tsetly

A tiny Tsetlin Machine in NumPy: little automata that learn rules you can read.

      forget ◀────────────────┼────────────────▶ remember
             1   2   3  …  N  │  N+1  …  2N

A Tsetlin Machine learns AND-rules over bits, like x0 ∧ ¬x1. Every rule is a team of Tsetlin automata, one per bit and one per negated bit. Each automaton walks along its states as the data rewards or penalises it, and once it crosses the middle, its bit joins the rule. Rules vote for or against each class, and the class with the most votes wins. That is the whole machine (Granmo, 2018).

Install

pip install tsetly

Learn XOR through noise

Twelve random bits, of which only the first two matter, and a fifth of the training labels flipped:

import numpy as np
from tsetly import TsetlinMachine

rng = np.random.default_rng(0)
X = rng.integers(0, 2, size=(5000, 12))
y = X[:, 0] ^ X[:, 1]
noisy = np.where(rng.random(5000) < 0.2, 1 - y, y)

tm = TsetlinMachine(clauses=8, T=15, s=3.9, seed=1)
tm.fit(X[:4000], noisy[:4000], epochs=20)

print(tm.score(X[4000:], y[4000:]))  # 1.0
print(tm)

class 0
   +2  ¬x0 ∧ ¬x1
   +2  x0 ∧ x1
   -2  x0 ∧ ¬x1
   -2  ¬x0 ∧ x1
class 1
   +3  ¬x0 ∧ x1
   +1  x0 ∧ ¬x1
   -3  ¬x0 ∧ ¬x1
   -1  x0 ∧ x1

It found XOR, ignored the ten distracting bits, and can tell you so.

API

License

MIT