跳转至

AWR Phase 14 Implementation Plan

For Claude: REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.

Goal: Add a narrow but package-usable AWR implementation for offline continuous-control training by reusing the current actor/value model lane and the package's discounted returns-to-go data processing.

Architecture: Reuse MLPIQLModel for tanh-Gaussian policy plus value regression, and make the packaged AWR path intentionally narrow: value learning from discounted returns-to-go, then behavior cloning weighted by exponentiated return advantages. Reuse the current offline dataset builder, reward processing, checkpoint, eval, predict, schedule, and early-stopping surfaces instead of creating a new trainer family.

Tech Stack: Python, PyTorch, Gymnasium, existing rl_training offline dataset, returns-to-go, and experiment infrastructure

Task 1: Freeze The Narrow AWR Scope

Files: - Create: docs/plans/2026-03-12-awr-phase14.md - Modify: README.md - Modify: docs/plans/2026-03-12-mainstream-rl-package-design.md - Modify: docs/plans/2026-03-12-rl-expansion-roadmap-design.md - Modify: docs/plans/2026-03-12-rl-yearly-sourcebook-design.md

Step 1: Freeze v1 boundaries

Document the first packaged AWR release as:

  • continuous Box actions only
  • flat vector observations only
  • offline dataset training only
  • single-process trainer only
  • discounted returns-to-go computed from the processed reward stream in v1
  • actor/value updates only, with no separate Q-critic path in this phase
  • no recurrent path
  • no image observations
  • no distributed runtime

Step 2: Record the package rationale

Explain that AWR is the next low-friction 2019 / 2020 batch-RL wave because:

  • it remains a recognizable actor-regression baseline in current offline / imitation libraries
  • it reuses the current actor/value mental model instead of requiring a new sequence or world-model stack
  • it productizes the package's returns-to-go processing for another shipped algorithm

Step 3: Keep test execution deferred

Record that tests are added but intentionally not executed until the user explicitly requests it.

Task 2: Add The AWR Learner

Files: - Create: src/rl_training/algorithms/awr.py - Modify: src/rl_training/algorithms/__init__.py

Step 1: Reuse the actor/value model family

Build AWR on MLPIQLModel and keep the learner narrow by using the policy head plus value head only in v1.

Step 2: Implement package-narrow AWR losses

Keep the package v1 behavior small and explicit:

  • regress the value head to discounted returns-to-go
  • estimate per-sample advantages as returns_to_go - value(obs)
  • optionally normalize those advantages before weighting
  • train the actor with exponentiated advantage-weighted behavior log-probabilities
  • expose knobs for beta and max_weight

Step 3: Expose public loss helpers

Add a readable awr_loss(...) helper and export AWR / AWRAlgorithm through the shared algorithms package.

Task 3: Add The Offline AWR Trainer

Files: - Create: src/rl_training/runtime/awr_trainer.py - Modify: src/rl_training/experiment/registry.py

Step 1: Reuse offline dataset and returns processing

Build the trainer on _infer_env_spaces(...) and _build_offline_dataset(...) from the current offline path, then derive discounted returns-to-go from the processed reward stream.

Step 2: Preserve shared controls

Keep support for:

  • eval_interval
  • early stopping callbacks
  • offline epoch / update budgets
  • learning-rate schedules
  • checkpoint save / resume

Step 3: Reuse standard evaluation and prediction

Evaluate and predict through the current deterministic actor helper path already used by IQL-family algorithms.

Task 4: Wire AWR Into The Package Surface

Files: - Modify: src/rl_training/api/algorithms.py - Modify: src/rl_training/api/__init__.py - Modify: src/rl_training/__init__.py - Create: configs/awr/pendulum.yaml - Create: src/rl_training/assets/configs/awr/pendulum.yaml - Modify: README.md

Step 1: Add the managed API entrypoint

Expose AWR through the root package and API namespaces.

Step 2: Ship a starter config

Add a packaged offline Pendulum-v1 config with narrow AWR defaults.

Step 3: Update package docs

Document AWR as a return-weighted offline actor/value baseline on the same dataset path.

Task 5: Add Unexecuted Coverage

Files: - Create: tests/test_awr_update.py - Create: tests/test_awr_trainer_smoke.py - Modify: tests/test_package_api_exports.py - Modify: tests/test_public_api.py - Modify: tests/test_experiment_manager.py - Modify: tests/test_checkpoint_workflows.py - Modify: tests/test_package_smoke.py - Modify: tests/test_cli.py

Step 1: Add learner-level coverage

Add a unit test for awr_loss(...), invalid beta / max_weight, and one update call.

Step 2: Add trainer smoke coverage

Add a small offline smoke test that checks checkpoint creation, returns-to-go metrics, and eval wiring.

Step 3: Extend package-surface expectations

Update public exports, managed API, checkpoint workflow, and packaged-config tests so AWR is treated as a shipped algorithm.

Step 4: Keep execution deferred

Do not run the tests until the user explicitly asks for test execution.