181–187 of 187 posts
Great, now next match against blue spot[0], starting even and adjusting the handicap each game. Blue spot is an adversarial ai and it's managed to beat average professionals on five handicaps, which is absolutely insane. [0]: https://codenamebluespot.com/
What does "on five handicaps" mean? The human received a 5-stone handicap?
Last time I read about these handicapped games in Go I saw that the model isn't trained to play from behind, so while in a normal game of Go, those 50 move Joseki's or any large move combo that's an 'even trade' will inevitably let the AI claw a lead at some point, the AI is already behind. It's similar in chess but there's a model specifically trained to play a knight down, and it's pretty cool to see the insane tac…
You can choose which piece it's missing. It's humbling to lose starting up a queen
Earlier quoted context omitted.
Yes, I can understand why absolute ratings are impossible to extract. However, since the population maintains some continuity over time (players gradually enter and then leave over time), would it not be possible to reconstruct relative ratings between players that didn't play during the same era?
Only if you were to assume that a player's prowess remains constant throughout their career, which we generally know to be false. (I'm completely inventing dates here) If Fisher played Kasparov in 1990 and Kasparov played Carlsen in 2020, you can only compare Carlsen to Fisher if you assume Kasparov's skill was about the same for this entire duration, which no one believes to be the case.
Chess requires you to be really sharp. Experience increases with age, but there's some cross-over point at which the decline in calculating speed is more important than increasing experience. Heck, I'm measurably better at chess tactics in the morning after a good night's sleep than in the evening after work.
I love this. I watched every game years ago between Lee Sedol, even though it was late at night. I love AI and I love Go. Off topic, but I wrote the first commercial Go program for the Apple II in the late 1970s.
Earlier quoted context omitted.
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 t…
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?
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)
Earlier quoted context omitted.
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…
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.
Earlier quoted context omitted.
This is a bit of a digression, but: It's an interesting question what (if anything) that larger dynamic range means . One thing you'll hear people say sometimes -- I've said it myself -- is that this shows that in some sense go is a "deeper" game than chess; there's more to know and understand, more variety of possible human skill. That might well be true. It certainly feels a more elegant game, and involves longer t…
Maybe the depth of a game is related to whether it will scale. Go is played on different board sizes and still works. If you make a Backgammon board bigger it would just be a slog and no real increase in tactical challenge. Chess rules dictate a set size of board, which I guess has been refined over time. If a Go board is made bigger or smaller it just adjusts the problem space, the rules and core of the game remain…