search

lightvector / KataGo

GTP engine and self-play learning in Go
https://katagotraining.org/
Other
3.47k stars 563 forks source link

How strong now? #162

Open bochen2027 opened 4 years ago

bochen2027 commented 4 years ago

How many stones can Katago latest net give a top pro? Assuming state of the art hardware used and normal timing

iopq commented 4 years ago

Maybe 3 is possible now. Yeonwoo loses at three stones, maybe with stronger hardware you can beat stronger opponents

https://www.youtube.com/watch?v=kEaFO2o-HfA

Friday9i commented 4 years ago

I think that with strong hardware (eg 2 RTX2080Ti and ponder), 3 stones and 0.5 komi would be quite fair against top Pros with medium/long time settings... And I even wonder if 4 stones and 0.5 komi would be possible with very strong hardware (eg 6 or 8 V100). Any news from Haylee (which played LZ 2 years ago, already!)? Or possibly Michael Redmond, that would be amazing? Or Yeonwoo (https://www.youtube.com/user/goingceo) could be challenged with a strong hardware?

lightvector commented 4 years ago

Keep in mind that at some point, more and more hardware will stop helping, as will more training. There's not really a guarantee that becoming stronger in even games makes you stronger in handicap games. At some point it should be expected to make you weaker in handicap games instead (e.g. you don't do a major moyo invasion that is the only real chance to win because you are 100% sure they will kill it, even if in the actual game the other player given their level would be only 10% to kill it). I would expect such a reversal to be true eventually to some degree even despite the asymmetric playout training.

lightvector commented 4 years ago

(in which case you need to actually invest into more and better ways of doing opponent modeling and specifically tuning for handicap game play than simply the asymmetric playout training).

Friday9i commented 4 years ago

I was not really convinced by your PDA approach in the beginning, but now it seems I'm more convinced than you ; -), you did an amazing job! Not sure regarding handi and improvement: as long as training still includes handi/PDA games, in the case with an invasion you describe above, the net will now that it is difficult and it has an opportunity against a weaker opponent, so it should try. What I find amazing is that training is only limited to 8x less visits and around 3 stones IIRC, and it allows the net to play extremely well with H6, and even reasonably well with H9! Anyway, we'll know more in the coming weeks / months with next nets, best wishes for the current run!