Metric E3
-> John Langford

 * Relate "metric MDPs" to the similarity score needed in
   instance-based methods. 
 * Relate the largest minimal alpha cover to the minimum alpha cover.
 * The covering notion is a convenient one, but fails to account for
   any global (or even analogical) structure in the state space.  Do
   you think the covering size is sufficiently small for real
   problems, or do we need a more sophisticated notion?  How do we
   capture the idea that walking to class is a little like walking to
   get your remote or marching in a graduation ceremony, but not
   identical?
 * How would this algorithm be applied to mountain car?