Reinforcement Learning

RA-PbRL introduces a provably efficient algorithm for risk-aware preference-based reinforcement learning, addressing the limitations of existing risk-neutral methods in applications demanding heighten…

QGFN boosts Generative Flow Networks (GFNs) by cleverly combining their sampling policy with an action-value estimate, creating controllable and efficient generation of high-reward samples.

Offline RL struggles with OOD action overestimation. QDQ tackles this by penalizing uncertain Q-values using a consistency model, enhancing offline RL performance.

This paper presents novel RL algorithms using multinomial logit function approximation, achieving O(1) computation and storage while nearly closing the regret gap with linear methods.

Provably efficient algorithms are introduced for interaction-grounded learning (IGL) with context-dependent feedback, addressing the lack of theoretical guarantees in existing approaches for personali…

OPT-AIL: Provably efficient adversarial imitation learning with general function approximation, achieving polynomial sample and interaction complexity, outperforming existing deep AIL methods.

This paper provides the first provable efficiency guarantees for practically-used RL algorithms leveraging privileged information, addressing limitations of previous empirical paradigms and opening ne…

Prospective Learning: a new framework enabling machines to learn effectively in dynamic environments where data distributions and goals shift over time.

PreTS algorithm efficiently identifies the most preferred policy in bandit problems with vector-valued rewards, achieving asymptotically optimal sample complexity.

Human-like choice modeling revolutionizes preference learning! A new tractable model, CRCS, significantly improves utility inference from human data, outperforming existing methods.

Preference Flow Matching (PFM) streamlines preference integration into pre-trained models using flow matching, overcoming fine-tuning limitations and enabling robust alignment with human preferences.

MGPO: Efficient online RL adaptation via prompt optimization of pre-trained multi-goal transformers.

Policy-Shaped Prediction (PSP) improves model-based reinforcement learning by focusing world models on task-relevant information, significantly enhancing robustness against distracting stimuli.

First-order policy optimization for robust average-cost MDPs achieves linear convergence with increasing step size and 0(1/ε) complexity with constant step size, solving a critical gap in existing res…

Boosting reinforcement learning, this paper introduces h-PMD, a novel algorithm enhancing policy mirror descent with lookahead for faster convergence and improved sample complexity.

Pessimistic Backward Policy for GFlowNets (PBP-GFN) tackles GFlowNets’ tendency to under-exploit high-reward objects by maximizing observed backward flow, enhancing high-reward object discovery and ov…

VPL: a novel multimodal RLHF personalizes AI by inferring user-specific latent preferences, enabling accurate reward modeling and improved policy alignment for diverse populations.

PASQL, a novel periodic agent-state Q-learning algorithm, significantly improves reinforcement learning in partially observable environments by leveraging non-stationary periodic policies to overcome …

PEAC: a novel unsupervised pre-training method significantly improves cross-embodiment generalization in reinforcement learning, enabling faster adaptation to diverse robots and tasks.

Boost continual reinforcement learning with Parseval regularization: maintaining orthogonal weight matrices preserves optimization, significantly improving RL agent training across diverse tasks.