Search

Parameter setting is a significant and demanding process in metaheuristics. It requires deep understanding of the considered algorithm as well as of the considered optimization problem. Thus, it has attracted ongoing scientific interest counting a considerable number of relevant approaches, which are mostly focused in offline and ad hoc procedures. Interestingly, the parameter tuners contain also their own parameters. It is a basic requirement that these parameters shall be easily tunable and mildly affecting the tuner's performance. The present work proposes an enhanced version of the recently proposed gradient-based parameter adaptation method, which offers an online general-purpose parameter tuning procedure that is applicable to any metaheuristic. The proposed method dynamically tunes the tuner's parameters by using the REINFORCE reinforcement learning paradigm. The method is demonstrated on the differential evolution algorithm for established test suites of variable dimensions as well as for the real-world application of detecting atomic clusters through the minimization of Lennard-Jones potentials. The obtained results are statistically analyzed and compared with state-of-the-art auto-tuning algorithms. The analysis reveals the promising performance of the proposed approach as well as its robustness under the considered experimental setting. [ABSTRACT FROM AUTHOR]

Abstract Metaheuristics have been established as effective solvers for challenging optimization problems. However, their performance is highly dependent on their parameter settings. For this reason, various parameter tuning techniques have been developed, spanning two major categories: online and offline techniques. Online techniques are based on performance data of the algorithm during its run, while offline techniques are based on preprocessing or historical performance data of the algorithm. Alternatively to these techniques, we propose a general online parameter adaptation method based on estimations of the algorithm's performance and gradient search in the parameter domain. The proposed method is demonstrated on Differential Evolution, a state-of-the-art metaheuristic for continuous optimization. Our experimental validation includes problems of low and high dimension as well as comparisons with distinguished adaptive algorithms. The obtained results suggest that the proposed approach is beneficial, relieving the user from the burden of proper parameterization. Graphical abstract Highlights • Metaheuristics can be effective solvers under proper parameterization. • We propose GPALS, a general online parameter adaptation method for metaheurtistics. • GPALS is based on approximate gradient search and line search in parameter domain. • The method is demonstrated on Differential Evolution algorithm for two test suites. • Significant benefits for the user and the algorithm are gained. [ABSTRACT FROM AUTHOR]

Evolutionary algorithms have been long established as an essential field of research in Computational Intelligence. Differential Evolution is placed among the most successful algorithms of this type. However, it has proved to be highly sensitive on its parameters. For this purpose, offline and online parameter-control methods have been proposed. Recently, a grid-based parameter adaptation procedure was introduced and successfully applied on Differential Evolution. Despite the generality of the method, the resulting algorithm was capable to compete with already tuned adaptive algorithms on high-dimensional test problems without any additional preprocessing. The present work extends the experimental study of this approach on the state-of-the-art CEC-2013 test suite. Two variants of the algorithm are considered and different initial conditions are tested to shed light on its performance aspects. Comparisons with other algorithms as well as between the proposed approaches are reported. The results verify the potential of the grid-based parameter adaptation method as a general-purpose alternative for parameter setting. [ABSTRACT FROM AUTHOR]