Reinforcement Learning

PAGAR: a novel semi-supervised IRL framework prioritizing task alignment over data alignment, leveraging expert demonstrations as weak supervision to derive task-aligned reward functions for improved …

Effective Metric-based Exploration Bonus (EME) enhances reinforcement learning exploration by using a robust metric for state discrepancy and a dynamically adjusted scaling factor based on reward mode…

Neural Episodic Control’s differentiable dictionary is shown to be a Universal Hopfield Network, enabling improved performance and a novel evaluation criterion.

Reinforcement learning with complex objectives made easy: This paper introduces an optimality-preserving translation to reduce problems with Linear Temporal Logic (LTL) objectives to standard average …

Provably efficient RL algorithms are designed to utilize immediate reward or transition information, significantly improving reward collection in unknown environments.

Boosting RL data efficiency for continuous control, this paper advocates Euclidean data augmentation using limb-based state features, significantly improving performance across various tasks.

Safe reinforcement learning is achieved via RL-AR, an algorithm that combines a safe policy with an RL policy using a focus module, ensuring safety during training while achieving competitive performa…

This paper pioneers a modular framework for reinforcement learning, addressing the challenge of learning under complex observations and simpler latent dynamics, offering both statistical and algorithm…

Reinforcement learning refines existing macro placements, enhancing chip design by improving power, performance, and area (PPA) metrics and integrating the often-overlooked metric of regularity.

Boost online finetuning of Decision Transformers by adding TD3 gradients, especially when pretrained with low-reward data.

RegQ: A novel regularized Q-learning algorithm ensures convergence with linear function approximation, solving a long-standing instability problem in reinforcement learning.

A novel regularized conditional diffusion model enables effective multi-task preference alignment in sequential decision-making by learning unified preference representations and maximizing mutual inf…

Memoroids and Tape-Based Batching revolutionize recurrent RL, enabling efficient processing of long sequences and improving sample efficiency by eliminating segmentation.

Reciprocators: AI agents that learn to cooperate by reciprocating influence, achieving prosocial outcomes in complex scenarios.

REBEL, a novel reinforcement learning algorithm, simplifies policy optimization by regressing relative rewards, achieving strong performance in language and image generation tasks with increased effic…

II-COS: a novel online sample selection method effectively controls individual and interactive constraints in real-time via predictive inference, improving efficiency and addressing various practical …

Recurrent Trace Units (RTUs) significantly enhance real-time recurrent learning in reinforcement learning, outperforming other methods with less computation.

Provably efficient randomized exploration in cooperative MARL is achieved via a novel unified algorithm framework, CoopTS, using Thompson Sampling with PHE and LMC exploration strategies.

First provably efficient randomized RL algorithms using multinomial logistic function approximation are introduced, achieving superior performance and constant-time computational cost.

This paper presents randomized algorithms with PAC bounds for solving inverse reinforcement learning problems in continuous state and action spaces, offering robust theoretical guarantees and practica…