Solving Multi-objective Optimization Problems Via Selective Inheritance Meta-Reinforcement Learning

An increasing number of machine learning algorithms are being applied to multi-objective optimization problems (MOPs), yielding promising results. However, many of these algorithms suffer from inefficient and low-quality task sampling methods, which degrade the quality of the Pareto Front (PF) calculated by the meta-model after several iterations. Our proposed algorithm integrates meta-learning techniques into the deep reinforcement learning (DRL) framework. To generate a high-quality and stable training dataset, we incorporate the concept of optimal genetic inheritance into the training process of the meta-model. The process involves: (1) saving the training data in a specified data structure; (2) selecting high-performance data based on loss values and storing them in a structure indexed by iteration number; (3) updating the data structure with each iteration; and (4) generating the training dataset based on iteration counts. Our algorithm demonstrates higher stability and better performance in computational experiments, effectively addressing multi-objective traveling salesman problems (MOTSP) and multi-objective vehicle routing problems with time windows (MOVRPTW).