Research map / Reinforcement learning / MDPs and planning
Safe Policy Improvement for POMDPs via Finite-State Controllers
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Constrain updates to retain the behavior policy on insufficiently sampled history-action pairs (SPIBB instead of unconstrained Basic RL) | Offline policy improvement on Tiger and Voicemail · Tiger; Voicemail | policy performance | Helped |
| Use SPIBB with a small threshold when the finite-history assumption is approximately satisfied | CheeseMaze · CheeseMaze | policy performance | Helped |
| Use unconstrained Basic RL on the estimated finite-history MDP rather than SPIBB | Offline policy improvement · CheeseMaze; Tiger; Voicemail | policy performance | Hurt |
| Increase the unknown-pair threshold N_wedge (ablation) | SPIBB · Tiger; Voicemail | reliability | Helped |
| Give the improved FSC more memory (ablation) | SPIBB · Tiger; Voicemail | reliability | Helped |
| Increase N_wedge without increasing the improved policy's memory (ablation) | SPIBB · Tiger; Voicemail when the finite-history assumption is violated | policy improvement over the behavior policy | No clear effect |
| Replace the randomized behavior policy with a deterministic improved policy while retaining memory size k=1 (ablation) | SPIBB and Basic RL · Partially observable settings with memoryless behavior policy | expected return | Hurt |
| Increase memory for the deterministic improved policy (ablation) | SPIBB · Voicemail with memoryless behavior policy | expected return | Helped |
Closest papers by meaning
- Model-based Reinforcement Learning for Confounded POMDPs (ICML 2024 · similarity 0.91)
- Scalable Policy-Based RL Algorithms for POMDPs (NeurIPS 2025 · similarity 0.91)
- Statistical Tractability of Off-policy Evaluation of History-dependent Policies in POMDPs (ICLR 2025 · similarity 0.90)
- Provable Representation with Efficient Planning for Partially Observable Reinforcement Learning (ICML 2024 · similarity 0.89)
- Multi-Environment POMDPs with Finite-Horizon Objectives (NeurIPS 2026 · similarity 0.89)
- Sequential Monte Carlo for Policy Optimization in Continuous POMDPs (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.