A Multi-agent Deep Reinforcement Learning-Based Collaborative Willingness Network for Automobile Maintenance Service

With the growth of maintenance market scale of automobile manufacturing enterprises, simple information technology is not enough to solve the problem of uneven resource allocation and low customer satisfaction in maintenance chain services. To solve this problem, this paper abstracts the automotive maintenance collaborative service into a multi-agent collaborative model based on the decentralized partially observable Markov decision progress (Dec-POMDP). Based on this model, a multi-agent deep reinforcement learning algorithm based on collaborative willingness network (CWN-MADRL) is presented. The algorithm uses a value decomposition based MADRL framework, adds a collaborative willingness network based on the original action value network of the agent, and uses the attention mechanism to improve the impact of the collaboration between agents on the action decision-making, while saving computing resources. The evaluation results show that, our CWN-MADRL algorithm can converge quickly, learn effective task recommendation strategies, and achieve better system performance compared with other benchmark algorithms.