applications. In M. Gendreau, & J.-Y. Potvin (Eds.), Hand-
book of Metaheuristics (pp. 129–168). Cham: Springer International Pub-
lishing. URL: https://doi.org/10.1007/978-3-319-91086-4_5. doi:10.
1007/978-3-319-91086-4_5.

T
Madhushini, N., Rajendran, C., & Deepa, Y. (2009). Branch-and-bound
algorithms for scheduling in permutation flowshops to minimize the sum

IP
of weighted flowtime/sum of weighted tardiness/sum of weighted flowtime
and weighted tardiness/sum of weighted flowtime, weighted tardiness and
weighted earliness of jobs. Journal of the Operational Research Society, 60 ,

CR
991–1004. doi:10.1057/palgrave.jors.2602642.

Martí, R., Panos, P., & Resende, M. (2017). Handbook of Heuristics. Hand-
book of Heuristics. Springer International Publishing. URL: https://books.
google.com.br/books?id=21UmrgEACAAJ.

US
Moscato, P. (1989). On evolution, search, optimization, GAs and martial arts:
toward memetic algorithms. Tech. Rep. California Inst. Technol.
AN
Pan, Q.-K., & Ruiz, R. (2012). Local search methods for the flowshop schedul-
ing problem with flowtime minimization. European Journal of Operational
Research, 222 , 31–43. doi:10.1016/j.ejor.2012.04.034.
Pan, Q.-K., & Ruiz, R. (2013). A comprehensive review and evaluation of per-
M

mutation flowshop heuristics to minimize flowtime. Computers & Operations

Research, 40 , 117–128. doi:10.1016/j.cor.2012.05.018.
Pan, Q.-K., Tasgetiren, M., & Liang, Y.-C. (2007). A discrete differential
ED

evolution algorithm for the permutation flowshop scheduling problem. (pp.

126–133). URL: https://www.scopus.com/inward/record.uri?eid=2-
s2.0-34548126517&doi=10.1145%2f1276958.1276976&partnerID=40&
md5=2ea4e1beec80cb71a6cbcb3700e6e93c. doi:10.1145/1276958.1276976.
PT

Pan, Q.-K., Tasgetiren, M. F., & Liang, Y.-C. (2008). A discrete differential evo-
lution algorithm for the permutation flowshop scheduling problem. Computers
& Industrial Engineering, 55 , 795–816. doi:10.1016/j.cie.2008.03.003.
CE

Pessoa, L. S., & Andrade, C. E. (2018). Heuristics for a flowshop scheduling

problem with stepwise job objective function. European Journal of Opera-
tional Research, 266 , 950–962. doi:10.1016/j.ejor.2017.10.045.
AC

Pessoa, L. S., Santos, A. C., & Resende, M. G. (2017). A biased random-key

genetic algorithm for the tree of hubs location problem. Optimization Letters,
11 , 1371–1384. doi:10.1007/s11590-016-1082-9.
Pinedo, M. L. (2016). Scheduling: theory, algorithms, and systems. Springer.

26
 ACCEPTED MANUSCRIPT

Prasetyo, H., Fauza, G., Amer, Y., & Lee, S. H. (2015). Survey on applications
of biased-random key genetic algorithms for solving optimization problems.
In Industrial Engineering and Engineering Management (IEEM), 2015 IEEE
International Conference on (pp. 863–870). IEEE.
Prasetyo, H., Putri, A. L., & Fauza, G. (2018). Biased random key genetic al-

T
gorithm design with multiple populations to solve capacitated vehicle routing
problem with time windows. In AIP Conference Proceedings (p. 020052). AIP

IP
Publishing volume 1977.
Reeves, C. R. (1995). A genetic algorithm for flowshop sequencing. Computers

CR
& operations research, 22 , 5–13. doi:10.1016/0305-0548(93)E0014-K.

Resende, M. G. C. (2012). Biased random-key genetic algorithms with appli-

cations in telecommunications. TOP , 20 , 130–153. URL: 10.1007/s11750-
011-0176-x. doi:10.1007/s11750-011-0176-x.

US
Ribas, I., Companys, R., & Tort-Martorell, X. (2013). An efficient iterated
local search algorithm for the total tardiness blocking flow shop problem.
International Journal of Production Research, 51 , 5238–5252.
AN
Rossi, F. L., Nagano, M. S., & Sagawa, J. K. (2017). An effective constructive
heuristic for permutation flow shop scheduling problem with total flow time
criterion. The International Journal of Advanced Manufacturing Technology,
90 , 93–107. doi:10.1007/s00170-016-9347-0.
M

Ruiz, R., & Stützle, T. (2007). A simple and effective iterated greedy algo-
rithm for the permutation flowshop scheduling problem. European Journal of
Operational Research, 177 , 2033–2049. doi:10.1016/j.ejor.2005.12.009.
ED

Ruiz, R., & Stützle, T. (2008). An iterated greedy heuristic for the sequence
dependent setup times flowshop problem with makespan and weighted tardi-
ness objectives. European Journal of Operational Research, 187 , 1143–1159.
doi:10.1016/j.ejor.2006.07.029.
PT

Stafford, E. (1988). On the development of a mixed-integer linear programming

model for the flowshop sequencing problem. Journal of the Opera-
tional Research Society, 39 , 1163–1174. URL: https://www.scopus.com/
CE

inward/record.uri?eid=2-s2.0-84974856861&doi=10.1057%2fjors.
1988.193&partnerID=40&md5=4ca71f80a7e97588b090c4ecb6c95854.
doi:10.1057/jors.1988.193.
AC

Stützle, T. (1998). Applying iterated local search to the permutation flow shop
problem. Technical Report Darmstadt, University of Technology.
Stützle, T., & Ruiz, R. (2017). Iterated greedy. In R. Martí, P. Panos, &
M. G. C. Resende (Eds.), Handbook of Heuristics (pp. 1–31). Cham: Springer
International Publishing. URL: 10.1007/978-3-319-07153-4_10-1. doi:10.
1007/978-3-319-07153-4_10-1.

27
 ACCEPTED MANUSCRIPT

Taillard, E. (1991). Robust taboo search for the quadratic assignment problem.
Parallel Computing, 17 , 443–455. doi:10.1016/S0167-8191(05)80147-4.
Tasgetiren, M. F., Kizilay, D., Pan, Q.-K., & Suganthan, P. (2017). Iter-
ated greedy algorithms for the blocking flowshop scheduling problem with
makespan criterion. Computers and Operations Research, 77 , 111 – 126.

T
doi:10.1016/j.cor.2016.07.002.
Tasgetiren, M. F., Pan, Q.-K., Suganthan, P. N., & Chen, A. H. (2011). A

IP
discrete artificial bee colony algorithm for the total flowtime minimization in
permutation flow shops. Information sciences, 181 , 3459–3475. doi:10.1016/

CR
j.ins.2011.04.018.

Tseng, L.-Y., & Lin, Y.-T. (2009). A hybrid genetic local search algorithm for the
permutation flowshop scheduling problem. European Journal of Operational
Research, 198 , 84–92. doi:10.1016/j.ejor.2008.08.023.
Tseng, L.-Y., & Lin, Y.-T. (2010).
US
A genetic local search algorithm
for minimizing total flowtime in the permutation flowshop schedul-
ing problem. International Journal of Production Economics, 127 ,
AN
121–128. URL: https://www.sciencedirect.com/science/article/pii/
S0925527310001830. doi:10.1016/j.ijpe.2010.05.003.
van de Velde, S. L. (1991). Minimizing the sum of the job completion times in
the two-machine flow shop by lagrangian relaxation. Ann. Oper. Res., 26 ,
M

257–268. URL: http://dl.acm.org/citation.cfm?id=115185.115197.

Wagner, H. M. (1959). An integer linear-programming model for machine
scheduling. Naval Research Logistics Quarterly, 6 , 131–140.
ED

Xu, J., Wu, C.-C., Yin, Y., & Lin, W.-C. (2017). An iterated local search for
the multi-objective permutation flowshop scheduling problem with sequence-
dependent setup times. Applied Soft Computing, 52 , 39–47.
PT

Xu, X., Xu, Z., & Gu, X. (2011). An asynchronous genetic local search algo-
rithm for the permutation flowshop scheduling problem with total flowtime
minimization. Expert Syst. Appl., 38 , 7970–7979. URL: http://dx.doi.org/
10.1016/j.eswa.2010.12.075. doi:10.1016/j.eswa.2010.12.075.
CE

Zhang, Y., Li, X., & Wang, Q. (2009). Hybrid genetic algorithm for permutation
flowshop scheduling problems with total flowtime minimization. European
Journal of Operational Research, 196 , 869–876. doi:10.1016/j.ejor.2008.
AC

04.033.

28
 ACCEPTED MANUSCRIPT

Appendix A. Parameter settings

BRKGA has a larger number of control parameters which makes difficult to

perform a full factorial design. Therefore, we used the iterated racing (Birattari
et al. (2010)) to tune the parameters. This method consists in sampling con-
figurations from a particular distribution, evaluating them using a statistical,

T
and refining the sampling distribution with repeated applications of F-Race.
We used the irace package (López-Ibáñez et al. (2016)), implemented in R, for

IP
such task.
We used a budget of 3,000 experiments in the tuning procedure over the
30 instances selected randomly. Each experiment was limited to one hour. To

CR
tune the BRKGA parameters, we used the following ranges: population size
multiplier px ∈ [1, 10]; percentage of elite individuals pe ∈ [10, 50]; percentage
of mutants introduced at each generation pm ∈ [10, 50]; probability on biased
crossover ρ ∈ [0.5, 0.8]; and number of iterations without improvement in the

US
best solution until population reset/shake (R) in [200, 2000]. The intensity of
the shaking (λ) and the local search iterations (LS) are in [1, 50]. Both R? and
R?? multipliers are in [2, 15]. Table A.1 shows the suggestion made by irace.
We picked the values of the first line of each table, rounded up to two digits for
AN
real values, and round to the next multiple of 10, in the case of integer values.

Table A.1: irace results for BRKGA variants.

px % pe % pm ρ LS λ R mR? mR??
M

9 0.30 0.23 0.55 10 10 1000 10 5

10 0.26 0.31 0.79 12 9 972 12 7
7 0.18 0.29 0.72 9 6 1208 11 4
ED

10 0.16 0.36 0.67 28 9 1309 6 5

8 0.20 0.30 0.74 12 4 1210 7 6
PT
CE
AC

29
 ACCEPTED MANUSCRIPT

Appendix B. Detailed results

Table B.2 depicts the full set of solutions compared to the best results known.
The first column shows the name of the instance. In bold, we highlighted
those instances that the approaches discussed in this paper improved from the
results reported in Fernandez-Viagas & Framinan (2017). The second column

T
depicts the best lower bounds found by CPLEX-24h, followed by the best-known
solution and the percentual gap between those solutions and the lower bounds,

IP
on the third and fourth columns, respectively. A dash (–) in “Lower bound” and
“Gap%” indicates that the optimal solution was found. The following columns
describe the percentage deviation of the best solution found by the algorithm

CR
compared to the best solution known, where BSCH(n) are the deviations from the
results reported on Fernandez-Viagas & Framinan (2017). A star (?) indicates
that the approach reached the solution. This table is also available in the CSV
format in the supplementary materials for download.

found for each algorithm.

Lower Best
US
Table B.2: Best results for each instance, and the percentage deviation of the best solution

BRKGA CPLEX CPLEX

AN
Instance Gap% BSCH(n) IGS ILS
bound solution SK-LS SK-NLS STD-LS STD-NLS 1h 24h

TA1 – 14033 – ? ? ? ? 0.18 ? ? ? ?

TA2 – 15151 – ? ? ? ? 1.76 ? ? ? ?
TA3 – 13301 – ? ? ? ? 2.71 ? ? ? ?
TA4 – 15447 – ? ? ? ? 0.84 ? ? ? ?
TA5 – 13529 – 0.72
M

? ? ? ? ? ? ? ?
TA6 – 13123 – ? ? ? ? 0.21 ? ? ? ?
TA7 – 13548 – ? ? ? ? ? ? ? ? ?
TA8 – 13948 – ? ? ? ? 0.62 ? ? ? ?
TA9 – 14295 – ? ? ? ? 0.38 ? ? ? ?
TA10 – 12943 – ? ? ? ? 0.22 ? ? ? ?
ED

TA11 – 20911 – ? ? ? ? 0.22 ? ? ? ?

TA12 – 22440 – ? ? ? ? 0.61 ? ? ? ?
TA13 – 19833 – ? ? ? ? ? ? ? ? ?
TA14 – 18710 – ? ? ? 0.07 0.44 ? ? ? ?
TA15 – 18641 – ? ? ? ? 0.52 ? ? ? ?
TA16 – 19245 – ? ? ? 0.02 1.81 ? ? ? ?
PT

TA17 – 18363 – ? ? ? ? 0.93 ? ? ? ?

TA18 – 20241 – ? ? ? ? 1.82 ? ? ? ?
TA19 – 20330 – ? ? ? ? 0.48 ? ? ? ?
TA20 – 21320 – ? ? ? ? 0.08 ? ? ? ?
TA21 32409.65 33623 3.74 ? ? ? ? 1.81 ? ? ? ?
TA22 30439.51 31587 3.77 ? ? ? ? 0.33 ? ? ? ?
CE

TA23 32695.65 33920 3.74 ? ? ? ? ? ? ? ? ?

TA24 30540.21 31661 3.67 ? ? ? ? 0.24 ? ? ? ?
TA25 32684.98 34557 5.73 ? ? ? ? 0.97 ? ? ? ?
TA26 32040.23 32564 1.63 ? ? ? ? 0.12 ? ? ? ?
TA27 32221.97 32922 2.17 ? ? ? ? 0.99 ? ? ? ?
TA28 31538.05 32412 2.77 ? ? ? ? 0.94 ? ? ? ?
AC

TA29 31926.00 33600 5.24 ? ? ? ? 0.76 ? ? ? ?

TA30 31477.20 32262 2.49 ? ? ? ? 0.83 ? ? ? ?
TA31 63425.08 64802 2.17 ? 0.11 0.41 0.45 0.51 0.06 0.06 0.11 0.17
TA32 66309.69 68051 2.63 ? 0.12 0.44 0.47 1.43 0.08 0.08 0.20 0.32
TA33 61567.20 63162 2.59 ? 0.23 0.44 0.44 0.44 0.29 0.29 0.44 0.34
TA34 66458.52 68226 2.66 ? 0.29 0.54 0.47 0.59 0.30 0.30 0.22 0.47
TA35 67629.56 69351 2.55 ? 0.11 0.25 0.16 0.75 0.13 0.13 0.30 0.22
TA36 65901.21 66841 1.43 ? 0.22 0.25 0.25 0.34 0.19 0.19 0.18 0.28
TA37 64867.45 66253 2.14 ? 0.04 0.57 0.66 1.15 0.01 0.01 0.12 0.19
Continue on next page. . .

30
 ACCEPTED MANUSCRIPT

Table B.2 (continued).

Lower Best BRKGA CPLEX CPLEX

Instance Gap% BSCH(n) IGS ILS
bound solution SK-LS SK-NLS STD-LS STD-NLS 1h 24h

TA38 62908.35 64332 2.26 0.04 0.06 0.07 0.07 0.49 0.04 0.04 0.12 0.25
TA39 61482.92 62981 2.44 ? 0.26 0.48 0.48 0.64 0.26 0.26 0.25 0.27
TA40 67123.46 68770 2.45 0.09 0.21 0.34 0.35 0.63 0.18 0.18 0.49 0.52

T
TA41 81457.49 87114 6.94 0.09 0.58 0.99 0.99 1.09 0.87 0.87 0.53 0.61
TA42 77769.23 82820 6.49 ? 0.16 1.01 1.01 1.01 0.91 0.91 0.50 0.39
TA43 75553.61 79931 5.79 0.17 0.38 0.90 0.90 0.90 0.89 0.89 0.53 0.50

IP
TA44 80360.36 86446 7.57 0.15 0.42 0.73 0.73 0.73 0.72 0.72 0.46 0.58
TA45 80971.24 86377 6.68 0.01 0.04 0.99 0.99 0.99 0.82 0.82 0.40 0.56
TA46 80746.51 86587 7.23 0.06 0.16 0.67 0.67 0.68 0.34 0.34 0.34 0.46
TA47 83445.20 88750 6.36 ? 0.09 0.87 0.87 1.12 0.54 0.54 0.49 0.72

CR
TA48 81293.29 86727 6.68 0.09 0.24 1.06 1.39 1.55 0.72 0.72 0.58 0.82
TA49 79599.33 85441 7.34 0.11 0.41 0.47 0.47 0.74 0.30 0.30 0.57 0.52
TA50 82597.34 87998 6.54 ? 0.34 0.91 0.91 0.95 0.34 0.34 0.40 0.59
TA51 110779.46 125831 13.59 ? 0.02 0.27 0.27 0.27 0.26 0.26 0.27 0.27
TA52 108234.50 119247 10.17 ? 0.15 0.52 0.52 0.57 0.44 0.44 0.08 0.25
TA53 105693.50 116459 10.19 ? 0.32 0.90 1.20 1.60 0.19 0.19 0.26 0.65
TA54 108682.29 120261 10.65 0.41 0.58 0.97 1.10 2.38 1.06 1.06 0.66 0.83
TA55
TA56
TA57
TA58
TA59
TA60
104864.01
106614.11
109049.19
108754.38
107458.13
110327.42
118184
120586
122880
122489
121872
123954
12.70
13.11
12.68
12.63
13.41
12.35
US?
0.11
0.17
0.03
0.11
0.17
0.38
0.31
0.31
0.27
0.23
0.37
0.86
1.06
0.83
1.31
0.54
0.97
1.03
1.06
0.83
1.53
0.57
1.06
2.22
1.07
1.63
2.42
1.26
1.24
0.62
0.43
0.54
0.67
0.42
0.71
0.62
0.43
0.54
0.67
0.42
0.71
0.50
0.42
0.45
0.29
0.36
0.49
0.67
0.46
0.74
0.60
0.54
0.59
AN
TA61 248085.57 253167 2.05 ? ? ? ? 0.03 ? ? 0.01 0.03
TA62 237414.97 241925 1.90 ? 0.07 0.07 0.07 0.07 0.05 0.05 0.07 0.07
TA63 233236.38 237832 1.97 0.17 0.38 0.40 0.40 0.41 0.32 0.32 0.39 0.40
TA64 223001.21 227522 2.03 ? 0.22 0.28 0.28 0.30 0.07 0.07 0.25 0.29
TA65 235251.37 240301 2.15 0.18 0.19 0.19 0.19 0.24 0.19 0.19 0.19 0.22
TA66 227625.84 232342 2.07 0.08 0.16 0.16 0.20 0.35 0.21 0.21 0.29 0.29
M

TA67 235913.31 240366 1.89 0.18 0.43 0.96 1.06 1.57 0.28 0.28 0.62 0.85
TA68 225998.50 230945 2.19 0.03 0.03 0.03 0.03 0.03 0.03 0.03 0.03 0.03
TA69 242577.86 247526 2.04 ? 0.06 0.06 0.06 0.06 0.06 0.06 0.06 0.06
TA70 238147.21 242933 2.01 0.10 0.26 0.26 0.26 0.26 0.22 0.22 0.26 0.24
TA71 283005.20 298385 5.43 0.41 0.77 1.21 1.23 1.36 1.09 1.09 0.87 1.14
TA72 259208.82 273674 5.58 ? 0.05 0.05 0.05 0.06 ? ? 0.05 0.05
ED

TA73 271813.41 288114 6.00 0.19 0.31 0.34 0.34 0.36 0.31 0.31 0.34 0.34
TA74 283451.60 301044 6.21 0.30 0.34 0.36 0.36 0.36 0.34 0.34 0.36 0.36
TA75 267821.37 284233 6.13 ? 0.02 0.02 0.02 0.02 ? ? 0.02 0.02
TA76 253800.63 269686 6.26 0.58 0.66 0.80 0.80 0.81 0.66 0.66 0.78 0.81
TA77 264572.14 279463 5.63 0.35 0.79 0.70 0.90 1.65 0.68 0.68 0.92 1.08
TA78 273381.62 290908 6.41 0.15 0.22 0.24 0.24 0.24 0.22 0.22 0.24 0.24
PT

TA79 286021.91 301970 5.58 0.31 0.43 0.43 0.45 0.52 0.42 0.42 0.45 0.45
TA80 276762.89 291283 5.25 0.16 0.19 0.34 0.34 0.35 0.21 0.21 0.29 0.34
TA81 322937.70 365463 13.17 0.83 1.15 1.28 1.30 1.36 1.23 1.23 1.17 1.30
TA82 331942.97 372449 12.20 0.16 0.29 0.31 0.31 0.31 0.29 0.29 0.31 0.31
TA83 329667.79 370027 12.24 0.63 0.79 1.08 1.09 1.20 0.96 0.96 1.05 1.12
TA84 332064.91 372393 12.14 0.66 0.98 1.12 1.08 1.19 1.11 1.11 0.97 1.12
CE

TA85 328577.91 368915 12.28 0.36 0.39 0.44 0.44 0.44 0.43 0.43 0.41 0.44
TA86 327315.43 370908 13.32 0.59 0.86 0.95 0.96 1.25 0.96 0.96 0.98 1.04
TA87 331099.33 373408 12.78 0.42 0.57 0.68 0.68 0.68 0.68 0.68 0.63 0.67
TA88 339644.64 384525 13.21 0.36 0.57 0.61 0.60 1.53 0.65 0.65 0.65 0.65
TA89 330702.73 374423 13.22 0.64 1.08 1.55 1.53 1.89 1.63 1.63 0.99 1.50
TA90 339754.76 379296 11.64 0.52 0.81 1.10 1.09 1.52 0.95 0.95 0.66 1.19
AC

TA91 998469.90 1041023 4.26 ? 0.04 0.05 0.05 0.14 0.03 0.03 0.05 0.06
TA92 988625.43 1028828 4.07 ? ? 0.01 0.01 0.01 0.01 0.01 0.01 0.01
TA93 1001693.83 1042357 4.06 ? 0.07 0.07 0.08 0.11 0.06 0.06 0.07 0.08
TA94 984205.61 1025564 4.20 ? 0.06 0.05 0.06 0.33 0.01 0.01 0.02 0.12
TA95 985397.30 1028963 4.42 ? 0.03 0.04 0.04 0.04 0.03 0.03 0.04 0.02
TA96 957895.26 998340 4.22 ? 0.05 0.06 0.06 0.09 0.06 0.06 0.06 0.06
TA97 1000477.26 1042570 4.21 ? ? ? ? 0.01 ? ? ? ?
TA98 1000307.85 1035915 3.56 ? 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01
TA99 975347.06 1015280 4.09 ? 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01
Continue on next page. . .

31
 ACCEPTED MANUSCRIPT

Table B.2 (continued).

Lower Best BRKGA CPLEX CPLEX

Instance Gap% BSCH(n) IGS ILS
bound solution SK-LS SK-NLS STD-LS STD-NLS 1h 24h

TA100 978537.50 1021865 4.43 ? 0.03 0.04 0.04 0.04 0.03 0.03 0.04 0.04
TA101 1102941.86 1219341 10.55 ? 0.13 0.20 0.19 0.37 0.02 0.02 0.15 0.18
TA102 1115381.71 1233161 10.56 ? 0.04 0.06 0.06 0.07 0.06 0.06 0.03 0.06

T
TA103 1130764.69 1259605 11.39 ? ? 0.01 0.01 0.02 0.01 0.01 0.01 0.01
TA104 1119826.64 1228027 9.66 ? ? ? ? ? ? ? ? ?
TA105 1105327.45 1215854 1.00 ? 0.02 0.09 0.08 0.33 0.08 0.08 0.06 0.06

IP
TA106 1108726.36 1218757 9.92 ? 0.04 0.04 0.04 0.06 0.04 0.04 0.04 0.03
TA107 1123430.72 1234330 9.87 ? ? ? ? ? ? ? ? ?
TA108 1123984.67 1240105 10.33 ? 0.02 0.02 0.02 0.04 0.02 0.02 0.02 0.02
TA109 1098798.57 1220058 11.04 ? ? 0.01 0.01 0.07 0.01 0.01 0.01 ?

CR
TA110 1127270.58 1235113 9.57 ? ? 0.02 0.02 0.03 0.01 0.01 0.02 0.02
TA111 6191253.65 6558109 5.93 ? ? ? ? 0.01 ? ? ? ?
TA112 6301558.79 6679339 6.00 ? ? ? ? ? ? ? ? ?
TA113 6261525.71 6624644 5.80 ? ? ? ? ? ? ? ? ?
TA114 6281508.02 6646006 5.80 ? ? ? 0.01 0.06 ? ? ? ?
TA115 6240001.85 6587110 5.56 ? 0.01 ? 0.03 0.04 0.01 0.01 0.01 0.01
TA116 6269214.53 6603291 5.32 0.01 0.01 0.01 0.01 0.02 0.01 0.01 0.01

US
?
TA117 6223303.96 6602685 5.67 ? ? ? ? ? ? ? ? ?
TA118 6277413.77 6629065 5.60 ? ? ? ? ? ? ? ? ?
TA119 6203828.97 6587638 6.19 ? ? ? 0.01 0.02 0.01 0.01 0.01 ?
TA120 6272934.31 6623849 5.59 ? ? ? 0.02 0.04 0.01 0.01 ? 0.01
AN
M
ED
PT
CE
AC

32
 ACCEPTED MANUSCRIPT

Minimizing flowtime in a flowshop scheduling problem

with a biased random-key genetic algorithm

Carlos E. Andradea , Thuener Silvab , Luciana S. Pessoab,∗

T
a AT&T Labs Research,

200 South Laurel Avenue, Middletown, NJ 07748 USA

b Department of Industrial Engineering, PUC-Rio,

IP
Rua Marquês de São Vicente, 225, Gávea - 22453-900 Rio de Janeiro, RJ, Brazil.

CR
Credit Author Statement

US
1. Carlos E. Andrade: Conceptualization, Methodology, Software, Valida-
tion, Formal Analysis, Investigation, Writing – Original Draft, Review &
Editing, Visualization, Supervision
AN
2. Thuener Silva: Conceptualization, Methodology, Software, Validation, In-
vestigation, Writing – Original Draft, Review & Editing
3. Luciana S. Pessoa: Conceptualization, Methodology, Software, Validation,
Investigation, Writing – Original Draft, Review & Editing, Supervision
M
ED
PT
CE
AC

∗ Corresponding
author. Tel: +55 (21) 3527-2180
Email addresses: [email protected] (Carlos E. Andrade),
[email protected] (Thuener Silva), [email protected] (Luciana S. Pessoa)

Preprint submitted to Expert Systems with Applications February 5, 2019