A deep reinforcement learning based solution for flexible job shop scheduling problem

Abstract

Flexible job shop Scheduling problem (FJSP) is a classic problem in combinatorial optimization and a very common form of organization in a real production environment. Traditional approaches for FJSP are ill-suited to deal with complex and changeable production environments. Based on 3D disjunctive graph dispatching, this work proposes an end-to-end deep reinforcement learning (DRL) framework. In this framework, a modified pointer network, which consists of an encoder and a decoder, is adopted to encode the operations to be scheduled according to the selected scheduling features. Then with the attention mechanism, an input is pointed as an action in each decoding step, and a recurrent neural network (RNN) is used to model the decoder network. To train the network to minimize the makespan, a policy gradient algorithm is applied to optimize its parameters. The trained model generates the scheduling solution as a sequence of consecutive actions in real-time without retraining for every new problem instance. Experimental results show that this method can obtain better performance than the classic heuristic rules when only one model is trained on all the test instances.