Reinforcement Learning

PaMoRL framework boosts model-based reinforcement learning speed by parallelizing model and policy learning stages over sequence length, maintaining high sample efficiency.

Leveraging simulation for real-world RL is often hampered by the sim-to-real gap. This paper shows that instead of directly transferring policies, transferring exploratory policies from simulation d…

Boost RL performance in large state spaces by efficiently learning a policy competitive with the best combination of existing base policies!

This paper proposes ‘generalized quasar-convexity’ to optimize problems with multiple probability distributions, offering adaptive algorithms with superior iteration complexities compared to existing …

Deep reinforcement learning optimizes automatic differentiation, achieving up to 33% improvement in Jacobian computation by finding efficient elimination orders.

This paper introduces OCR-CFT, a novel method for general offline-to-online RL, achieving stable and efficient performance improvements by addressing evaluation and improvement mismatches through opti…

Optimal Top-Two Algorithm solves best arm identification problem with improved efficiency and computational cost, achieving asymptotic optimality.

A new algorithm for multi-fidelity best-arm identification achieves asymptotically optimal cost complexity, offering significant improvements over existing methods.

Dope: Efficient algorithms optimize human preference elicitation for learning to rank, minimizing ranking loss and prediction error with absolute and ranking feedback models.

Tri-BBAI & Opt-BBAI achieve optimal asymptotic and near-optimal non-asymptotic sample & batch complexities in batched best arm identification.

Opponent Modeling with In-context Search (OMIS) leverages in-context learning and decision-time search for stable and effective opponent adaptation in multi-agent environments.

Opponent modeling based on subgoal inference (OMG) outperforms existing methods by inferring opponent subgoals, enabling better generalization to unseen opponents in multi-agent environments.

POWR: a novel RL algorithm using operator world models and policy mirror descent achieves global convergence with improved sample efficiency.

OPERA: A new algorithm intelligently blends multiple offline policy evaluation estimators for more accurate policy performance estimates.

This paper introduces efficient approximate posterior sampling for contextual bandits using diffusion model priors, improving Thompson sampling’s performance and expressiveness.

This paper presents a novel two-level online control algorithm that learns to control continuous-time linear systems under adversarial disturbances, achieving sublinear regret.

New reinforcement learning model clarifies the role of information structure in partially-observable sequential decision-making problems, proving an upper bound on learning complexity.

This paper provides novel algorithms and matching lower bounds for multi-armed bandits with expert advice and contextual linear bandits, resolving open questions and advancing theoretical understandin…

This paper tackles the ‘curse of horizon’ in off-policy evaluation for partially observable Markov decision processes (POMDPs) by proposing novel coverage assumptions, enabling polynomial estimation e…

A novel algorithm, TIE, optimally teaches a family of linear behavior cloning learners, achieving instance-optimal teaching dimension while providing efficient approximation for larger action spaces.