A Bidding-Based Deep Reinforcement Learning Approach for Multi-Agent Job Shop Scheduling Problem
提出一种基于竞价的深度强化学习方法，通过智能选择器和竞标者协调消费者偏好，实现多智能体作业车间调度的高效资源分配，实验表明在大规模实例中能获得高社会福利。
In demand-driven personalized production, multi-agent job shop scheduling problem plays a pivotal role. Balancing the private preferences between consumers poses a challenge in achieving efficient resource allocation. A bidding-based deep reinforcement learning approach is proposed to generate a consensus schedule. An intelligent selector and an intelligent bidder (IB) are designed for each consumer to perform operation selection and determine the corresponding bid price, respectively. A job shop agent is established to collect bids and allocate resource through winner determination. To assist the IB to learn the correlation between consumers and the multi-agent job shop scheduling environment without revealing preferences, a graph neural network is adopted to extract observations. A reward function based on critic value guides the negotiation process among IBs. Extensive computational experiments demonstrate that the proposed approach achieves high social welfare outcomes. It outperforms in handling large-scale instances with more than 11 consumers and 20 jobs per consumer.