Evacuation method for multi-agent based on deep reinforcement learning
作者
Lingjun Liu,Youdong Ding
出处
期刊:日期:2024-03-15卷期号:: 1534-1539被引量:2
标识
DOI:10.1109/iaeac59436.2024.10503701
摘要
This paper proposes a MIXPPO (Mixed Proximal Policy Optimization) method based on deep reinforcement learning to address the evacuation task of multiple agents in a dynamic environment. In this work, we integrating four modules: observation data normalization, advantage function normalization, learning rate decay and orthogonalization of the weight matrix for the policy network to improve the Proximal Policy Optimization (PPO) algorithm. These improvements enable the model to generate safe and effective evacuation strategies. The experiments indicate that in simulated scenarios, our approach can effectively guide multiple agents to move towards exits while avoiding collisions with each other, offering a novel solution to the multi-agent evacuation problem.