Research on Multi-Agent Task Allocation and Path Planning Based on Pri-MADDPG

In this paper, we aim to develop a reinforcement learning (RL) based algorithm for the task allocation and path planning problem of multi-agent systems where all agents autonomously head to task points with obstacle avoidance. To address the challenge of slow convergence speed and insufficient reward setting when using traditional RL methods, the named Pri-MADDPG algorithm based on prioritized experience replay is proposed. By integrating task allocation and path planning problem, we first construct a framework for multi-agent reinforcement learning training by designing essential elements including appropriate observation space, action space, and reward functions. Then a prioritized experience replay method, in which the value network loss is employed for the priority evaluation, is utilized to enhance policy learning performance. A reward mechanism is further improved through taking into consideration of both global task objectives and individual objectives. To verify the effectiveness of Pri-MADDPG algorithm, experiments are finally carried out with the well-designed reward mechanism. The results demonstrate that all agents can autonomously accomplish task allocation with smooth and highly safe trajectories while achieving faster convergence speed, better stability, and superior performance.