Research map / Reinforcement learning / Visual and representation RL
Programmatically Interpretable 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 |
|---|---|---|---|
| Replace a neural policy with a sketch-constrained programmatic policy | DRL driving policy · CG-Speedway-1 | reward: 118.39 -> 115.32 | Hurt |
| Replace a neural policy with a sketch-constrained programmatic policy | DRL driving policy · Aalborg | reward: 71.23 -> 54.91 | Hurt |
| Use a sketch-constrained programmatic policy instead of a neural policy | TORCS driving policy · CG-Speedway-1 | steering action standard deviation: 0.5981 -> 0.1312 | Helped |
| Use a sketch-constrained programmatic policy instead of a neural policy | TORCS driving policy · Aalborg | steering action standard deviation: 0.9008 -> 0.2483 | Helped |
| Transfer the learned programmatic policy to unseen tracks rather than use the trained DRL policy | TORCS policy trained on CG-Speedway-1 · CG track 2 and E-Road | race completion | Helped |
| Transfer the learned programmatic policy to unseen tracks rather than use the trained DRL policy | TORCS policy trained on Aalborg · Alpine 2 and Ruudskogen | race completion | Helped |
| Use the programmatic policy instead of the neural policy under blocked sensors | TORCS driving policy · CG-Speedway-1 and Aalborg with RPM and TrackPos blocked | distance raced (meters): 21 -> 1976; 17 -> 200; 71 -> 1477; 20 -> 287 | Helped |
| Remove conditional branching from the policy sketch (ablation) | NDPS · CG-Speedway-1 | reward: 115.25 vs 115.32 | No clear effect |
| Remove conditional branching from the policy sketch (ablation) | NDPS · Aalborg | reward: 52.81 -> 54.91 | Hurt |
| Remove input augmentation (ablation) | NDPS · TORCS tracks | race completion | Hurt |
| Remove sketch guidance (ablation) | NDPS · TORCS tracks | race completion | Hurt |
| Use SMT-based rather than Bayesian parameter optimization (ablation) | NDPS · Acrobot, CartPole, and MountainCar | reward: -127.21 -> -84.16; 143.21 -> 183.15; -143.86 -> -108.06 | Helped |
Closest papers by meaning
- Programmatic Reinforcement Learning without Oracles (ICLR 2022 · similarity 0.91)
- Learning to Synthesize Programs as Interpretable and Generalizable Policies (NeurIPS 2021 · similarity 0.91)
- Synthesizing Programmatic Reinforcement Learning Policies with Large Language Model Guided Search (ICLR 2025 · similarity 0.88)
- π-Light: Programmatic Interpretable Reinforcement Learning for Resource-Limited Traffic Signal Control (AAAI 2024 · similarity 0.86)
- ORCAID: Oblique Rule-Based Continuous-Action Interpretation for Deep RL Policies (NeurIPS 2026 · similarity 0.86)
- Prompted Policy Search: Reinforcement Learning through Linguistic and Numerical Reasoning in LLMs (NeurIPS 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.