sf_transfer_recovery
Run a bounded successor-features plus GPI transfer recovery experiment for changing linear-reward RL tasks.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- sf_transfer_recovery
- description
- Run a bounded successor-features plus GPI transfer recovery experiment for changing linear-reward RL tasks.
02 资源文件 2
03 SKILL.md 正文
Successor-Feature Transfer Recovery Harness
Use this skill to validate a reduced, mechanism-faithful recovery of Successor Features for Transfer in Reinforcement Learning. It should be used after successor-feature and GPI utilities exist and before claiming the paper mechanism has been recovered.
Inputs
- Attempt directory containing
module_plan.jsonandenvironment/runtime_handoff.json. - Generated skill root containing
successor_feature_modelandgeneralized_policy_improvement. - Soft-mode permission for a reduced/proxy experiment when full four-room or MuJoCo recovery is not feasible.
Outputs
recovery/recovery_result.jsonwithmean_transfer_advantage.recovery/logs/generated_data_item.jsondescribing the gridworld and reward weights.recovery/logs/generated_skill_invocations.jsonproving generated modules were imported/called.recovery/logs/training_trace.jsonrecording exact dynamic-programming/source-policy construction traces.
Workflow
- Construct a small deterministic shared-dynamics gridworld with feature-bearing terminal goals.
- Define source reward weights and held-out transfer weights.
- Derive source policies using exact value iteration for each source reward.
- Compute successor features for each source policy using the generated successor-feature module.
- Reweight source successor features for each held-out reward and apply generated GPI selection.
- Evaluate GPI and a restricted single-source baseline on each held-out reward.
- Save numeric returns, mechanism checks, and validation-ready artifacts.
Validation
Run:
python scripts/run_sf_transfer_recovery.py --attempt-dir <attempt_dir> --skills-root <generated_skills_root>
python tests/test_recovery_harness.py
Limitations
This is a reduced/proxy recovery. It does not reproduce the paper's four-room continuous-navigation or MuJoCo experiments, but it preserves the core mechanism of shared dynamics, changing linear rewards, successor-feature reweighting, and GPI transfer.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).