Drew Bagnell

dbagnell@ri.cmu.edu

Research Scientist

My Website

Biography

I grew up in Palm Bay, Florida. I attended the University of Florida and studied electrical engineering while working for Keith Doty in robotics. After graduating and spending a short time as a hardware and software designer in the telecom industry, I joined the robotics institute to study machine learning. My thesis work involves theoretical and practical issues in applying machine learning to make decisions.

Research Interests

I'm a new faculty member at Carnegie Mellon University in Robotics. I am interested in "closing the loop" on complex systems; that is, I am interested in designing algorithms that allow systems to observe their own operation and improve performance. My belief is that the border land between planning, control and computational learning is particularly rich with research challenges and potential to make real, immediate impact on applications. I'm particularly interested in systems for which we can obtain at best a partial model. To this end, I'm excited about extending research tools that come from information theory, statistics, control theory, statistical physics and optimization. At the moment, I am particularly focused on two areas in machine learning. First I am working on applications of learning and decision making applied to mobile robotics. Second, I am interested in developing rich, structured probabilistic models that are appropriate for both making and learning decisions.

Tags

Markov Decision Processes, Reinforcement Learning

Papers

• Policy Search by Dynamic Programming (2003)
  Dynamic programming techniques can make direct policy search computationally and sample efficient.
• Covariant Policy Search (2003)
  A simple algorithm leads to very fast, covariant policy search.
• Autonomous Helicopter Control using Reinforcement Learning Policy Search Methods (2001)
  We consider the difficult control problem of learing to fly an autonomous helicopter using limited observational data of its dynamics. To that end, we develop policy search techiques that perform well on average with respect to dynamic consistent with our
• Solving Uncertain Markov Decision Problems (2001)
  Finding good policies in uncertain models