Uncovering Reward Goals in Distributed Drone Swarms Using Physics-Informed Multiagent Inverse Reinforcement Learning
提出一种物理信息多智能体逆强化学习方法，从观测数据推断无人机集群的控制目标函数和网络拓扑，帮助理解集群行为并用于经验推断和模仿学习。
The cooperative nature of drone swarms poses risks in the smooth operation of services and the security of national facilities. The control objective of the swarm is, in most cases, occluded due to the complex behaviors observed in each drone. It is paramount to understand which is the control objective of the swarm, whilst understanding better how they communicate with each other to achieve the desired task. To solve these issues, this article proposes a physics-informed multiagent inverse reinforcement learning (PI-MAIRL) that: 1) infers the control objective function or reward function from observational data and 2) uncover the network topology by exploiting a physics-informed model of the dynamics of each drone. The combined contribution enables to understand better the behavior of the swarm, whilst enabling the inference of its objective for experience inference and imitation learning. A physically uncoupled swarm scenario is considered in this study. The incorporation of the physics-informed element allows to obtain an algorithm that is computationally more efficient than model-free IRL algorithms. Convergence of the proposed approach is verified using Lyapunov recursions on a global Riccati equation. Simulation studies are carried out to show the benefits and challenges of the approach.