Choosing a Workflow

Ex-Fuzzy exposes several modeling paths. Choose the smallest workflow that matches the uncertainty, interpretability, and performance needs of the task.

Estimator Choices

Goal	Recommended entry point	Notes
Standard fuzzy rule classifier	ex_fuzzy.BaseFuzzyRulesClassifier	Good default for most supervised classification tasks.
Interpretable fuzzy regressor	ex_fuzzy.BaseFuzzyRulesRegressor	Type-1 regression with crisp Takagi-Sugeno or fuzzy Mamdani consequents.
Mine rules before optimization	ex_fuzzy.RuleMineClassifier	Useful when frequent, high-confidence patterns should seed the model.
Greedy fuzzy rule tree with native evidential uncertainty	ex_fuzzy.FERL	Returns belief, plausibility, ignorance, and prediction sets without a calibration split.
Coverage-aware predictions	ex_fuzzy.ConformalFuzzyClassifier	Produces prediction sets after calibration on held-out data.
Lower-level rule mining	ex_fuzzy.rule_mining	Use when you need direct control over support, confidence, lift, or rule depth.

Fuzzy Set Types

Type	Use when	Tradeoff
Type-1	You need a simple, fast, interpretable baseline.	Least expressive, easiest to inspect.
Interval Type-2	Membership boundaries are uncertain or noisy.	More expressive, higher computational cost.
General Type-2	You are doing advanced uncertainty modeling.	Most complex and usually best reserved for research workflows.

Backend Choices

backend="pymoo"

Default CPU backend. Use it for reproducibility, checkpoint support, and smaller or medium-sized datasets.

backend="evox"

GPU-oriented backend powered by EvoX and PyTorch. It runs on CUDA when available and otherwise falls back to CPU. Classification and Type-1 regression objectives are evaluated in memory-aware batches.

Minimal Example

from ex_fuzzy import BaseFuzzyRulesClassifier
from sklearn.datasets import load_iris
from sklearn.model_selection import train_test_split

X, y = load_iris(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.33, random_state=0, stratify=y
)

clf = BaseFuzzyRulesClassifier(nRules=10, nAnts=4, backend="pymoo")
clf.fit(X_train, y_train, n_gen=30, pop_size=30)

print(clf.score(X_test, y_test))

Regression Example

from ex_fuzzy import BaseFuzzyRulesRegressor
from sklearn.datasets import make_regression
from sklearn.model_selection import train_test_split

X, y = make_regression(n_samples=500, n_features=5, random_state=0)
X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.25, random_state=0
)

reg = BaseFuzzyRulesRegressor(
    nRules=20,
    nAnts=3,
    consequent_type="crisp",
    backend="evox",
)
reg.fit(X_train, y_train, n_gen=30, pop_size=50)

print(reg.score(X_test, y_test))
print(reg.optimization_device_)  # "cuda" or "cpu"

Reproducibility

Evolutionary fitting is stochastic. For experiments, report the data split, random seeds, population size, number of generations, backend, and fuzzy set type. For benchmark tables, run multiple seeds and summarize the distribution instead of relying on one run.

FERL’s fixed-partition mode is deterministic. Learned-split FERL uses bootstrap cut estimates; set random_state and report learned_n_boot for reproducible experiments.