Confessions of a Masked MDP
In reinforcement learning, one presumes the action set to be known ahead of time, and, thus, one must know if any state-action pairs are not permissible
ProofPaper ↗
Key points
- properties, we study how to combine the search for an optimal policy with the learning of the mask itself, under a linear MDP model in which each attempted
- optimistic optimization and show that it is minimax optimal, which we show is also computationally efficient.
- We propose an alternative, which is also computationally efficient, which plans its exploration strategy.
- formal interpretation as an analogue of information-directed-sampling for deterministic feedback, which may be of independent interest.
Sources (1)
- [1]Confessions of a Masked MDParXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 20, 03:22 PM
In reinforcement learning, one presumes the action set to be known ahead of time, and, thus, one must know if any state-action pairs are not permissible
properties, we study how to combine the search for an optimal policy with the learning of the mask itself, under a linear MDP model in which each attempted
Extractive summary: sentences quoted from the sources.
Before this
- Oct 20, 2026[AINews] DeepSeek DeepInfer: Open Source SOTA Inference running 155T toks/day/cluster making $22m/day at ~80% margin for a 81 day payback
- Oct 20, 2026SAPD: Learning from Agent Experience without Environment Interaction
- Oct 20, 2026Do Diffusion LLMs Scale? Testing Mercury 3 + Decide
- Oct 19, 2026Policy Search in Code Space for Multi-Agent Pathfinding
- Oct 19, 2026Long-Horizon Multi-Object Rearrangement via Distilled Tree Search over Continuous Actions
- Oct 18, 2026Outcome-Anchored Credit Reallocation for Agentic Knowledge Graph Question Answering