Petri Net Modeling and Multi-Agent Dynamic Path Planning for AGV Scheduling System

In this paper, a modeling method of Petri net is proposed, which models AGV and its environment as Petri net and joins the monitoring place to prevent AGV conflict. On this basis, a non-cooperative multi-agent deep reinforcement learning framework is proposed based on the deep deterministic strategy gradient algorithm. The AGV path planning problem is regarded as a partially observable Markov decision process and the neural network is trained based on the corresponding reinforcement learning. In addition, the positive and negative feedback triggered by Petri Net is added into the reward function to reduce congestion in AGV system, and the junction node is treated as an agent instead of an AGV, which reduces the state space of the system and improves the generalization ability of the neural network. The experiment was carried out in AnyLogic simulation software, and the effectiveness and feasibility of the path planning were verified through the control experiment with or without Petri net constraint in the reward function.