Part of the Special ECE Seminar Series
Modern Artificial Intelligence
Discovering an agent's controllable latent state
John Langford, Microsoft
Given rich sensors (like cameras), how can you learn what an agent can do? This is challenging in highly stateful environments where accomplishing tasks requires complex plans, since randomly exploring may never discover some capabilities of the agent. It turns out however that some of the core principles of discovering controllable latent state are straightforward, enabling us to discover the controllable latent state after a few thousand steps even in environments where random exploration is not viable. Furthermore, this forms a good foundational representation for optimizing reward functions.
John Langford studied Physics and Computer Science at the California Institute of Technology, earning a double bachelor’s degree in 1997, and received his Ph.D. from Carnegie Mellon University in 2002. Since then, he has worked at Yahoo!, Toyota Technological Institute, and IBM‘s Watson Research Center. He is also the primary author of the popular Machine Learning weblog, hunch.net and the principle developer of Vowpal Wabbit. Previous research projects include Isomap, Captcha, Learning Reductions, Cover Trees, and Contextual Bandit learning.