Reinforcement Learning

Madoc: A novel multi-agent framework calibrates RL policies for new environments using limited offline data, achieving superior performance in various locomotion tasks.

SeqComm, a novel multi-level communication scheme, tackles multi-agent coordination by leveraging asynchronous decision-making and a two-phase communication process for improved efficiency and theoret…

LoRa-PI: a model-free RL algorithm learns and exploits low-rank MDP structures for order-optimal sample complexity, achieving ε-optimal policies with O(poly(A)) samples.

New metric, λ-discrepancy, precisely detects & mitigates partial observability in sequential decision processes, significantly boosting reinforcement learning agent performance.

DrilDICE robustly tackles covariate shift in offline imitation learning by using a stationary distribution correction and a distributionally robust objective, significantly improving performance.

Minimax-optimal, computationally efficient algorithms are proposed for distributionally robust offline reinforcement learning, addressing challenges posed by function approximation and model uncertain…

IMAX-PPO: A novel multi-agent RL algorithm leveraging imitation learning to predict opponent actions, achieving superior performance in complex games.

MetaCURL: First algorithm for non-stationary Concave Utility Reinforcement Learning (CURL), achieving near-optimal dynamic regret by using a meta-algorithm and sleeping experts framework.

Provable near-optimality in meta-RL is achieved using a novel bilevel optimization framework and universal policy adaptation algorithm.

Meta-DT: Offline meta-RL masters unseen tasks via conditional sequence modeling and world model disentanglement, showcasing superior few-shot and zero-shot generalization.

Meta-Controller: A novel few-shot behavior cloning framework enables robots to generalize to unseen embodiments and tasks using only a few reward-free demonstrations, showcasing superior few-shot gene…

MADPO, a novel MARL framework, uses mutual policy divergence maximization with conditional Cauchy-Schwarz divergence to enhance exploration and agent heterogeneity in sequential updating, outperformin…

MEow, a novel MaxEnt RL framework, achieves superior performance by unifying policy evaluation and improvement steps, enabling exact soft value function calculation without Monte Carlo approximation.

Boosting diffusion model sample quality, especially with few steps, is achieved via a novel maximum entropy inverse reinforcement learning approach, jointly training the model and an energy-based mode…

CoWorld: a novel model-based RL approach tackles offline visual RL challenges by using online simulators as testbeds, enabling flexible value estimation & mitigating overestimation bias for effective …

Maia-2: A unified model for human-AI alignment in chess, coherently captures human play across skill levels, significantly improving AI-human alignment and paving the way for AI-guided teaching.

MADIFF: Offline multi-agent learning uses attention-based diffusion models to achieve effective coordination and teammate modeling, outperforming existing methods.

Logarithmic Smoothing enhances pessimistic offline contextual bandit algorithms by providing tighter concentration bounds for improved policy evaluation, selection and learning.

CINDERELLA: a new algorithm achieves state-of-the-art no-regret bounds for continuous RL problems by exploiting local linearity.

Greedy algorithms for linear contextual bandits achieve poly-logarithmic regret under the novel Local Anti-Concentration condition, expanding applicable distributions beyond Gaussians and uniforms.