Research map / Reinforcement learning / Visual and representation RL
Compositional Policy Learning in Stochastic Control Systems with Formal Guarantees
Đorđe Žikelić, Mathias Lechner, Abhinav Verma, Krishnendu Chatterjee, Thomas A Henzinger
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 |
|---|---|---|---|
| compose verified edge policies via an abstract graph instead of verifying the whole task directly (ablation) | CLAPS · Stochastic Nine Rooms (0,0)->(2,2) | verified reach-avoid probability: Fail -> 33.0% | Improved |
| use multiplicative RASM bound instead of additive RASM bound, lambda=10 (ablation) | CLAPS bound · 9-Rooms | reach-avoid probability lower bound: 0.9 -> 0.91820 | Improved |
| use multiplicative RASM bound instead of additive RASM bound, lambda=100 (ablation) | CLAPS bound · 9-Rooms | reach-avoid probability lower bound: 0.99 -> 0.99866 | Improved |
| use multiplicative RASM bound instead of additive RASM bound, lambda=1000 (ablation) | CLAPS bound · 9-Rooms | reach-avoid probability lower bound: 0.999 -> 0.99999 | Improved |
| use multiplicative bound to certify 99.9% reach-avoidance with smaller lambda (ablation) | CLAPS bound · 9-Rooms | lambda required: lambda 1000 -> 100 | Improved |
| use systematic graph decomposition instead of manual simplest decomposition via the bottom right room (ablation) | CLAPS · Stochastic Nine Rooms | verifiability | Improved |
| train a single PPO policy directly from initial to target states (ablation) | PPO + Policy+RASM · Stochastic Nine Rooms | verification | Hurt |
Closest papers by meaning
- Learning Control Policies for Stochastic Systems with Reach-Avoid Guarantees (AAAI 2023 · similarity 0.91)
- Compositional Reinforcement Learning from Logical Specifications (NeurIPS 2021 · similarity 0.90)
- Regret-Free Reinforcement Learning for Temporal Logic Specifications (ICML 2025 · similarity 0.86)
- Run-Time Task Composition with Safety Semantics (ICML 2024 · similarity 0.85)
- Assume-Guarantee Reinforcement Learning (AAAI 2024 · similarity 0.85)
- Absolute State-wise Constrained Policy Optimization: High-Probability State-wise Constraints Satisfaction (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.
