Selecting Genetic Operators to Maximise Preference Satisfaction in a
Workforce Scheduling and Routing Problem
Haneen Algethami
1
, Dario Landa-Silva
1
and Anna Mart
´
ınez-Gavara
2
1
School of Computer Science, ASAP Research Group, The University of Nottingham, Nottingham, U.K.
2
Estad
´
ıstica y Investigaci
´
on Operativa, Universidad de Valencia, Valencia, Spain
Keywords:
Genetic Operators, Constraints Satisfaction, Scheduling and Routing Problem, Home Health Care.
Abstract:
The Workforce Scheduling and Routing Problem (WSRP) is a combinatorial optimisation problem that in-
volves scheduling and routing of workforce. Tackling this type of problem often requires handling a consider-
able number of requirements, including customers and workers preferences while minimising both operational
costs and travelling distance. This study seeks to determine effective combinations of genetic operators com-
bined with heuristics that help to find good solutions for this constrained combinatorial optimisation problem.
In particular, it aims to identify the best set of operators that help to maximise customers and workers pref-
erences satisfaction. This paper advances the understanding of how to effectively employ different operators
within two variants of genetic algorithms to tackle WSRPs. To tackle infeasibility, an initialisation heuristic
is used to generate a conflict-free initial plan and a repair heuristic is used to ensure the satisfaction of con-
straints. Experiments are conducted using three sets of real-world Home Health Care (HHC) planning problem
instances.
1 INTRODUCTION
The workforce scheduling and routing problem
(WSRP) involves scheduling and routing of work-
force to visit customers at different locations in order
to complete a set of tasks or activities. The problem
arises in real-world scenarios, such as home health
care, security guard routing and rostering, mainte-
nance personnel scheduling among other worker al-
location problems (Castillo-Salazar et al., 2016).
The WRSP is a combination of two combinato-
rial optimisation problems, personnel scheduling and
routing, which are known to be NP-hard problems
(Lenstra and Kan, 1981). The scheduling aspect al-
locates workforce to customers in order to fulfil work
demands as well as satisfying their preferences. The
routing aspect requires generating routes for workers
to visit customers across various locations and within
given time windows. Researchers have reported that
real-world instances of the WSRP are large and diffi-
cult to solve (Mısır et al., 2015; Castillo-Salazar et al.,
2016). Hence, there is a need to develop efficient al-
gorithms to solve this type of problem.
Preliminary work evaluated a set of genetic oper-
ators within a steady-state genetic algorithm applied
to a few instances of a real-world home health care
(HHC) problem (Algethami and Landa-Silva, 2015).
That work produced evidence that some operators ob-
tain better results than others when used within the
steady-state genetic algorithm for WSRP scenarios.
The present paper conducts a more comprehensive
study in order to achieve a deeper understanding of
the behaviour and performance of the various genetic
operators when applied to the WSRP.
The aim in this paper is to identify the best combi-
nation of genetic operators for each WSRP instance,
in order to maximise the satisfaction of customers and
workers preference constraints. Twelve genetic oper-
ators are considered in different combinations. The
two genetic algorithms (GA) applied in this study are
a steady state GA and a generational GA.
2 RELATED WORK
The routing component of the WSRP is related to
variants of the classical vehicle routing problem
(VRP) and in particular to the vehicle routing problem
with time windows (VRPTW) (Toth and Vigo, 2014).
Many GA applications have been used to tackle VRP
including hybrid approaches incorporating heuristic
methods and problem-specific operators to avoid pre-
416
Algethami H., Landa-Silva D. and Martà nez-Gavara A.
Selecting Genetic Operators to Maximise Preference Satisfaction in a Workforce Scheduling and Routing Problem.
DOI: 10.5220/0006203304160423
In Proceedings of the 6th International Conference on Operations Research and Enterprise Systems (ICORES 2017), pages 416-423
ISBN: 978-989-758-218-9
Copyright
c
2017 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
mature convergence of the GA (Prins, 2004; Chang
and Chen, 2007). In addition, the study by (Prins,
2004) suggested the best genetic components for an
efficient GA to tackle VRP problems. According to
that study, order crossover (OX) is the most suitable
operator for VRP-like problems.
Genetic algorithms (GAs) have been effective in
providing good solutions relatively quickly, partic-
ularly when addressing real-world scheduling prob-
lems (Kotecha et al., 2004; Aickelin and Dowsland,
2004). It has been argued that this success is a result
of the GAs capability to solve different segments of a
problem simultaneously (Rothlauf, 2003).
A number of studies have applied GAs to real-
world problems where scheduling and routing are
combined. Examples include (Cowling et al., 2006;
Mutingi and Mbohwa, 2014). In those works, the fo-
cus has been on algorithm design in order to obtain
good solutions. However, well-known operators and
repair heuristics were used to deal with infeasibility
issues. Far too little attention has been given to in-
troducing new genetic operators to reduce the overall
cost. To date, the impact of selecting compatible op-
erators for tackling WSRP instances has not yet been
investigated.
The focus of this paper is not to produce the
most competitive genetic algorithm, but to advance
the understanding of how different combinations of
genetic operators perform when tackling preference
constraints in instances of the WSRP. The problem
instances used in this study were also tackled in
(Laesanklang and Landa-Silva, 2016; Pinheiro et al.,
2016). This work seeks to identify effective combina-
tions of genetic operators for tackling preference con-
straints in WSRP instances to then inform the design
of competitive GAs to tackle this difficult problem.
3 PROBLEM DESCRIPTION
A WSRP solution is a daily plan of visits, i.e. a set
of workers W = {w
1
,w
2
,... ,w
|W |
} assigned to per-
form a set of tasks T = {t
1
,t
2
,. .. ,t
|T |
} for customers
at different locations. The assignment of a worker to
travel to a customer location in order to perform a
task is called a visit. Thus, x
w
i, j
is a binary decision
variable that indicates if a path connects two nodes
(visit i and visit j) or not. The assignment x
w
i, j
= 1
means that worker w travels from visit i to visit j,
thus w makes both visits. For visit j, if x
w
i, j
= 1 then
y
j
r
j
1 where r
j
is the number of workers required
for visit j and y
j
is an integer decision variable indi-
cating the number of unsatisfied assignments, hence
wW
iT
jT
x
w
i, j
+ y
j
= r
j
.
This paper tackles a home health care (HHC) plan-
ning problem, in which workers are nurses, doctors,
health carers, etc., and customers are patients receiv-
ing health care at their home. Several features have
been identified as important in solutions to HHC sce-
narios, such as distance travelled and customers’ and
workers’ requirements and preferences (Mankowska
et al., 2014). A good quality plan for an HHC plan-
ning problem should have a low operational cost as
well as assigning workforce. Thus, a solution requires
all tasks to be assigned while satisfying some require-
ments. That is, assigning tasks according to workers’
skills and avoiding time conflicts in respect to work-
ers’ time and area availability. A time conflict oc-
curs when a worker is assigned to visits overlapping in
time. Additional preferences include workers prefer-
ring to work in certain geographical areas, customers
requiring workers with special skills or preferring cer-
tain workers to perform a task.
Table 1 lists WSRP objectives and constraints
considered here. See (Laesanklang and Landa-Silva,
2016) for details of the MIP model of this WSRP.
Note that in (Laesanklang and Landa-Silva, 2016),
unassigned visits constraint is considered as a soft
constraint. However, here this is a hard constraint,
hence all visits must be assigned. Additionally, in this
paper, time-conflict constraint is introduced, while
the study by (Laesanklang and Landa-Silva, 2016)
avoided conflicts. The decomposition method divided
a problem into sub-problems, then available workers
were updated so that no conflicting assignments ex-
ists.
A solution S is a set of assignments to workers in
order to make visits. The objective function includes
the operational cost and the penalty cost. The oper-
ational cost is the accumulated cost d
i, j
+ p
w
j
, where
d
i, j
is the distance travelled between visit i to visit j
and p
w
j
is the cost of assigning worker w to visit j, i.e.
wages plus journey costs for all workers, as calculated
by the service provider in our HHC scenarios.
The penalty cost is the accumulated penalty for
the violations on constraints. An assignment can be
written as a tuple x
w
i, j
,y
j
,a
w
j
,ψ
w
j
,θ
w
j
,τ
w
j
. Where, a
w
j
is
the arrival time of a worker w to the location of a visit
j. The assignment is also composed of binary deci-
sion variables indicating an assignment of worker w to
visit j with violations on area availability (ψ
w
j
), time
availability (θ
w
j
) and conflicting assignments (τ
w
j
).
The non-satisfaction of preferences is also in-
cluded in the penalty cost. There are three types
of preferences including preferred worker-customer
pairing, worker’s preferred region and customer’s pre-
ferred skills. There is a degree of satisfaction for these
preferences when assigning a worker w to a task j and
Selecting Genetic Operators to Maximise Preference Satisfaction in a Workforce Scheduling and Routing Problem
417
Table 1: Objectives and constraints in the WSRP.
Objectives Hard constraints Soft constraints
Minimise the operational cost Assign all visits Respect workers area availability
Minimise the penalty cost Respect visit time (No time-conflicts) Respect workers time availability
Respect max working time per week Assign preferred workers to visits
Respect min working time per week Assign preferred workers with a specific skill
Assign qualified workforce Assign workers to preferred areas
is given by ρ
w
j
which has a value that ranges between
0,3
. For each assignment, the satisfaction value for
each preference ranges between
0,1
, from not sat-
isfied to satisfied. The satisfaction level is reverted to
a penalty by subtracting it from the full satisfaction
score, which is 3r
j
for a visit j.
f (S) = λ
1
wW
iT
jT
(d
i, j
+ p
w
j
)x
w
i, j
+ λ
2
wW
iT
jT
(3r
j
ρ
w
j
)x
w
i, j
+ λ
3
wW
jT
(ψ
w
j
+ θ
w
j
)
+ λ
4
jT
y
j
+ λ
5
wW
jT
τ
w
j
(1)
The best solution should have: the least operational
cost and the least penalty cost. A weighted sum
is proposed to combine the objectives into a single
scalar value (Pinheiro et al., 2016; Algethami et al.,
2016). The objective function is written as in equation
(1), where weights λ
1
,. .. ,λ
5
are defined to establish
priority between objectives (more about the weights
used here later in the paper).
4 ALGORITHMS AND
OPERATORS
This paper investigates the behaviour of various
genetic operators when tackling instances of the
WSRP by analysing their performance within rela-
tively straightforward implementations of two varia-
tions of GAs. A simple solution representation al-
lows direct implementation of genetic operators on
the genotype (Rothlauf, 2003). Thus, a direct repre-
sentation scheme is used for chromosome encoding.
A vector of integers of length equal to the number of
visits, |T|, represents a one-day plan. Hence, all visits
are assigned. Indexes of the chromosome correspond
to the set of tasks T , for example the i
th
gene in the
chromosome means the corresponding visit with in-
dex i T . In order to increase the possibility of ob-
taining a feasible initial plan, indexes in the chromo-
some are associated to visits in non-decreasing order
of visit start time. In this way, index 1 is for the visit
with the earliest start time and index |T| is for the visit
with the latest start time. For each visit in the vector,
a worker w is selected at random from W . A worker w
may undertake more than one visit, and some workers
may not be utilised as a part of a particular one-day
plan.
4.1 Genetic Algorithms (GA)
Two GAs are implemented in this study: a steady-
state genetic algorithm (SSGA) and a generational ge-
netic algorithm (GenGA) (Vavak and Fogarty, 1996).
Such relatively simple algorithms were selected in
order to analyse the emergent behaviour and perfor-
mance of the operators on a straightforward GA im-
plementation.
Initially, a time conflict reduction (TCR) opera-
tor is applied to each individual in the initial popu-
lation in order to reassign visits and reduce the num-
ber of time conflicts. After that, the evolutionary pro-
cess for each GA is executed as follows. For the
SSGA, there is only one population P of size M dur-
ing the whole evolutionary process, where parents are
selected and the offspring is inserted; thus, no gen-
erations required. Two parents i, j are selected by
tournament selection from the parent lists L
1
and L
2
,
each of size M/2. To create a parent list, six different
individuals are selected at random from P and split
into two groups of three; the best individual of one
group is added to L
1
and the best individual of the
other group is added to L
2
. This process is repeated
until the two lists of parents are complete. Then, for
i, j from 1, .. ., M/2, parent i in L
1
and parent j in L
2
are combined through crossover, producing two off-
spring. The next step is to apply mutation operator.
A mutation operator is applied with some probability
to the generated offspring. That is, if the mutation is
applied to an individual k, the mutated individual k
0
replaces k, regardless of the objective function value.
The recombination plus mutation process is imple-
mented on the two parents; the best two individuals
out of the two parents and the two children are added
into P so that the population size remains constant.
For the GenGA a new population is created at the
ICORES 2017 - 6th International Conference on Operations Research and Enterprise Systems
418
start of each generation. The recombination plus mu-
tation process is repeated M/2 times until the new
population P
0
is complete. At this point, individuals in
P
0
are sorted in non-decreasing order of their fitness.
The best 10% of solutions found are never removed
from the population. However, the worst 10% of in-
dividuals in P
0
are replaced by randomly generated
individuals to introduce diversity onto the population.
After this, the new population replaces the old one,
i.e. P = P
0
. Then, the WSR operator (described be-
low) is applied onto infeasible individuals within the
population based on hard constraints violations shown
in Table 1. Finally, population P
0
is passed to the next
generation.
4.2 Repair Operators
A time conflict reduction (TCR) operator works as
follows. Each pair of visits, i and j, are compared
to identify any time conflicts, i.e. the same worker w
being assigned to the two visits at the same time. If
there is a time conflict, worker w is replaced in visit
j by another worker w
0
, selected from the list of pre-
ferred workers for visit j, if that list exists, or selected
at random otherwise. Because TCR is applied only
once on an individual, by removing one pair of visits
at a time, it cannot ensure that all time conflicts are
removed, but it does reduce their number.
The worker suitability repair (WSR) operator
seeks to improve the suitability of workers for each
visit, and works as follows. For each visit in an in-
dividual, the assigned worker is checked against the
skills requirements, maximum hours constraints and
time conflicts (i.e. the hard constraints listed in Ta-
ble 1). If the worker does not satisfy these require-
ments, the operator aims to find another worker who is
feasible for that visit. If no such worker can be found,
the operator leaves the original worker in place. Thus,
the WSR operator cannot ensure that all visits have a
suitable worker, but it does improve the overall as-
signment with respect to the constraints.
4.3 Genetic Operators
The aim of this study is to select the best configuration
of crossover and mutation operators that can tackle
the WSRP within the SSGA and GenGA. Twelve op-
erators are implemented, ten well-known operators
plus two cost-based operators tailored for the problem
tackled here.
Ten well-known operators were chosen after a lit-
erature survey of operators applied in WSRP-related
problems (Algethami and Landa-Silva, 2015). The
operators selected are divided into two groups. One
Algorithm 1: Cost-based uniform crossover (CBUX).
Require: parent individuals p
1
and p
2
Ensure: offspring individuals o
1
and o
2
1: o
1
,o
2
new empty individual
2: for i 1 to |T | do
3: Let w
i
1
be worker assigned to visit i p
1
4: Let w
i
2
be worker assigned to visit i p
2
5: if w
i
1
and w
i
2
are both available for visit i then
6: o
1
o
1
w
i
1
7: o
2
o
2
w
i
2
8: else
9: if one of w
i
1
or w
i
2
, called w
i
, is available for
visit i then
10: o
1
o
1
w
i
11: o
2
o
2
w
i
12: else
13: o
1
o
1
w
i
2
14: o
2
o
2
w
i
1
15: end if
16: end if
17: end for
group has five scheduling operators: single-point
crossover (1PX), uniform crossover (Mitchell, 1998),
two-point crossover (2PX) (Hartmann, 1998), half-
uniform crossover (HX) (Eshelman, 1991) and ran-
dom swap mutation (RSM) (Cicirello and Cernera,
2013). The other group has ve routing operators:
order crossover (OX) (Zheng and Wang, 2003), cy-
cle crossover (CX) (Oliver et al., 1987), partially
matched crossover (PMX) (Zhu, 2000), inversion mu-
tation (IM) (Eshelman, 1991) and scramble mutation
(SM) (Cicirello and Cernera, 2013).
Two cost-based operators operators have been
purposely designed to improve the satisfaction of soft
constraints in the WSRP, even at the expense of hav-
ing a larger total cost in the solution.
One of these operators is a cost-based uniform
crossover (CBUX) shown in Algorithm 1. Cost-based
crossovers have been applied in the literature to pro-
duce improved results by restricting mating to the fea-
sible region only (Kotecha et al., 2004). This CBUX
operator works as follows. Each position i for the two
parents, corresponding to the worker assigned to visit
i, is processed one at a time (line 2). The availabil-
ity, in terms of time and area, of the worker assigned
to visit i is examined for each parent. If both parents
have an available worker in that position, their gene is
copied to the corresponding offspring (lines 5–7). If
only one of the parents has an available worker in that
position, that gene is copied to both offspring (lines
9–11). If no parent has an available worker in that
position, offspring 1 gets the gene from parent 2 and
offspring 2 gets the gene from parent 1 (lines 13–14).
Selecting Genetic Operators to Maximise Preference Satisfaction in a Workforce Scheduling and Routing Problem
419
Algorithm 2: Cost-based mutation (CBM).
Require: individual k
1: Choose a random mutation point i k
2: Let w
i
be the worker assigned to visit i
3: Let
i
be the list of preferred workers for i
4: if w /
i
then
5: Choose a random worker w
0
i
6: Replace w with w
0
in k for visit i
7: end if
A cost-based mutation (CBM) is shown in Algo-
rithm 2. This operator seeks to ensure that workers
assigned to visits are among those considered as pre-
ferred workers for that visit. As part of the input data
in the problem instances considered here, a list of pre-
ferred workers is given for each visit as defined by the
patients. Then, for position i in the individual, CBM
tries to assign one of the preferred workers for that
visit i, only if the one already assigned is not a pre-
ferred worker (lines 4–6).
5 EXPERIMENTS AND RESULTS
Table 2: Parameter settings used in the experiments.
Parameter Settings
Population size M 100
Crossover operators 1PX, 2PX, UX, HX, PMX, OX, CX, CBUX
Crossover rate P
c
10%, 50%, 100%
Mutation operators RSM, IM, SM, CBM
Mutation Rate P
m
1%, 10%, 30%
Running time 5 minutes
Experiments were conducted to evaluate the perfor-
mance of different operators on real-world WSRP
scenarios. The GAs described in Section 4 were im-
plemented with different algorithm configurations as
stated in Table 2. Values for mutation and crossover
rates are taken from previous parameter tuning exper-
iments.
The best suitable combinations of operators, each
combination is one crossover operator with one muta-
tion operator, might be later embedded in a more ef-
ficient approach. To this end, the experimental study
focused on comparing the performance of the vari-
ous genetic operator combinations aimed at satisfying
customers’ and workers’ preference constraints. Nev-
ertheless, one of the major issues in the random di-
rect representation is allowing infeasible individuals
throughout the search process, so that the end result
can have individuals with high number of constraint
violations. Thus, reducing hard constraint violations,
such as skills required, maximum hours requirements
and time conflicts, is required to maintain feasibility
in WSRP solutions. As explained in Section 4, the
WSR mechanism is applied to individuals that present
hard constraint violations. WSR implementation oc-
curs in two stages of the GA: after the TCR, and then
again after the mutation operator in the optimisation
process.
For each GA, 32 mutation-crossover combina-
tions with 9 different rates were applied. Thus 288×2
algorithm configurations and each one was executed 8
times, all seeded with the same initial population. The
best cost solution was obtained from each set with the
same amount of computation time. The implementa-
tion was in Java running on a PC with I7 four-core
processor with hyper-threading enabled and 16GB of
RAM.
5.1 Problem Instances
Problem instances from three UK real-world HHC
scenarios are used as instances of WSRP. The
instances data and weights used here (blue set-
ting) are available at https://drive.google.com/open?
id=0B2OtHr1VocuSNGVOT2VSYmp6a2M. In this
study, three scenarios were used with 7 problem in-
stances each, for a total of 21 instances. Table 3 shows
the main features of each problem instance.
Scenario A instances are considered the smallest,
while instances in scenario B are larger. Problem
instances in scenario C are very different to the in-
stances in the other 6 scenarios in that the number of
workers is much larger than the number of visits.
Table 3: Features of the WSRP instances.
Instance A1 A2 A3 A4 A5 A6 A7 Mean
Number Visits 31 31 38 28 13 28 13 26
Number Workers 23 22 22 19 19 21 21 21
Number Areas 6 4 5 4 4 8 4 5
Instance B1 B2 B3 B4 B5 B6 B7 Mean
Number Visits 36 12 69 30 61 57 61 47
Number Workers 25 25 34 34 32 32 32 31
Number Areas 6 5 7 5 8 8 7 7
Instance C1 C2 C3 C4 C5 C6 C7 Mean
Number Visits 177 7 150 32 29 158 6 80
Number Workers 1037 618 1077 979 821 816 349 814
Number Areas 8 4 7 8 6 11 6 7
5.2 Performance of Operators
The first set of experiments was designed to select
the best combination of the operators that maximise
customer/worker requirements and preferences sat-
isfaction. To do so, all operators listed in Table 2
were examined by statistical analysis to determine
their performance. There are four mutation operators
(RSM, IM, SM, CBM) and eight crossover operators
divided into three different groups: routing crossovers
ICORES 2017 - 6th International Conference on Operations Research and Enterprise Systems
420
Table 4: Performance Comparison of Crossover Operators Grouped by Category (Routing, Scheduling, Cost-Based) Under a
Mutation Operator.
Mutation RSM SM IM CBM
SSGA
Crossover OX UX CBUX OX UX CBUX OX UX CBUX OX 1PX CBUX
Dev. 0.09% 1.66% 4.99% 0.07% 1.23% 4.90% 0.04% 1.47% 4.24% 0.66% 1.37% 3.23%
#Best 0.71 0.19 0.00 0.76 0.14 0.00 0.86 0.05 0.00 0.48 0.38 0.05
Score 0.90 0.57 0.57 0.93 0.60 0.12 0.98 0.50 0.17 0.76 0.60 0.29
GenGA
Crossover PMX UX CBUX PMX 1PX CBUX PMX 1PX CBUX PMX UX CBUX
Dev. 0.87% 0.70% 8.98% 2.28% 1.17% 9.82% 2.90% 1.92% 9.65% 0.65% 3.12% 5.19%
#Best 0.38 0.38 0.14 0.43 0.38 0.14 0.29 0.52 0.19 0.57 0.19 0.14
Score 0.67 0.74 0.24 0.74 0.69 0.24 0.67 0.74 0.31 0.81 0.50 0.33
(OX, CX, PMX), scheduling crossovers (1PX, 2PX,
UX, HX) and cost-based crossovers (CBUX). Each
crossover was combined with one mutation at a time
for a total of 32 combinations, however only the best
performing crossover operators from each group is
presented in Table 4. The GA was executed for 5 min-
utes using the highest rate values, i.e. P
c
= 100% and
P
m
= 30% to ensure that the operators were utilised.
The following three metrics were used to measure
the performance of the combinations of operators:
Dev. average percentage deviation from the best pref-
erence value (the three preferences satisfaction value
of all the configurations applied). Best fraction of in-
stances in a set for which a configuration matches the
best preference value. This performance metric is ab-
solute and can be compared across existing results in
different tables. Score fraction of the instances for
which the current method produces better solutions
than the other configurations, i.e. ‘win’. This score is
calculated as ((q × (p 1)) r)/(q × (p 1)), where
p is the number of configurations compared, q is the
number of problem instances, and r is the number of
instances in which the p 1 competing configurations
find a better result. Hence, the best score value is
1, when r = 0, and the worst score value is 0, when
r = q × (p 1). This is a relative measure of per-
formance. Hence, these values are only meaningful
within one table and not across different tables.
The results presented in Table 4 are the perfor-
mance metrics values that correspond to each of the
four mutation operators, including the CBM operator
proposed in this study. These values are calculated
based on the average preferences satisfaction values
for each run. The crossover categories are: the best
routing operator, the best scheduling operator and the
CBUX operator proposed in this study. Thus, each
mutation operator has three comparable crossover op-
erators values, and the best crossover out of the three
is highlighted in bold. The aim is to identify the best
crossovers for each mutation with respect to the pref-
erences value by grouping crossovers based on their
category, thus mutation operators are not comparable
in this table.
Table 5: Performance Comparison Between the Best Com-
binations of Operators ( Mutation - Crossover ).
Procedur Dev. #Best Score
SSGA
RSM-OX 0.01% 0.52 0.87
SM-OX 0.03% 0.29 0.82
IM-OX 0.15% 0.00 0.50
CBM-OX 0.23% 0.10 0.45
GenGA
RSM-UX 0.92% 0.57 0.84
SM-PMX 7.57% 0.05 0.43
IM-1PX 2.73% 0.24 0.68
CBM-PMX 18.36% 0.05 0.25
For SSGA, OX provides the best scores, with the
highest number of best values and the lowest devia-
tion when combined with all mutation operators. For
GenGA, UX provides the best scores for RSM with
the highest number of the best values and the low-
est deviation. Even though UX is selected as the first
competing crossover, PMX obtained the same number
of the best values for RSM; the winning crossovers
are considered in the next overall comparison. Addi-
tionally, 1PX provides the best score value with the
highest number of best solutions and lowest on devi-
ations among the compared crossovers for IM, while
PMX provides the best score value for both SM and
CBM, with the lowest deviation obtained for CBM
only.
Table 5 shows a comparison between the cho-
sen combinations of operators (mutation - crossover)
from Table 4. The aim is to identify the best combina-
tion for each GA by using the same performance mea-
surement matrices explained above. The best combi-
nation is highlighted in bold.
The results indicate that RSM–UX and RSM–UX
obtain the highest score and the smallest deviation
value among all methods, with the maximum frac-
tions of the best solutions of 0.87 and 0.84 for SSGA
and GenGA respectively. Interestingly, cost-based
methods failed to achieve good results in compari-
son to the generic operators. This result might be due
to the search space restrictions that led to infeasible
areas. However, when combined with more generic
Selecting Genetic Operators to Maximise Preference Satisfaction in a Workforce Scheduling and Routing Problem
421
Table 6: Results of the best f (S) produced by different combinations of operators, crossover probabilities P
c
and mutation
probabilities P
m
.
SSGA GenGA
Instance Procedure P
c
P
m
f (S) C pt(s) Procedure P
c
P
m
f (S) Cpt(s)
A
1 SM-PMX 0.5 0.3 5.1 176.6 RSM-PMX 1 0.3 3.5 180.8
2 SM-OX 1 0.3 4.4 184.4 SM-UX 1 0.3 2.8 190.7
3 SM-UX 1 0.3 6.1 240.8 SM-1PX 1 0.3 3.3 212.3
4 SM-UX 0.5 0.3 2.3 159.9 SM-UX 1 0.3 1.4 114.5
5 SM-1PX 1 0.1 3.2 50.7 SM-HX 1 0.3 2.4 52.4
6 SM-2PX 1 0.1 4.4 198.6 RSM-HX 1 0.3 3.6 109.0
7 SM-2PX 1 0.1 4.2 32.7 RSM-HX 1 0.3 3.7 83.2
B
1 SM-UX 1 0.3 2.1 239.7 RSM-PMX 1 0.3 1.7 255.0
2 SM-OX 1 0.1 2.4 50.7 RSM-HX 1 0.3 1.8 14.7
3 RSM-PMX 0.5 0.3 2.8 292.9 RSM-PMX 0.5 0.3 1.9 274.1
4 SM-2PX 1 0.3 2.9 75.3 SM-2PX 1 0.3 2.1 138.0
5 RSM-UX 0.5 0.3 3.8 243.5 RSM-2PX 1 0.3 2 243.2
6 RSM-UX 1 0.3 2.5 226.1 RSM-2PX 0.5 0.3 1.7 229.2
7 RSM-UX 0.5 0.3 3.2 231.8 RSM-1PX 0.5 0.3 1.9 288.2
C
1 CBM-OX 1 0.1 5454.5 285.0 CBM-OX 1 0.3 159418.6 304.8
2 SM-HX 1 0.01 4.8 15.0 SM-HX 1 0.3 3.2 152.7
3 RSM-2PX 1 0.3 3270.7 299.4 SM-OX 1 0.3 82582 293.9
4 RSM-OX 1 0.01 22.2 210.0 IM-HX 1 0.3 17.6 269.8
5 SM-PMX 1 0.1 20.1 172.6 IM-HX 1 0.3 16 267.3
6 CBM-PMX 1 0.01 20776.5 266.6 CBM-OX 1 0.01 94335.9 297.8
7 SM-UX 1 0.3 4.9 0.7 SM-UX 1 0.3 4.3 15.9
operators, in the case of CBM, they generate more di-
verse individuals that led up to high deviation among
all mutations.
5.3 Computational Results for Different
Instance Sizes
Table 6 presents the best objective values obtained for
all instances. The columns under SSGA and GenGA
provide the best values obtained for each GA under
the stopping criterion for each combination. All val-
ues are averaged and only the best values are pre-
sented. The remaining column C pt shows the compu-
tation time where the best value is found in seconds.
Two issues were considered to compile this table: the
GAs performance on each instance and the best per-
forming combinations/settings under each GA with
the minimum computation time.
It appears that GenGA provides better results
than SSGA on 85.71% of all instances. The best-
performing operators under the methods applied are
PMX, UX and HX across all instances when com-
bined with RSM and SM. However, CBM managed
to obtain some of the best results, especially for sce-
nario C instances. This indicates that problem domain
knowledge needs to be incorporated in operators for
more complicated instances. The average computa-
tion times for the best solution found for all instances
are as follows: SSGA, and GenGA are 173.954 s and
189.88 s respectively. In terms of convergence speed,
both SSGAs used here converged earlier to a local
minima, with poor results in problem sets A and B.
For problem set C, more computation-time provided
better results when using GenGA.
Despite the fact that the cost values still need to
be improved, this study has helped to understand the
performance of various combinations of genetic op-
erators executed with different probability rates and
implemented on simple steady-state and generational
GAs.
6 CONCLUSION
This paper has investigated the suitability of a set of
genetic operators when applied within a steady-state
genetic algorithm (SSGA) and a generational genetic
algorithm (GenGA) to tackle the workforce schedul-
ing and routing problem (WSRP). Twelve operators
were considered in this study including two operators
incorporating problem domain knowledge, and ten
well-known operators (three mutation operators and
ICORES 2017 - 6th International Conference on Operations Research and Enterprise Systems
422
seven crossover operators) from the literature. From
the experimental results, existing operators such as
RSM and UX perform the best. Future research will
look at investigating the performance of the repair op-
erators, parameter setting of the operators and the de-
sign of an improved evolutionary approach informed
by the better understanding achieved in this paper.
REFERENCES
Aickelin, U. and Dowsland, K. A. (2004). An indirect ge-
netic algorithm for a nurse-scheduling problem. Com-
puters & Operations Research, 31(5):761 – 778.
Algethami, H. and Landa-Silva, D. (2015). A study of ge-
netic operators for the workforce scheduling and rout-
ing problem. In 11th Metaheuristics International
Conference (MIC 2015), pages 1–11.
Algethami, H., Pinheiro, R. L., and Landa-Silva, D. (2016).
A genetic algorithm for a workforce scheduling and
routing problem. In 2016 IEEE Congress on Evolu-
tionary Computation (CEC), pages 927–934.
Castillo-Salazar, J. A., Landa-Silva, D., and Qu, R. (2016).
Workforce scheduling and routing problems: litera-
ture survey and computational study. Annals of Oper-
ations Research, 239(1):39–67.
Chang, Y. and Chen, L. (2007). Solve the vehicle routing
problem with time windows via a genetic algorithm.
Discrete and continuous dynamical systems supple-
ment, pages 240–249.
Cicirello, V. A. and Cernera, R. (2013). Profiling the
distance characteristics of mutation operators for
permutation-based genetic algorithms. In Boonthum-
Denecke, C. and Youngblood, G. M., editors, FLAIRS
Conference, Florida. AAAI Press.
Cowling, P., Colledge, N., Dahal, K., and Remde, S. (2006).
The trade-off between diversity and quality for multi-
objective workforce scheduling. In Proceedings of
the 6th European Conference on Evolutionary Com-
putation in Combinatorial Optimization, EvoCOP’06,
pages 13–24. Springer-Verlag.
Eshelman, L. J. (1991). The CHC adaptive search algorithm
: How to have safe search when engaging in nontradi-
tional genetic recombination. Foundations of Genetic
Algorithms, pages 265–283.
Hartmann, S. (1998). A competitive genetic algorithm for
resource-constrained project scheduling. Naval Re-
search Logistics (NRL), 45(7):733–750.
Kotecha, K., Sanghani, G., and Gambhava, N. (2004).
Genetic algorithm for airline crew scheduling prob-
lem using cost-based uniform crossover. In Manand-
har, S., Austin, J., Desai, U. B., Oyanagi, Y., and
Talukder, A. K., editors, AACC, volume 3285 of Lec-
ture Notes in Computer Science, pages 84–91, Kath-
mandu, Nepal. Springer.
Laesanklang, W. and Landa-Silva, D. (2016). Decomposi-
tion techniques with mixed integer programming and
heuristics for home healthcare planning. Annals of
Operations Research, pages 1–35.
Lenstra, J. K. and Kan, A. H. G. (1981). Complexity of
vehicle routing and scheduling problems. Networks,
11(2):221–227.
Mankowska, D., Meisel, F., and Bierwirth, C. (2014). The
home health care routing and scheduling problem with
interdependent services. Health Care Management
Science, 17(1):15–30.
Mısır, M., Smet, P., and Vanden Berghe, G. (2015). An
analysis of generalised heuristics for vehicle routing
and personnel rostering problems. Journal of the Op-
erational Research Society, 66(5):858–870.
Mitchell, M. (1998). An Introduction to Genetic Algo-
rithms. The MIT Press, Cambridge, MA, USA.
Mutingi, M. and Mbohwa, C. (2014). Health-care staff
scheduling in a fuzzy environment: A fuzzy genetic
algorithm approach. In Conference Proceedings (DFC
Quality and Operations Management). International
Conference on Industrial Engineering and Operations
Management.
Oliver, I. M., Smith, D. J., and Holland, J. R. C. (1987). A
study of permutation crossover operators on the travel-
ling salesman problem. In Proceedings of the Second
International Conference on Genetic Algorithms and
their application, pages 224–230, Hillsdale, NJ, USA.
L. Erlbaum Associates Inc.
Pinheiro, R. L., Landa-Silva, D., and Atkin, J. (2016).
A variable neighbourhood search for the workforce
scheduling and routing problem. In Advances in Na-
ture and Biologically Inspired Computing, pages 247–
259. Springer, Pietermaritzburg, South Africa.
Prins, C. (2004). A simple and effective evolutionary algo-
rithm for the vehicle routing problem. Computers &
Operations Research, 31(12):1985 – 2002.
Rothlauf, F. (2003). Representations for genetic and evo-
lutionary algorithms. Studies in Fuzziness and Soft
Computing, 104:9–32.
Toth, P. and Vigo, D. (2014). The vehicle routing problem,
volume 18. Siam.
Vavak, F. and Fogarty, T. C. (1996). Comparison of steady
state and generational genetic algorithms for use in
nonstationary environments. In Evolutionary Compu-
tation, 1996., Proceedings of IEEE International Con-
ference on, pages 192–195. IEEE.
Zheng, D.-Z. and Wang, L. (2003). An effective hy-
brid heuristic for flow shop scheduling. The Interna-
tional Journal of Advanced Manufacturing Technol-
ogy, 21(1):38–44.
Zhu, K. Q. (2000). A new genetic algorithm for VRPTW.
In Proceedings of the International Conference on Ar-
tificial Intelligence. Citeseer.
Selecting Genetic Operators to Maximise Preference Satisfaction in a Workforce Scheduling and Routing Problem
423