Resilient Output Formation-Containment Tracking of Heterogeneous Multi-Agent Systems: A Learning-Based Framework Using Dynamic Data

This paper investigates a resilient output formation-containment tracking (FCT) problem for heterogeneous multi-agent systems (MASs) under unknown dynamics and uncertainties. A learning-based control framework using online dynamic data is proposed with three hierarchical phases. First, fully distributed observers for agents with various types of objectives are presented under a directed graph. The estimations of tracking reference and time-varying formation are coordinated in terms of both dynamics and states. Second, dynamic data filters based on the internal model principle and partial observations are introduced to reconstruct the MASs information and formulate a virtual tracking system, where the reinforcement learning (RL) technique is applied. Based on two proposed off-policy schemes, the RL algorithm is adapted to a hybrid form under the dynamic data. An ideal tracking controller is uniformly learned and essential dynamics are extracted from the same data. Third, the integrated resilient output FCT controller is further derived using previous learning results. The adaptive neural networks and compensation functions are utilized in a data-driven manner to address unknown faults and uncertainties. The integration of filtering, estimation, and learning broadens a more general control framework than existing results. Finally, validations are demonstrated by numerical simulations.