Modeling observation in intelligent agents : knowledge and belief

Abstract

In this paper, a method is presented for controlling autonomous agent behavior by filtering the agent's input. Without such filtering, the agent is allowed to have exact knowledge of the state of its domain, resulting in a pattern of performance that is unrealistic and consistently successful. However, filtering that knowledge into beliefs is a way of making it possible for the agent to be unsuccessful some of the time. That is, if the agent is working from beliefs, and the beliefs happen to be wrong, then the agent may not reach its goal at that particular instant. An application for this method--control of an autonomous combat force in a simulation system--is developed and demonstrated in this paper. The algorithm for generating beliefs about battlefield events models the information-gathering system of a combat force. However, this model attempts to simulate the results of the information-gathering system, and not the cognitive or perceptive processes contained in such a system.