Multi-agent reinforcement learning-aided evolutionary algorithm for a many-objective distributed hybrid flow shop scheduling problem

Abstract

The distributed hybrid flow shop scheduling problem (DHFSP) is common in real-world production environments and is typically constrained by factors such as time-of-use electricity tariffs, due dates, and worker assignments. The complexity of these factors often requires decision-makers to consider multiple optimization objectives simultaneously. Consequently, this paper investigates a many-objective DHFSP and establishes the corresponding mathematical model. To efficiently solve the model, a multi-agent reinforcement learning-aided evolutionary algorithm (MRLEA) is developed. In MRLEA, a grid-based evolutionary framework is introduced to enhance the selection pressure of non-dominated solutions while maintaining their diversity. Meanwhile, an intelligent local search process based on multi-agent group decision-making is integrated into the evolutionary framework to improve convergence speed and overcome local optima. Specifically, each agent corresponds to an optimization objective and must engage in collaborative reinforcement learning with other agents to adaptively assign local search strategies for each solution. After executing the local search, the rewards for agents that make crucial contributions to the final decision are updated using the state-of-the-art Reward Centering technique. Additionally, a dominance judgment is made between the old and new solutions in the corresponding objective dimensions of these agents to retain the more promising solutions. Comprehensive experiments are conducted to validate the superior performance of MRLEA in addressing the proposed many-objective DHFSP.