Scaling Life-long Learning
The Faculty of Informatics is pleased to announce a seminar given by Richard Sutton
DATE: Monday, September 3rd 2012
PLACE: University of Lugano, room A31, Red building (Via G. Buffi 13)
As Moore's law continues its inexorable march, we will soon have the ability to build the computational hardware needed to support strong artificial intelligence. It is less clear when we will have the requisite ideas and algorithms of the corresponding software, but we should be preparing for it now. In this spirit, I focus in this talk on the problem of learning from a long interaction between a robot and its environment as the computational resources are limited but scale as predicted by Moore's law. Conventional model-free reinforcement learning (RL) is poorly suited to this problem because it utilizes very little computation, and conventional model-based RL is poorly suited because it requires far too much. The solution may be to use model-free methods as building blocks of a scalable approach to model-based RL, as we are doing in my laboratory. I will describe recent work designing robots and algorithms that scale gracefully to continual real-time learning about arbitrarily complex environments.
Richard S. Sutton is a professor and iCORE chair in the department of computing science at the University of Alberta. He is a fellow of the Association for the Advancement of Artificial Intelligence and co-author of the textbook Reinforcement Learning: An Introduction. Before joining the University of Alberta in 2003, he worked in industry at AT&T and GTE Labs, and in academia at the University of Massachusetts. He received a PhD in computer science from the University of Massachusetts in 1984 and a BA in psychology from Stanford University in 1978. Rich's research interests center on the learning problems facing a decision-maker interacting with its environment, which he sees as central to artificial intelligence. He is also interested in animal learning psychology, in connectionist networks, and generally in systems that continually improve their representations and models of the world.
HOST: Prof. Jürgen Schmidhuber