PySATL Expert

Pysatl-expert is experimental expert system designed for the automatic identification of the distribution law of a random variable.

Currently, the system uses an ensemble approach based on bootstrap aggregation and a heuristic decision-making strategy.

---

Features

• Modular architecture: Easily add new distributions, goodness-of-fit criteria, and selection strategies.
• Domain pre-validation: Automatically filters distributions based on their theoretical support.
• Bootstrap analysis: Estimating the confidence level of identification by repeatedly resampling the source data.
• ML-Ready: The system generates standardized feature vectors ready for training machine learning models.

---

Architecture

• core/: Abstract interfaces and base classes of the system.
• distributions/: Implementations of statistical distributions (Normal, Exponential, Weibull).
• criteria/: Wrappers over statistical tests.
• models/: Reports and feature vectors.
• strategy/: The logic for making the final decision.

---

Decision strategy

The identification algorithm is driven by the HeuristicStrategy, which relies on three core principles:

1. Score Unification

Statistical criteria return results on different scales, the strategy aligns them to a single standard -- the Penalty Score.

• The lower the score - the better
• Metrics where "higher is better" are inverted
• Metrics with low variability are scaled so that their contribution is comparable to more sweeping criteria.

2. Complexity Penalty

• If two distributions describe the data with equal accuracy, the simpler model (with fewer parameters) is preferred.

3. Bootstrap Consensus

• The final decision is made based on a vote based on the results of multiple bootstrap iterations. Confidence of the system is calculated as the percentage of votes cast for the winning allocation.

Note: The current implementation of HeuristicStrategy is a baseline solution, with parameters selected empirically.

The main goal of this strategy is to validate the pipeline and test the hypothesis that the selected set of features is sufficient to separate distribution classes. Thanks to the modular architecture, future plans include replacing the heuristic algorithm with a trainable ML classifier (e.g., Random Forest), which will be able to automatically identify nonlinear relationships between statistical metrics without manually adjusting weights.

---

Example

import numpy as np
from pysatl_expert.pipeline import DistributionPipeline
from pysatl_expert.core.pipeline_components import PipelineComponents
from pysatl_expert.distributions.normal import NormalDistribution
from pysatl_expert.distributions.exponential import ExponentialDistribution
from pysatl_expert.criteria.selectors.simple_selector import SimpleCriterionSelector
from pysatl_expert.strategy.heuristic_strategy import HeuristicStrategy
from pysatl_expert.models.feature_extractor import FeatureExtractor

# 1. Assembling the system components
components = PipelineComponents(
    distributions=[NormalDistribution(), ExponentialDistribution()],
    criterion_selector=SimpleCriterionSelector(),
    strategy=HeuristicStrategy(),
    feature_extractor=FeatureExtractor()
)

# 2. Initializing the pipeline
pipeline = DistributionPipeline(components)

# 3. Running identification (with 100 bootstrap iterations)
data = np.random.normal(loc=5, scale=2, size=150)
report = pipeline.identify_best(data, n_bootstraps=100)

print(report)