Integrating Reinforcement Learning and Optimization Task: Evaluating an Agent to Dynamically Select PSO Communication Topology.

A recent study catalogued hundreds of meta-heuristics proposed over the past three decades in Swarm Intelligence (SI) literature. This scenario makes it difficult for the practitioner to choose the most suitable meta-heuristic (RL) for a specific problem. This paper shows that Reinforcement Learning could be a powerful tool for SI. First, we describe a Reinforcement Learning environment to solve an optimization problem. Then, we investigate the usage of Proximal Policy Optimization to dynamically set the Particle Swarm Optimization topology accordingly to the simulation states. Our RL proposal reached competitive fitness values, even when evaluated in non-trained scenarios. In addition, we show the actions’ distribution by simulation in the Rastrigin. The paper demonstrates how RL could be integrated to improve meta-heuristics capabilities, opening new research paths where RL will be used to improve meta-heuristics or select them accordingly to their strengths.