← All threads

RL Regret Bounds

Theoretical analysis of regret minimization in reinforcement learning and games, including Bayesian and worst-case bounds.

5 papers

Where this stands

The written synthesis of this thread is for subscribers. Subscribe.

How this thread developed

  1. Oct 2026 · KAIST

    Unmodified posterior sampling achieves minimax regret in reinforcement learning

    Shows unmodified posterior sampling achieves minimax Bayesian regret in finite-horizon tabular MDPs.

  2. 2 further papers

    Oct 2026 · University of Cambridge, University of Sydney

    Reinforcement learning can stably trap itself in low-return policies via representation superposition

    Identifies representation superposition as a mechanism trapping RL in low-return policies, offering a new source of regret.

  3. Oct 2026 · University of Texas at Austin, University of British Columbia

    Multi-Fidelity Control Variates Stabilize Data-Scarce Reinforcement Learning

    Extends multi-fidelity policy gradients to actor-critic methods (MFPG-PPO), stabilizing data-scarce RL with provable cross-fidelity control variates.

3 of 5 papers shown