The problem of unsupervised learning in an unmodeled agent of an unmodeled environment is one of the hard problems in intelligent robotics, but it is a reasonable description of what human infants do. I will present a summary of some initial work I have done in this area--a Brain-Emulating Cognition and Control Architecture (BECCA).

A BECCA-driven agent bootstraps a model of itself and its environment through two simple algorithms: S-Learning and Context-Based Similarity.  In S-Learning, sequences of experiences provide the basis for future predictions and command selection.  Context-Based Similarity uses those sequences to form abstract concepts, dramatically reducing the dimensionality of the learning problem.  Implementations of BECCA in simulation and in hardware will be given as illustrative examples.