Resumen
The multi-objective optimization (MOO) of complex systems remains a challenging task in engineering domains. The methodological approach of applying MOO algorithms to simulation-enabled models has established itself as a standard. Despite increasing in computational power, the effectiveness and efficiency of such algorithms, i.e., their ability to identify as many Pareto-optimal solutions as possible with as few simulation samples as possible, plays a decisive role. However, the question of which class of MOO algorithms is most effective or efficient with respect to which class of problems has not yet been resolved. To tackle this performance problem, hybrid optimization algorithms that combine multiple elementary search strategies have been proposed. Despite their potential, no systematic approach for selecting and combining elementary Pareto search strategies has yet been suggested. In this paper, we propose an approach for designing hybrid MOO algorithms that uses reinforcement learning (RL) techniques to train an intelligent agent for dynamically selecting and combining elementary MOO search strategies. We present both the fundamental RL-Based Hybrid MOO (RLhybMOO) methodology and an exemplary implementation applied to mathematical test functions. The results indicate a significant performance gain of intelligent agents over elementary and static hybrid search strategies, highlighting their ability to effectively and efficiently select algorithms.