Multi-component Maintenance Optimization: an Approach combining Genetic Algorithm and Multiagent Reinforcement Learning

Abstract

The maintenance strategy optimization of the systems with intermediate buffers is a typical maintenance optimization. As the number of components in the system is increased, the state space and action space of the maintenance optimization of the manufacturing system with buffer inventory increase exponentially. Multi-agent reinforcement learning is an effective method to optimize the maintenance decision making of large multi-component system. However, as the number of agents increases, the reward function of multi-agent reinforcement learning tends to be complicated, and each agent will receive a noisy reward signal, so it is difficult for multi-agent reinforcement learning to converge to the optimal strategy. Considering the excellent global optimization ability of genetic algorithm, this paper adopts genetic algorithm as the central unit to guide the decision-making of each agent, and establishes a bilateral interaction mechanism between multi-agent reinforcement learning and genetic algorithm, through which both genetic algorithm and multi-agent reinforcement learning can learn the solutions provided by the other party. Numerical research results show that the proposed method is superior to multi-agent reinforcement learning and genetic algorithm in terms of solution quality.