Integral reinforcement learning-based optimal pursuit-evasion control for multi-QUAVs: A zero-sum game approach

Abstract

This paper investigates the optimal pursuit-evasion control (PEC) under zero-sum game for multiple quadrotor unmanned aerial vehicles (multi-QUAVs) with unknown dynamics, aiming to capture an evader QUAV (EQUAV). First, the pursuit-evasion error dynamics are constructed based on multi-pursuer QUAVs (multi-PQUAVs) and an EQUAV. Second, within the framework of a zero-sum game, adversarial strategies are designed for both the multi-PQUAVs and the EQUAV. By minimizing the cost function, the multi-PQUAVs aim to minimize the pursuit-evasion error, while the EQUAV seeks to maximize pursuit-evasion error. Third, an actor-critic neural network (NN) based on integral reinforcement learning (IRL) is developed to optimize the adversarial strategies of both the multi-PQUAVs and the EQUAV while updating their strategies toward the optimal approximate solution. The stability analysis demonstrates that the pursuit-evasion error, critic NN weight error, PQUAVs’ actor NN weight error, and EQUAV’s actor NN weight error are uniformly ultimately bounded. Finally, simulations validate the effectiveness and adaptability of the IRL-based optimal PEC algorithm under zero-sum game.