Eficient Configuration of Optimization Algorithms Marcelo de Souza Univ. do Estado de Santa Catarina Univ. Federal do Rio Grande do Sul marcelo.desouza@udesc.br Marcus Ritt Instituto de Informática Univ. Federal do Rio Grande do Sul marcus.ritt@inf.ufrgs.br Manuel López-Ibáñez Universidad de Málaga University of Manchester manuel.lopez-ibanez@uma.es ABSTRACT We propose a set of capping methods to speed-up the automatic con- fguration of optimization algorithms. First, we build a performance envelope based on previous executions of known confgurations, which defnes the minimum required performance for new confg- urations. Then, we use the performance envelope to evaluate new confgurations, stopping poor performers early. We propose difer- ent methods to aggregate previous executions into a performance envelope, and evaluate them on several confguration scenarios. The proposed methods produce solutions of the same or better qual- ity as confguring without capping, but reduce the efort required to confgure optimization algorithms. This manuscript for the Hot-Of-the-Press track at GECCO 2022 is based on the article łCapping methods for the automatic confguration of optimization algorithmsž, published in Computers & Operations Research [2]. CCS CONCEPTS · Computing methodologies → Search methodologies; · Math- ematics of computing → Combinatorial optimization; · Applied computing → Operations research. KEYWORDS Automatic algorithm confguration, parameter tuning, capping ACM Reference Format: Marcelo de Souza, Marcus Ritt, and Manuel López-Ibáñez. 2022. Efcient Confguration of Optimization Algorithms. In Genetic and Evolutionary Computation Conference Companion (GECCO ’22 Companion), July 9ś13, 2022, Boston, MA, USA. ACM, New York, NY, USA, 2 pages. https://doi.org/ 10.1145/3520304.3534078 1 INTRODUCTION The algorithm confguration task aims at fnding parameter con- fgurations that optimize the expected performance of a target algorithm on a set of problem instances. Automatic confguration methods eliminate potential biases of manual search approaches and make the confguration process reproducible. A widely used algorithm confgurator is irace [4], which iteratively samples a set of confgurations and evaluates them on a subset of the problem instances using racing. Confgurations that perform statistically worse than others are eliminated during the racing procedure. The Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for proft or commercial advantage and that copies bear this notice and the full citation on the frst page. Copyrights for third-party components of this work must be honored. For all other uses, contact the owner/author(s). GECCO ’22 Companion, July 9ś13, 2022, Boston, MA, USA © 2022 Copyright held by the owner/author(s). ACM ISBN 978-1-4503-9268-6/22/07. https://doi.org/10.1145/3520304.3534078 Previous executions (performance profles) Aggregation methods Profle-based envelope  max Area-based envelope Figure 1: Overview of the capping methods. surviving (elite) confgurations are used to update the probabilistic models and guide the sampling of new confgurations in subsequent iterations. The confguration process is costly, since several parameter con- fgurations are evaluated on diferent problem instances. When confguring decision algorithms, the confguration time can be reduced by determining a bound on the running time of new con- fgurations based on the time used by the best confguration found so far. Existing capping methods based on this idea [3, 5] are not suitable when confguring optimization algorithms, since the per- formance of the algorithm is given by the value of the best found solution. Hence, we propose a set of capping methods for confgur- ing optimization algorithms, which assume an anytime behavior and compute a performance envelope based on previous executions. The envelope is used to evaluate new executions, stopping poor performers early. 2 PROPOSED CAPPING METHODS Figure 1 shows the main ideas behind the proposed capping meth- ods. We represent an execution by its performance profle, i.e. the cost of the best found solution over the execution time (or any other measure of computational efort). Before evaluating a confguration on a given instance, the performance profles of previous executions on that instance are aggregated into the performance envelope. In the profle-based approach, the envelope is a performance profle that defnes the maximum allowed solution cost (considering a minimization problem) for each value of time. In the area-based approach, the area below the performance profle of each previous execution is computed, and these area values are aggregated into a maximum allowed area  max for new executions. For both profle- and area-based approaches, we propose two strategies for envelope construction. In the adaptive strategy, all non-capped previous executions are considered, and the (profle- or area-based) envelope is determined based on a user-defned aggressiveness parameter. We build an envelope that would cap the desired amount of the previous non-capped executions. After each 17