Innovative Realization of Quantitative Goals in BDI Agents via Partial Utility Functions

Abstract

Agents play an important role in high level artificial intelligence in such areas as distributed decision support, robot control, computer games, etc. Currently, the most popular high-level agent architectures are based on the belief-desire-intention (BDI) model. BDI agents are usually specified in modal logic. This is efficient for defining event goals. However, defining quantitative goals can be very difficult in many popular formalisms. In this paper we propose a method for expressing quantitative goals by associating partial utility functions with agent’s goals. We propose a modified BDI agent architecture which is loosely based on fuzzy logic. In this architecture, approximation of partial derivatives of those functions enables us to use gradient based optimization algorithms in the intention reconsideration step to weight some action specializations. Using the proposed approach allows us to easily combine quantitative and event goals, and consider them all while planning. This paper also describes a simple language which can be used to elegantly describe generic action libraries in accordance to the proposed model.