12:00 Wed 8 Nov 1995, WeH 7220

	    Using Testing to Iteratively Improve Training:
	       Evaluating What Has Been Learned So Far

			     Peter Stone

Active Learning techniques are those that influence the training set
in order to improve or speed up learning.  In a previous RL chat, I
presented results on the task of learning to shoot a moving ball in
Robotic Soccer.  In a limited scenario, an agent learned to score 95%
of the time.  However, increasing the range of the scenario (using a
larger instance space) makes the task significantly more difficult,
thus requiring more sophisticated ML techniques.  In this talk, I
present a general Active Learning paradigm and show how it can be
applied to the robotic soccer task.  Confidence in past learning is
measured by analyzing performance on successive test sets drawn from
an identical distribution of instances.