Research map / Reinforcement learning / MDPs and planning
Mo' States Mo' Problems: Emergency Stop Mechanisms from Observation
Samuel Ainsworth, Matt Barnes, Siddhartha Srinivasa
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| Add e-stops based on expert visitation to Q-learning | Q-learning · FrozenLake-v0 | sample efficiency | Improved |
| Add e-stops based on expert visitation to Q-learning | Q-learning · FrozenLake-v0 | asymptotic reward | No clear effect |
| Add e-stops based on expert visitation | One-step actor-critic · FrozenLake-v0 | sample efficiency | Improved |
| Add e-stops using expert-derived state bounds | DDPG · Inverted pendulum | convergence speed | Improved |
| Add e-stops using expert-derived state bounds | DDPG · Inverted pendulum | convergence reliability | Improved |
| Add e-stops using expert-derived state bounds | DDPG · HalfCheetah-v3 | reward | Improved |
| Increase the fraction of low-visitation states replaced by e-stops (ablation) | Value iteration · FrozenLake-v0 | convergence speed | Improved |
| Increase the fraction of states removed until states crucial to reaching the goal are pruned (ablation) | Value iteration with e-stops · FrozenLake-v0 | optimal policy reward | Hurt |
| Replace the least-visited half of states with e-stops (ablation) | Q-learning · FrozenLake-v0 | asymptotic reward | No clear effect |
| Use e-stops learned from a suboptimal expert instead of an optimal expert (ablation) | LearnedEStop · FrozenLake-v0 | e-stop policy performance | No clear effect |
| Construct the support set from fewer expert roll-outs (ablation) | LearnedEStop · FrozenLake-v0 | e-stop policy performance | No clear effect |
Closest papers by meaning
- The Courage to Stop: Overcoming Sunk Cost Fallacy in Deep Reinforcement Learning (ICML 2025 · similarity 0.86)
- Reinforcement Learning with Lookahead Information (NeurIPS 2024 · similarity 0.85)
- Reinforcement Learning with Multi-Step Lookahead Information Via Adaptive Batching (NeurIPS 2026 · similarity 0.85)
- The Power of Resets in Online Reinforcement Learning (NeurIPS 2024 · similarity 0.85)
- Optimizing Test-Time Compute via Meta Reinforcement Finetuning (ICML 2025 · similarity 0.85)
- Exploring Starts Are Not Enough: Counterexamples and a Fix for Monte Carlo Exploring Starts (NeurIPS 2026 · 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.
