Reducing the Memory Footprint of Temporal Difference Learning over Finitely Many States by Using Case-Based Generalization