Research
Currently, my research involves building scalable systems for ML pracitioners that will fulfill the Sky Computing vision. We are open source. This research involves virtualizing GPUs to scale DL training to trillions of parameters and designing learnable scheduling policies for migrating jobs across different and regions clouds (incl. on-premise).
Previously,my masters's and undergraduate research primarily focused on practical problems and applications for reinforcement learning (RL), including NLP, query optimization for databases, and video streaming.
|
|
Balsa: Learning a Query Optimizer Without Expert Demonstrations
Zongheng Yang,
Wei-lin Chiang,
Frank Luan,
Gautam Mittal,
Michael Luo,
Ion Stoica
Special Interest Group on Management of Data (SIGMOD), 2022
Arxiv |
Video |
Code
An end2end query optimizer trained via deep RL that exceeds the query-performance of expert solvers by up to 2.8x.
|
|
MESA: Offline Meta-RL for Safe Adaptation and Fault Tolerance
Michael Luo,
Ashwin Balakrishna,
Brijen Thananjeyan,
Suraj Nair,
Julian Ibarz,
Jie Tan,
Chelsea Finn,
Ion Stoica,
Ken Goldberg
Neural Information Processing Systems (NeurIPS) Safe Control Workshop, 2021
Website |
Arxiv |
Video |
Code
Safe RL algorithm that meta-learns from offline datasets to safely adapt to unseen environments.
|
|
IMPACT: Importance Weighted Asynchronous Architectures with Clipped Target Networks
Michael Luo,
Jiahao Yao,
Richard Liaw,
Eric Liang,
Ion Stoica
International Conference on Learning Representations (ICLR), 2020
Website |
Arxiv |
Video |
Code
An algorithm for distributed reinforcement learning that tunes the tradeoff between distributed data collection and learning sample efficiency to optimize for training speed by combining the sample efficiency of PPO and the data throughput from IMPALA.
|
|
Connecting Context-specific Adaptation in Humans to Meta-learning
Rachit Dubey*,
Erin Grant*,
Michael Luo*,
Karthik Narasimhan,
Thomas L. Griffiths
Preprint.
Arxiv |
Code
We introduce a framework for using contextual information about a task to guide the initialization of task-specific models before adaptation to online feedback, which leads to faster adaptation to online feedback than that of zero-shot multitask approaches.
|
|
Recovery RL: Safe Reinforcement Learning with Learned Recovery Zones
Brijen Thananjeyan*,
Ashwin Balakrishna*,
Suraj Nair,
Michael Luo,
Krishnan Srinivasan,
Minho Hwang,
Joseph E. Gonzalez,
Julian Ibarz,
Chelsea Finn,
Ken Goldberg
International Conference on Robotics and Automation (ICRA), 2021
Website |
Arxiv |
Video |
Code
An algorithm for safe reinforcement learning which utilizes a set of offline data to learn about constraints before policy learning and a pair of policies which separate the often conflicting objectives of task directed exploration and constraint satisfaction to learn contact rich and visuomotor control tasks.
|
|
Distributed Reinforcement Learning is a Dataflow Problem
Eric Liang*,
Zhanghao Wu*,
Michael Luo,
Sven Mika,
Ion Stoica
Neural Information Processing Systems (NeurIPS), 2021.
Arxiv |
Video |
Code
We propose RLFlow, a hybrid actor-dataflow programming model for distributed RL, that leads to highly composable and performant implementations of RL algorithms, which results to faster training and significant code reductions.
|
|
Accelerating Quadratic Optimization with Reinforcement Learning
Jeffrey Ichnowski,
Paras Jain,
Bartolomeo Stellato,
Goran Banjac,
Michael Luo,
Francesco Borrelli,
Joseph E. Gonzalez,
Ion Stoica,
Ken Goldberg
Neural Information Processing Systems (NeurIPS), 2021.
Website |
Arxiv |
Video |
Code
An intelligent application of RL that tunes the parameters of existing Quadratic Program (QP) solvers and improving solving times by up to 3x.
|
|
Discovering Non-monotonic Autoregressive Orderings with Variational Inference
Xuanlin Li*,
Brandon Trabucco*,
Dong Huk Park,
Yang Gao,
Michael Luo,
Sheng Shen,
Trevor Darrell
International Conference on Learning Representations (ICLR), 2021.
Website |
Arxiv |
Video |
Code
We propose the first domain-independent unsupervised / self-supervised learner that discovers high-quality autoregressive orders through fully-parallelizable end-to-end training without domain-specific tuning.
|
|
LazyDAgger: Reducing Context Switching in Interactive Robot Imitation Learning
Ryan Hoque,
Ashwin Balakrishna,
Brijen Thanajeyan,
Carl Putterman,
Michael Luo,
Daniel Seita,
Daniel S. Brown,
Ken Goldberg
Conference on Automation Science and Engineering (CASE), 2021.
Website |
Arxiv |
Video
An algorithm for interactive imitation learning that learns to minimize human context switching through sustained interventions and maintains the same supervisor burden for prior algorithms.
|
|
AlphaGarden: Learning Seed Placement and Automation Policies For Polyculture Farming with Companion Plants
Yahav Avigal, Anna Deza, William Wong, Sebastian Oehme, Mark Presten, Mark Theis, Jackson Chui, Paul Shao, Huang Huang, Atsunobu Kotani, Satvik Sharma,
Michael Luo,
Stefano Carpin, Joshua Viers, Stavros Vougioukas, Ken Goldberg
International Conference on Robotics and Automation (ICRA), 2021
Website |
Paper |
Code
We investigate different seed placement and pruning algoritms in a polyculture garden simulator to jointly maximize diveristy and coverage of various plants types.
|
|
SkyPilot: A Broker for Sky Computing
Core contributor; Developed Sky Storage, Sky On-premise, and multi-node core features.
|
|
Ray/RLlib
Core contributor; Created distributed model-free, model-based, and meta-learning RL algorithms on RLlib, including APPO/IMPACT, MAML, MBMPO, and Google Dreamer.
|
|
CS 189: Introduction to Machine Learning
Teaching Assistant: Fall 2019
CS 162: Operating Systems and System Programming
Reader: Fall 2018
|
|