Deep Reinforcement Learning for Multiobjective Optimization

This article proposes an end-to-end framework for solving multiobjective optimization problems (MOPs) using deep reinforcement learning (DRL), that we call DRL-based algorithm (DRL-MOA). The idea of decomposition is adopted to decompose the MOP into a set scalar subproblems. Then, each subproblem modeled as neural network. Model parameters all subproblems are optimized collaboratively according neighborhood-based parameter-transfer strategy and DRL training algorithm. Pareto-optimal solutions can be directly obtained through trained neural-network models. Specifically, traveling salesman problem (MOTSP) solved in this DRL-MOA method by modeling Pointer Network. Extensive experiments have been conducted study various benchmark methods compared with it. It found once model available, it scale newly encountered no need retraining model. simple forward calculation network; thereby, iteration required always reasonable time. proposed provides new way means DRL. has shown characteristics, example, strong generalization ability fast speed comparison existing optimizations. experimental results show effectiveness competitiveness terms performance running