Seminar on Computational Learning and Adaptation


 
A Connectionist Approach to Reinforcement Learning for Robotic Control: The Advantages of Indexed Partitioning

Maria L. Gini
Department of Computer Science and Engineering, University of Minnesota
Minneapolis, MN
gini@tc.umn.edu

We explore the use of a connectionist-learning system designed to allow the application of reinforcement learning to robotic control. In particular, we compare direct and indexed partitioning methods and find that indexed partitioning has advantages in time complexity, space complexity, learning speed (measured in trials), and success rate. As application domain we chose the problem of learning to back a truck and trailer rig to a target location by steering the front wheels of the truck. We present extensive simulation results and results from runs on a real robot which learns on-line. We describe some of the subtle difficulties we faced in transferring our algorithm from simulation to a real robot. In simulation we obtained a good success rate when using a uniform distribution for the initial configuration of the truck and trailer rig. This proved to be very difficult to achieve with the real robot, and required us to build a robot with a self-positioning mechanism which is capable of achieving any desired initial configuration.


Date: Thurs., May 18

Time: 4:15-5:30PM

Place: Cordura 100


Return to the seminar schedule