Usages of Action Models in Learning

		Joseph O'Sullivan

I shall take as a starting point for this discussion that
whereas various research has shown how is it _possible_ to
learn relatively simple strategies from very little initial
knowledge, theoretical and emperical results indicate that
such simple learning approaches will require unrealistic
amounts of training data to learn significantly more complex
functions.  

I shall review the various approachs towards scaling up
learning, but will concentrate on techniques by which
previously learned knowledge in the form of "action models"
are used to reduce the need for new data in subsequent
learning.  

Two papers will ground this discussion;

o Mahadevan, "Enhancing Transfer in Reinforcement Learning
  by Building Stochastic Models of Robot Action", ML92 

o Mitchell, O'Sullivan and Thrun, "Explanation-Based
  Learning for Mobile Robot Perception", submitted to Robot
  Learning Workshop at ML94. (draft version).

These papers approach action models from two different
perspectives in the domain of robot learning, and so provoke 
interesting comparisons (and this talk!).

Note: These papers are available online in
      /afs/cs/user/josullvn/rltalk/{mahadevan, rldraft}.ps
      Peruse at will.