Synthesizing efficient agents from partial programs

Abstract

Partial programs are useful for incremental specifications of an agent's behavior in dynamic environments. Unfortunately, executing arbitrary partial programs involves theorem proving, which can be computationally intractable. This paper presents an algorithm for generating efficient implementations of agent behavior from a subclass of partial programs. The algorithm employs a completion strategy that systematically reduces the number of previous percepts being considered. As a result, the compiled structure corresponds to a finite agent with minimal perceptual histories so that its actions can be decided within a bounded time during execution.

The author would like to thank John Woodfill, Narinder Singh, Nils Nilsson, Michael Genesereth, and Benjamin Grosof for helpful discussions on the ideas presented in this paper.