Research map / Reinforcement learning / Visual and representation RL
Local Explanations for 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 |
|---|---|---|---|
| Explain policies with policy-dynamics-based meta-states and strategic states instead of action-grouped Viper-D explanations | Minipacman EAT and HUNT policies · Minipacman user study | human prediction accuracy | Helped |
| Explain policies with policy-dynamics-based meta-states and strategic states instead of action-grouped Viper-D explanations | Minipacman EAT and HUNT policies · Minipacman user study | explanations labeled Unclear: less than 5% vs more than 25% | Helped |
| Approximate out-path counts by sampling states from other meta-states (ablation) | SSX · Minipacman | computation time for out-path counts: approximately 50% | Helped |
| Approximate out-path counts by sampling states from other meta-states (ablation) | SSX · Minipacman | primary strategic-state Pacman position stability: less than one space on average | No clear effect |
| Increase the local approximation size N (ablation) | SSX · Minipacman HUNT | strategic-state position faithfulness: N=3 vs N=4: 0.70; N=3 vs N=5: 0.92; N=3 vs N=6: 1.14 | Hurt |
| Perturb the initial board by randomly removing 3 pieces of food (ablation) | SSX · Minipacman HUNT | Minipacman position stability: 0.63 | No clear effect |
| Perturb the initial board by randomly removing 3 pieces of food (ablation) | SSX · Minipacman HUNT | Ghost position stability: 1.61 | Mixed |
Closest papers by meaning
- Generation of Policy-Level Explanations for Reinforcement Learning (AAAI 2019 · similarity 0.90)
- Explainable Reinforcement Learning via Model Transforms (NeurIPS 2022 · similarity 0.90)
- Explaining Reinforcement Learning Agents through Counterfactual Action Outcomes (AAAI 2024 · similarity 0.89)
- ORCAID: Oblique Rule-Based Continuous-Action Interpretation for Deep RL Policies (NeurIPS 2026 · similarity 0.88)
- RICE: Breaking Through the Training Bottlenecks of Reinforcement Learning with Explanation (ICML 2024 · similarity 0.87)
- AMAGO-2: Breaking the Multi-Task Barrier in Meta-Reinforcement Learning with Transformers (NeurIPS 2024 · similarity 0.87)
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.