Abstract:

In this paper,a collaborative dynamic target assignment and path planning algorithm based on Multi-Agent Deep Reinforcement Learning (MADRL) is proposed to address the cooperative path planning of multi-CUAV for multi-dynamic target point monitoring.Specifically,the communication system environment for multi-CUAV-to-ground is modeled,and the weighted sum of the total completion time and total interruption time of multi-CUAV minimization optimization is formulated which is subject to the communication connectivity constraints with the cellular network,the collision avoidance constraints,etc.Then the formulated optimization problem is discretized and transformed into the Markov game.To tackle the formulated combinatorial optimization problem,Deep Deterministic Policy Gradient (DDPG) series algorithms are proposed to effectively handle the cooperative and competitive behaviors among multi-CUAV.Meanwhile,a shared experience pool is introduced to improve the stability and learning efficiency of the proposed algorithm.The feasibility,effectiveness,and robustness of the proposed MADRL framework are demonstrated by numerical simulations,which indicates that the multi-agent twin delayed deep deterministic policy gradient algorithm (MATD3) exhibits a superior performance for cooperative target assignment and path planning with dynamic target locations.

Key words: unmanned aerial vehicle, reinforcement learning, cellular-connected unmanned aerial vehicle, multi-agent deep reinforcement learning, path planning, dynamic target assignment