Reinforcement Learning

Learned optimizer OPEN tackles RL’s non-stationarity, plasticity loss, and exploration using meta-learning, significantly outperforming traditional and other learned optimizers.

C-GAIL stabilizes Generative Adversarial Imitation Learning by applying control theory, resulting in faster convergence, reduced oscillation, and better expert policy matching.

BRAID: A novel, conservative fine-tuning method surpasses offline design optimization by cleverly combining generative diffusion models with reward models, preventing over-optimization and generating …

Equivariant Graph Neural Networks boost multi-agent reinforcement learning by improving sample efficiency and generalization, overcoming inherent exploration biases.

BRO (Bigger, Regularized, Optimistic) achieves state-of-the-art sample efficiency in continuous control by scaling critic networks and using strong regularization with optimistic exploration.

Lifelong learning in linear bandits gets a boost! A new algorithm, BOSS, achieves low regret without the usual ‘task diversity’ assumption, opening doors for more realistic sequential multi-task lear…

BECAUSE: a novel algorithm for generalizable offline model-based reinforcement learning that leverages bilinear causal representation to mitigate objective mismatch caused by confounders in offline da…

New algorithms conquer adversarial low-rank MDPs, improving regret bounds for unknown transitions and bandit feedback.

This paper introduces ‘bandits with ranking feedback,’ a novel bandit variation providing ranked feedback instead of numerical rewards. It proves instance-dependent cases require superlogarithmic reg…

The Confidence-Rated Bandits with Abstentions (CBA) algorithm significantly improves reward bounds for prediction with expert advice by strategically leveraging an abstention action.

Longer context in RL boosts generalization but slows down learning; this paper reveals the crucial tradeoff and offers theoretical insights.

AUF-MICNS: A novel sequential method efficiently solves the avoiding undesired future problem by dynamically updating influence relations in non-stationary environments while minimizing action costs.

PASPO: a novel autoregressive policy optimization method for constrained allocation tasks guarantees constraint satisfaction and outperforms existing methods.

This paper presents a novel unified framework for deriving information-theoretic lower bounds for bandit learnability, unifying classical methods with interactive learning techniques and introducing a…

Reinforcement learning agents achieve emergent cultural accumulation by balancing social and independent learning, outperforming single-lifetime agents.

This work introduces PDOA, an offline adaptation framework for constrained multi-objective RL, using demonstrations instead of manually designed preferences to infer optimal policies while satisfying …

This paper provides novel theoretical analyses of utility functions in MORL, characterizing preferences and functions guaranteeing optimal policies.

Adaptive algorithm achieves tight regret bounds for infinitely many-armed bandits under generalized rotting constraints, addressing the challenge of decreasing rewards over time.

Amortized sampling from complex posteriors using diffusion models is achieved via a novel data-free learning objective, Relative Trajectory Balance (RTB). RTB’s asymptotic correctness is proven, offe…

Large-scale transformers achieve grandmaster-level chess play via supervised learning on a new 10M game benchmark dataset, demonstrating impressive generalization beyond memorization.