Solving the flow-shop scheduling problem with human factors and two competing agents with deep reinforcement learning

Abstract

With the development of globalized production, customer satisfaction has become a critical concern for companies. Multi-agent models have gained attention, aiming to meet customer demands while reducing costs. The bi-agent flow-shop scheduling problem (BAFSP) in the manufacturing industry is the main focus of this article, where agents represent different customers. The objective of BAFSP is to minimize the total weighted makespan for agents. The BAFSP introduces various position-dependent learning effects. Furthermore, each task has an independent release date, aligned with real production scenarios. A deep reinforcement learning (DRL) approach based on the transformer architecture is proposed to address the BAFSP with learning effects (BAFSP-LE). Experimental results demonstrate that the proposed method outperforms other commonly used heuristics. To explore the feasibility of combining DRL with metaheuristics, the proposed DRL method is used as an initial solution generator. The method improves the performance of these metaheuristics within the same computation time.