PhD Student at CMU & Student Researcher at Google Brain
beysenba @cs.cmu.edu, @ben_eysenbach
Bio: I'm a PhD student in the Machine Learning Department at Carnegie Mellon University and a student researcher in Google Brain. I am co-advised by Ruslan Salakhutdinov and Sergey Levine. My PhD is supported by the National Science Foundation (GFRP) and the Hertz Fellowship. Previously, I was a Resident at Google Brain. I studied math and computer science at MIT.
Research summary: My research has focused on designing better RL algorithms.
- Data-driven control [1, 2]
- Safety and robustness [1, 2, 3]
- Unsupervised RL and skill learning [1, 2]
- Planning and inference [1, 2] ... and demonstrating that these algorithms work on real robots [1, 2]
Research opportunities: I am usually looking for students to help with research projects both during the semester and over the summer. If you are interested, please send me an email. I especially encourage students from underrepresented groups to reach out.
- Two recent papers were accepted to NeurIPS 2021: Replacing Rewards with Examples: Example-Based Policy Search via Recursive Classification (oral) Robust Predictable Control (spotlight). I received a "Outstanding Reviewer Award" for reviewing for NeurIPS this year.
- Excited to release a few new papers!
- Mismatched No More: Joint Model-Policy Optimization for Model-Based RL (with Sasha Khazatsky): We propose a single objective for jointly training a model and policy for model-based RL, such that updates to either component improves a lower bound on expected return.
- The Information Geometry of Unsupervised Reinforcement Learning: We prove that unsupervised skill learning algorithms, such as DIAYN, are not optimal for learning all reward-maximizing policies. I think this work is interesting because it suggests that other, yet-uninvented, skill learning algorithms may be much better for preparing to solve new tasks.
- Recurrent Model-Free RL is a Strong Baseline for Many POMDPs (with Tianwei Ni): Simply using recurrent architecture allows standard model-free RL algorithms to perform very well on meta-RL, robust RL, and other sorts of POMDPs. We discover a few key tricks that make this possible, and release code so that future work can include this very strong baseline.
See Google Scholar for a complete and up-to-date list of publications.
Assorted Blog Posts
© 2021 Ben Eysenbach