Hello! I am a NYC-based ML research engineer. You can reach me at email@example.com
- Project Nabla (post TBD): imitation learning for the 1v1 game Super Smash Bros Melee, trained on 20k+ replays. The agent displays human-like behavior and skills such as edge-guarding and recovery. It regularly wins against a CPU Level 9 and takes stocks off of bad human players (me). If you have lots of CPUs and are interested in contributing, please let me know :)
- Robust Policies via Mid-Level Visual Representations: when using image-based deep RL for hard robotics tasks like pick and place, using embeddings from unsupervised pretraining can dramatically speed up sample complexity and improve robustness. We investigate many standard vision objectives and how they perform across various tasks. Paper accepted to CoRL ‘20
- Theory and Application of Bonus-based Exploration in RL: drawing some connections between Random Network Distillation and LinUCB, we propose a bonus for the deep RL setting that recovers LinUCB in the linear case. In practice, it outperforms RND on a maze environment. Master’s thesis.
2021+: SWE @ google.
2020 - 2021: MS in EECS at UC Berkeley, advised by Prof. Jiantao Jiao.
2017 - 2020: BA in CS at UC Berkeley
Pursuing a multi-agent deep RL research project in my free time. Also tentatively working on self-play for Project Nabla, but am bottlenecked by time/compute right now.
subscribe via RSS