Does Learning Mean a Decrease in Entropy?

Shannon introduced his concept of entropy as a measure of the uncertainty as to the outcome of some event. If this is the case, then learning about a system should be reflected by a real decrease in the entropy of the variable that is the target of learning. I examined a robot that learned to navigate in a simple arena, and analyzed the entropy of the variables for action, state, and reward. Entropy does indeed decrease in the initial stages of learning, but that the introduction of new options complicates the scenario and can in fact lead to a subsequent increase in entropy. I examine the implications from this finding and propose a model for general learning.