Research map / Reinforcement learning / Policy gradient and actor-critic
First Order Constrained Optimization in Policy Space
Yiming Zhang, Quan Vuong, Keith Ross
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 |
|---|---|---|---|
| use FOCOPS instead of CPO | safe RL on speed-limited robots · Ant-v3 | reward return: 1406.0 -> 1830.0 | Improved |
| use FOCOPS instead of CPO | safe RL on speed-limited robots · Humanoid-v3 | reward return: 3952.7 -> 4529.7 | Improved |
| use FOCOPS instead of PPO-L | safe RL on speed-limited robots · Humanoid-v3 | reward return: 471.3 -> 4529.7 | Improved |
| use FOCOPS instead of TRPO-L | safe RL on speed-limited robots · HalfCheetah-v3 | reward return: 1621.59 -> 1612.2 | No clear effect |
| use FOCOPS instead of TRPO-L | safe RL on speed-limited robots · Hopper-v3 | reward return: 750.3 -> 1953.4 | Improved |
| use FOCOPS instead of TRPO-L | safe RL on speed-limited robots · HalfCheetah-v3 | cost return: 164.93 -> 152.36 | Improved |
| use FOCOPS instead of PPO-L | circle tasks · Ant-Circle | reward return: 637.4 -> 965.9 | Improved |
| use FOCOPS instead of CPO | circle tasks · Humanoid-Circle | reward return: 671.0 -> 1106.1 | Improved |
| evaluate trained agent on unseen seeds | FOCOPS vs other constrained algorithms · Hopper-v3 generalization | reward return: 2749.9 vs 2446.2 | Hurt |
| evaluate trained agent on unseen seeds | FOCOPS vs other constrained algorithms · Ant-v3 generalization | reward return: 1721.4 -> 1934.9 | Improved |
| set nu_max to +infinity instead of 2 (ablation) | FOCOPS · speed-limit tasks | performance: 0.3% degradation | No clear effect |
| use lambda other than 1.5 (ablation) | FOCOPS · speed-limit tasks | performance: 7.4% worse | Hurt |
Closest papers by meaning
- Absolute State-wise Constrained Policy Optimization: High-Probability State-wise Constraints Satisfaction (NeurIPS 2026 · similarity 0.89)
- Constrained Reinforcement Learning Under Model Mismatch (ICML 2024 · similarity 0.88)
- Last-Iterate Global Convergence of Policy Gradients for Constrained Reinforcement Learning (NeurIPS 2024 · similarity 0.87)
- Deterministic Policies for Constrained Reinforcement Learning in Polynomial Time (NeurIPS 2024 · similarity 0.87)
- Reward Penalties on Augmented States for Solving Richly Constrained RL Effectively (AAAI 2024 · similarity 0.87)
- Near-Optimal Sample Complexity for Online Constrained MDPs (NeurIPS 2025 · similarity 0.86)
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.
