Research map / Reinforcement learning / Visual and representation RL
Detecting Rewards Deterioration in Episodic Reinforcement Learning
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 |
|---|---|---|---|
| Use covariance-weighted UDT instead of Mean, CUSUM, or Hotelling | Fixed-policy reward monitoring · Pendulum, HalfCheetah, and Humanoid deterioration scenarios | detection rate and time until detection | Helped |
| Use UDT, PDT, or MDT for increased control cost and additive noise | Fixed-policy reward monitoring · Pendulum, HalfCheetah, and Humanoid | detection rate: 100% | Helped |
| Use Mean rather than the proposed tests | Fixed-policy reward monitoring · Pendulum, control cost increased from 110% to 300% | detection rate: 110% to 300% | Hurt |
| Use two lookback horizons rather than only a short horizon | Sequential degradation tests · HalfCheetah, +10% control cost | detection | Helped |
| Use a short lookback horizon rather than only a long horizon | Sequential degradation tests · HalfCheetah, +30% control cost | time until detection | Helped |
| Exploit covariance structure rather than ignore time-step variability | UDT and PDT · HalfCheetah, individual tests | detection performance with additional samples | Helped |
| Use UDT when some reward changes are positive and others negative | UDT · Pendulum, longer pole | detection robustness | Hurt |
| Use PDT instead of UDT to focus on a subset of time-steps | PDT · HalfCheetah, increased gravity | detection robustness | Helped |
| Use PDT when time-step weights differ extremely | PDT · Pendulum, modified length and mass | detection robustness | Hurt |
| Run Mean, Hotelling, and PDT in parallel as MDT | MDT · Pendulum modified length and mass and other tested scenarios | detection robustness | Helped |
| Bootstrap whole episodes for sequential false-alarm control | BFAR · Pendulum, HalfCheetah, and Humanoid under H0 | false alarm rate: 5% | No clear effect |
| Use covariance-aware UDT rather than a simple mean test | UDT · Uniform degradation in episodic signals | asymptotic power: G^2 = 1 + Σ w_ij(λ_i−λ_j)^2 | Helped |
Closest papers by meaning
- Episodic Return Decomposition by Difference of Implicitly Assigned Sub-trajectory Reward (AAAI 2024 · similarity 0.87)
- The Distributional Reward Critic Framework for Reinforcement Learning Under Perturbed Rewards (AAAI 2025 · similarity 0.86)
- DARLING: Detection Augmented Reinforcement Learning with Non-Stationary Guarantees (NeurIPS 2026 · similarity 0.85)
- Offline RL with Observation Histories: Analyzing and Improving Sample Complexity (ICLR 2024 · similarity 0.85)
- Reinforcement Learning with Segment Feedback (ICML 2025 · similarity 0.85)
- Episodic Novelty Through Temporal Distance (ICLR 2025 · similarity 0.85)
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.