This paper described a system that transfers the results of prior learning to significantly speed up reinforcement learning on related tasks. Vision processing techniques are utilized to extract features from the learned function. The features are then used to index a case base and control function composition to produce a close approximation to the solution of a new task. The experiments demonstrated that function composition often produces more than an order of magnitude increase in learning rate compared to a basic reinforcement learning algorithm.
The author would like to thank Rob Holte for many useful discussions and help in preparing this paper. This work was in part supported by scholarships from the Natural Sciences and Engineering Research Council of Canada and the Ontario Government.