Instance-based state identification...
-> Haym
* What's yet another term for "perceptual aliasing"?
* Imagine a "delayed tiger" domain in which the decision maker is told
  the door that contains the tiger (1 or 2), then a random number of
  steps pass in which the decision maker receives a random observation
  (a or b, equally likely), then the decision maker gets the go ahead
  and must choose 1 or 2.  How well would a fixed history window
  approach work for this domain?  Nearest sequence memory?  Perceptual 
  distinctions?  How could we change NSM to work perfectly on this
  task?
* How does using a k-nearest neighbor algorithm allow the learner to
  make more precise distinctions given more data?
* How does McCallum deal with saving *all* experience?  What are some
  advantages of this approach?
* When does k-nearest neighbor run into trouble?