[Purposes] This paper investigates the effectiveness of multi-agent reinforcement learning algorithms for maintenance optimization of multi-component production system. The feasibility of applying domain knowledge of maintenance optimization in reinforcement learning is also studied. [Methods] The maintenance decision making process of the production system was modeled as a Markov decision process (MDP), which was solved by a shaped reward distributed Q-learning（SR-DQL）algorithm. The domain knowledge of maintenance optimization was introduced into reinforcement learning by designing parameters of agents and reward shaping. [Findings] The proposed methods were validated using a production system with five production units and four inventory buffers. The proposed SR-DQL algorithm had a 6% ehancement of average revenuse comparing with the commonly used Q-learning. SR-DQL also outperformed distributed Q-learning and deep reinforcement learning algorithms. [Conclusions] The SR-DQL algorithm can effectively deal with the maintenance optimization problem of large-scale production systems, and reward shaping can improve the performance of the reinforcement learning algorithm.