On the Development of Tools for Modelling Dynamic Beliefs Based on Past Data

Abstract

In order to develop effective ubiquitous computing systems, we often need to predict an agent’s behaviour based on past data. One way to do this is to maintain a model of what the agent believes at any point in time, as well as a mechanism for changing the beliefs to incorporate new information. In the knowledge representation community, this process is captured through formal belief revision operators. In this paper, we assume that we are monitoring the behaviour of an agent that uses a belief revision operator to incorporate new information; but we do not know exactly which operator is being used. Given past data about the beliefs of the agent, we propose two approaches for predicting future changes in belief. In the first approach, we simply search for all revision operators consistent with the data. In the second approach, we use machine learning to predict if a certain formula will be believed based on past data. We describe work in progress on prototype software to experiment with both approaches, and discuss when each is appropriate. We argue that modelling the dynamic beliefs of an agent in this way can be a useful component of a software system tasked with predicting behaviour when new information is received.