Monday 20 September 1993, 3:00, WeH 4601


	Discovering Skills in Reinforcement Learning
		by S. Thrun and A. Schwartz

In this informal and incredibly prelimnary chat I will investigate the
automatic identification of skills in reinforcement learning scenarios.
A skill collapses a whole sequence (field) of actions into a single
action, thus reducing the number of decisions to be made tremendously.
Skills appear to be useful if the learner wants to structure the
learning tasks at hand.  I will present a half-baked approach to the
automatic discovery of skills that co-occur in a multitude of related
reinforcement learning tasks. Initial results in a discrete navigation
domain are presented. I will relate the proposed technique to previous
work on learning skills (such as Singh, Dayan, Kaelbling, Lin,
Tenenberg/Whitehead) and point out future research directions.


Sebastian