Learning-based design of fault-tolerant cooperative multi-agent systems

This paper addresses fault-tolerance of multi-agent systems. Modeling the behaviors of each agent in the system as a local automaton and the system behaviors as the composition of them, the fault-tolerance property requires that the system can satisfy a global regular language specification prior to as well as after the occurrence of a fault. In particular, this paper concerns with sensor and actuator failures occurring in individual agents, in which the former are captured as loss of observability of a local event of a failed agent, and the latter are modeled as a total loss of an event from an agent. Should either failure occur, the corresponding local supervisors need to be redesigned and the team needs to be reconfigured as the original task decomposition scheme and controllers would fail to satisfy the specification. For such a pursuit, we propose frameworks for addressing sensor and/or actuator failure tolerance by incorporating a learning-based supervisor synthesis approach and control reconfiguration mechanism in the face of possible failures. Illustrative examples are also presented to show the frameworks´ effectiveness.