Research map / Reinforcement learning / Policy gradient and actor-critic
Intentional Updates for Streaming Reinforcement Learning
Arsalan Sharifnassab, Mohamed Elsayed, Kris De Asis, A. Rupam Mahmood, Richard S Sutton
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 ObGD update with intentional updates (StreamAC -> Intentional AC) | streaming actor-critic · MuJoCo Ant-v4 | average return: 4898 -> 5513 | Improved |
| Replace ObGD update with intentional updates | streaming actor-critic · MuJoCo Humanoid-v4 | average return: 4920 -> 6227 | Improved |
| Replace ObGD update with intentional updates | streaming actor-critic · MuJoCo Walker2d-v4 | average return: 3590 -> 3515 | No clear effect |
| Intentional TD instead of StreamTD(lambda) | critic under fixed policy · MuJoCo value prediction | normalized RMS error: StreamTD up to 2.54x error of Intentional TD | Improved |
| Add trace-aware sigma term to step size (vs naive eta/<rho z, z>) (ablation) | Intentional AC · MuJoCo | average return | Improved |
| Add diagonal RMSProp scaling (ablation) | Intentional AC · MuJoCo | average return | Improved |
| Add adaptive delta clipping (ablation) | Intentional AC · MuJoCo | average return | No clear effect |
| Keep SparseInit (ablation) | Intentional AC · MuJoCo Humanoid-v4 | average return: 2503 -> 6227 | Improved |
| Keep SparseInit (ablation) | StreamAC · MuJoCo Humanoid-v4 | average return: 610 -> 4920 | Improved |
| Keep LayerNorm (ablation) | Intentional AC · MuJoCo Ant-v4 | average return: 2323 -> 5513 | Improved |
| Intentional scaling vs alpha=1 on effective update ratio (ablation) | Intentional AC critic · MuJoCo | effective update ratio: 2.99 -> 1.84 | Improved |
| Intentional scaling vs alpha=1 on effective update ratio (ablation) | Intentional AC policy · MuJoCo | effective update ratio: 1.83 -> 2.61 | Hurt |
Closest papers by meaning
- Squeezing More from the Stream : Learning Representation Online for Streaming Reinforcement Learning (ICML 2026 · similarity 0.88)
- Deep Policy Gradient Methods Without Batch Updates, Target Networks, or Replay Buffers (NeurIPS 2024 · similarity 0.86)
- Stable Gradients for Stable Learning at Scale in Deep Reinforcement Learning (NeurIPS 2025 · similarity 0.85)
- Learning to Trust Bellman Updates: Selective State-Adaptive Regularization for Offline RL (ICML 2025 · similarity 0.85)
- PRECISE: SDE-Consistent Stochastic Sampling for RL Post-Training of Flow-Matching Models (NeurIPS 2026 · similarity 0.85)
- Iteratively Refined Behavior Regularization for Offline Reinforcement Learning (NeurIPS 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.
