I had an idea that I /think/ would be useful, perhaps extremely so, so can I get a short assesment on viability? Basically, what might happen if every training generation, you cloned the network and feed each copy very nearly the same seed, and added an incentive for those cards be reasonably to significantly different from each other?
I think it might completely change what the network uses long term storage for, but that also means that none of the existing settings would be anywhere near correct. One incredibly useful thing it might be able to incentivise is the creation of smaller card dumps, because it could grind down the chance of making bad cards further than otherwise. (60% chance of a good card giving a chance at full marks 3/5 times vs 60%^5 or 0.046656% chance at full marks)
Oh, uh, the clones would probably be temporal rather than physical unless you were experimenting with miniature networks.
I think it might completely change what the network uses long term storage for, but that also means that none of the existing settings would be anywhere near correct. One incredibly useful thing it might be able to incentivise is the creation of smaller card dumps, because it could grind down the chance of making bad cards further than otherwise. (60% chance of a good card giving a chance at full marks 3/5 times vs 60%^5 or 0.046656% chance at full marks)
Oh, uh, the clones would probably be temporal rather than physical unless you were experimenting with miniature networks.