Takagi-Sugeno (TSK) inference

TSK consequents are crisp functions of the inputs, so there is no defuzzification: the output is the firing-weighted average of the rule consequents.

A consequent may be:

• a number — zero-order (Sugeno constant);
• a mapping {"const": b0, "x": b1, ...} — first-order linear in the inputs;
• any callable f(**inputs) -> float.

import fuzzytool as fz

x = fz.Variable("x", (0, 10), terms=["small", "large"])

sys = fz.TSK()
sys.rule(x["small"], 0.0)                       # zero-order
sys.rule(x["large"], {"const": 1.0, "x": 1.0})  # first-order: 1 + x
sys.rule(x["small"], lambda x: x**2)            # arbitrary callable

sys(x=10.0)    # -> 11.0

TSK systems are typically faster and differentiable, which makes them the basis for trainable neuro-fuzzy models (ANFIS, on the roadmap).