Knowledge-Based Anytime Computation

Abstract

This paper describes a real-time decision-making model that combines the expressiveness and flexibility of knowledge-based systems with the real-time advantages of anytime algorithms. Anytime algorithms offer a simple means by which an intelligent system can trade off computation time for quality of results. Previous attempts to develop knowledge-based anytime algorithms failed to produce consistent, predictable improvement of quality over time. Without performance profiles, that describe the output quality as a function of time, it is hard to exploit the flexibility of anytime algorithms. The model of progressive reasoning that is presented here is based on a hierarchy of reasoning units that allow for gradual improvement of decision quality in a predictable manner. The result is an important step towards the application of knowledge-based systems in time-critical domains.