Using deep learning in CGOL search

[pcallahan]

This may be a naive question, since I'm far behind the state of the art in CGOL search, and can only watch with amazement as oscillators like the recent p41 are found, but I'm curious if any of the newer methods are applying deep learning or similar approaches in guiding their searches. I've seen hints of it, and even a thread on tensorflow, but that looked connected to using GPUs to carry out conventional CGOL operations.

Some historical context:

When I was working on ptbsearch in 1996-97, I was able to get some results from a conventional backtracking search, but this required terminating damaged catalysts fast, unless specifically introduced as transparent blocks with greatly increasing running time. It also required placing them blindly at locations where they were very unlikely to be useful, since I had no better way to determine this than to run them, even if only a small fraction of expanding frontiers had any chance of doing more than destroying them.

Around the same time and unbeknownst to me, David Buckingham was building the first set of Herschel conduits without the benefit of a backtracking search (I asked him later) but largely through his own deep intuition. Presumably, he could watch the expanding frontier of a pattern and make a guess at where it was worth placing a catalyst. Even so, he was just as limited as my search was in selecting from a library of potential catalysts.

In short, what would it take to have a search program that could explore the addition of catalysts and other parts exhaustively, but had the computational equivalent of David Buckingham looking at the pattern and actually coming up with a shortlist of what is worth exploring? Even better, it should be able to try possibilities that are unknown to human intuition, but could be trained by generating a set of trials.

Two questions:

• How much is this even worth doing? E.g. maybe CGOL is so sensitive to small variations that enumerating an exact set of small interactions makes more sense than training to recognize ones that are roughly similar.
• Are some of the successful search programs already doing this?

My old intuition would have been that recognizing patterns by machine learning would not be useful. I.e., you can recognize a picture of a cat using a variety of identifiable features, but to recognize an expanding CGOL frontier that miraculously causes a block to vanish and reappear in the same place 30 steps later can only be determined by carrying out the test.

However, it could still be useful in guiding search, i.e. recognizing patterns that have a higher probability (though still very small) of resulting in a transparent block reaction. Using generative AI, it might even be possible to try out many of these and test them, resulting in a higher probability of finding new transparent block reactions that an exhaustive enumeration.

[hotdogPi]

My oscillator search program has parameters set manually (mainly active region, maximum offset, number of generations both before and after the spark, "pattern too big" threshold, and "pattern stabilized" detection which incorrectly includes moving gliders since they have constant population), and it's possible that some algorithm could determine them automatically (e.g. if I search a 31×31 area for the second most common active region, it would decide to search within 23×23 for the sixth most common — or maybe 17×23 if it tends to expand on one axis more than the other).

[pcallahan]

My oscillator search program has parameters set manually ... and it's possible that some algorithm could determine them automatically

That's a good point that I hadn't considered. Whether or not machine learning can recognize useful patterns intuitively, it can probably help refine tuning parameters, at least if there are too many to handle exhaustively.