Research map / Reinforcement learning / Policy gradient and actor-critic
Probabilistic Constrained Reinforcement Learning with Formal Interpretability
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes AWaVO, an intrinsically interpretable constrained RL framework that casts constrained RL as Wasserstein variational inference with an adaptive sliced Wasserstein distance and distributional critic, claiming a Θ(1/√T) global convergence rate and probabilistic decision interpretation on simulated and real quadrotor tasks.
Problem
Interpreting reward functions and optimal policies in reinforcement learning, including convergence guarantees, training transparency, and decision explanations, for safety-critical constrained sequential decision-making.
Approach
Reframes constrained RL as variational inference under an augmented probabilistic graphical model, minimizing an adaptive generalized sliced Wasserstein distance (A-GSWD) between a variational action distribution and the optimality likelihood; the actor sets hypersurface parameters and the distributional critic supplies the variational distribution.
Key results
- Claims a global convergence rate of Θ(1/√T) for AWaVO with formal proofs under stated assumptions.
- Empirically, the convergence rate is reported as within Θ(1/√T) < C_rate ≤ Θ(1/T^1.2) on average; CRPO performs slightly better than AWaVO in the simulated drone task without disturbances.
- In simulation AWaVO is compared against PaETS, TRPO-IPO, PCPO and CRPO over 10 seeds; in the real quadrotor, AWaVO slightly outperforms PCPO.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AWaVO instead of PCPO | quadrotor tracking control · Real-world quadrotor tasks | performance | Helped |
| Use AWaVO instead of CRPO | simulated drone task without disturbances · Simulated Drone | performance | Hurt |
| Use CRPO instead of PaETS | simulated tasks · Acrobot, Cartpole, Walker, and Drone | performance | Helped |
| Use AWaVO instead of PaETS | simulated tasks · Acrobot, Cartpole, Walker, and Drone | performance | Helped |
| Use PCPO instead of TRPO-IPO | simulated tasks · Acrobot, Cartpole, and Walker | performance | Helped |
| Use TRPO-IPO instead of PCPO | simulated tasks · Drone | performance | Helped |
Surprising findings
- CRPO shows comparable or slightly better convergence than AWaVO in simulation, yet the authors favor AWaVO for interpretability.
- PCPO outperforms TRPO-IPO on Acrobot, Cartpole and Walker, while TRPO-IPO outperforms PCPO on the drone tasks.
Limitations
- Trustworthiness of the critic-based posterior is unresolved; no confidence intervals established for the Bayesian network outputs.
- No human-centered validation of the interpretations; policy and regulatory issues not considered.
- Scalability to larger, more complex real-world domains is not formally analyzed.
Experimental setup
Datasets: OpenAI Gym GUARD
Benchmarks: Acrobot Cartpole Walker Drone real quadrotor flight tasks
Metrics: performance convergence rate constraint satisfaction
Baselines: PaETS TRPO-IPO PCPO CRPO
Builds on: CRPO GSWD ASWD distributional RL CPO
Closest papers by meaning
- Polynomial-Time Approximability of Constrained Reinforcement Learning (ICML 2025 · similarity 0.86)
- Last-Iterate Global Convergence of Policy Gradients for Constrained Reinforcement Learning (NeurIPS 2024 · similarity 0.86)
- Deterministic Uncertainty Propagation for Improved Model-Based Offline Reinforcement Learning (NeurIPS 2024 · similarity 0.86)
- DARLING: Detection Augmented Reinforcement Learning with Non-Stationary Guarantees (NeurIPS 2026 · similarity 0.86)
- Multi-Step Likelihood-Ratio Correction for Reinforcement Learning with Verifiable Rewards (NeurIPS 2026 · similarity 0.86)
- Bayesian Exploration Networks (ICML 2024 · 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.