Candidate hypotheses · none proven · you decide

Engram Hypotheses

The idea hiding between two sources nobody read together, proposed by a machine, killed-or-kept by a jury, shown with its sources.

← frontier
ecology-evolution × machine learning · unresolved

Does a neuroevolution algorithm do best when it obeys biology's mutation-load law?

deleterious mutation loadneuroevolution performance · via mutation rate

Population genetics says a population's load of harmful mutations scales as U·d/s, the mutation supply over the strength of selection. A neuroevolution algorithm sets its mutation rate by hand (here, 0.03). The system proposed these are the same knob: the network should evolve best when the selection pressure it imposes is tuned to that fixed mutation rate, holding the harmful-mutation load at the optimum a biological population settles into.

If a neuroevolution run and a biological population obey the same load law, then sweeping the algorithm's selection strength against its fixed mutation rate should trace the U·d/s curve, with peak performance at the load minimum. Does that hold quantitatively, or is the shared 'mutation rate' just a coincidence of vocabulary?

It read full-text papers on mutation load, selection, and neuroevolution, grounded the facts, and re-asked the jury. The quantitative link was plausible, but the facts did not pin the biological and algorithmic mutation rates to one mechanism, so it stayed open.

Is this a real connection or a coincidence of shared words? The facts above are grounded in the sources; the leap between them is what is unproven. Make the case, or settle it with a reference.