Solving Traveling Salesman Problems Using Parallel Environments in Reinforcement Learning

Shaohua Hu, Ming Zhu, Tenghai Qiu, Zhiqiang Pu, Xiaolin Ai

Keywords: Traveling Salesman Problems, Parallel Environments, Reinforcement Learning

TL;DR: Solve Traveling Salesman Problems Using Parallel Environments in Reinforcement Learning

Abstract: The traveling salesman problems (TSP) are NP-hard and difficult to solve since the search space increases significantly with problem size. Reinforcement learning (RL) is a promising method for its powerful search abilities with the help of GPUs. However, sampling speed is a bottleneck since training requires a large number of samples, and current methods ignore this issue. We propose to use GPU-based parallel environments to increase the sampling speed in RL; furthermore, we use powerful neural networks transformers with self-attention to enhance the policy so that the long-distance topology can be learned. The experimental results demonstrate the advantages of GPU-based parallel environments in terms of sampling speed and objective values. The code is available at: \href{https://github.com/zhumingpassional/RLSolver}{https://github.com/zhumingpassional/RLSolver}.

Submission Number: 17