Multi-Agent Swarm Control Based on LSTM-MATD3

Firstly, this study constructed a typical multi-agent simulation task scenario for autonomous obstacle avoidance, swarm formation, and target point arrival tasks of multiple unmanned aerial vehicles. Secondly, based on the Multi Agent Double Delay Deep Deterministic Policy Gradient Algorithm (MATD3), a long short term memory Recurrent Neural Network (LSTM) is introduced to solve partially observable reinforcement learning problems, enhancing the learning ability of time series sample data and improving the efficiency of multiple drones performing obstacle avoidance and clustering tasks as a group in complex environments. In addition, inspired by the idea of artificial potential field method, a dense reward function was designed to enable the agent to receive positive feedback even when the optimal solution was not found. Finally, in the simulation experiment, this study designed a global reward function and a maximum collision penalty, which will enable the agent to fully explore the map and greatly accelerate network training, with certain application prospects.