search

ymgaq / AQ

A Go game engine. Download here: http://github.com/ymgaq/AQ/releases
GNU General Public License v3.0
324 stars 70 forks source link

Is Google's new paper helpful to improve the strength of AQ? #27

Closed gogoai closed 6 years ago

gogoai commented 7 years ago

The paper introduces AlphaGo Zero, the latest evolution of AlphaGo, the first computer program to defeat a world champion at the ancient Chinese game of Go. Zero is even more powerful and is arguably the strongest Go player in history. Previous versions of AlphaGo initially trained on thousands of human amateur and professional games to learn how to play Go. AlphaGo Zero skips this step and learns to play simply by playing games against itself, starting from completely random play. In doing so, it quickly surpassed human level of play and defeated the previously published champion-defeating version of AlphaGo by 100 games to 0.

ymgaq commented 7 years ago

In the paper, they describe the training amount for the final version as

Over the course of training, 29 million games of self-play were gener­ated.

The method is too heavy to handle with a normal GPU. From my experience, the network in the paper takes about 4 msec for each position with my machine (i.e. GTX 1080).

Therefore, even if the average move count of self-play games is 250, it takes

29,000,000[games] * 250[positions] * 1600[playouts] * 0.004[sec]
= 46,400,000,000[sec] = 1471[year]

If you wait 1500 years, I will develop it.

gogoai commented 7 years ago

I always thought the new paper had advanced methods, I see, thank you

aTan-aka-Xellos commented 7 years ago

@ymgaq is it realistik to build and use some distributed resources provided by go community? It achived AlphaGo Lee lvl in 3 days, so it kinda 150 years on usual GPU?) If involve 100+ people to contribute their PC time... or I just dreaming and it basically impossible to achive? :)

P.S. what is your opinion of influence of this paper to existing/coming soon bots strength? (Especially for avaliable for public, not like FineArt which has resources but not public)

Thank you.

higherdim commented 6 years ago

Leela Zero has done some excellent work using distributed computing resources. That should be a good way to overcome the heavy resource requirement problem.

ymgaq commented 6 years ago

I agree with you, and am looking forward to the success of the LeelaZero project.