Research map / Continual learning / Online and replay-based
Building a Subspace of Policies for Scalable Continual Learning
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 |
|---|---|---|---|
| Use an adaptively extended policy subspace instead of FT-N's separate stored models | Continual RL on Brax HalfCheetah · Brax HalfCheetah scenarios | Performance: 1.16±0.20 -> 1.27±0.27 | Helped |
| Use an adaptively extended policy subspace instead of FT-N's separate stored models | Continual RL on Brax Ant · Brax Ant scenarios | Performance: 0.97±0.20 -> 1.11±0.17 | Helped |
| Use an adaptively extended policy subspace instead of FT-N's separate stored models | Continual RL on Brax Humanoid · Brax Humanoid scenarios | Performance: 0.65±0.46 -> 1.76±0.19 | Helped |
| Use an adaptively extended policy subspace instead of PackNet | Continual RL on Brax HalfCheetah · Brax HalfCheetah scenarios | Performance: 0.85±0.14 -> 1.27±0.27 | Helped |
| Use an adaptively extended policy subspace instead of PackNet | Continual RL on Brax Ant · Brax Ant scenarios | Performance: 1.08±0.21 -> 1.11±0.17 | Helped |
| Use an adaptively extended policy subspace instead of PackNet | Continual RL on Brax Humanoid · Brax Humanoid scenarios | Performance: 0.96±0.21 -> 1.76±0.19 | Helped |
| Use an adaptively extended policy subspace instead of PackNet | CW10 · CW10 | Performance: 0.83 ± 0.02 -> 0.81 ± 0.06 | No clear effect |
| Use an adaptive number of policy heads instead of one per task | CSP on CW10 · CW10 | # Heads: 10 -> 5.3 ± 1.6 | Helped |
| Select policies with Monte Carlo rollouts rather than the learned critic (ablation) | CSP on Brax HalfCheetah · Brax HalfCheetah scenarios | Performance: 1.27±0.27 -> 1.88±0.19 | Helped |
| Select policies with Monte Carlo rollouts rather than the learned critic (ablation) | CSP on Brax Ant · Brax Ant scenarios | Performance: 1.11±0.17 -> 1.24±0.07 | Helped |
| Select policies with Monte Carlo rollouts rather than the learned critic (ablation) | CSP on Brax Humanoid · Brax Humanoid scenarios | Performance: 1.76±0.19 -> 1.98±0.22 | Helped |
| Increase the threshold for extending the subspace (ablation) | CSP · Brax threshold analysis | Performance: performance is still above $70\%$ | Hurt |
Closest papers by meaning
- Self-Composing Policies for Scalable Continual Reinforcement Learning (ICML 2024 · similarity 0.88)
- Continual Reinforcement Learning by Planning with Online World Models (ICML 2025 · similarity 0.87)
- Online Continual Learning for Interactive Instruction Following Agents (ICLR 2024 · similarity 0.87)
- Autonomous Option Invention for Continual Hierarchical Reinforcement Learning and Planning (AAAI 2025 · similarity 0.86)
- Continual Knowledge Adaptation for Reinforcement Learning (NeurIPS 2025 · similarity 0.86)
- Knowledge Retention in Continual Model-Based Reinforcement Learning (ICML 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.