An Actor-Critic Based Path Planning Method for the Traveling Salesman Problem

This paper introduces AC-TSP, an Actor-Critic reinforcement learning model designed to solve the Traveling Salesman Problem (TSP). The model employs an encoderdecoder architecture, where the encoder leverages 1D convolutional layers (Conv1D) and positional encoding to extract node features, and the decoder integrates a gated recurrent unit (GRU) with an attention mechanism to generate feasible routes sequentially. Although trained solely on instances with 50 nodes, AC-TSP exhibits strong generalization capabilities across varying problem sizes. Compared to conventional baseline algorithms, AC-TSP consistently delivers high-quality solutions at multiple scales while significantly enhancing computational efficiency.