Efficient Resource-Bounded Reasoning in AT-RALPH

Abstract

Anytime algorithms have attracted growing attention in recent years as a key mechanism for implementing models of bounded rationality. The main problem, however, as with planning systems in general, is the integration of the modules and their interface with the other components of the system. We have implemented a prototype of ATRALPH (Anytime Rational Agent with Limited Performance Hardware) in which an offline compilation process together with a runtime monitoring component guarantee the optimal allocation of time to the anytime algorithms. The crucial meta-level knowledge is kept in the anytime library in the form of conditional performance profiles. These are extensions of an earlier notion of performance description – they characterize the performance of each elementary anytime algorithm as a function of run-time and input quality. This information, used by the compiler to produce the performance profile of the complete system, is also used by the runtime system to measure the value of computation and monitor the execution of the top-level procedure in the context of a particular domain. The result is an efficient and cheap meta-level control for real-time decision making that separates the performance components from the schedule optimization mechanism and automates the second task.