lightvector
commented
4 years ago Via Google translate, I get:
katago do a training for star array
Recently, it was discovered that the official website of Star array is conducting targeted training for katago. I think it is possible to conduct a training for Star Array (web version) to achieve a Star Array with the same configuration. Dual card 2080ti, can win 3X
If this automatic translation is not terribly wrong, then I would guess you are claiming that Golaxy is attempting to specifically train an adversarial agent to exploit KataGo and win more games than an agent of its strength normally would win already. And that you are suggesting that KataGo should attempt to do the same back to Golaxy? I of course don't know what your evidence/sources are for this claim, but if they are indeed doing as you claim, then it would sound to me like they're wasting their time.
Targeted training against someone you view as a competitor is the kind of thing you could do if you really want to win some particular tournament against that competitor. But there aren't any tournaments right now so there is little point? You could also do it if you plan to launch some sort of silly PR campaign, like "hey look, our bot wins X% of the time this other bot is so weak". But I'd guess that almost any fixed bot or small set of bots can be beaten > 90% of the time if you can play hundreds of thousands of games and turn the full force of the AlphaZero-like loop towards optimizing against the networks that you wish to exploit. It would be interesting from a research perspective to see how far you can push this of course, but also we already know bots have lots of holes in their understanding, the fact that you can do such a thing would not be so impressive.
Go players know this too - one way to beat your opponent is to learn trick plays that specifically you think the opponent will mess up on, but in the long run you get better by practicing the honest best plays that work well against even the strongest possible resistance you can find, and that you think will work well against everyone. Improving yourself is how you approach "the divine move", rather than tricking one specific opponent.
So again, assuming I haven't misread your message (automatic translation is of course often unreliable!) - I don't think KataGo would want to spend compute doing this kind of thing. It would be interesting research still, but costly, and KataGo's goal is to simply be a good and free engine for anyone to use and be as good as itself instead of trying to beat particular other bot. And unlike closed-source bots, KataGo also is happy to share ideas and algorithms instead of keeping them secret, because that is how you help everyone improve in the future.
Harder-Run
sente361
hope366
LangChuang
poptangtwe
最近发现，星阵官网在对katago进行针对性训练。我觉得可以进行一次针对星阵（网页版）的训练，达到可以超过同配置的星阵。双卡2080ti，可以赢3X