Research map / Reinforcement learning / Policy gradient and actor-critic
LipsNet: A Smooth and Robust Neural Network with Adaptive Lipschitz Constant for High Accuracy Optimal Control
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 MLP actor with LipsNet-L (local Lipschitz via MGN) | TD3 actor on DMControl · DMControl Cartpole/Reacher/Cheetah/Walker, no noise | action fluctuation ratio: Cartpole 0.04→0.01; Reacher 2.07→0.01; Cheetah 1.08→0.94; Walker 1.89→0.93 | Helped |
| Replace MLP actor with LipsNet-L | TD3 actor on DMControl · DMControl, no noise | total average return: Cartpole 805→831; Reacher 981→983; Cheetah 816→822; Walker 926→945 | No clear effect |
| Replace MLP actor with LipsNet-G (global Lipschitz) (ablation) | TD3 actor on DMControl · DMControl Cartpole and Cheetah, no noise | total average return: Cartpole 805→691; Cheetah 816→702 | Hurt |
| Replace MLP actor with LipsNet-L | TD3 actor on DMControl · DMControl with observation noise | action fluctuation ratio: Cartpole 0.58→0.17; Reacher 2.41→0.04; Cheetah 1.13→1.08; Walker 2.02→1.21 | Helped |
| Replace MLP actor with LipsNet-L | TD3 actor on DMControl · DMControl with observation noise | total average return: Cartpole 763→823; Reacher 972→978; Cheetah 813→818; Walker 911→929 | Helped |
| Replace MLP actor with LipsNet-L | TD3 on Gym humanoid · Gym humanoid | action fluctuation ratio: 0.83→0.66 (about 20% reduction) | Helped |
| Replace MLP actor with LipsNet-L | TD3 on Gym humanoid · Gym humanoid | total average return: 4968→5266 | Helped |
| Replace MLP actor with LipsNet-L | SPIL on vehicle trajectory tracking · sine curve, observation noise 1e-2 | action fluctuation ratio: 0.224→0.022 | Helped |
| Replace MLP actor with LipsNet-L | INFADP on double integrator · double integrator, no noise | total average return and action fluctuation ratio: return -78.4→-58.7; fluctuation 0.34→0.12 | Helped |
| Add reward penalty on consecutive action difference (sparse reward) (ablation) | TD3 on DMControl Cartpole · Cartpole sparse reward | action fluctuation ratio: 0.04 (MLP) → 0.27 at alpha 0.01, 0.21 at 0.1, 0.02 at 1 (return collapses to 13 at 1) | Hurt |
| Replace MLP actor with layer-wise spectral norm (MLP-SN) (ablation) | INFADP on double integrator · double integrator, no noise | action fluctuation ratio: 0.34→0.13 (return -78.4→-65.9) | Helped |
Closest papers by meaning
- LipsNet++: Unifying Filter and Controller into a Policy Network (ICML 2025 · similarity 0.94)
- ODE-based Smoothing Neural Network for Reinforcement Learning Tasks (ICLR 2025 · similarity 0.90)
- Improve Robustness of Reinforcement Learning against Observation Perturbations via l∞ Lipschitz Policy Networks (AAAI 2024 · similarity 0.89)
- A Dynamical System Perspective for Lipschitz Neural Networks (ICML 2022 · similarity 0.89)
- A provable control of sensitivity of neural networks through a direct parameterization of the overall bi-Lipschitzness (NeurIPS 2024 · similarity 0.85)
- Enhancing Certified Robustness via Block Reflector Orthogonal Layers and Logit Annealing Loss (ICML 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.