Novel Dynamic Decomposition-Based Multi-objective Evolutionary Algorithm Using Reinforcement Learning Adaptive Operator Selection (DMOEA/D-SL)

Abstract

Within the multi-objective (static) optimizationfield, various works related to the adaptive selection ofgenetic operators can be found. These include multiarmedbandit-based methods and probability-basedmethods. For dynamic multi-objective optimization,finding this type of work is very difficult. The maincharacteristic of dynamic multi-objective optimization isthat its problems do not remain static over time; on thecontrary, its objective functions and constraints changeover time. Adaptive operator selection is responsible forselecting the best variation operator at a given timewithin a multi-objective evolutionary algorithm process.This work proposes incorporating a new adaptiveoperator selection method into a Dynamic MultiobjectiveEvolutionary Algorithm Based onDecomposition algorithm, which we call DMOEA/D-SL.This new adaptive operator selection method is basedon a reinforcement learning algorithm called State-Action-Reward-State-Action Lambda or SARSA (λ).SARSA Lambda trains an Agent in an environment tomake sequential decisions and learn to maximize anaccumulated reward over time; in this case, select thebest operator at a given moment. Eight dynamic multiobjectivebenchmark problems have been used toevaluate algorithm performance as test instances. Eachproblem produces five Pareto fronts. Three metrics wereused: Inverted Generational Distance, GeneralizedSpread, and Hypervolume. The non-parametricstatistical test of Wilcoxon was applied with a statisticalsignificance level of 5% to validate the results.