Improved Deep Reinforcement Learning Algorithm Based on Pointer Network for Traveling Salesman Problem
Traveling salesman problem is a classic problem in combinatorial optimization. The development of deep rein-forcement learning provides a new way to solve this problem. In the deep reinforcement learning algorithm based on the point-er network for the traveling salesman problem, the encoders of the strategy network and the value network both employ the complex long short-term memory network structure, which leads a long training time to the large-scale traveling salesman problem. Considering the independence of the position order among the input nodes, this paper modifies the recurrent neural network of the encoder in the pointer network and replaces the long short-term memory network of encoders in the strategy network and the value network with the one-dimensional convolutional neural network. An improved deep reinforcement learning algorithm based on the pointer network is proposed, which reduces the training time by 12％to 15％compared with the original model on the same scale of resolving the problem. The experimental results verify the effectiveness of the im-proved algorithm.
国家科技期刊平台
NSTL
万方数据
