Perceptual Distinctions
-> Peng
* What's another term for "perceptual aliasing"?
* How does learning to make predictions provide an improved state
  representation?
* How could you use model-based RL in this setting?
* What is the Baum-Welch algorithm being used for in this work?
* What test does Chrisman do to determine when new states are needed
  in the model?
* What is the "exacerbated exploration" problem?
* What is the "extended concealment of crucial features" problem?
* What is the "oversplitting" problem?
* What problem underlies the "utile distinctions" conjecture?  Would
  it be more or less severe when observations are actual sensory data
  (images, say)?