MULTI-AGENT PATH PLANNING BASED ON IMPROVED ASYNCHRONOUS ACTOR-CRITIC AGENT ALGORITHM

Existing approaches often rely on centralized or reactive planning methods, which can become inefficient or cause deadlocks in complex environments. There is a clear need for a process that handles these challenges while maintaining real-time performance and robustness. We propose a multi-agent path planning method based on the improved Asynchronous Actor-Critic Agent (A3C) algorithm. First, we enhance the Actor-Critic neural network within the A3C framework by integrating it with the VGG network to develop a fully decentralized strategy. Additionally, we improved the reinforcement learning rewards and penalties, allowing agents to perform real-time reactive path planning with partially visible information, exhibiting implicit coordination. Second, we use the sequence information input characteristics and long-term memory capabilities of Long Short Term Memory (LSTM) neural networks, allowing the neural network to gain "long-term memory" and enhance network learning speed. The LSTM neural network replaces the deep neural networks (DNN) in the original A3C network. We tested this approach in various environments with different sizes, agent quantities, and obstacle densities. Finally, we quantified the results based on average planning length, average planning time, and success rates over 100 tests. The experimental results show that the proposed method significantly improves the success rate and efficiency of multi-agent path planning in noisy and uncertain environments.