Reinforcement Learning

Hierarchical Programmatic Option framework (HIPO) uses human-readable programs as options in reinforcement learning to solve long, repetitive tasks with improved interpretability and generalization.

GUIDE: Real-time human-shaped AI agents achieve up to 30% higher success rates using continuous human feedback, boosted by a parallel training model that mimics human input for continued improvement.

Generative Trajectory Augmentation (GTA) significantly boosts offline reinforcement learning by generating high-reward trajectories using a conditional diffusion model, enhancing algorithm performance…

Generative world models enhance multi-agent decision-making by simulating trial-and-error learning, improving answer accuracy and explainability.

GDPO: A novel method optimizes graph diffusion models for any objective using reinforcement learning, achieving state-of-the-art performance in diverse graph generation tasks.

HEPO, a novel constrained optimization method, consistently surpasses heuristic-trained policies in reinforcement learning by ensuring policy improvement over heuristics, regardless of heuristic quali…

GCPO: a novel on-policy goal-conditioned reinforcement learning framework tackles limitations of existing HER-based methods by effectively addressing multi-goal Markovian and non-Markovian reward prob…

Goal Reduction with Loop-Removal accelerates Reinforcement Learning (RL) and accurately models human brain activity during goal-directed learning by efficiently deriving subgoals from distant original…

Restless multi-armed bandits with global rewards (RMAB-G) are introduced, extending the model to handle non-separable rewards and offering novel index-based and adaptive policies that outperform exist…

LLMs guided by Monte Carlo Tree Search generate precise, efficient Python code as world models for model-based reinforcement learning, significantly improving sample efficiency and inference speed.

CP3ER, a novel consistency policy with prioritized proximal experience regularization, significantly boosts sample efficiency and stability in visual reinforcement learning, achieving state-of-the-art…

This paper introduces two novel algorithms, achieving optimal regret in generalized linear contextual bandits despite limited policy updates, a significant advancement for real-world applications.

GP-TopK: A novel contextual bandit algorithm uses Gaussian processes with a Kendall kernel for efficient & accurate top-k recommendations, even with limited feedback.

This paper delivers non-asymptotic accuracy bounds for confidence intervals in linear stochastic approximation, leveraging a novel multiplier bootstrap method.

Functional Bilevel Optimization tackles the ambiguity of using neural networks in bilevel optimization by minimizing the inner objective over a function space, leading to scalable & efficient algorith…

TTCT translates natural language constraints into effective training signals for safe reinforcement learning, enabling agents to learn safer policies with lower violation rates and zero-shot transfer …

First oracle-free, computationally tractable algorithms for provably convergent multivariate distributional RL are introduced, achieving convergence rates matching scalar settings and offering insight…

SMG (Separated Models for Generalization) enhances visual RL generalization by disentangling task-relevant and irrelevant visual features via cooperative reconstruction, achieving state-of-the-art per…

FlexPlanner: Deep reinforcement learning solves flexible 3D floorplanning, improving wirelength and alignment significantly.

Bayesian best-arm identification algorithm achieves near-optimal sample complexity by incorporating an early-stopping criterion.