An Multi-Agent Reinforcement Learning-Based Adaptive Multi-Objective Evolutionary Algorithm

Multi-objective evolutionary algorithms (MOEAs) have demonstrated significant success in solving multi-objective optimization problems (MOPs). However, their performance is highly sensitive to the configuration of algorithmic hyperparameters, which often requires manual tuning and domain expertise. To address this limitation, we propose a multi-agent reinforcement learning (MARL)-based multi-objective evolutionary algorithm based on decomposition (MMOEA/D) that enables joint adaptive configuration of multiple hyperparameter types, including weight vector, neighborhood size, and operator. The proposed method formulates the dynamic selection of hyperparameters as a fully observable multi-agent Markov decision process (MMDP). Each agent independently learns to adjust a specific type of hyperparameter based on a carefully designed state representation. A Long Short-Term Memory (LSTM) network is employed as the core of each agent’s Q-network to model temporal dependencies and facilitate more informed decision-making. The proposed MMOEA/D is evaluated on a set of widely used multi-objective benchmark problems, and its performance is compared with several existing methods. Experimental results demonstrate that MMOEA/D exhibits strong competitiveness.