Selected Papers and Projects

We introduce a new RL framework that is flexible to implement offline, online, off-policy and on-policy RL algorithms. We also release several RL algorithms with that framework.

This paper addresses the problem of policy selection in domains with abundant logged data, but with a very restricted interaction budget.

We proposes a novel offline RL approach that achieves SOTA results and based on the principle of selective imitation.

We develop a SOTA RL agent for Starcraft II using multi-agent RL, self-play and imitation learning.

We propose a novel gating mechanism that makes the training of gated RNNs much easier.

We propose a novel imitation learning algorithm called R2D3 that can learn to solve sparse reward and hard-exploration problems with sparse rewards.

We designed an offline RL algorithm that can learn from demonstrations without any rewards.

We show that the popular reinforcement learning (RL) strategy of estimating the state-action value (Q-function) by minimizing the mean squared Bellman error leads to a regression problem with confounding, the inputs and output noise being correlated. Hence, direct minimization of the Bellman error can result in significantly biased Q-function estimates.

We overcome the challenges of Offline RL, we introduce Regularized Behavior Value Estimation (R-BVE). Unlike most approaches, which use policy improvement during training, R-BVE estimates the value of the behavior policy during training and only performs policy improvement at deployment time

We propose a novel co-operative Multi-agent RL algorithm that is stable.

We propose GrTXL that introduces gating mechanism which stabilizes the learning of transformers for RL.