Strategic Control of Reactive Behavior in Intelligent Agents

Abstract : A series of experiments was conducted to evaluate the real-time performance of a proposed agent architecture. The architecture is a blackboard architecture whose key features include: distribution of perception, action, and cognition among parallel processes; limited-capacity I/O buffers with best-first retrieval and worst-first overflow; dynamic control planning; dynamic focus of attention; and a satisficing algorithm for the execution cycle. The experiments focus on the architecture's satisficing algorithm for the execution cycle, which is the unit-process of all reasoning and critical to real-time performance. The experiments focus on the architecture's satisficing algorithm for the execution cycle, which is the unit-process of all reasoning and critical to real-time performance. The execution cycle has three steps: (1) notice possible operations; (2) choose the best operation with respect to the current control plan; and (3) execute the chosen operation. Because the executed operations can change the agent's control plan, this cycle allow the agent to dynamically construct and modify plans that control its own behavior. The problem with this cycle in a real-time context is the unbounded time associated with step 1.