In th Tencent Weiqi Tournament Li Qincheng 9p beat Ke Jie in what I presume was the human final, and then played AI FineArt taking 2 stones (I don't know if white still had full komi). Rather surprisingly to me, he won convincingly in a game of a messy sprawling fight.

Very interesting. Handicap games are really too hard for bots.

"Too hard" might be the wrong way to put it.

The handicap stones change the opening phase of the game. If you like, think of it as a closely related but slightly different game << just as playing by the ancient rules starting with four stones placed is a slightly different game >>

A neural net has to learn (be trained) to evaluate a function.

Because low handicap go is closely related to no handicap go, I do not think a lot of training would be needed. But it would be necessary.

High handicap go is more different a game. Again training would be needed.

I don't think we can conclude too much from 1 game about the difficulty of top bots beating top pros on 2 stones. FineArt beats plenty of pros on Fox giving them 2 stones (I remember Ke Jie lost online on 2 stones pretty quickly ages ago). Probably a top pro playing for big money in a serious tournament is a tougher challenge than the average pro FineArt beats on Fox. Here I think the problem was FineArt got embroiled in an early disadvantageous fight in which the right goal for the human was fairly clear (keep having enough liberties to keep white dead and don't die yourself) and within the reading ability of a human so FineArt didn't have the opportunity to display its superior strategic judgement and slowly catch up and then outfight in more even positions. I expect Li Qincheng will be aware of the relative strengths and weaknesses of FineArt so this was presumably no accident and well done to him for steering the game in this direction.

@mike Novack
Do you mean this AI was not trained, or not trained enough for a 2 stones handicap? How so?

The winrate change on W pass may hint the difficulty/difference of "2" stones (1 effective handicap stone) games. I think they are still in the realm of ordinary games (after a misplayed corner or the like). The next stone seems more likely to push a bot out of its practice.

I think this conversation may be rather skewed. At least in what I saw, Black still gives 7.5 komi, so the difference is not actually 2 stones. But (if that is correct) I don't see why they didn't use the normal version of playing even but with White giving komi - maybe they just wanted to see how AI copes with handicaps?

This game was an exhibition with the winner of the 4th Tencent tournament. Li beat Ke Jie to get to the final. He then had to play the exhibition game on the same day of the final, I believe, and since time limits in both events were 2h each, he must have been to some extent physically "handicapped."

This is not true. Li's game against Ke Jie was on 20th November. The AI game was on 21st.

What do YOU mean by trained enough? Training a neural net is specific to the function it is being trained to evaluate. A very large amount of training in "facial recognition" (something neiral nets can be trained to do) is not going to help it play go.

What I am saying is that handicap go is slightly different than go and more different the more stones. I am NOT saying that its training for go (level go) is not relevant.

Look, take a step back and consider HUMANS playing go. How did the introduction of komi change how people played? How do handicap stones affect how go is played? You wouldn't expect humans to play the same way without komi as with komi or with handicap stones involved. Now think, HOW did the human know to do that? Don't you think that the neural net would have to learn the difference too? << would need training for THAT >>

Sure that's exactly what I am asking. If there is no specific training for playing a 2 stones then why?

Another question is how to train bot to play handicap games. Obviously white is going to lose if both side play the best move according to the same algorithm. Then there is nothing to be learned. The hard part is the goal of learning now becomes exploring the particular weakness of the opponent.

How to train isn't limited to THIS problem. Ladders come to mind also (and ladder training is MORE than simply seeing whether ladders work --- there is the matter of ladder breakers too. Especially since in high level games the ladders are mostly "potential ladders" --- plays made to threaten a ladder).

How do the strong AI programs handle resignation? When AlphaGo burst on the scene I asked if it would resign immediately if it were to play white giving a handicap. I was somewhat asking as a joke, but I still wonder what happens to the winrate for white in a handicap game. Would all moves for white have way below 50% winrate but we would see the losing winrate move closer and closer to become over 50%? Again, how do these AIs decide to resign?

Typically you set a winrate resignation threshold e.g. 5 or 10%, so if the winrate drops below this the bot resigns (most LZ self-play training games use 5). Obviously you want to set the threshold above the starting winrate for a handicap game (Elfv1 as white giving 2 stones with 7.5 komi is down at 0.4% already, LZ 191 is 12%). Based on its resign behaviour I think the FineArt bot that gives 2 handicaps on Fox also has some logic like if the bot hasn't reduced the winrate deficit by some amount by move 100 it resigns, i.e am I making progress at catching up? No -> Resign. This is just using the vanilla bots designed for even games, there are also modifications of LZ that make it handle handicap games better with things like dynamic komi adjustments.