Reinforcement Learning

LocalOMD: Adaptive OMD in extensive-form games achieves near-optimal sample complexity by using fixed sampling and local updates, reducing variance and generalizing well.

SeeX, a novel bi-level optimization framework, effectively tackles the challenge of exploration in visually cluttered environments by training a separated world model to extract relevant information a…

MUN: a novel goal-directed exploration algorithm significantly improves world model reliability and policy generalization in sparse-reward goal-conditioned RL, enabling efficient navigation across div…

CurrMask: a novel curriculum masking paradigm for offline RL, achieving superior zero-shot and fine-tuning performance by dynamically adjusting masking schemes during pretraining, enabling versatile s…

AI agents learn to balance helpfulness and self-preservation using empathy to gauge social relationships and guide reward sharing.

A novel Decomposition-based Policy Learning (DPPL) method optimally balances short-term and long-term rewards, even with interrelated objectives, by transforming the problem into intuitive subproblems…

Learn deep Successor Features (SFs) directly from pixels, efficiently and without representation collapse, using a novel, simple method combining TD and reward prediction loss!

DDiffPG: A novel actor-critic algorithm learns multimodal policies from scratch using diffusion models, enabling agents to master versatile behaviors in complex tasks.

Learning against adaptive adversaries in Markov games is hard, but this paper shows how to achieve low policy regret with efficient algorithms by introducing a new notion of consistent adaptive advers…

First-ever sublinear regret & constraint violation bounds achieved for infinite horizon average reward CMDPs with general policy parametrization using a novel primal-dual policy gradient algorithm.

AI agent MINIMO learns to generate challenging mathematical conjectures and prove them, bootstrapping from axioms alone and self-improving in both conjecture generation and theorem proving.

AI agents efficiently learn Nash equilibria in adversarial team Markov games using a novel learning algorithm with polynomial complexity, resolving prior limitations.

Contrastive Trajectory Representation (CTR) boosts multi-agent reinforcement learning by learning distinguishable agent trajectories using contrastive loss, thus improving performance significantly.

Latent Plan Transformer (LPT) solves long-term planning challenges in reinforcement learning by using latent variables to connect trajectory generation with final returns, achieving competitive result…

Humans autonomously select goals based on both observed and latent learning progress, impacting goal-conditioned policy learning.

New CRL algorithms guarantee global convergence, handle multiple constraints and various risk measures, improving safety and robustness in AI.

Novel optimistic RL algorithm using kernel methods achieves no-regret performance in the challenging infinite-horizon average-reward setting.

KALM: Knowledgeable agents learn complex tasks from LLMs via offline RL using imaginary rollouts, significantly outperforming baselines.

Kaleidoscope: Learnable Masks for Heterogeneous MARL achieves high sample efficiency and policy diversity by using learnable masks for adaptive partial parameter sharing.

Iteratively Refined Behavior Regularization boosts offline reinforcement learning by iteratively refining the reference policy, ensuring robust and effective control policy learning.