Forbes and Andre paper

Notes:
 Continuous state and action problem
 Main justification for memory based: avoid catastrophic interference
 Uses a model as well (did others do this?)
 contributions
  value updating algorithm
  instance averaging
  automatic dimension scaling
 Claim: Instance-based methods less prone to forgetting than
 	parameterized function approximator.
  How can we test this?
 How does this paper differ from the Smart/Kaelbling paper?
  * updates neighbors, too
  * kernel regression
  * gradient-based update
  * limit density of examples
  * toss those that contribute little accuracy

(at section 1.4)