GGTAN: Graph Gated Talking-Heads Attention Networks for Traveling Salesman Problem

Traveling Salesman Problem (TSP) is one of the most typical NP-hard combinatorial optimization problems with a variety of real-life applications. In this paper, we propose a Graph Gated Talking-Heads Attention Networks (GGTAN) trained with reinforcement learning (RL) for tackling TSP. GGTAN can learn characteristic structure information better by introducing talking-heads attention mechanism and a gated convolutional sub-network, which make hidden information moving across between attention heads and control each attention head's importance respectively, unlike recently proposed models which use attention mechanism for solving TSP. Experimental results on TSP up to 100 nodes demonstrate that our model obtains shorter tour lengths than other learning-based methods under the same solve strategy for problem instances of fixed graph sizes, and achieves better generalization on variable graph sizes compared with recent state-of-the-art models on the optimality gap.