Research map / Reinforcement learning / Visual and representation RL
Combining Behaviors with the Successor Features Keyboard
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 CSFA instead of USFA or MSFA point-estimate successor features | SF&GPI in Playroom · Find and Place tasks | GPI performance | Helped |
| Remove categorical representation (scalar estimates, shared estimator) (ablation) | CSFA · Playroom long-horizon tasks | GPI performance | Hurt |
| Do not share the categorical estimator across cumulants (independent) (ablation) | CSFA · Place near task | GPI performance | Hurt |
| Pass gradients from Q-learning to the task encoder (no stop grad) (ablation) | CSFA · Place tasks | GPI performance | Hurt |
| Do not bound task encoding norm ||w|| to unit sphere (ablation) | CSFA · Playroom tasks | GPI performance | Hurt |
| Pass gradients to task encoder from SF approximator (ablation) | CSFA · Playroom tasks | GPI performance | Hurt |
| Decrease number of bins in the pmf (ablation) | CSFA · Playroom full curriculum | GPI performance | Hurt |
| Decrease number of residual blocks in cumulant network (ablation) | CSFA · Playroom long-horizon tasks | GPI performance (%): 100 -> 90 | Hurt |
| Use SFK instead of Distral and MTRL on longer task combinations | Transfer with curriculum · Put x 3 | Samples to reach performance: 100+ fewer samples | Helped |
| Use SFK without a curriculum of shorter tasks | SFK vs Distral · Transfer tasks, no curriculum | Frames to reach same performance: 200+ million frames faster | Helped |
| Replace CSFA with USFA inside SFK (ablation) | SFK · Transfer tasks | Transfer performance | Hurt |
| Sample directly in task-encoding space instead of Bernoulli coefficients (ablation) | SFK · Transfer tasks | Transfer speed | Hurt |
Closest papers by meaning
- Composing Task Knowledge With Modular Successor Feature Approximators (ICLR 2023 · similarity 0.89)
- Learning Successor Features the Simple Way (NeurIPS 2024 · similarity 0.85)
- SF-DQN: Provable Knowledge Transfer using Successor Feature for Deep Reinforcement Learning (ICML 2024 · similarity 0.85)
- The Master Key Hypothesis: Unlocking Cross-Model Capability Transfer via Linear Subspace Alignment (NeurIPS 2026 · similarity 0.83)
- Reward-Aware Proto-Representations in Reinforcement Learning (NeurIPS 2025 · similarity 0.83)
- Constructing an Optimal Behavior Basis for the Option Keyboard (NeurIPS 2025 · similarity 0.83)
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.