I am interested in applying learning to robotics applications. In particular, I am interested in transfer learning, so that knowledge learned from one problem can be applied to similar but new problems. This is particularly important in robotics, because training experience is much more important (due to imperfect models of the real world) while also being much harder to obtain (compared to working with purely virtual entities, for example in simulators).
My work in transfer learning hinges on the concept of using feature-based state representation. The fundamental problem of transfering knowledge is to represent the knowledge in the right way. In order to transfer knowledge between problems, the knowledge has to be represented in a way that's meaningfull across problems, even if this representation is not necessarily useful for getting high performance out of any single problem.
I have done some work of transfering knowledge in the context of dynamic programming. In order to avoid the computational complexity of dynamic programming, I am currently exploring different policy representations that use discrete trajectories. This has been published at ICRA 2006 (see publications). While this work is in the Marble Maze domain, I am now extending it to the Little Dog domain.
Some past research:
Before coming to CMU, I did a Bachelor's and Master's degree at McGill University. My Master's advisor was Doina Precup and my master's thesis was called "Automated Discovery of Options in Reinforcement Learning" (PDF).
Before working on my Master's degree, I also worked with Prof. Tom Shultz and Francois Rivest in the Laboratory for Natural and Simulated Cognition on a Java Applet that implements and compares regular Backpropagation neural networks with Cascade Correlation neural networks.
|2006-07-10||MADE BY MARTIN|