Context-based policy search: transfer of experience across problems
  ----------------------------_______________________________________

  Leonid Peshkin {pesha@ai.mit.edu} 
  Harvard Center for Artificial Intelligence 
  Dworkin Bld. 134, Cambridge, MA 02138, USA

  Edwin D. de Jong {edwin@cs.brandeis.edu
  Computer Science Department 
  Brandeis University, Waltham, MA 02454-9110


Appears in proceedings of ICML-02 workshop on Development of Representations 
The current version and related work available from  
                              http://www.eecs.harvard.edu/~pesha/papers.html


Abstract:

An important question in reinforcement learning is how generalization 
may be performed. This problem is especially important if the learning 
agent receives only partial information about the state of the environment.
Typically, the bias required for generalization is chosen by the
experimenter. Here, we investigate a way for the learning method to
extract bias from learning one problem and apply it in subsequent problems.
We use a gradient-based policy search method, and look for controllers that
consist of a context component and an action component. Empirical results on
a two-agent coordination problem are reported. It was found that learning a
bias made it possible to address problems that were not solved otherwise.