A Novel Solution to JSPs Based on Long Short-Term Memory and Policy Gradient Algorithm
Jing Ren, Chengkai Ye, F. Yang
Abstract
Based on long short-term memory (LSTM) and policy gradient algorithm, this paper proposes a novel solution to the job-shop scheduling problems (JSPs). Firstly, two LSTM networks with identical structures were established, serving as the encoding and decoding networks, respectively. Next, a pointer network was introduced to determine the job with the highest priority in the current state, creating a job sequence. Another neural network (NN) was constructed to evaluate the current job sequence. The evaluation results were taken as the baseline of the policy gradient algorithm for reinforcement learning. Then, the job sequence was optimized and updated by gradient descent. The effectiveness of our method was demonstrated through contrastive experiments on benchmark problems.