Neural Network Based Learning Methods For Solving Single-Machine Scheduling Problem

As one of the most important scheduling models,the single-machine scheduling problem has a wide practical background in real life.In some cases,many multi-machine scheduling and complex scheduling problems can be often decomposed into several single-machine scheduling problems.So it is deserved for further developing efficient scheduling algorithm to improve production efficiency.Many single-machine scheduling problems have been proved to be NP-hard.Based on its good adaptability,stability,self-learning ability and generalization ability,neural network based learning methods have effective applications in many fields.In this paper,the learning algorithms based Pointer Networks framework are selected to investigate the problem of single machine scheduling.Both supervised learning and reinforcement learning policy are used to test the effectiveness,learning ability and the generalization ability of the algorithm.Experimental simulation shows that the proposed algorithm can provide a better approximate optimal solution for the single machine scheduling problem,and that it remains good adaptability on the large-scale problem in combination with the proposed rolling optimization strategy.Summarily,the works in this thesis mainly include three aspects as follows:?A supervised learning model is developed based on Pointer Networks with LSTM network to model the single machine scheduling problem,and the Lagrange relaxation method is applied to build the data set required for training.Numerous experimental simulations are conducted to verify the effectiveness,learning ability and the generalization ability of the algorithm.?To improve the generalization ability of the algorithm on large scale scheduling problem,a rolling optimization strategy is proposed in combination with the supervised learning model proposed in the previous chapter,which decomposes the large-scale scheduling problem into a small scale scheduling problem.Simulation results shows that the generalization ability on large scale scheduling problem has been greatly improved of the modified algorithm.?Considering the supervised learning algorithm relies heavily on the quality of training data,we further introduce the reinforcement learning policy and establishes an unsupervised learning model for the single-machine scheduling problems.Preliminary experiment simulations show that improvements can be achieved compared with the supervised model.