Deep Reinforcement Learning Based Dynamic Scheduling of Random Arrival Tasks in Cloud Manufacturing

Abstract

Compared with the stable orders of traditional manufacturing, cloud manufacturing (CMfg) fulfilled with masses of random orders, so the CMfg server needs an algorithm with low time and space complexity to prevent the server from crashing due to excessive instantaneous data. Besides, the random changes of manufacturing resources and service must be considered when establishing a scheduling model for CMfg. To solve this problem, we propose an adaptive Deep Q-Networks (ADQN) method with a resizable network that converts cloud manufacturing scheduling problems with multiple objectives into specific reinforcement learning goal and can adapt to changing environments. Our experimental results show that ADQN is comparable to other real-time scheduling methods, the average subtask completion time and the standard deviation of occupation obtained by ADQN keep at a low level.