[PatrickHenkel1]

Phase 6 of the 1.0.0 refactoring, the last phase: Run replaces the Logger, models can be compared across several training runs each, the documentation is rewritten and the version is 1.0.0. After this merge, refactoring_v1.0.0 can go to main.

Run instead of the Logger

physXAI.utils.logging is removed (Logger, get_full_path, create_full_path):

• Run.save(name, prep, model, td) writes everything about a trained model:

  • {name}_preprocessing.json
  • {name}_model.json
  • {name}_training_data.json (metrics, horizon metrics, history)
  • the model file
  • {name}.html (the report)

  It uses the same file layout as AgentLib-MPC; save_for_agentlib now calls it.

• Saving in pipeline: pipeline / online_pipeline save only with an explicit save_path. There is no default directory anymore. model.trained_model holds the model of the last pipeline.

• Logging: messages go through the physXAI logger from the standard library.

  • set_log_level('warning') replaces Logger.print_level.
  • The Keras progress bars follow the log level.
  • If the application configures logging itself, the messages are handed to it instead of being printed.

• Feature selection: save_result(result, run) writes into a run directory.

Model comparison (physXAI.evaluation.comparison)

• repeat_training(path, prep, model, seeds=5, run=run, name='ann') trains a model with several seeds and saves each run to <run>/<name>/seed_<seed>/.

  • By default the seed changes both the data split and the weight initialization (vary=('split', 'init')).
  • Models trained with the same seeds share the same splits.

• ModelComparison collects the runs:

  • ModelComparison.load(run.path) treats each sub directory as a model.
  • It also accepts a list of directories, {label: dir}, or training data held in memory.
  • Runs saved by physXAI 0.x can be loaded too.

• Tables:

  • runs(): seeds, model class, training time.
  • metrics() and horizon().
  • scores(): one row per run.
  • summary(): per model the number of runs, mean, std, min, median and max.

• The errors over the horizon are compared on the horizon that all runs reach. This makes models with different outputs comparable, e.g. Change(T) against the trajectory of an RNN.

• Plot:

  • box plots of the single step RMSE, one panel per output (each point is one run);
  • RMSE and bias over the horizon: median and min–max band (or IQR);
  • box plots of the mean RMSE over the horizon and of the RMSE at its end;
  • a summary table.

• New executables/bestest_hydronic_heat_pump/TAir_compare.py: LinReg, ANN, a modular model with recursive training and an LSTM, 3 seeds each, about 2 min.

  Horizon RMSE reaTZon_y (test, K)	mean	std	min	max
  linreg	0.33	0.09	0.24	0.42
  ann	0.30	0.09	0.20	0.37
  modular_recursive	0.26	0.12	0.16	0.39
  rnn	0.29	0.18	0.10	0.47

  The spread across seeds is larger than the differences between the models. With 14 days of BOPTEST data, the test set of each RNN run holds only a single window.

Documentation

• User guide: getting started, expressions, data pipeline and split, models, multi step evaluation and training, recurrent models, evaluation and comparison, AgentLib-MPC export.
• Migration guide 0.x → 1.0: a table mapping old to new calls, and TAir_sim_linreg rewritten for 1.0 (vT inside the model with ModularANN(solver='ols'), and the column alternative).
• API reference for all modules (expr, io, training, splitting, windows, comparison, …). The pages for removed modules are gone.
• README updated (example, model types, structure).
• mkdocs build --strict passes. To get there, four docstrings that named parameters missing from their signatures were fixed, and warnings about missing types are switched off.

CI and cleanup

• New workflow quality.yml: ruff check and mkdocs build --strict for PRs. gh-pages.yml also builds with --strict.
• Ruff findings fixed in the legacy files (imports, whitespace, an undefined type name, ambiguous names). regression_prototype.py is excluded as experimental.
• test_coverage.py: each test writes into its own tmp_path run, so the tests no longer depend on each other through the shared Logger directory (26 integrate modular anns #37).
• Version 1.0.0; CHANGELOG entry [1.0.0] – 2026-11-06.

Tests

• Unit tests: 134 passed, 1 skipped (the RNN contract test against AgentLib main). New tests:
  • comparison: saving, seeds and splits, loading equals the in-memory runs, summary, plot and saving, a single step model and an RNN on a common horizon, runs of 0.x;
  • Run.save;
  • logging and Keras verbosity.
• AgentLib-MPC main: the test_physXAI_plugin tests (3 passed) and simple_mpc_nn_physXAI.run_example(testing=True) pass.
• Executables: all executables, including TAir_compare.py, pass the smoke run. TAir_compare.py also ran with the full number of epochs.

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.

An organization owner can choose the repository organization as the billing source for reviews requested by members with a paid Copilot license. Learn about Copilot code review billing.