Integrating Memory-Based Perturbation Operators into a Tabu Search
Algorithm for Real-World Production Scheduling Problems
Manuel Schlenkrich
1,2
, Michael B
¨
ogl
2
, Anna Gattinger
2
, Ionela Knospe
2
and Sophie N. Parragh
1
1
Johannes Kepler University Linz, Altenbergerstraße 69, 4040 Linz, Austria
2
RISC Software GmbH, Softwarepark 32a, 4232 Hagenberg im M
¨
uhlkreis, Austria
Keywords:
Real-World Production Scheduling, Tabu Search, Perturbation, Destroy and Repair Operator, Elite Solutions.
Abstract:
Production scheduling problems, arising in real-world use cases, are characterized by a very large number
of operations and complex constraints. In order to handle such problems, practical solution approaches need
to be generic enough, to capture all relevant restrictions, while being able to calculate good solutions in a
short amount of time. Metaheuristic methods, especially combinations of trajectory and population-based
approaches, are promising techniques to meet this criterion. In this work, we develop a framework for de-
riving and integrating memory-based perturbation operators into a highly flexible Tabu Search algorithm for
scheduling problems, in order to enhance its overall performance. The perturbation operators are inspired
by evolutionary algorithms and collect valuable solution information during the Tabu Search procedure via
an elite solution pool. This information is used in a destroy-and-repair step integrated into the Tabu Search
procedure, aiming to preserve promising solution structures. We investigate several parameters and perform
computational experiments on job-shop benchmark instances from literature, as well as on a real-world indus-
try use case. Integrating the developed memory-based perturbation operators into the Tabu Search algorithm
leads to significant performance improvements on the real-world problem. The benchmark evaluations demon-
strate the robustness of the approach, when dealing with sensitive parameters.
1 INTRODUCTION
The rise of Industry 4.0 across companies in vari-
ous sectors, accompanied by the introduction of dig-
ital twins and automated decision making, has cre-
ated a number of challenging planning problems and
increased the need for practical solution approaches.
Among those planning problems, scheduling of com-
plex production systems, as they occur in semicon-
ductor manufacturing, textile processing or the chem-
ical industry, is one of the most demanding ones.
In the literature several different approaches to
tackle production scheduling problems are avail-
able, such as the use of mixed integer linear pro-
gramming, constraint programming, machine learn-
ing techniques or metaheuristic methods, to name the
most popular choices. Even though a broad range of
solution methods exists, the availability of approaches
for problems of practical size and complexity is lim-
ited (Schlenkrich and Parragh, 2023). While some
methods are rather restricted in the size of the prob-
lems, that they can solve in reasonable time, meta-
heuristic solution methods are often able to capture
complicated constraints arising in real-world settings
and can provide good solutions in a short amount of
time.
In their work on real-life scheduling problems
with rich constraints and dynamic properties, B
¨
ogl
et al. (2021) give a profound overview of the most
important aspects, that need to be considered when
developing scheduling tools for real industry cases.
They present a Tabu Search (TS) framework, mak-
ing use of an activity list-based representation of the
schedule, following the approach of Moumene and
Ferland (2009). The TS framework is embedded in
the production planning and execution software envi-
ronment of our partner company and is currently used
as the scheduling tool. The algorithm allows to han-
dle very large problem instances up to tens of thou-
sands of operations on a large variety of machines.
However, with increasing problem size, also the re-
quired computation time to reach a solution of ac-
ceptable quality rises. Thus, enhancing the existing
solution method, to counteract the growing computa-
tional burden, is of utmost importance and therefore
the main goal of this work.
Schlenkrich, M., Bögl, M., Gattinger, A., Knospe, I. and Parragh, S.
Integrating Memory-Based Perturbation Operators into a Tabu Search Algorithm for Real-World Production Scheduling Problems.
DOI: 10.5220/0012271900003639
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 13th International Conference on Operations Research and Enterprise Systems (ICORES 2024), pages 213-220
ISBN: 978-989-758-681-1; ISSN: 2184-4372
Proceedings Copyright © 2024 by SCITEPRESS Science and Technology Publications, Lda.
213
For this purpose, we extend the TS of B
¨
ogl et al.
(2021), by deriving and integrating memory-based
perturbation (MBP) operators to foster diversification
within the trajectory-based metaheuristic method. We
introduce a framework for deducing perturbation op-
erators that modify solution candidates on the basis
of a set of elite solutions. Promising solution pat-
terns, that are shared among the elite solutions, are
conserved, while the remaining part of the solution
candidate is largely modified, in order to diversify the
investigated solutions in the search space.
The main inspiration for our approach is com-
ing from other available works, that investigate diver-
sification strategies in different metaheuristic proce-
dures, such as Adaptive Large Neighborhood Search
(ALNS) (Ropke and Pisinger, 2006). It is an exten-
sion of Large Neighborhood Search (LNS) (Shaw,
1997). LNS and ALNS respectively, use large mod-
ifications that can rearrange a solution candidate by
up to 30% or 40% to explore various regions of the
search space. A similar core idea is used in Iterated
Local Search (ILS), see e.g., Braune et al. (2007) for
an application in the scheduling context. ILS is a
trajectory-based metaheuristic method, that alternates
intensification and diversification phases, via small
and large step modifications respectively. The large
modifications are reached by performing perturbation
steps, that oftentimes rely on random destroy and re-
pair operators.
While both approaches, ALNS and ILS, perform
intensification and diversification steps iteratively,
there also exist procedures that aim to foster these
principles in two consecutive phases. In the work of
Lunardi et al. (2021), four metaheurstics are tailored
to a complex scheduling problem. The authors pro-
pose to combine different methods in order to create
more powerful approaches and use the solution of one
method as the starting point for a second one.
In our work we make use of the destroy and re-
pair concept similar to ALNS and ILS. In contrast to
ALNS however, we do not perform several large mod-
ifications of the solution candidate in a row, but use it
as a one-time perturbation operator, if the Tabu Search
procedure seems to be trapped in a local minimum. In
contrast to ILS, we do not randomly destroy parts of
the solution, but aim at preserving promising parts of
the solution candidate, by collecting structural infor-
mation on the best solutions found so far during the
TS.
In the following, we first present the investigated
scheduling problem in Section 2. Then, in Section
3 we describe our solution approach and Section 4
summarizes the numerical experiments. Finally, in
Section 5 we conclude our work and provide future
research directions.
2 PROBLEM DESCRIPTION
Within this section we describe the investigated pro-
duction scheduling problem, which has also been ad-
dressed by B
¨
ogl et al. (2021). Since we propose an
extension to their approach, we tackle the same un-
derlying problem.
The task is to schedule a set of n jobs on a set
of k machines, such that a given objective function
is optimized. Each job j
i
consists of ˆn
i
operations
{o
i1
, . . . , o
i ˆn
i
}. While machines have different at-
tributes concerning their availability to process one or
more operations at each given time, also operation-
specific restrictions need to be considered.
For each operation to be scheduled, the following
data is available: An operation cannot be scheduled
before its associated earliest start time. It can finish
later than its latest finish time, but this might influ-
ence the objective function value, in the case that ac-
tivity tardiness is considered. Each operation must
be assigned to a machine out of the set of compatible
machines. For each of the compatible machines the
consumed capacity, production duration and produc-
tion costs are defined. After production the teardown
duration for each activity needs to be respected. Setup
times for activities are both machine- and sequence-
dependent and need to be considered. For each pro-
duction setup there is a machine dependent setup cost.
Any operation has a set of predecessors, which must
be finished before the activity itself can start, and a
set of successors, which can only start after the ac-
tivity itself is finished. These sets are not restricted
to a single predecessor and successor, but can contain
multiple ones. Also, these sets may contain opera-
tions that belong to other jobs than the activity itself.
For activities that have a direct precedence relation, an
overlap can be defined. Waiting times have to be con-
sidered after an operation is finished, however without
occupying the assigned machines. Also, transporta-
tion times arise, when operations are transferred to
another machine. Transportation times are machine
dependent and are provided in a transportation time
matrix. Additional machines might be needed in or-
der to perform an operation. In this case an additional
operation is created that needs to be scheduled simul-
taneously. In case that a partial schedule is already un-
der execution, current start and end times of execution
can be specified. Finally, several objective functions
can be defined. Popular choices are minimizing the
makespan, total weighted tardiness on activity level
(TWT) or total weighted order tardiness (TWOT).
ICORES 2024 - 13th International Conference on Operations Research and Enterprise Systems
214
3 MEMORY-BASED
PERTURBATION
FRAMEWORK
TS is a trajectory-based metaheuristic method, itera-
tively evaluating solution candidates in order to find
near-optimal solutions for optimization problems,
such as production scheduling problems. Neighbor-
hood operations, which are small modifications of the
current solution, are applied to generate new candi-
dates. The term ”tabu” in the name of the method
refers to a list of forbidden neighborhood moves,
the so called tabu list, which helps the algorithm
avoid getting trapped in local optima. Tabu Search’s
strength lies in its ability to intensify the search for
better solutions, rather than diversifying the search
space. This can be advantageous in promising areas
of the solution space, however might delay the search
procedure in the case that the algorithm is stuck in an
unfavorable area. In order to support diversification
of trajectory based metaheuristics, several works pro-
pose combinations with population based approaches,
see e.g. Lunardi et al. (2021).
We foster this concept of combining trajectory and
population-based metaheuristics, but instead of se-
quentially performing different methods we directly
integrate principles of evolutionary algorithms into an
operator used within the TS framework. The idea is
to diversify solution candidates, as soon as the speed
of improvement falls below a certain level, meaning
that the TS is not able to make use of its strong in-
tensification mechanism. Perturbation operators are a
common tool to diversify solution candidates. They
perform a large modification of the current candidate,
moving to another area in the solution space, which
is possibly more promising. The problem with com-
mon perturbation operators is, however, that much
of the effort already invested by the TS algorithm is
lost, because the structure of the candidate is modi-
fied randomly. To counteract this issue, we develop
a memory-based perturbation framework, that aims at
preserving promising solution structures of the cur-
rent solution, while only destroying non-promising
parts in order to reach diversification. Information
about promising structures is extracted from an elite
solution pool, which is collected during the TS proce-
dure. The solution pool is then evaluated by means of
a configurable evaluation criterion, that detects shared
properties or structures among the elite solutions. Af-
ter that a parameterizable acceptance criterion de-
cides, which of the solution characteristics are shared
among enough of the elite solutions, in order to be
classified as promising properties, that should be con-
served for later iterations within the TS. In the fol-
lowing, the unpromising part of the candidate is re-
moved, while the promising part is preserved, using
the destroy-and-repair concept. After rearranging the
removed part, the newly generated solution candidate
is diversified, without destroying the most promising
solution patterns.
Algorithm 1 presents the TS procedure using a
first improvement implementation. It includes the
mechanism to trigger a perturbation step, which could
be either a random perturbation or one of the proposed
MBP operators. The remainder of the work will focus
on the perturbation step, while keeping the core com-
ponents and parameters of the TS unchanged.
Data: Problem data, Tabu Search parameters,
Perturbation parameters
Result: Best found solution
initial solution s construction heuristic;
incumbent solution s
s;
best solution s
s;
nonImprovingIterations = 0;
while termination criterion not met do
bestOb jective = ;
nonImprovingIterations + +;
while maxNeighborhoodSize not reached
do
generate next non tabu neighbor ˆs of
s
;
if ob jective( ˆs) bestOb jective then
¯s ˆs;
bestOb jective ob jective(¯s);
end
if ob jective( ˆs) ob jective(s
) then
nonImprovingIterations = 0;
break;
end
end
s
¯s;
update tabu list;
if ob jective(s
) < ob jective(s
) then
s
s
;
end
if nonImprovingIterations
perturbationTriggerLimit then
s
PerturbationOperator(s
);
nonImprovingIterations = 0;
end
end
Return s
as best found solution;
Algorithm 1: Tabu Search procedure in the ”first improve-
ment” implementation including perturbation steps.
Integrating Memory-Based Perturbation Operators into a Tabu Search Algorithm for Real-World Production Scheduling Problems
215
Figure 1: The Memory-Based Perturbation Framework.
There are several possibilities to construct a per-
turbation operator, that serves the purpose described
above. We have developed a framework, which can be
used to derive different operators, based on four im-
portant components, namely the design of the elite so-
lution pool, choice of a suitable evaluation criterion,
definition of an acceptance level and procedure for the
destroy-and-repair phase. Figure 1 presents the devel-
oped MBP framework, which is further described in
the following sections.
Elite Solution Pool. The elite solution pool is a col-
lection of solution candidates obtained during the ex-
ecution of the TS. Since the goal is to gather infor-
mation on promising solution structures, the solution
candidates selected for the elite solution pool are sup-
posed to be of adequate quality. This could, for ex-
ample, mean that every newly best found solution is
added to the pool. Another option would be to select
all newly found incumbent solutions, which are the
solution candidates used by the TS to generate neigh-
boring solutions. On top of the initial selection crite-
rion for the elite solution pool, also a time-dependent
criterion can be defined, restricting the elite solutions,
based on the TS iteration in which they have been
added to the set. This could either mean assigning
higher relevance to more recent candidates or delet-
ing earlier and therefore more outdated candidates.
Evalution Criteria. Once the relevant elite solution
pool is defined, a criterion for evaluating the selected
solution candidates need to be specified. This crite-
rion determines the solution properties, that should be
compared among the elite solutions in order to deduce
promising patterns among them, meaning that several
solutions of good quality share the same characteris-
tic. There are numerous possibilities to define such
an evaluation criterion, e.g., to compare the machine
index of activities, the activity list index or the over-
lap of ranges around activity starting times. Through
the highly modular approach, this criterion can be ex-
changed effortless, leading to different MBP opera-
tors.
Overlap Acceptance Level. After evaluating the
solutions in the elite solution pool according to the
chosen evaluation criteria, an overlap acceptance level
needs to be specified. This acceptance level deter-
mines, if the chosen properties are shared among
enough elite solutions, to be defined as promising so-
lution patterns. These acceptance levels can again
be defined and replaced in a highly flexible manner,
leading to different operators. A very strict accep-
tance level would be to categorize solution patterns
as promising, only if they are shared among all solu-
tions in the elite solution pool. This criterion could be
moderated by only demanding an overlap on a frac-
tion of the elite solutions. An approach to keep the
size of the preserved part of the solution stable, is to
always categorize a fixed fraction of the solution as
promising, namely the one with the highest overlap
among the elite solutions. We will later refer to this
approach by using a so called ”destroy percentile”,
meaning that the percentile of the solution candidate
with the lowest overlap will be destroyed, while the
rest is preserved.
Destroy-and-Repair Phase. In the last phase of the
MBP, the gathered information about high quality so-
lutions and their shared promising structures is used
to preserve those structures in the current solution,
while diversifying the remaining parts. The so called
destroy phase removes all activities, that do not meet
the selected overlap acceptance criteria from the cur-
rent solution. Those removed activities are then in-
serted into the perturbated candidate, using a con-
figurable insertion mechanism, that could either be
based on activity properties or inserts randomly. The
result of the repair phase, is a new diversified solu-
tion candidate, that still contains the most promising
patterns, identified in the elite solution pool.
4 NUMERICAL STUDY
In order to evaluate the impact of the proposed ap-
proach on the performance of the TS algorithm, which
in its base version is currently used in the planning
software of the company, we derive two types of MBP
operators. The first one uses the activity list index
evaluation criterion, meaning that elite solutions are
compared concerning the positions of each activity in
the encoded representation of the solution, namely the
activity list. The second operator uses the machine
index evaluation criterion, meaning that shared prop-
erties are evaluated by means of the position of each
ICORES 2024 - 13th International Conference on Operations Research and Enterprise Systems
216
activity on the machines. Both operators store the new
best found solutions within the elite solution pool and
use the destroy percentile acceptance criterion.
We first test the TS algorithm on job-shop
scheduling benchmark instances from the literature,
where optimal solutions are known. Note, that these
problems represent a simplified version of the real-
world scheduling problem that is described in Sec-
tion 2. See e.g., Adams et al. (1988) for more infor-
mation about the job-shop scheduling problem. It is
not our goal to compete with solution approaches and
solvers, that are tailored to the job-shop scheduling
problem, such as the CP Optimizer (Laborie et al.,
2018). For our experiments on the benchmark in-
stances, we performed optimization runs with a CPU
time limit of 15 minutes per instance, which allows at
most 10,000 to 50,000 iterations per instance, due to
the overhead of the software framework. The aim of
the experiments on the benchmark instances is to bet-
ter understand the sensitivity of the approach towards
parameters, such as the number of non-improving it-
erations before a perturbation is triggered or the size
of the destroy percentile. In order to put the obtained
results into perspective, we compare them to the re-
cent approach of Yuan et al. (2023), who use deep
reinforcement learning to learn effective policies for
job-shop scheduling problems. We have chosen a
learning-based approach as a comparison, because in
the face of the complexity of our real-world schedul-
ing problem, this technique would be a reasonable al-
ternative to the currently used algorithm. In contrast,
integrating a constraint programming solver into the
software environment of the company was not a fea-
sible option.
In a second step, we apply the TS algorithm to
a real-world industry use case, which was provided
by the partner company, in order to assess its capa-
bility to improve solution quality under realistic con-
ditions. For all following experiments we chose an
equal base setting for the TS, such as employed con-
struction heuristics and available neighborhood oper-
ators. The length of the tabu list is 100 and the maxi-
mum neighborhood size for the operators is set to 300,
to name a few of the most important parameters. All
experiments were carried out on an Intel(R) Xeon(R)
Gold 5315Y CPU @3.20GHz with 8GB RAM.
4.1 Job-Shop Benchmark Results
The job-shop benchmark instances consist of two sep-
arate test sets, with different objective functions. For
instances of the first set, the makespan should be min-
imized, while for instances of the second test set, the
total weighted tardiness (TWT) is to be minimized.
The makespan test set consists of 153 instances and
consists of five subsets, taken from different works in
the literature, namely abz from Adams et al. (1988), la
from Lawrence (1984), orb from Applegate and Cook
(1991), swv from Storer et al. (1992) and ta from Tail-
lard (1993). The number of jobs ranges from 10 to
100, while the number of machines ranges from 5 to
20, making the largest instance have 2000 operations.
The TWT test set consists of 63 instances and is di-
vided into four subsets, namely abz, la, mt and orb,
which are all taken from Singer and Pinedo (1998).
All instances consist of 10 jobs on 10 machines, re-
sulting in 100 operations each.
As a base version we tested the TS algorithm with-
out any perturbation of the solution. In a next step,
we implemented a random perturbation operator with
a fixed destroy percentile. This random operator re-
moves a randomly chosen part of the solution can-
didate and inserts it back randomly. For our exper-
iments we have chosen a destroy percentile of 30%,
which turned out to be a good choice for random per-
turbation in earlier tests. A very sensitive parame-
ter is the number of non-improving steps, after which
the perturbation operator is called. For our numerical
study we choose to investigate the settings 10 and 15
for this parameter. We have chosen these values, as
we have seen that choosing a smaller value than 10
results in a search procedure, in which the solution is
destroyed too often, and on the other hand, for val-
ues larger than 15 the solutions are very similar to the
base version, since the perturbation is rarely called.
It must be noted that the best setting for this parame-
ter very much depends on the problem instances and
that intensive parameter tuning is not always possi-
ble in practice. Perturbation operators that perform
well, even for suboptimal choices of this parameter
are therefore favorable.
In a next step we performed experiments using
the TS with implementations of the MBP operators.
The first derived operator uses the activity list index
evaluation criterion. We performed test runs with pa-
rameter NonImproving set to 10 and 15 in order to
compare the behavior to the random perturbation set-
ting. Since we have observed similar results for the
MBP operator in both settings, we decided to use the
slightly better performing parameter setting of 15 for
the final experiment, which was running the TS with
the MBP operator, that uses the machine index eval-
uation criterion. For each of the derived perturbation
operators we investigate three settings for the destroy
percentile, namely 10%, 30% and 50%.
Table 1 presents the results for the makespan test
instances, displaying the percentage deviation from
the optimal makespan, averaged over the respective
Integrating Memory-Based Perturbation Operators into a Tabu Search Algorithm for Real-World Production Scheduling Problems
217
instance group. For the makespan instances we ad-
ditionally report the results obtained by Yuan et al.
(2023). Table 2 shows the results for the TWT in-
stances, reporting the absolute deviation of the opti-
mal tardiness in minutes, again averaged over the re-
spective instance group.
Integrating the MBP operators into the TS leads to
improvements on both instance sets, compared to the
base version. The average deviation from the opti-
mal makespan over all 153 test instances was reduced
from 12,7% to 10,5%. In contrast, the random per-
turbation with the NonImproving parameter set to 15
led to nearly no improvement. With this parameter
set to 10, the random perturbation even led to a worse
makespan than the base version. This indicates that
choosing the trigger level for the perturbation too low,
in combination with large uncontrolled modifications,
can harm the search procedure by jumping to another
search area, without exploiting the full potential of the
current region. Among the MBP operators, however,
we observe rather stable results, meaning that even for
a high number of perturbation calls, solution quality
stays adequate. This is a practically highly relevant
aspect, since those parameters cannot always be tuned
in a real-world setting, where time is scarce.
On the second benchmark test set the absolute de-
viation of the TWT from the optimal value averaged
over all 63 instances was reduced from 194 minutes
to 182 minutes by applying the MBP operators. Also,
for the TWT instances the results show, how sensitive
the search procedure is to the NonImproving param-
eter. For a wrong choice of the parameter, as can be
seen in the columns for a value of 10 in the random
perturbation case, the results drastically deteriorate.
Again, this can be reasoned by the fact, that solution
candidates are modified too drastically, before new
better solutions can be found by the neighborhood
operators. This test set, however, also demonstrates,
that for good choices of the NonImproving parame-
ter, even the random perturbation can lead to better
results, as it can be seen for the choice of 15.
4.2 Real-World Industry Case
Finally, we have the possibility to evaluate the pro-
posed perturbation operators on a real-world industry
case with 905 operations on 67 machines. It includes
all the aspects and constraints described in Section
2. Activity deadlines, as well as order deadlines are
available in the data, meaning that we can evaluate ac-
cording to the two objectives, that are relevant for the
company, namely total weighted tardiness (on activ-
ity level) and total weighted order tardiness (on or-
der level). For each of the two objectives we per-
form test runs, using the TS without perturbation as
the base version, with random perturbation, as well
as with MBP operators using the activity list index
and the machine index evaluation criteria. For the
perturbation operators we again test 10%, 30% and
50% destroy percentiles. We have seen, that for the
real-world use case strong diversification after a small
number of unsuccessful iterations is even more crucial
than for the benchmark instances, due to the smaller
overall number of possible iterations. Therefore, we
evaluate the values 2 and 10 for the number of non-
improving steps before perturbation. We perform 100
iterations per approach and objective, resulting in a
runtime of approximately 1 hour for each of the ex-
periments. Tables 3 and 4 present the evaluation re-
sults on the real-world industry case for the TWT and
TWOT objective, respectively. A more detailed rep-
resentation of the search procedure and the objective
value paths, reached with the best parameter settings
for each approach, can be found in Figure 2 for the
TWT case and in Figure 3 for the TWOT case. For all
perturbation approaches the best evaluated parameter
for the destroy percentile was 10%. In the TWT case,
the best found setting for the non-improving parame-
ter was 2 for MBP with the activity list index criterion,
while it was 10 for random perturbation and MBP
with the machine index criterion. For the TWOT ob-
jective on the other hand, a value of 2 performed best
for the MBP with the machine index criterion, while
10 was superior for MBP with the activity list index
criterion. For random perturbation both 2 and 10 gave
the same result.
First of all, we see that random perturbation is not
effective for the real-world use case. Destroying a
random part of the solution makes it difficult to re-
cover quickly and can even lead to worse results than
using no perturbation at all. For the TWT objective
function, TS with MBP using the machine index crite-
rion leads to a solution that is 26% better than the base
version. TS with MBP using the activity list index cri-
terion is 29% better. Figure 2 shows that the four ap-
proaches perform equally until iteration 60. The base
approach without perturbation is from this point stuck
in a local minimum. Random perturbation cannot re-
cover from its large modifications, which destroyed a
large part of the previously achieved solution quality.
Both MBP operators, however, continue to find new
best solutions by applying more controlled perturba-
tion steps, preserving the best solution properties.
In the TWOT case, the ranking of the two MBP
operators is switched, with the machine index crite-
rion performing best. It leads to an improvement of
40% over the base version, while TS using MBP with
activity list index leads to a slightly lower improve-
ICORES 2024 - 13th International Conference on Operations Research and Enterprise Systems
218
Table 1: Evaluation results for the MBP operators on the makespan test set: percentage deviation of the optimal makespan
averaged over instance group.
Perturbation None (Base) Random MBP - Activity list index MBP - Machine index DRL
1
NonImproving 10 15 10 15 15
Destroy percentile 30% 30% 10% 30% 50% 10% 30% 50% 10% 30% 50%
abz 13.4% 17.9% 14.0% 12.5% 11.3% 10.8% 13.0% 11.8% 13.3% 13.1% 12.6% 12.5% 14.7%
la 3.3% 7.5% 3.4% 2.6% 2.1% 2.6% 2.7% 2.1% 2.3% 2.5% 2.5% 2.8% 10.9%
orb 7.1% 11.3% 4.4% 4.4% 3.5% 3.7% 5.8% 3.6% 4.2% 4.7% 6.1% 5.0% 20.9%
swv 19.3% 21.1% 17.8% 16.0% 16.4% 17.4% 16.1 16.2% 17.6% 17.7% 17.4% 16.6% 22.4%
ta 16.2% 18.5% 16.6% 13.5% 14.1% 15.3% 14.0% 13.9% 15.1% 14.1% 14.1% 14.9% 18.3%
Average 12.7% 15.6% 12.5% 10.5% 10.6% 11.5% 10.8% 10.5% 11.4% 11.0% 11.0% 11.4% 16.8%
1
Yuan et al. (2023)
Destroy percentile = fraction of the solution that is destroyed in the perturbation step, DRL = Deep Reinforcement Learning, MBP = Memory-Based Perturbation, NonImprov-
ing = Number of non-improving Tabu Search steps before perturbation is triggered.
Table 2: Evaluation results for the MBP operators on the TWT test set: total difference to the optimal tardiness in minutes
averaged over instance group.
Perturbation None (Base) Random MBP - Activity list index MBP - Machine index
NonImproving 10 15 10 15 15
Destroy percentile 30% 30% 10% 30% 50% 10% 30% 50% 10% 30% 50%
abz 66.7 253.8 37.7 58.2 63.8 31.2 66.7 59.3 63.7 63.7 68.8 60.7
la 80.7 173.7 69.0 82.6 80.7 74.4 76.9 74.2 73.4 77.9 78.3 78.0
mt 165.7 197.0 185.3 175.33 162.3 175.3 153.0 175.3 175.3 175.3 168.0 175.3
orb 340.0 576.7 311.9 336.2 333.9 328.3 319.7 332.5 317.7 331.9 327.5 326.7
Average 194.6 355.2 175.7 193.4 191.5 183.9 183.6 188.3 182.1 190.1 187.9 187.6
Destroy percentile = fraction of the solution that is destroyed in the perturbation step, MBP = Memory-Based Perturbation, NonImproving = Number of
non-improving Tabu Search steps before perturbation is triggered.
Table 3: Evaluation results in minutes for the real-world use
case, displayed for the total weighted tardiness (TWT) after
100 iterations per approach.
NonImproving 2 10
Destroy percentile 10% 30% 50% 10% 30% 50%
No perturbation (Base) 107,572
Random perturbation 146,217 146,217 146,217 107,572 107,572 107,572
MBP-Activity list index 76,548 146,217 146,217 107,572 107,572 107,572
MBP-Machine index 102,315 82,663 146,217 79,967 107,572 107,572
Table 4: Evaluation results in minutes for the real-world
use case, displayed for the total weighted order tardiness
(TWOT) after 100 iterations per approach.
NonImproving 2 10
Destroy percentile 10% 30% 50% 10% 30% 50%
No perturbation (Base) 50,328
Random perturbation 50,328 50,328 50,328 50,328 50,328 50,328
MBP-Activity list index 45,027 50,328 50,328 31,322 50,328 50,328
MBP-Machine index 29,898 48,133 30,900 48,133 39,615 50,328
ment of 37%. As it can be seen in Figure 3, also for
this objective all approaches perform equally during
the early phases of the search procedure, however di-
verge in later iterations.
5 CONCLUSIONS
We have proposed a framework to derive perturbation
operators that preserve promising structures of solu-
tion candidates, based on an elite solution pool, col-
lected during a TS procedure. After presenting a va-
riety of possibilities to configure the parameters for
these perturbation operators, we have implemented
two variants and performed numerical experiments on
job-shop benchmark instances from the literature, as
well as on a real-world industry use case. In sev-
eral test runs, we have compared the performance
Figure 2: Objective value path over 100 Tabu Search itera-
tions for the real-world industry case with the objective to
minimize total weighted tardiness (TWT).
Figure 3: Objective value path over 100 Tabu Search itera-
tions for the real-world industry case with the objective to
minimize total weighted order tardiness (TWOT).
of the TS before and after the integration of the de-
rived perturbation operators. We see that the TS in-
cluding the proposed perturbation steps is competi-
Integrating Memory-Based Perturbation Operators into a Tabu Search Algorithm for Real-World Production Scheduling Problems
219
tive with learning-based techniques, such as deep re-
inforcement learning. Due to their ability to capture
the complexity of the problem at hand, those meth-
ods would have been an implementable alternative to
the currently used TS algorithm. However, our re-
sults support the choice for the enhancement of our
metaheuristic approach. Moreover, experiments on
benchmark instances demonstrate how sensitive ran-
dom perturbation operators react to crucial parame-
ters, such as the number of non-improving solutions
before triggering a perturbation. Concerning stability,
the MBP operators have shown a significantly better
behavior.
Most notably, the integration of the MBP oper-
ator leads to significant improvements on the real-
world case study. For both evaluated objective func-
tions, TWT and TWOT, the TS with MBP opera-
tor outperformed the base version without perturba-
tion steps, as well as the random perturbation ap-
proach. In the TWT case, activity tardiness was re-
duced by 29%, while in the TWOT case order tar-
diness was reduced by 40%. This shows, that espe-
cially on large instances the mechanism to preserve
promising solution structures within the perturbation
steps has the potential to increase overall solution per-
formance. Since the presented framework relies on
several different parameters that can be analyzed, we
plan to further investigate the impact of other factors
on the overall solution performance. We will in more
detail evaluate the influence of the neighborhood size,
which determines how many neighbors are evaluated
by the neighborhood operators of the TS. Also other
neighbor selection strategies, such as best improve-
ment, will be compared to the current algorithmic set-
ting. Moreover, we plan to derive additional MBP
operators from the proposed framework, investigating
other evaluation criteria, that could be based, e.g., on
activity starting times, or more sophisticated repair-
mechanisms.
ACKNOWLEDGEMENTS
The research was supported by the Austrian Research
Promotion Agency (FFG) (grant # FO999907709).
This support is gratefully acknowledged.
REFERENCES
Adams, J., Balas, E., and Zawack, D. (1988). The Shifting
Bottleneck Procedure for Job Shop Scheduling. Man-
agement Science, 34(3):391–401.
Applegate, D. and Cook, W. (1991). A Computational
Study of the Job-Shop Scheduling Problem. ORSA
Journal on Computing, 3(2):149–156.
B
¨
ogl, M., Gattinger, A., Knospe, I., Schlenkrich, M., and
Stainko, R. (2021). Real-life scheduling with rich con-
straints and dynamic properties an extendable ap-
proach. Proceedings of the 2nd International Confer-
ence on Industry 4.0 and Smart Manufacturing (ISM
2020), 180:534–544.
Braune, R., Wagner, S., and Affenzeller, M. (2007).
Optimization Methods for Large-Scale Production
Scheduling Problems. In Moreno D
´
ıaz, R., Pichler, F.,
and Quesada Arencibia, A., editors, Computer Aided
Systems Theory EUROCAST 2007, pages 812–819,
Berlin, Heidelberg. Springer Berlin Heidelberg.
Laborie, P., Rogerie, J., Shaw, P., and Vil
´
ım, P. (2018).
IBM ILOG CP optimizer for scheduling. Constraints,
23(2):210–250.
Lawrence, S. (1984). Resouce constrained project schedul-
ing : an experimental investigation of heuristic
scheduling techniques (supplement). Graduate School
of Industrial Administration, Carnegie-Mellon Uni-
versity.
Lunardi, W. T., Birgin, E. G., Ronconi, D. P., and Voos, H.
(2021). Metaheuristics for the online printing shop
scheduling problem. European Journal of Opera-
tional Research, 293(2):419–441.
Moumene, K. and Ferland, J. A. (2009). Activity list
representation for a generalization of the resource-
constrained project scheduling problem. European
Journal of Operational Research, 199(1):46–54.
Ropke, S. and Pisinger, D. (2006). An adaptive large neigh-
borhood search heuristic for the pickup and delivery
problem with time windows. Transportation Science,
40(4):455 – 472.
Schlenkrich, M. and Parragh, S. N. (2023). Solving large
scale industrial production scheduling problems with
complex constraints: an overview of the state-of-the-
art. 4th International Conference on Industry 4.0 and
Smart Manufacturing, 217:1028–1037.
Shaw, P. (1997). A new local search algorithm provid-
ing high quality solutions to vehicle routing problems.
APES Group, Dept of Computer Science, University of
Strathclyde, Glasgow, Scotland, UK, 46.
Singer, M. and Pinedo, M. (1998). A computational study
of branch and bound techniques for minimizing the to-
tal weighted tardiness in job shops. IIE Transactions,
30(2):109–118.
Storer, R. H., Wu, S. D., and Vaccari, R. (1992). New
Search Spaces for Sequencing Problems with Appli-
cation to Job Shop Scheduling. Management Science,
38(10):1495–1509.
Taillard, E. (1993). Benchmarks for basic scheduling prob-
lems. European Journal of Operational Research,
64(2):278–285.
Yuan, E., Cheng, S., Wang, L., Song, S., and Wu, F.
(2023). Solving job shop scheduling problems via
deep reinforcement learning. Applied Soft Computing,
143:110436.
ICORES 2024 - 13th International Conference on Operations Research and Enterprise Systems
220