Research map / Reinforcement learning / Policy gradient and actor-critic
Overcoming The Spectral Bias of Neural Value Approximation
Ge Yang, Anurag Ajay, Pulkit Agrawal
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 |
|---|---|---|---|
| Replace the MLP value approximator with FFN | Value approximation on the Toy MDP · Toy MDP | epochs to fit: 2000 -> 400 | Improved |
| Use a 12-layer MLP instead of the 4-layer MLP (ablation) | Supervised value approximation on the Toy MDP · Toy MDP | fit quality | No clear effect |
| Train the 4-layer MLP for 2000 rather than 400 epochs (ablation) | Value approximation on the Toy MDP · Toy MDP | fit quality: 400 epochs -> 2000 | Improved |
| Use FFN for the critic while retaining an MLP actor (ablation) | SAC · DeepMind Control Suite | learning performance | Improved |
| Use FFN only for the actor and retain an MLP critic (ablation) | SAC · DeepMind Control Suite | learning performance | No clear effect |
| Reduce the Fourier feature-to-input ratio below 40 (ablation) | FFN with SAC · Quadruped-run | learning performance | Hurt |
| Vary the bandwidth parameter (ablation) | FFN with SAC · Walker-run and Quadruped-run | learning performance | Mixed |
| Remove the target network (ablation) | FFN with SAC · Finger-turn and Quadruped-walk | learning performance | No clear effect |
| Remove the target network (ablation) | FFN with SAC · Cheetah-run, Walker-run, and Quadruped-run | learning performance | Hurt |
| Remove the target network (ablation) | MLP with SAC · DeepMind Control Suite | learning performance | Hurt |
| Increase the replay ratio (ablation) | MLP with SAC · Quadruped-run | learning stability | Hurt |
| Increase the replay ratio (ablation) | MLP with SAC · Walker-run | learning performance | Improved |
Closest papers by meaning
- Kernel Value Regression in Offline Reinforcement Learning (NeurIPS 2026 · similarity 0.89)
- Stable Offline Value Function Learning with Bisimulation-based Representations (ICML 2025 · similarity 0.87)
- A Unifying View of Linear Function Approximation in Off-Policy RL Through Matrix Splitting and Preconditioning (NeurIPS 2025 · similarity 0.86)
- Kernel-Based Function Approximation for Average Reward Reinforcement Learning: An Optimist No-Regret Algorithm (NeurIPS 2024 · similarity 0.86)
- Efficient Off-Policy Learning for High-Dimensional Action Spaces (ICLR 2025 · similarity 0.85)
- Neural Network Approximation for Pessimistic Offline Reinforcement Learning (AAAI 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.
