跳转至

XQL Phase 11 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 XQL implementation for offline continuous-control training by reusing the current IQL runtime lane and replacing the value update with the packaged extreme value loss.

Architecture: Reuse MLPIQLModel, the existing offline dataset path, and the current checkpoint / eval / predict plumbing instead of inventing a new runtime family. Implement XQL as an IQL-adjacent learner with the same actor / critic update shape, plus a small value-loss helper that supports a stable Gumbel-rescale objective and a package-default fallback to the familiar expectile path when needed.

Tech Stack: Python, PyTorch, Gymnasium, existing rl_training offline dataset and experiment infrastructure

Task 1: Freeze The Narrow XQL Scope

Files: - Create: docs/plans/2026-03-12-xql-phase11.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 XQL release as:

  • continuous Box actions only
  • flat vector observations only
  • offline dataset training only
  • single-process trainer only
  • IQL-style actor / critic updates with an XQL extreme value update
  • no recurrent path
  • no image observations
  • no online fine-tuning runtime in this phase

Step 2: Record the package rationale

Explain that XQL is the next low-friction offline wave because:

  • the official implementation is built on top of the IQL codebase
  • it extends the current actor / critic / value decomposition instead of creating a new model family
  • it adds a recognizable 2023 offline baseline without forcing a world-model or sequence stack

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 XQL Learner

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

Step 1: Reuse the existing IQL model family

Build XQL on MLPIQLModel and keep the same narrow continuous offline assumptions as IQL.

Step 2: Implement the packaged extreme value loss

Keep the package v1 behavior small and explicit:

  • reuse the current IQL TD critic update
  • reuse the current advantage-weighted actor regression update
  • replace the value update with a stable XQL-style Gumbel-rescale loss over min(Q1, Q2) - V
  • keep a vanilla_value_loss / expectile fallback for a narrow compatibility path
  • expose knobs for loss_temperature, max_value_diff_exp, and max_advantage_weight

Step 3: Expose public loss helpers

Add readable xql_loss(...), xql_value_loss(...), and gumbel_rescale_loss(...) helpers and export XQL / XQLAlgorithm through the shared algorithms package.

Task 3: Add The Offline XQL Trainer

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

Step 1: Reuse the offline dataset stack

Build the trainer on _infer_env_spaces(...), _build_offline_dataset(...), and _evaluate_iql_policy(...) from the current offline path.

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 with the current deterministic IQL helper and expose package prediction through checkpoint workflows.

Task 4: Wire XQL 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/xql/pendulum.yaml - Create: src/rl_training/assets/configs/xql/pendulum.yaml - Modify: README.md

Step 1: Add the managed API entrypoint

Expose XQL through the root package and API namespaces.

Step 2: Ship a starter config

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

Step 3: Update package docs

Document XQL as an IQL-adjacent offline follow-on with an extreme value loss on the same offline data path.

Task 5: Add Unexecuted Coverage

Files: - Create: tests/test_xql_update.py - Create: tests/test_xql_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 gumbel_rescale_loss(...), metric keys from xql_loss(...), and one update call.

Step 2: Add trainer smoke coverage

Add a small offline smoke test that checks checkpoint creation and eval wiring.

Step 3: Extend package-surface expectations

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

Step 4: Keep execution deferred

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