CONVOLUTION TWO-DIMENSIONAL REINFORCEMENT LEARNING BASED JOB-SHOP SCHEDULING

摘要

At present, how to use the reinforcement learning method to solve the NP complete problem is a discussion hotspot. This paper proposes a convolution two-dimensional reinforcement learning method (CTRLM) to effectively improve the neural network performance. This method makes it possible to introduce weight sharing and deep scheduling information extracting in reinforcement learning framework. Thereby enhancing the accuracy and speed of the scheduler. This transformation method can effectively explore the coupling relationship. The experiment result shows that a trained network with CTRLM can achieve the 90% of the optimal solution with extreme speed. The effect is obviously better than the same type of ANN-RL method and the other traditional method.