Reference: Collinot, A. & Hayes-Roth, B. Real-Time Control of Reasoning: Experiments with Two Control Models. KSL, March, 1990.
Abstract: An intelligent agent must identify and perform logically correct actions in response to external events, and it must perform them at appropriate times. The top-level objective of such an agent would be to maximize (or ensure a lower bound on) the value of some global utility function that integrates the values of its responses to events, weighted by the importance of those events, over time. In this paper, we focus on four properties that might facilitate an agent's achievement of its global utility objective: selectivity, responsiveness, robustness, and scalability. We assume a very general agent architecture, and we focus on its reasoning component. As opposed to a best- next control model we propose a satisficing control model for the reasoning process. We have conducted preliminary experiments to test the following hypotheses: the satisficing model provides good selectivity, responsiveness, robustness, and scalability (both when measured against the best-next model and when measured in absolute terms); therefore, the satisficing model provides a high global utility for the agent's performance. The results of our experiments confirm our hypotheses.