Hacker News Reader

Comment by dlevine | Hacker News Reader

Apparently 2 stones is a huge advantage. An estimate is that the computer is roughly 4-600 ELO stronger on an even match.

Also, the human played a strategy tailored to that huge initial advantage. He said that the AI did not handle this particularly well, and played high probability moves instead of trying to lure him into a mistake.

Also, even though this was the best Go engine, it was not running on a supercomputer, and had a relatively limited amount of time per move.

So, this was an important victory for a human, but not a sign that humans are now stronger than AIs at Go.

View on HN

> Also, the human played a strategy tailored to that huge initial advantage. He said that the AI did not handle this particularly well, and played high probability moves instead of trying to lure him into a mistake.

Yeah, katago's training is not really focused at all on handicap games, because it's by nature learning from even games against similar-strength opponents.

It doesn't have specific training from playing in a way to exploit a weaker player. In a handicap game you have to give your opponent opportunities to fuck up if you want to play optimally.

If a move loses 0.0005 points if the opponent plays optimally, katago won't play it even if there's ~zero chance a weaker player would play it right.

There have been go AIs that tried to train more directly on uneven opponents, one called "sai" comes to mind, but katago has huge advantages otherwise and won out over the others (for very good reason, it's a great project).

While AlphaGo originally only had win rate as a metric, modern Go AIs have more knobs, including an evaluation of "complexity".

Just stating this off the top of my head so I could be misremembering, but I heard that the KataGo settings used were tweaked to favor complexity. This was most apparent in Game 1 which Shin Jinseo lost, where the AI had an unusual opening. However, the last game was quite plain leading me to wonder whether that setting was present in the last game (or at all).

Ah this is interesting. Essentially the idea is that the compute can try and move into positions that it can evaluate but humans might have trouble evaluating because of the board state's complexity?

By the way that is exactly what humans do when playing with white in high stone handicap games. They place their stones all around the board, start little fights everywhere and wait for the weaker player to misread or misevaluate something. Suddenly two or three fights merge in a one sided larger one and part of the handicap is gone.

That's the basic idea, yeah.

When playing white in handicap games, you want to make your opponent uncomfortable.

Play moves where the simple/safe/obvious move is just a little bit bad. Force them to choose between complex fights or a slow death of 100 slightly suboptimal moves.

It feels really wrong to defend like 10 times in a row, so if you make them do that they'll lash out at the wrong time and you can take advantage.

You also want to look for moves where...even if their best response means it's even or a little bit worse for you, there's ~reasonable responses where you win out or it goes complex.

A lot of the time it's not even crazy complex fights, it's more just situations where the judgement of what is more points is difficult.

(Note: most of this stops applying as strongly if it's a teaching game, which most handicap games are, there you have other considerations besides winning)

There are tweaks to move choice like that, but it's all evaluation time. None of that happens in training. Zero times in the RL loop does katago see a two stone game against a weaker player.

You can kind of tweak towards play this metric or that, but it's not the same.

Maybe not to the extent you mentioned earlier (I definitely agree that it's not "focused" on this) but there is definitely asymmetric playout training in the training data (see comments on lightvector/KataGo issues #39 and #162 mentioning it for example), that is presumably how they got the tweak for playoutDoublingAdvantage (i.e. actually having a few thousand of the millions upon millions of training games be games where playouts have been doubled for advantage).

Yeah, thanks, I forgot that existed. That definitely weakens my point quite a bit.

It's still not _quite_ the same thing because an advantage in playouts is not a great model for how a stronger/weaker player dynamic actually works, but it is something for sure, way more than the nothing that I said exists.

> So, this was an important victory for a human, but not a sign that humans are now stronger than AIs at Go.

Another way to look at this: Go's handicap system gives us a genuinely interesting metric for the distance between a human and a machine at this specific game. Instead of just "computers beat humans" we get a quantified gap.

The question of whether machines or humans are stronger is moot, isn't it?

In any intellectual contest between human and machine, all the machine winning implies is that the endeavor is algorithmic.

The machine can be given practically unlimited memory and compute; we consider it cheating if the human would use memory aids. The machine could be implemented as many agents cooperating; we'd think it's not right if thousands of humans collaborated to face the machine, etc.

So statements like "not a sign that humans are now stronger than AIs at Go" are pretty meaningless, IMO

> not a sign that humans are now stronger than AIs at Go

Totally wrongheaded, actually, since the computer gave the human a 2 stone advantage from the start.

I'm sorry, but Elo is not an acronym for Electronic Light Orchestra. You don't write ELO, but simply Elo.

Reply on news.ycombinator.com