Use Taguchi Orthogonal Arrays as an optional search criteria (PR attached)

[bigattichouse]

Instead of 100 iterations overnight, we can search 81 (or 243, or more) parameters in an orthogonal array using the Taguchi method. This means we can iterate much faster when searching for changes with higher signal to noise rations (SNR/dB). So, instead of just 100 random walks, you could search the space of 3-6X the coverage of your current overnight runs.

I've created a sample change that uses my POSIX-style C taguchi program/lib: https://github.com/bigattichouse/taguchi I realize that you prefer stand-alone projects, I'm sure it could be implemented in native python (there are a few external libraries). I just wanted to show you the possibility and see what you thought.

I've been using this, and my BluePrint (https://github.com/bigattichouse/blueprint ) experiment/model prompts to make designing experiments in my home lab much more fruitful, and as soon as I saw your post (and the flurry of posts around it) I realized that the method might drastically improve the efficiency of your search.

I'll make a pull request associated with this issue in a bit, feel free to ignore it, but I'd be interested in discussion about it, as I think the idea could really help, even if you re-implement it.

• Mike @ BigAtticHouse

[bigattichouse]

It needs some love, I really only wanted to present the idea with some code to back it up. I think your system could really benefit from the technique, even if it doesn't specifically use what I've done.