A Bayesian approach to attention control and concept abstraction

Abstract

Representing and modeling knowledge in the face of uncertainty has always been a challenge in artificial intelligence. Graphical models are an apt way of representing uncertainty, and hidden variables in this framework are a way of abstraction of the knowledge. It seems that hidden variables can represent concepts, which reveal the relation among the observed phenomena and capture their cause and effect relationship through structure learning. Our concern is mostly on concept learning of situated agents, which learn while living, and attend to important states to maximize their expected reward. Therefore, we present an algorithm for sequential learning of Bayesian networks with hidden variables. The proposed algorithm employs the recent advancements in learning hidden variable networks for the batch case, and utilizes a mixture of approaches that allows for sequential learning of parameters and structure of the network. The incremental nature of this algorithm facilitates gradual learning of an agent, through its lifetime, as data is gathered progressively. Furthermore inference is made possible, when facing a large corpus of data that cannot be handled as a whole. © Springer-Verlag Berlin Heidelberg 2007.