Large State-Space Stochastic Control
Speaker: Bob Givan , Purdue University ECE
Date: June 11 2002
I will discuss and evaluate two quite different approaches to finding good policies in stochastic control problems with extremely large state spaces.
• I present two novel sampling techniques for deriving control policies, contrast these with two known techniques, and present results evaluating all four sampling approaches against three network-control problems.
• I present a machine-learning method for logically specified control problems (e.g., probabilistic STRIPS problems). This approach leverages solutions to small problem instances (i.e., those with few domain objects) as training data to learn a policy that generalizes well to large problem instances. Key to the method is a policy language bias based on a concept language over the domain definition predicates, similar to recent work by Martin&Geffner (KR-00). We show results for familiar probabilistic STRIPS domains such as logistics and enriched blocks-world problems.
Bob Givan received his BS degree in Mathematics and Biology from Stanford University in 1987, and his PhD degree in Computer Science from MIT in 1996. After a one year postdoc at Brown University, he became Assistant Professor of Electrical and Computer Engineering at Purdue University, in lovely West Lafayette, Indiana, where he continues to this day.
See other events that are part of AI Colloquium Series Spring 2001
See other events happening in June 2002