Motivations for the Development of a Multi-objective Algorithm
Configurator
Nguyen Thi Thanh Dang and Patrick De Causmaecker
K.U. Leuven KULAK, CODeS, Etienne Sabbelaan 53, 8500 Kortrijk, Belgium
Keywords: Automated (Automatic) Algorithm Configuration, Multi-objective Optimization.
Abstract: In the single-objective automated algorithm configuration problem, given an algorithm with a set of
parameters that need to be configured and a distribution of problem instances, the automated algorithm
configurator will try to search for a good parameter configuration based on a pre-defined performance
measure. In this paper, we point out two motivations for the development of a multi-objective algorithm
configurator, in which more than one performance measure are considered at the same time. The first
motivation is a parameter configuration case study for a deterministic single machine scheduling algorithm
with two performance measures: minimization of the average running time and maximization of the total
number of optimal solutions. The second one is the configuration problem for non-exact multi-objective
optimization algorithms. In addition, a discussion of solving approach for the first motivating problem is
also presented.
1 INTRODUCTION
The automated algorithm configuration problem,
sometimes called parameter tuning, can be
informally stated as follows: given an algorithm A
(target algorithm) with a list of parameters, a set of
problem instances B and a performance measure for
the evaluation of A's quality on B; the automated
algorithm configurator (or simply configurator) will
try to find a good parameter configuration of A, i.e.,
an assignment of A's parameters into specific values,
in such a way that the performance measure's value
on B is optimized. The problem instance set B could
be generated from a problem's distribution and its
size is usually chosen to be large to well represent
the problem. Since this set is large, the exact
evaluation of the performance measure over B often
incurs very high computational costs. Therefore, the
configurator normally deals with a subset of B,
called training instance set, during the configuring
procedure. To avoid over-fitting, this training set
certainly should be sufficiently large and widely
spread over the problem distribution, then, the final
parameter configuration obtained is tested on
another problem instance set, named test instance
set, to evaluate the efficiency of the configurator.
The test set could be a subset of B, or even the same
as B.
The automated algorithm configuration problem
could be considered as an optimization problem, in
which the search space is the set of possible
parameter configurations and the objective function
is the performance measure over the training
instance set. This optimization problem presents
three challenges. The evaluation of each solution's
quality is often very expensive. This evaluation
could be stochastic if the target algorithm is not
deterministic. And finally, different types of
variables (target algorithm's parameters) could exist
in the same configuration problem. To be more
specific, an algorithm's parameter could be
continuous, integer or categorical. A parameter
could also be conditional, i.e., its activation could
depend on specific values of some other parameter.
The automated algorithm configuration problem
has received large attention in the last decade. A vast
number of solving approaches, originating from
various fields, have been proposed in recent years.
According to Stutzle et al., 2013, they could be
classified into four groups: approaches from the
experimental design community, sequential
statistical testing techniques, model-based
optimization approaches and metaheuristics. Several
other works have also been devoted to a literature
review of the algorithm configuration topic, such as
Hoos, 2012. Among these methods, irace (López-
328
Dang N. and De Causmaecker P..
Motivations for the Development of a Multi-objective Algorithm Configurator.
DOI: 10.5220/0004925203280333
In Proceedings of the 3rd International Conference on Operations Research and Enterprise Systems (ICORES-2014), pages 328-333
ISBN: 978-989-758-017-8
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
Ibánez et al., 2011, a statistical testing technique),
Gender-based Genetic Algorithm (Ansótegui et al.,
2009), ParamILS ((Hutter et al., 2009), a
metaheuristic based on Iterated Local Search) and
SMAC (Hutter et al 2011, a model-based
optimization method) have shown remarkable results
on several hard algorithm configuration problems
with large numbers of parameters. As an example,
ParamILS has been used to configure the CPLEX
solver with 76 parameters and has achieved a
speedup ratio up to 50 over CPLEX’s default
parameter configuration as well as configurations
got from CPLEX’s automatic tuning tool (Hutter et
al 2010). Several other academic applications of
ParamILS have also been reported on the website of
its authors: http://www.cs.ubc.ca/labs/beta/Projects/
ParamILS/applications.html.
Solving approaches for the automated algorithm
configuration problem so far are single-objective,
i.e., only one performance measure is considered
during the configuring stage. In this paper, we
describe two motivations for the development of a
multi-objective algorithm configurator, in which
multiple performance measures are considered at the
same time. The first one is an algorithm
configuration case study for a single machine
scheduling algorithm, with the requirement of
balancing between target algorithm’s robustness and
speed, the second one is the case when the target
algorithm is a non-exact multi-objective
optimization algorithm.
Please note that there are some works in the
evolutionary computation community that are
dedicated to a Multi-Objective Tuning Problem
(MOTP), such as the tuner BONESA proposed in
Smit et al., 2011. However, MOTP is different from
the multi-objective algorithm configuration
considered in this paper. In MOTP, the problem
instance set is just a limited and small set of
continuous optimization functions, so that it does not
have to deal with the problem of configuration’s
expensive evaluation. The main goal of MOTP is
trying to balance algorithm configurations’
performance over all optimized functions, i.e.,
finding robust algorithm configurations, while the
multi-objective algorithm configuration problem
focus on solving the problem of over-fitting: trying
to deal with a quite large training instance set.
This paper is organized as follows: section 2
describes the first motivation and a proposal of
solving approach, section 3 presents the second
motivation and section 4 draws some conclusions
and future works.
2 THE FIRST MOTIVATION
2.1 Problem Description
The target algorithm considered in this problem is
SiPSi, a single machine scheduling algorithm
proposed in Tanaka et al., 2012. The algorithm and
relevant problem instances are downloadable from
Tanaka’s website (http://turbine.kuee.kyoto-
u.ac.jp/~tanaka/index-e.html). SiPSi is applied on
different single machine scheduling instances with
various sizes (the number of jobs). This algorithm is
single-objective, exact and deterministic. It has 25
parameters, including 14 continuous parameters, 7
integer parameters and 4 categorical parameters.
Among them, 11 parameters are conditional.
In this case study, a pre-defined cut-off time is
specified for each problem size. The running of
SiPSi on a problem instance is stopped when either
an optimal solution is found or the running time
reaches the relevant cut-off time. The goal of the
algorithm configuration task is to find parameter
configurations that have small running time and
large number of optimally solved instances over the
whole instance set. The specification of cut-off time
for each problem size is two-fold: first, when some
researcher develops an algorithm for solving a well-
known problem, some so-far-state-of-the-art method
for that problem usually exists and the researcher
wants to “beat” the performance of this state-of-the-
art algorithm, then problem-size-dependent cut-off
time is an appropriate choice; second, if we are
interested in configuring algorithm for very large
problem’s sizes, e.g., from 500 to thousands of jobs,
a configuring procedure directly deals with a
training instance set of such large sizes is
impossible, in that case, problem instance set with
smaller sizes in which cut-off time is gradually
increased according to problem size can be helpful,
since we can try the configuring procedure on this
alternative problem instance set, in the hope that the
obtained parameter configuration is robust enough to
work well on even unseen and larger problem
instance sizes.
During Tanaka’s manual tuning procedure, he
observed a possible trade-off between the two
mentioned performance measures: “A safe
parameter configuration performs well for many
types of instances, but the algorithm becomes slow,
on the other hand, a tuned set of parameters
improves the speed of the algorithm, but the
performance deteriorates considerably (and
sometimes the algorithm fails to find an optimal
solution) for some specific instances due to
MotivationsfortheDevelopmentofaMulti-objectiveAlgorithmConfigurator
329
parameter sensitivity”. Such an example is shown in
table 1. In this experimental example, the first
parameter configuration, dubbed c
default
, is the
currently default configuration of SiPSi. The second
one, denoted by c
tuned
, is a configuration obtained by
using ParamILS (Hutter et al., 2009), with a
weighted aggregation function of the two
performance measures. Results are tested on 5975
problem instances. From table 1, we can see the
trade-off between the total average running time
over all problem sizes and the total number of
optimally solved instances. The total number of
optimal solutions gained from c
default
is smaller than
c
tuned
, but the average running time results of c
default
are better and also statistically different throughout
the non-parametric pairwise Wilcoxon test with
significance level 0.05.
Table 1: Summarized performance of two configurations.
#optimal
solutions
#total average
runtime (s)
c
default
5963
603.2
c
tune
d
5970
693.1
From this analysis, the motivation for a multi-
objective automated algorithm configurator becomes
obvious. As being seen in table 1, when we consider
the two performance measures’ values over the
whole test set, incomparable parameter
configurations do exist. Therefore, result obtained
from such a multi-objective algorithm configurator
is a Pareto-approximation set of parameter
configurations. The algorithm designer will do a
manually deeper analysis on the resulting
configurations, such as consideration of average
running time on each problem sizes, the minimum
and maximum running time, the average optimality
gap, etc, to choose the final best algorithm
configuration for future usage.
2.2 Solving Approach
In this part, we discuss our initial idea for a solution
approach. Taking advantage of remarkable
techniques currently available for single-objective
algorithm configuration problem, we try to adapt
them into the multi-objective context. Among the
four groups of single-objective configurators, we
decide to focus on the group of metaheuristics, due
to their significant results on several hard algorithm
configuration problems.
In order to design a solving approach for this
case study, we firstly discuss the question of which
metaheuristic should be chosen. An analysis of
challenges when applying that metaheuristic to our
multi-objective algorithm configuration problem as
well as our proposed solution approach is presented.
Finally, addition of preference information into our
configurator is mentioned.
2.2.1 Choosing a Metaheuristic
Since we are solving an algorithm configuration
problem, chosen metaheuristic should be parameter-
independent as much as possible. We keep this
condition in mind during the search for an
appropriate multi-objective metaheuristic approach.
Multi-objective metaheuristics can be divided
into three major groups: aggregation-based, Pareto-
dominance-based and indicator-based methods
(Basseur et al., 2012). Among them, we decide to
focus on the aggregation-based group. The reason
for our choice is the impossibility of using adaptive
capping strategies in the search procedure when the
two later groups are used. We explain capping
strategies in the next sub-section.
In aggregation-based approaches, all objective-
functions are integrated into a single function
throughout some scalarization technique, in which
various sets of weight for each objective function are
considered sequentially, each set will guide the
search towards a point of optimal Pareto front. The
main advantage of these approaches is that if solving
methods for single-objective version of the
considered problem are already available, we can re-
use them to solve the multi-objective problem in a
fairly straight forward manner. However, these
techniques have a major drawback when the single-
objective algorithm requires expensive
computational cost, since each run of this algorithm
just aims to only one point in the true Pareto front.
Algorithms in this group usually require parameters
for specifying running time amount for each
scalarization. Algorithms’ performance is influenced
by these parameters. To avoid such a parameter
sensitivity, we choose the Adaptive Anytime Two-
Phase Local Search (AA-TPLS) proposed in
Dubois-Lacoste et al 2011, in which these running
time amount values are adaptively chosen based on
information collected from previous runs.
The underlying single-objective algorithm
configurator used in AA-TPLS could be either
ParamILS or GGA, since such a choice is
independent of the general framework of AA-TPLS.
2.2.2 Adaptation: Challenges and Solutions
In this subsection, we firstly describe a group of
techniques, named adaptive capping, for solving the
ICORES2014-InternationalConferenceonOperationsResearchandEnterpriseSystems
330
expensive evaluation problem; then, we discuss
challenges faced in adaptations of the chosen
metaheuristics into our configuration problem, as
well as our initial ideas for solving them.
a) Capping Techniques. The most difficult
challenge in solving algorithm configuration
problems is the expensive evaluation of an algorithm
configuration’s quality, due to the usual large size of
the training instance set. In ParamILS by Hutter et al
2009, special techniques, named adaptive capping
methods, have been developed to reduce the
computational effort required for this task. Given
two parameter configuration c
1
and c
2
, a problem
instance set S and a performance measure f that
needs to be minimized, assume that c
1
has been run
on S already and its performance value is known.
Now we want to compare c
1
and c
2
and discard the
worse one. The idea of adaptive capping is to try to
not run c
2
on the whole set S, but stopping earlier at
some instance according to information collected
before.
Here we briefly describe the two adaptive
capping techniques proposed by Hutter et al 2009,
namely trajectory-preserving and aggressive
capping, and propose some additional adaptive
capping methods to improve the reduction of
computational cost.
Trajectory-preserving: Let us assume that the
performance measure f is the average running
time over the whole instance set, N is the total
number of problem instances in S, t
1
and t
2
are
the total running time of c
1
and c
2
, respectively,
on S (t
1
= f
1
x N). When we run c
2
on S, if the
total running time value so far is larger than t
1
,
we can stop the run immediately and conclude
that c
2
is worse than c
1
, instead of running c
2
on
the whole set S.
Aggressive capping: this is a heuristic version of
trajectory-preserving capping, in which the total
running time t
best
of the best configuration so far
c
best
is saved and T x t
best
will serve as an upper
bound for the running of other configurations.
Here T is a predefined ratio (T = 2 in ParamILS’s
setting). When a configuration c
i
is run on S, if t
i
got so far exceeds the upper bound, c
i
’s run is
stopped.
Additional capping methods: In our case study,
the target algorithm is deterministic. We do not
have to deal with the difficulty of stochastic
evaluation. Hence, another capping idea is that
for each parameter configuration c and each
problem instance p, the performance of c on p is
saved when its running time exceeds a pre-
defined threshold (we do not save all cases for
reasons of memory size). This information is
useful when c is revisited. Moreover, from this
information, we can also roughly predict
performance of a configuration c
i
on p if c
i
’s
neighbors have been run on p so far. The
predicted value could be a weighted sum of c
i
’s
neighbors’ performance values on p, whereas the
weight is smaller for larger dissimilarity between
p and the neighbor. For a good prediction, it is
reasonable to have an upper bound for the
dissimilarity between c
i
and its considered
neighbors.
b) Adaptation of AA-TPLS. First, we briefly
describe general ideas of the AA-TPLS in
Dubois-Lacoste et al 2011. The AA-TPLS
algorithm has two phases. In the first phase, the
available single-objective algorithm is separately
applied on a randomly chosen objective function
to conduct an initial solution for the next phase.
(Application on all objective functions is also
possible. In this case, the best solution according
to the first scalarization in phase 2 will be taken.)
Then, in the second phase, a sequence of
scalarizations will be called; the single-objective
algorithm is used to solve the problem at each
scalarization. Solution obtained from a
scalarization will be used as the initial solution
for its succeeding scalarization. Non-dominated
solutions are also saved in an archive. The set of
weight for each objective function in each
scalarization is adaptively determined based on
solutions from previous scalarization to
guarantee the coverage the true-Pareto front.
The challenge when applying AA-TPLS to our
problem lies in the algorithm’s sequential
implementation. Each step in phase 2 just provides
only one solution, so that if the single-objective
algorithm requires too large running time, it might
be impossible to get a reasonable number of Pareto-
optimal solutions. In the literature of algorithm
configuration techniques, for problem with large
number of parameters (more than 24), the
appropriate amount of each configurator’s run
should be at least 10 hours. Therefore, the sequential
implementation of AA-TPLS should be modified in
order to deal with such an expensive computational
running time. Here we propose a parallelization of
AA-TPLS’s phase 2 as depicted in Figure 1. We still
keep a part of the sequential property due to its
important role: adapting weight values in
scalarizations based on information from previous
steps to ensure the diversity of the obtained
approximation front. The new scheme of phase 2 has
MotivationsfortheDevelopmentofaMulti-objectiveAlgorithmConfigurator
331
k sequential steps. At each step, n different
scalarizations are run in parallel. In order to guide
the search better, we can let them cooperate with
each other: after some time interval, the best
solutions obtained so far of every parallel
scalarization are put into a global archive, then each
scalarization will check this archive for a better
solution than the best one found by itself, based on
its currently used set of weight, if such a solution
exists, it will be updated into the scalarization
searching process.
It is worth noting that besides the usage of
capping strategies mentioned in previous part,
performance value (for each objective function) of
the best solution found so far on every run problem
instances will also be saved, and the order of
training instances considered in every parallel
scalarization at the same step are fixed. Thanks to
that, when the best configuration c from some
scalarization s
i
r
is considered for entering some other
scalarization s
j
r
, the computational cost of evaluating
c according to the s
j
r
’s set of weight will be lower,
since we just need to run c on problem instances in
s
i
r
that have not been saved yet.
Besides, order of instances in the training set also
influences ParamILS’s or GGA’s results. Since at
each iteration, just a subset of the training set is used
for configuring in order to save computational cost,
this subset is extended gradually after time, so that
the more iterations the algorithm runs, the more
exact the estimation of performance measure value
on the whole training set is. Therefore, in this AA-
TPLS, to deal with such a variance, order of training
instances are randomly shuffled after each step of
phase 2.
Figure 1: Parallel scheme of AA-TPLS’s phase 2.
2.2.3 Integration of Preference Information
Trade-off between two considered performance
measures is possible. For example, given two
configurations c
1
and c
2
, if the percentage of
corresponding optimal solutions are respectively
90% and 70%, while the total average running time
of c
1
is just some second less than c
2
, c
1
is certainly
preferred. Consequently, extreme points towards the
average running time axis are not interesting and
could be ignored. Moreover, configurations with
small optimality gap for hard unsolved instances are
also preferable. Integration of such preference
information into the configuration algorithm is
useful to reduce the computational effort and make a
better matching with algorithm designer’s desire.
We are in presently collaborating with Tanaka, the
author of the target algorithm SiPSi, to identify
characteristics of this integration.
3 THE SECOND MOTIVATION
In this section, we explain the second algorithm
configuration problem supporting the need for a
multi-objective algorithm configurator: the case
when the target algorithm is a non-exact Multi-
Objective Optimization Algorithm (MOOA).
Given a MOOA configuration c and problem
instance p, the result obtained from a run of c on p is
an approximation set of the optimal Pareto front. In
order to compare two MOOA configurations based
on p, we can use different performance assessment
methods, including unary indicators, binary
indicators, statistical testing on multiple-run and
attainment function analysis. We refer to Ziztler et
al., 2008 for a comprehensive description. Among
these methods, the unary indicator approaches, in
which a real-value is assigned to each approximation
set to define a total order in the objective function
space, are the ones that usually require the lowest
computational cost. However, each indicator has its
own preference and might bias towards some
specific parts the true Pareto front. As suggested in
Knowles et at., 2006: “Each unary indicator is based
on different preference information - therefore using
them all will provide more information than using
just one”. Indeed, different unary indicators could
give inconsistent conclusion, as observed in the
experimental results of Jaeggi et al 2008. Moreover
preference of some unary indicators have been
theoretically shown, as e.g. in the analysis of
hypervolume indicator in Auger et al., 2012 and of
the R2 indicator in Brockhoff et al., 2012. Hence,
the configuration task for a MOOA could be done as
follows: first, a multi-objective algorithm
configurator is used to find a set of good parameter
configurations based on a set of unary indicators as
performance measures; then, a post-processing step
starts, in which a final best algorithm configuration
is decided based on more expensive assessment
methods, e.g., statistical testing over several runs
with different indicators.
Scalarization
1
1
Scalarization
n
1
Scalarization
1
k
Scalarization
n
k
Step 1
Step k
.
.
.
ICORES2014-InternationalConferenceonOperationsResearchandEnterpriseSystems
332
4 CONCLUSIONS
In this paper, we present two algorithm
configuration problems in which more than one
performance measures should be considered at the
same time, leading to a multi-objective algorithm
configurator. Development of the solving approach
is still in progress. We are currently focusing on the
first problem: balancing between algorithm
configuration’s robustness and speed, while the
second one is reserved for our future work. Although
several single-objective configurators with
significant performance have been proposed in the
literature, the application in a multi-objective
context is quite challenging. However, we believe
that such an effort is worthwhile. The consideration
of more than one algorithm performance measure
during the automated configuration process and the
postponement of the final choice to a deeper analysis
in a post-processing phase will give more flexibility
to the algorithm designer. Indeed, the definition of a
good algorithm configuration in practice usually
depends on various performance perspectives.
ACKNOWLEDGEMENTS
The authors would like to thank Professor Thomas
Stutzle for his valuable comments. This work is
supported by the Belgian Science Policy Office
(BELSPO) in the Interuniversity Attraction Pole
COMEX. (http://comex.ulb.ac.be) and by Research
Foundation Flanders (FWO).
REFERENCES
Ansótegui, C., Sellmann, M., & Tierney, K. (2009). A
gender-based genetic algorithm for the automatic
configuration of algorithms. In Principles and
Practice of Constraint Programming-CP 2009 (pp.
142-157). Springer Berlin Heidelberg.
Basseur, M., Zeng, R. Q., & Hao, J. K. (2012).
Hypervolume-based multi-objective local search.
Neural Computing and Applications, 21(8), 1917-
1929.
Brockhoff, D., Wagner, T., & Trautmann, H. (2012). On
the Properties of the R2 Indicator. In Proceedings of
the fourteenth international conference on Genetic and
evolutionary computation conference (pp. 465-472).
ACM.
Dubois-Lacoste, J., López-Ibáñez, M., & Stützle, T.
(2011). Improving the anytime behavior of two-phase
local search. Annals of mathematics and artificial
intelligence, 61(2), 125-154.
Hoos, H. H. (2012). Automated algorithm configuration
and parameter tuning. In Autonomous Search (pp. 37-
71). Springer Berlin Heidelberg.
Hutter, F., Hoos, H. H., Leyton-Brown, K., & Stützle, T.
(2009). ParamILS: an automatic algorithm
configuration framework. Journal of Artificial
Intelligence Research, 36(1), 267-306.
Hutter, F., Hoos, H. H., & Leyton-Brown, K. (2010).
Automated configuration of mixed integer
programming solvers. In Integration of AI and OR
Techniques in Constraint Programming for
Combinatorial Optimization Problems (pp. 186-202).
Springer Berlin Heidelberg.
Hutter, F., Hoos, H. H., & Leyton-Brown, K. (2011).
Sequential model-based optimization for general
algorithm configuration. In Learning and Intelligent
Optimization (pp. 507-523). Springer Berlin
Heidelberg.
Jaeggi, D. M., Parks, G. T., Kipouros, T., & Clarkson, P.
J. (2008). The development of a multi-objective tabu
search algorithm for continuous optimisation
problems. European Journal of Operational Research,
185(3), 1192-121
Knowles, J., Thiele, L., & Zitzler, E. (2006). A tutorial on
the performance assessment of stochastic
multiobjective optimizers. Tik report, 214, 327-332.
López-Ibánez, M., Dubois-Lacoste, J., Stützle, T., &
Birattari, M. (2011). The irace package, iterated race
for automatic algorithm configuration. IRIDIA,
Université Libre de Bruxelles, Belgium, Tech. Rep.
TR/IRIDIA/2011-004.
Smit, S. K., & Eiben, A. E. (2011). Multi-problem
parameter tuning using BONESA. In Artificial
Evolution (pp. 222-233).
Stützle, T., & López-Ibáñez, M. (2013, July). Automatic
(offline) configuration of algorithms. In Proceeding of
the fifteenth annual conference companion on Genetic
and evolutionary computation conference companion
(pp. 893-918). ACM.
Tanaka, S., & Fujikuma, S. (2012). A dynamic-
programming-based exact algorithm for general
single-machine scheduling with machine idle time.
Journal of Scheduling, 15(3), 347-361.
Zitzler, E., Knowles, J., & Thiele, L. (2008). Quality
assessment of pareto set approximations. In
Multiobjective Optimization (pp. 373-404). Springer
Berlin Heidelberg.
MotivationsfortheDevelopmentofaMulti-objectiveAlgorithmConfigurator
333