基于人工势场法的无人机集群突防博弈研究OA

Targeting the challenges of high-dimensional continuous decision non-convergence,low exploration efficiency,and insufficient policy robustness in multi-to-multi red-blue quadrotor swarm zero-sum penetration games within three-dimensional airspace,this paper proposes a Multi-Agent Deep Deterministic Policy Gradient(MADDPG)solution framework integrating artificial potential field priors with opponent strategy prediction.First,the three-degree-of-freedom UAV kinematics are embedded into a complete-information differential game,designing a"mission-threat-cooperation"three-tier reward structure,and introducing differentiable potential field energy to transform sparse terminal rewards into dense gradient signals,achieving explicit representation of the"seeking-advantage-avoiding-disadvantage"prior.Second,a potential field-guided hybrid exploration mechanism is constructed,online modulating Ornstein-Uhlenbeck process(OU)noise using potential energy directions,and offline smoothing target Q-values with potential field regularization,improving sample utilization and suppressing overestimation.Furthermore,a lightweight opponent strategy predictor is integrated,introducing a meta-game term into the Actor gradient,enabling red-team policy updates to simultaneously minimize opponent expected payoffs,proactively disrupting enemy decision consistency and accelerating convergence to Nash equilibrium.Simulation results demonstrate that the proposed method achieves stable win rates exceeding 90%in 2v2 and 4v4 dense confrontations,systematically induces blue team to generate redundant accelerations and energy dissipation,continuously creates spatial-temporal gaps to complete collision-free penetration,significantly outperforming MADDPG without prediction,validating the framework's scalability,real-time performance,and robustness in multi-to-multi zero-sum games.