Example 16: ANFIS trained with sezgi¶

The same model class trained without gradients by sezgi's jDE, compared with Jang's hybrid rule on a sine curve.

Needs an extra

This example uses the sezgi bridge (pip install "bulanik[sezgi]"). The documentation build installs it, so the output below is real.

Source¶

examples/16_anfis_with_sezgi.py

"""Train an ANFIS without gradients (sezgi's jDE) and compare with Jang's hybrid rule.

Needs the sezgi extra: pip install "bulanik[sezgi]".
"""

import math

from bulanik.anfis import ANFIS
from bulanik.sezgi import train_anfis

X = [[2 * math.pi * i / 49] for i in range(50)]
y = [math.sin(x[0]) for x in X]

start = ANFIS(1, 3, ranges=[(0, 2 * math.pi)])  # 3 gaussian MFs -> 3 rules, 12 parameters
before = start.get_params()
result = train_anfis(start, X, y, budget=20000, seed=1)
print(f"sezgi RMSE {result.rmse:.4f} ({result.evals_used} evaluations)")

hybrid = ANFIS(1, 3, ranges=[(0, 2 * math.pi)])
history = hybrid.fit(X, y, epochs=10)
print(f"hybrid RMSE {history[-1]:.4f} (10 epochs)")

assert result.rmse < 0.1
assert start.get_params() == before  # train_anfis trains a copy
assert abs(result.anfis.rmse(X, y) - result.rmse) < 1e-12

Output¶

Run when this page is built:

sezgi RMSE 0.0595 (20000 evaluations)
hybrid RMSE 0.0041 (10 epochs)

What to notice¶

train_anfis treats the parameters as one vector and spends a fixed budget of evaluations.
On this smooth problem the hybrid rule reaches a much lower error with far less work; metaheuristics shine when gradients are unavailable.

Go further¶

Run it yourself: python examples/16_anfis_with_sezgi.py.