New Go-playing trick defeats world-class Go AI–but loses to human amateurs | Hacker News Reader

New Go-playing trick defeats world-class Go AI–but loses to human amateurs

arstechnica.com

16 pointsredbell7 comments

Reminds me of the dota games played against OpenAI Five during their last experiment. Top level human players were able to adopt an ad-hoc adversarial policy of their own by playing in unorthodox ways to begin to reliably defeat the AI team towards the end of the games despite majority 99%+ getting spanked. Article linked below mentions the player who probably had the best policy, 10 win streak against the bots. https://www.player.one/open-ai-five-99-4-winrate-vs-humans-d...

I'm just about to fall asleep, so possible I'm misreading this, but it seems like this is just a bug in the scoring step, not an actual "trick" to defeat the Go AI, right? Any human player scoring this game would just say the black stones in white's territory are dead and that would be that. If black and white were played by humans and they disagreed, then they would just keep playing the game out, they wouldn't assume black had won. It seems like they are just discovering their program does not mark stones as dead correctly.

I think you're right, but the implication is still that the AI often learns very specific strategies, whereas humans understand the game rules and then learn how to play within those rules.

The AI, despite its large training set, did not learn the rules of the game properly, as shown by this scoring mistake. And I do think the implications for other AIs failing spectacularly are quite notable. We've seen this before with AIs being able to find the tiniest of tumors, but then failing on a tumor that fills half the screen.

> whereas humans understand the game rules

I'm not so sure it's that simple. The "understanding" a humans have is build upon knowing thousands of strategies. Does a person understand chess if he knows the rules but has never played or seen an actual game? I would say no. In some way I argue the AI understands Go better then any human since it mostly defeats humans. That there are strategies that work well against it doesn't mean it doesn't understand the game. Humans miss obvious things all thr time too.

It's kind of weird: looking at the board in the article as a Go player, I would say the game is over and white won. But they're using a ruleset (Tromp-Taylor) which is commonly used in computer Go but almost never between humans, precisely because it requires you to play past the point where it's completely obvious that the game is decided.

So it's not clear that humans understand the rules and computers do not: humans play a slightly different game than the rules the AI is being evaluated against.

I think the AI did learn the game correctly though. It's reasonable for white to pass assuming the black stones in its territory are dead. It's reasonable for black to pass assuming that it cannot make its stones in white's territory live. The only unreasonable thing is the scoring which incorrectly counts the dead black stones as alive.

I believe that human players would also pass in this situation.

At first it was just an adversarial game of Go, and then came the grey goo.

Reply on news.ycombinator.com