Deep Reinforcement Learning Driven Strategy Design for Distributed Consensus Black-Box Evolutionary Optimization

Abstract

With the development of communication networks and smart terminals, distributed systems are crucial in critical domains like intelligent transportation and industrial internet. However, traditional distributed evolutionary algorithms struggle to address optimization demands in complex dynamic environments, as they are constrained by static parameter configurations and rigid communication topologies. To solve this, this paper introduces a novel mechanism where each agent autonomously learns its own evolutionary strategy via reinforcement learning (RL), enabling the automatic selection of learning targets and the adaptive adjustment of their weights. This mechanism is seamlessly integrated into the Multi-Agent Swarm Optimization with Internal and External Learning (MASOIE) framework, giving rise to RL-MASOIE. Experiments on heterogeneous benchmarks show RL-MASOIE outperforms the original MASOIE and existing black-box distributed algorithms in average fitness, with excellent robustness under diverse network topologies.