Path Planning for Multi-Agent System in Games via Behavior Prediction

In this paper, the path planning problem for the unmanned aerial vehicles based on the multi-agent system in games is addressed, under conditions of obstacle avoidance, collision avoidance, dynamical model uncertainties, and input constraints. In the game, there exist friendly agents and opposing agents, and friendly agents can catch up with the opposing agents by path planning. A long short term memory model by an attention mechanism is proposed to predict the behavior of the opposing agents, assisting the subsequent path planning for friendly agents. A value function involving a constrained control integral item is constructed to transform the path planning problem into an optimal control problem with input constraints based on the Hamilton-Jacobi-Bellman equation. An integral reinforcement learning method and a Bellman update equation based on the historical data without the knowledge of model parameters are designed to solve the Hamilton-Jacobi-Bellman equation. Simulation results are provided to verify the effectiveness of the proposed method.