The first ANFIS training step is merged
The engine can now fit the output side of its fuzzy rules. Here is what that covers, what it does not do yet, and what else changed this week.
What merged
PR #114 merged on 15 October 2026. It adds the first half of ANFIS training: fitting each rule’s linear output by ridge least squares.
- Both rule types. The fit works on the type-1 path and on the interval type-2 path (Nie-Tan).
- Checked against planted answers. Tests build a 3 × 3 × 3 rule base with known outputs, and the fit recovers them on both paths.
- Your units in, your units out. Inputs are standardised for the fit and mapped back exactly to raw units.
What it does not do yet
- No fuzzy-set training. The second half of ANFIS, which moves each set’s centre and width, is the next piece of work.
- No run on the archive. Training needs 1,170 NC-045 distinct device states; 868 NC-069 are archived so far.