Reinforcement Learning | Artificial Intelligence Institute for Advances in Optimization

Reinforcement Learning (RL) has been brought to the forefront of ML due to recent breakthroughs in computer games, navigation and object manipulation. These successes are primarily due to a new breed of RL algorithms that use off-policy learning with function approximation (e.g., using Deep Q-learning), and build off existing paradigms such as actor-critic. Despite these high-profile empirical successes, some basic questions about their convergence and sample complexity, statistical computational trade-offs, connections to optimization theory, and fundamental need for exploration are still unanswered. The Institute investigates these underlying core themes and use them to guide research into three directions: Off-Policy RL, Safe RL, and Representation Learning for RL.

Publications

Related work

Reinforcement Learning from Optimization Proxy for Ride-Hailing Vehicle Relocation. Enpeng Yuan, Wenbo Chen, and Pascal Van Hentenryck. Journal of Artificial Intelligence Research (JAIR), 75, 985-1002, November, 2022.
See the Future through the Void: Active Pre-Training with Successor Features,
Hao Liu, Pieter Abbeel.
In the proceedings of the International Conference on Machine Learning (ICML), Virtual, July 2021.
State Entropy Maximization with Random Encoders for Efficient Exploration,
Younggyo Seo, Lili Chen, Jinwoo Shin, Honglak Lee, Pieter Abbeel, Kimin Lee.
In the proceedings of the International Conference on Machine Learning (ICML), Virtual, July 2021.
arXiv 2102.09430
Self-Supervised Policy Adaptation during Deployment,
Nicklas Hansen, Yu Sun, Pieter Abbeel, Alexei A. Efros, Lerrel Pinto, Xiaolong Wang.
In the proceedings of the 7th International Conference on Learning Representations (ICLR), Virtual, April 2021.
arXiv 2007.04309
Mutual Information-based State-Control for Intrinsically Motivated Reinforcement Learning,
Rui Zhao, Yang Gao, Pieter Abbeel, Volker Tresp, Wei Xu. In the proceedings of the 7th International Conference on Learning Representations (ICLR), Virtual, April 2021. arXiv 2103.08107
coupling Representation Learning from Reinforcement Learning,
Adam Stooke, Kimin Lee, Pieter Abbeel, Michael Laskin.
In the proceedings of the International Conference on Machine Learning (ICML), Virtual, July 2021.
arXiv 2009.08319
A decentralized policy gradient approach to multi-task reinforcement learning.S. Zeng, A. Anwar, T. Doan, A. Raychowdhury, and J. Romberg. arxiv:2006.04338, June 2020.
In-field performance optimization for mm-wave mixed- signal doherty power amplifiers: a bandit approach. S. Xu, F. Wang, H. Wang, and J. Romberg. IEEE Trans. Circuits and Systems, to appear, 2020.
Primal: Pathfinding via reinforcement and imitation multi-agent learning. G. Sartoretti, J. Kerr, Y. Shi, G. Wagner, S. Kumar, S. Koenig, and H. Choset. IEEE Robotics and Automation Letters, 4(3):2378– 2385, 2019.