When do differences matter? On-line feature extraction through cognitive economy

For an intelligent agent to be truly autonomous, it must be able to adapt its representation to the requirements of its task as it interacts with the world. Most current approaches to on-line feature extraction are ad hoc; in contrast, this paper derives principled criteria for representational adequacy by applying the psychological principle of cognitive economy to reinforcement learning. The criteria are principled because they are based on an analysis of the amount of reward the agent forfeits when it generalizes over states. This analysis shows that action-value errors are sometimes irrelevant, and that the agent may optimize its performance with limited cognitive resources by grouping together states whose differences do not matter in its task. The paper presents an algorithm based on this analysis, incorporating an active form of Q-learning and partitioning continuous state-spaces by merging and splitting Voronoi regions. The experiments illustrate a new methodology for testing and comparing representations by means of learning curves. Results from the puck-on-a-hill task demonstrate the algorithm's ability to learn effective representations, superior to those produced by some other, well-known, methods.