Direct Policy Search Method in Fault Tolerant Autonomous Systems

This work is concerned with a new type of realtime reconfigurable control systems that is based on the use of a multi-agent system. To this end, two stages have been examined in the context of decision making; the fault detection and identification (FDI) stage and the reconfiguration stage (RC). The agent based FDI detects that a fault has occurred. It then further diagnoses the situation. The RC stage follows this by adapting or changing the control architecture to accommodate the fault. The agent based problem is to synchronize or integrate these two stages in the overall structure of a control system on real world applications. The multi-agent architecture proposed in this paper has several advantages in terms of modularity, reliability, ability to learn and achieve overall higher robustness over past “single software” methods. Specifically, this paper concentrates on one of the agents introduced, namely the Reconfiguration agent. Tests on a simulation of an example system have been carried out to demonstrate this new organisation of reconfigurable control systems.