(please see the main page for schedule information)
A map is a key component for a mobile robot. Maps at their core allow the robot to answer three questions: (1) "where have I been?" (2) "where am I now?" and (3) "how do I get where I want to go?" A huge body of robotics research assumes their existence, and another large body of research tries to build them. But building maps can be time consuming, manually intensive, and require expert knowledge in the form of detailed models of the robot's motion and sensor apparatus. In this talk I will show how maps can be learned directly from the robot's subjective experience of sensations and actions, without any models. I'll introduce a new algorithm, Action Respecting Embedding (ARE), inspired by kernel-based dimensionality reduction techniques. ARE extracts a low dimensional representation of data that also respects the provided action labelling. The resulting subjective map explicitly encodes the robot's trajectory (answering question one), and I'll show how it can be used for both planning (question three) and localization (question two). Although originally conceived in the context of mobile robots, ARE is a general technique for extracting representations from a sequence of observations and actions.
Michael Bowling is a professor at the University of Alberta in Edmonton, His research focuses on the intersections of machine learning, games, and robotics. He has been actively involved in the emerging field of multiagent learning and a long-time participant in the RoboCup robot soccer initiative. He is now particularly excited about opponent modelling in poker, machine learning for commercial computer games, and robots learning representations from experience.