Resumen

BRAMBILA-HERNANDEZ, José Alfredo et al. Novel Dynamic Decomposition-Based Multi-Objective Evolutionary Algorithm Using Reinforcement Learning Adaptive Operator Selection (DMOEA/D-SL). Comp. y Sist. [online]. 2024, vol.28, n.2, pp.739-749.  Epub 31-Oct-2024. ISSN 2007-9737.  https://doi.org/10.13053/cys-28-2-5018.

Within the multi-objective (static) optimization field, various works related to the adaptive selection of genetic operators can be found. These include multi-armed bandit-based methods and probability-based methods. For dynamic multi-objective optimization, finding this type of work is very difficult. The main characteristic of dynamic multi-objective optimization is that its problems do not remain static over time; on the contrary, its objective functions and constraints change over time. Adaptive operator selection is responsible for selecting the best variation operator at a given time within a multi-objective evolutionary algorithm process. This work proposes incorporating a new adaptive operator selection method into a Dynamic Multi-objective Evolutionary Algorithm Based on Decomposition algorithm, which we call DMOEA/D-SL. This new adaptive operator selection method is based on a reinforcement learning algorithm called State-Action-Reward-State-Action Lambda or SARSA (λ). SARSA Lambda trains an Agent in an environment to make sequential decisions and learn to maximize an accumulated reward over time; in this case, select the best operator at a given moment. Eight dynamic multi-objective benchmark problems have been used to evaluate algorithm performance as test instances. Each problem produces five Pareto fronts. Three metrics were used: Inverted Generational Distance, Generalized Spread, and Hypervolume. The non-parametric statistical test of Wilcoxon was applied with a statistical significance level of 5% to validate the results.

Palabras llave : Adaptive; operator; selection; dynamic; multi-objective; optimization.

        · texto en Inglés     · Inglés ( pdf )