A Bayesian Language for Cumulative Learning

Contents

A cumulative learning agent is one that learns and reasons as it interacts with the world. The Bayesian paradigm provides a natural framework for cumulative learning as an agent can use its observations and prior models to reason about a particular situation, and also learn posterior models. Cumulative learning requires a rich, first-order representation language in order to handle the variety of situations an agent may encounter. In this paper, I present a new Bayesian language for cumulative learning, called IBAL. This language builds on previous work on probabilistic relational models, and introduces the novel feature of observations as an integral part of a language. The key semantic concept is a scope, which consists both of models and observations. The meaning of a scope is the knowledge state of an agent. The language is declarative, and knowledge states can be composed in a natural way. In addition to presenting a language, this paper also presents an EM based learning algorithm called functional EM for learning IBAL models.