Reinforcement Learning

Offline RL’s performance often lags behind imitation learning, but this paper reveals that policy learning and generalization, not value function learning, are often the main bottlenecks.

Decision Mamba (DeMa) outperforms Decision Transformer (DT) in offline RL trajectory optimization with 30% fewer parameters in Atari and a quarter in MuJoCo, demonstrating the efficacy of Mamba’s line…

Offline imitation learning achieves surprisingly strong performance, matching online methods’ efficiency under certain conditions, contradicting prior assumptions.

New multi-agent imitation learning algorithm (MIFQ) leverages inverse soft Q-learning and factorization for stable, efficient training, achieving state-of-the-art results on challenging benchmarks.

Successive Concept Bottleneck Agents (SCoBots) improve reinforcement learning by integrating interpretable layers, enabling concept-level inspection and human-in-the-loop revisions to fix misalignment…

Hierarchical Human Knowledge-guided MARL (hhk-MARL) framework accelerates large-scale multi-agent training by integrating suboptimal human knowledge, significantly improving performance and scalabilit…

IsCiL: a novel adapter-based continual imitation learning framework that efficiently adapts to new tasks by incrementally learning and retrieving reusable skills.

MERIT-IRL: First in-trajectory IRL framework learns reward & policy incrementally from ongoing trajectories, guaranteeing sub-linear regret.

Boosting AI generalization: CENIE framework quantifies environment novelty via state-action coverage, enhancing unsupervised environment design for robust generalization.

Deep RL agents often suffer from instability due to the ‘chain effect’ of value and policy churn; this paper introduces CHAIN, a novel method to reduce this churn, thereby improving DRL performance an…

Linear ensemble sampling achieves a state-of-the-art regret bound of Õ(d³/²√T) with a logarithmic ensemble size, closing the theory-practice gap in linear bandit algorithms.

New modeling of bid space in multi-unit uniform price auctions achieves regret of Õ(K4/3T2/3) under bandit feedback, improving over prior work and closing the gap with discriminatory pricing.

This paper significantly improves Bayes regret bounds for hierarchical Bayesian bandit algorithms, achieving logarithmic regret in finite action settings and enhanced bounds in multi-task linear and c…

C-Procgen reveals implicit curriculum in Procgen’s multi-level training, showing learning shifts gradually from easy to hard contexts.

This paper introduces seq-NMF, a novel method for unsupervised subtask discovery in reinforcement learning that leverages selection variables to enhance generalization and data efficiency.

This study introduces a novel framework for individualized reinforcement learning, guaranteeing the identifiability of latent factors influencing state transitions and providing a practical method for…

Hybrid RL algorithms achieve sharper error/regret bounds than existing offline/online RL methods in linear MDPs, improving sample efficiency without stringent assumptions on behavior policy quality.

CODA: A novel method for solving contextual goal-oriented problems with offline datasets, using unlabeled trajectories and context-goal pairs to create a fully labeled dataset, outperforming other bas…

Low reward variance drastically improves contextual bandit regret, defying minimax assumptions and highlighting the crucial role of eluder dimension.

CATY-IRL: A novel, provably efficient algorithm solves Inverse Reinforcement Learning’s scalability issues for large state spaces, improving upon state-of-the-art methods.