A Deep Reinforcement Learning Approach to the Flexible Flowshop Scheduling Problem with Makespan Minimization

Recent work has demonstrated the efficiency of deep reinforcement learning (DRL) in making optimization decisions in complex systems. Compared with other DRL algorithms, the proximal policy optimization (PPO) has higher stability and lower complexity. The typical flexible flowshop scheduling problem (FFSP) with identical parallel machines is an NP-hard problem. This paper is the first case to utilize PPO to solve the problem with makespan minimization. The particular state, action and reward function are designed for the FFSP to follow the Markov property. The efficiency of PPO is evaluated on the wafer pickling instance and random instances with different scales. The results show that PPO can always provide satisfactory solutions within a reasonable computational time.