An Information Theory Based Anticipation Architecture

We introduce in this paper information theoretic tools that can be used in an animat for constructing an internal predictive model. We first explore how information theoretic quantities such as causal entropy can be used for extracting time-dependance between stimuli, allowing for an estimation of the expected delay between two given stimuli. We also provide tools for identifying relevant contingency relationships in the flux of perceptions. For both types of relationship, we also define confidence values based on information quantities. These tools are used together for on-line construction of a predictive model of the environment. We propose an anticipation-based architecture for animats that uses this predictive model for action-decision according to reward expectation. The proposed architecture takes into account the problems of learning sequences of actions and identifying non-occurring expected stimuli. It also incorporates a simple and computationally efficient saliency filtering. We illustrate the effectiveness of the tools proposed with preliminary results about their predictive efficiency when embedded into the proposed architecture. We conclude by discussing the issues addressed by the proposed model, those that are left open and future investigations.