Multiagent Virtual Community Drill Approach Based on Asynchronous Advantage Actor-Critic Paradigm

In an era of increasingly severe population aging, community emergency rescue has become more and more important. Simultaneously, the widespread application of wearable sensors makes it easier to locate emergency responders. Therefore, the rapid planning of appropriate rescue paths is a crucial issue. In addition, unlike human behavior patterns, the movement of vehicles is more complex, involving two types of continuous actions: driving direction and speed, both of which are easily influenced by road traffic conditions. In this occasion, a virtual community drill for emergency rescue approach is put forward, specifically, this approach utilizes virtual community scenario to simulate the community traffic environment in the real world and discretizes the community environment and the action space of vehicles. Meanwhile, a multi-agent path planning method based on Asynchronous Advantage Actor-Critic (A3C) paradigm is proposed, which is consisted of the main decision model and the sub-decision model. In addition, a complete experience rollout buffer is presented. Subsequently, the results clearly demonstrate that the heterogeneous parallel A3C network architecture and the holistic road evaluation method undoubtedly beneficiate to faster and more stable model convergence. In comparison to Deep Q-Network (DQN) and traditional path planning methods, the proposed approach exhibits significant advantages in terms of relevant evaluation metrics.