Skip to Main Content
A class of beliefs-desires-intentions (BDI) agent architecture is presented for control systems with a high degree of autonomy. The architecture contains agents for modelling, controller optimization, implementation and to monitor performance. The global convergence of performance of the agent system is proven under three mild assumptions. Relevant features of the agent structure are competing modellers and controllers. The benefit is an enhanced ability to learn new plant dynamics of varying complexity and controller adaptation. The new family of control agent architectures is called cautiously optimistic, a name to reflect the most important property of the new architecture: modelling results are applied with caution for control but current models are accepted until measurements do not contradict them with a margin. A cautiously optimistic control agent (COCA) is proven to have converging performance to a nearly optimal performance for stationary dynamics of a real plant under fairly general assumptions.