A Graph Pointer Network-Based Multi-Objective Deep Reinforcement Learning Algorithm for Solving the Traveling Salesman Problem

Traveling Salesman Problems (TSPs) have been a long-lasting interesting challenge to researchers in different areas. The difficulty of such problems scales up further when multiple objectives are considered concurrently. Plenty work evolutionary algorithms has introduced solve multi-objective TSPs with promising results, and the deep learning reinforcement surging. This paper introduces graph pointer network-based (MODGRL) algorithm for TSPs. MODGRL improves an earlier algorithm, called DRL-MOA, by utilizing network learn graphical structures Such improvements allow be trained on small-scale TSP, but can find optimal solutions large scale NSGA-II, MOEA/D SPEA2 selected compare DRL-MOA. Hypervolume, spread coverage over Pareto front (CPF) quality indicators were assess algorithms’ performance. In terms hypervolume indicator that represents convergence diversity Pareto-frontiers, outperformed all competitors three well-known benchmark problems. findings proved MODGRL, improved network, indeed performed better, measured indicator, than DRL-MOA other algorithms. comparable leading group, indicator. Although better both them just average regarding evenness CPF remind performance measure Pareto-frontiers from perspectives. Choosing well-accepted suitable one’s experimental design is very critical, may affect conclusions. Three also experimented extra iterations, validate whether iterations affected results show NSGA-II greatly Spread indicators. raise fairness concerns comparisons using fixed stopping criteria algorithms, which appeared many others. Through these lessons, we concluded hypervolumne, urge fair designs comparisons, order derive scientifically sound