Real-time production scheduling using a deep reinforcement learning-based multi-agent approach

Abstract

In the real-time scheduling (RTS) research field, it has been shown that employing multiple dispatching rules (MDRs) for the components in a flexible manufacturing system will improve production performance much more than a single dispatching rule (SDR). To fulfill the goal of Industry 4.0 in production control and improve production performance, this study deploys a deep reinforcement learning-based multi-agent (DRLBMA) approach for real-time scheduling. The proposed approach uses the MDRs strategy by integrating two main methodologies: an off-line learning module and a Deep Q-learning-based multi-agent module. The proposed method employs a two-level self-organizing map (SOM) to determine the system’s states. The proposed methodology determines the best MDRs decision. The approach is applied to a case study of a smart manufacturing system. The results of the proposed method are compared to different scheduling strategies, such as reinforcement learning (RL)-based real-time scheduling, two-level self-organized map (SOM), and continuous rescheduling using a single dispatching rule, such as earliest due date (EDD), shortest processing time (SPT), and longest processing time (LPT). The findings reveal that, in terms of the total weighted tardiness, throughput, and mean cycle time performance criteria, the proposed DRLBMA-based real-time scheduling is more efficient than these scheduling strategies.