Reinforcement Learning

Researchers enhanced Generative Flow Network training by introducing variance-reducing control variates for divergence-based learning objectives, accelerating convergence and improving accuracy.

Offline RL agents often fail in real-world scenarios due to unseen test states. SCAS, a novel method, simultaneously corrects OOD states to high-value, in-distribution states and suppresses risky OOD …

LOLIPOP: A novel algorithm achieving near-optimal regret for offline contextual Markov Decision Processes (CMDPs) using only O(H log T) offline density estimation oracle calls.

This paper introduces Offline Behavior Distillation (OBD) to synthesize compact expert behavioral data from massive sub-optimal RL data, enabling faster policy learning.

DARAIL, a novel algorithm, tackles off-dynamics reinforcement learning by combining reward modification with imitation learning to transfer a learned policy from a source to a target domain. This app…

Model-free policy gradient methods using occupancy functions are developed for online and offline RL, achieving computational efficiency and handling arbitrary data distributions.

OASIS, a novel data-centric approach, shapes offline data distributions toward safer, higher-reward policies using a conditional diffusion model, outperforming existing offline safe RL methods.

Normalize-and-Project (NaP) boosts reinforcement learning by stabilizing layer normalization, preventing plasticity loss, and enabling effective learning rate control.

AI models struggle with changing data; this paper introduces Soft Resets, a novel learning approach that uses an adaptive drift to gracefully guide parameters toward initialization, improving adaptabi…

A novel stochastic bandit algorithm using soft tree ensemble models achieves lower cumulative regret than existing ReLU-based neural bandit algorithms, offering a constrained yet effective hypothesis …

Deep RL agents trained under non-stationarity suffer performance collapse due to representation degradation; this work reveals this in PPO and introduces Proximal Feature Optimization (PFO) to mitigat…

AI agents learn better with well-designed training environments. This paper reveals flaws in current environment-selection methods and introduces Sampling for Learnability (SFL), a new approach that …

NeuralSolver: A novel recurrent solver efficiently and consistently extrapolates algorithms from smaller problems to larger ones, handling various problem sizes.

NEORL: Novel nonepisodic RL algorithm guarantees optimal average cost with sublinear regret for nonlinear systems!

Near-optimal dynamic regret is achieved for adversarial linear mixture MDPs with unknown transitions, bridging occupancy-measure and policy-based methods for superior performance.

This paper presents near-optimal sample complexity bounds for solving distributionally robust reinforcement learning problems with general Lp norms, showing robust RL can be more sample-efficient than…

New distributional RL algorithm (DCFP) achieves near-minimax optimality for return distribution estimation in the generative model regime.

New algorithm, POAM, excels at multi-agent cooperation by adapting to diverse and changing teammates in dynamic scenarios.

This paper introduces efficient algorithms, MR-NaS and DBMR-BPI, for identifying optimal policies across multiple reward functions in reinforcement learning, achieving competitive performance with the…

In multi-agent imitation learning, achieving regret equivalence is harder than value equivalence; this paper introduces novel algorithms that efficiently minimize the regret gap under various assumpti…