Research map / Reinforcement learning / Policy gradient and actor-critic
Noisy Derivative-Free Optimization With Value Suppression
Hong Wang, Hong Qian, Yang Yu
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 |
|---|---|---|---|
| add value suppression to SRACOS (VS vs MPS baseline) | SRACOS · Ackley n=100 sigma=0.1 | true function value (lower better): 2.95 -> 0.93 | Improved |
| add value suppression to SRACOS (VS vs MPS baseline) | SRACOS · Sphere n=1000 sigma=1 | true function value (lower better): 97.16 -> 72.41 | Improved |
| add value suppression to SRACOS with single positive solution (VS+NO MPS vs NO MPS) (ablation) | SRACOS · Sphere n=100 sigma=1 | true function value (lower better): 15.41 -> 7.41 | Improved |
| use multiple positive solutions (MPS) instead of one (ablation) | SRACOS · Ackley n=100 sigma=0.1 | true function value (lower better): 3.69 -> 2.95 | Improved |
| add re-evaluation, threshold selection, or both to SRACOS | SRACOS · Sphere n=100 sigma=1 | true function value (lower better): 15.41 (NO MPS) -> 24.07 (REEVAL), 24.88 (TS), 20.65 (REEVAL+TS) | Hurt |
| add sampling (10 samples) to SRACOS | SRACOS · Sphere n=100 sigma=1 | true function value (lower better): 15.41 -> 6.53 | Improved |
| value suppression vs MPS baseline on OpenAI Gym | SRACOS · HalfCheetah-v1 | accumulated reward: 1388.27 -> 1924.60 | Improved |
| value suppression vs MPS baseline on OpenAI Gym | SRACOS · LunarLander-v2 | accumulated reward: -187.70 -> 80.40 | Improved |
| add extra Gaussian noise (1x noise level) | VS · HalfCheetah-v1 | accumulated reward: 1924.60 -> 1286.29 | Hurt |
| change max non-update iterations u from 500 to 100 or 1000 (ablation) | VS · OpenAI Gym tasks | accumulated reward: HalfCheetah-v1 1924.60 (u=500) -> 1554.27 (u=100), 1773.89 (u=1000) | No clear effect |
Closest papers by meaning
- Dynamic Anisotropic Smoothing for Noisy Derivative-Free Optimization (ICML 2024 · similarity 0.83)
- Stochastic Zeroth-Order Optimization Under Heavy-Tailed Noise (NeurIPS 2026 · similarity 0.83)
- Colored Noise in PPO: Improved Exploration and Performance through Correlated Action Sampling (AAAI 2024 · similarity 0.83)
- How Free is Parameter-Free Stochastic Optimization? (ICML 2024 · similarity 0.82)
- Is Noise Conditioning Necessary for Denoising Generative Models? (ICML 2025 · similarity 0.82)
- Derivatives of Stochastic Gradient Descent in parametric optimization (NeurIPS 2024 · similarity 0.82)
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.
