Feature for ML by faux66 · Pull Request #4 · PySATL/pysatl-expert

faux66 · 2026-04-14T19:18:59Z

1. Mathematical Foundation Expansion

Implemented 8 continuous probability distributions: Normal, Exponential, Weibull, Gamma, LogNormal, Student's t, Beta, and Uniform

2. Update GenericCriterion

Implemented Dynamic Parameter Injection: The adapter now automatically maps distribution fitting results from scipy.stats (e.g., shape, mu) to specific engine attributes required by pysatl-criterion (e.g., a, loc, df) using a comprehensive alias map.
Introduced Signature Introspection: Using the inspect module, the system now determines if a statistic requires cdf_vals. This allows for Lazy Evaluation, calculating the theoretical CDF only when explicitly requested by the test engine

3. Automated Discovery (DynamicCriterionSelector)

Developed selection mechanism that dynamically scans the pysatl-criterion library at runtime. This eliminates manual class mapping.

4. Feature for ML (FeatureVector)

Developed selection mechanism that dynamically scans the pysatl-criterion library at runtime. This eliminates manual class mapping and allows the system to automatically support new statistical tests as they are added to the external library.
Implemented Missing Value Encoding: If a test is mathematically inapplicable (e.g., negative values for Exponential) or fails, it is assigned a value of -1.0.

…lity

f1i3g3

Выглядит круто

…chicalExpertModel

…pendendency

faux66 · 2026-08-10T18:14:48Z

Overview

This Pull Request introduces the core Machine Learning Expert System for pysatl-expert. It replaces the legacy empirical heuristics with a data-driven 2-stage Hierarchical Random Forest Classifier (HierarchicalExpertModel) powered by a high-performance SQLite Critical Value Cache (CVCache and MLStrategy).

The system automatically identifies empirical sample distributions across 8 candidate laws (Normal, Exponential, Weibull, Uniform, Student, Gamma, Beta, LogNormal) with over 95% accuracy, supporting sample sizes N between 30 and 1000 and location shifts (loc != 0).

Architectural Breakdown & Core Components

1. 2-Stage Hierarchical Classifier (`HierarchicalExpertModel`)

Stage 1 (Family Model): Predicts distribution families (Family 1, Family 2, Family 3) obtained via Ward hierarchical clustering.
Stage 2 (Sub-Family Models): Dedicated Random Forest classifiers predict the exact probability distribution within the identified family.
Feature Selection: Automatically selects top 30 Stage 1 features and top 20 Stage 2 intra-family features during model fitting to optimize accuracy and execution speed.

2. Fast Critical Value Cache and ML Strategy (`CVCache`, `MLStrategy`)

Microsecond Critical Value Lookup: CVCache preloads compressed header and zstd-compressed float32 blobs from expert_cv_database.sqlite and performs O(1) linear interpolation (np.interp) over sample size N.
Feature Binarization: Binarizes raw Goodness-of-Fit (GoF) test values (1.0 if statistic <= CV_95%, else 0.0) to form robust decision boundaries.

3. Feature Space & Dynamic Criteria Selector (`FeatureVector`, `DynamicCriterionSelector`)

Fixed 192-Dimensional Feature Vector: Aggregates 8 descriptive sample statistics (min, max, sample_size, skew, kurtosis, coef_of_variation, relative_iqr, entropy) and 184 binarized GoF test scores.
Single Canonical Blacklist: Consolidates computationally heavy or missing database criteria (bhs, kl_int, kl_sup, cq*, rs, ahs, hp, graph metrics) into FeatureVector.BLACKLIST to guarantee complete schema synchronization across criteria selectors and feature extractors.

4. Facade API and Visual Report Generator (`app.py`, `visual_report.py`)

Facade Function: evaluate_sample(data) provides a top-level single-function interface that automatically wires pipeline components, fits parameters, executes identification, and formats output.
Visual Diagnostics: Generates 4-panel PNG report figures containing Empirical Histogram vs Fitted PDF, ECDF vs Theoretical CDF, Q-Q Plot, and Candidate Distribution Rankings.

5. Workspace Reorganization

scripts/ Directory: Contains workflow utilities including generate_ml_dataset.py (multiprocessed 16,000-sample dataset generator), cluster_distributions.py (Ward hierarchical clustering script), and train_model.py (model training script).

f1i3g3 · 2026-08-10T18:54:13Z

Документация же такая же, как в experiment?

f1i3g3 · 2026-08-12T00:17:26Z

+            for target in potential_targets:
+                if hasattr(self.engine, target):
+                    setattr(self.engine, target, p_value)
+                    break


Здесь установит первый попавшийся, так и должно быть?

f1i3g3 · 2026-08-12T00:22:59Z

-        "lilliefors",
-        "cramer_von_mises",
-        "gini_index",
-        "moran_test",


Сунуть в БД в будущем? Если релизить в прод.

f1i3g3 · 2026-08-12T00:36:14Z

+        self.stage1_model = RandomForestClassifier(
+            n_estimators=200, max_depth=12, random_state=42, n_jobs=2
+        )
+        self.stage1_model.fit(X_df[self.stage1_features], y_families)


Кажется, здесь снова обучается на всех данных. Так и должно быть?

f1i3g3 · 2026-08-12T00:38:21Z

+
+
+class CVCache:
+    """Thread-safe cache for critical values queried from SQLite database.


threading.Lock не увидел.

f1i3g3 · 2026-08-12T00:41:18Z

Точно кластеризация следует за цветами на дендрограмме?

f1i3g3 · 2026-08-12T00:45:26Z

Пометка на потом: починить criterion.

faux66 added 5 commits April 14, 2026 21:13

feat: new dependencies

38ec361

feat: add new distributions

2cf0cdb

refactor: update generic for criterion compatibility

0babe03

feat: implement selector for automated criteria selection

0401a3d

refactor: update class with fixed-size FeatureVector for ML compatibi…

749fc64

…lity

f1i3g3 assigned faux66 Apr 14, 2026

f1i3g3 added the enhancement New feature or request label Apr 14, 2026

f1i3g3 self-requested a review April 14, 2026 21:19

f1i3g3 reviewed Apr 14, 2026

View reviewed changes

alex98247 and others added 14 commits June 3, 2026 17:26

Update dependency to pysatl-criterion

8ccb594

feat: support location shift for semi-bounded distributions

916a91a

docs: adopt Google Style docstrings

617f9ff

docs: update docs

5d194c1

docs: update docs

5b37a69

refactor: remove legacy strategy and selector

2955e1c

feat: add MLStrategy and CVCache for fast work with local DB

cfba758

feat: add family clustering script, generated json-config, and Hierar…

ac92eb8

…chicalExpertModel

feat: add dataset generator and model training script

94ac882

feat: add CLI apllication and visual report generator

01e25f4

refactor: optimize FeatureVector, dynamic criteria selector, clean de…

9f53d6a

…pendendency

fix: test and pyproject

c2002da

fix: ci

36810fa

fix: deps verisons

08e49be

f1i3g3 self-requested a review August 10, 2026 18:53

f1i3g3 approved these changes Aug 12, 2026

View reviewed changes

faux66 requested a review from alex98247 August 13, 2026 13:43

Navigation Menu

Uh oh!

Feature for ML - #4

Feature for ML#4
faux66 wants to merge 19 commits into
mainfrom
feat/mlstrategy

faux66 commented Apr 14, 2026

Uh oh!

f1i3g3 left a comment

Uh oh!

faux66 commented Aug 10, 2026

Uh oh!

f1i3g3 Aug 10, 2026

Uh oh!

f1i3g3 Aug 12, 2026

Uh oh!

f1i3g3 Aug 12, 2026

Uh oh!

f1i3g3 Aug 12, 2026

Uh oh!

f1i3g3 Aug 12, 2026

Uh oh!

f1i3g3 Aug 12, 2026

Uh oh!

f1i3g3 Aug 12, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants



		class CVCache:
		"""Thread-safe cache for critical values queried from SQLite database.

Uh oh!

Conversation

faux66 commented Apr 14, 2026

Uh oh!

f1i3g3 left a comment

Choose a reason for hiding this comment

Uh oh!

faux66 commented Aug 10, 2026

Overview

Architectural Breakdown & Core Components

1. 2-Stage Hierarchical Classifier (HierarchicalExpertModel)

2. Fast Critical Value Cache and ML Strategy (CVCache, MLStrategy)

3. Feature Space & Dynamic Criteria Selector (FeatureVector, DynamicCriterionSelector)

4. Facade API and Visual Report Generator (app.py, visual_report.py)

5. Workspace Reorganization

Uh oh!

f1i3g3 Aug 10, 2026

Choose a reason for hiding this comment

Uh oh!

f1i3g3 Aug 12, 2026

Choose a reason for hiding this comment

Uh oh!

f1i3g3 Aug 12, 2026

Choose a reason for hiding this comment

Uh oh!

f1i3g3 Aug 12, 2026

Choose a reason for hiding this comment

Uh oh!

f1i3g3 Aug 12, 2026

Choose a reason for hiding this comment

Uh oh!

f1i3g3 Aug 12, 2026

Choose a reason for hiding this comment

Uh oh!

f1i3g3 Aug 12, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

1. 2-Stage Hierarchical Classifier (`HierarchicalExpertModel`)

2. Fast Critical Value Cache and ML Strategy (`CVCache`, `MLStrategy`)

3. Feature Space & Dynamic Criteria Selector (`FeatureVector`, `DynamicCriterionSelector`)

4. Facade API and Visual Report Generator (`app.py`, `visual_report.py`)