Something must have gone wrong then, the AI is supposed to hang a queen if it expects the human to hang a queen, it wouldn't even try to win games.carldaman wrote: ↑Sun Oct 06, 2019 10:45 amDietrich Kappe may have attempted something like that with his BadGyal and EvilGyal nets, but they eventually became too strong for most people.
https://github.com/dkappe/leela-chess-w ... d-Networks
The unsuccessful one was like this, using a zero approach machine learning with a genetic algorithm that would try random inputs and would award how far they got in the level, so new generations generally improve but it's too slow.Laskos wrote: ↑Sun Oct 06, 2019 10:50 amThe goals are different, but is the first one SL and the second one zero approach RL solely, or the latter is SL + RL?Ovyron wrote: ↑Sun Oct 06, 2019 9:17 am
At least, this worked for Super Mario Kart (where the AI trying to imitate the human outperformed the NN trying to win races, with a fraction of the effort. Even when the human wasn't an expert in the game.)
The successful approach was this one, a recurrent neural network that doesn't even play the game at first, it is shown footage of the play by the human, and it tries to guess what will be the next input (even if the input is suboptimal.) The supervised learning had the problem that if the NN saw a situation it hadn't seen before, it'd break, so the human had to record more footage where he'd intentionally get into those situations, so the network could learn how to get out of them. And after it was done by just emulating the human it could drive and win the gold cup by itself (passing the turing test, as I couldn't really tell the difference between the human driving and the AI driving, which would fulfill BrendanJNorman's wish if it worked for chess.)