"saikyou no igo zero" (Crazy Stone)

[EdLee]

Hi Marcel,

日本語Windows® 7/8/8.1/10 (32/64bit)

I hope there will be an iOS version as well.

That'd be nice.

a win rate of 95% against the previous version.

Very roughly speaking, is it fair to say it's at least 1-stone better than the previous version ?
( Curious about its winrate giving 2 stones to the previous version. )

[paK0]

1-stone better than the previous version ?
( Curious about its winrate giving 2 stones to the previous version. )

Probably more.

At dan ranks 100 ELO is about one rank, and 95% win rate is around 350(?) so it should be three to four stones stronger.

All the values are rough, someone with more love for math than me might give you a more accurate answer.

[jlt]

95% winrate between similar programs A and B is not the same as 95% winrate between program A and a human with the same Elo rank as program B.

[Elom]

Denniss Hassabis said Alphago performs better against weaker versions of itself than it does against weaker humans. It's likely the same for all AI.

[Bill Spight]

Denniss Hassabis said Alphago performs better against weaker versions of itself than it does against weaker humans. It's likely the same for all AI.

The self play trap?

[hyperpape]

I imagine that this could be true of people, if we could freeze someone's play and have them play against themselves. A significant part of what you learn is identifying your own bad patterns. It just might be the case that AI is more predictable (and learns primarily through self-play) so it is even better at exploiting its past selves' weaknesses.

[Bill Spight]

I imagine that this could be true of people, if we could freeze someone's play and have them play against themselves. A significant part of what you learn is identifying your own bad patterns. It just might be the case that AI is more predictable (and learns primarily through self-play) so it is even better at exploiting its past selves' weaknesses.

Well, with self play, that's what it is trained to do. Perhaps diminishing returns kick in, in terms of points. That is, even with the same win rate difference, the wins get smaller and smaller. But against other players with different strengths and vulnerabilities, at some point those small refinements do not produce the same win rate differences.

[Bas]

Unfortunately there is no information if a GPU is supported, only CPU requirements are mentioned.

[Elom]

The self play trap?

But I don't think developers would use others engines to showcase the strength of their own.

I imagine that this could be true of people, if we could freeze someone's play and have them play against themselves. A significant part of what you learn is identifying your own bad patterns. It just might be the case that AI is more predictable (and learns primarily through self-play) so it is even better at exploiting its past selves' weaknesses.

There may be some human examples .

If a lone NN network is a hundred percent savant and zero percent seer, couldn't that both either pronounce or diminish the effect of doing well in self-play? I'm sure search balances things out, but not quite so much that pure bots can spot ladders .

[Phobos]

Who do I have to kill for a version for linux?

[Gomoto]

There is a version for android.

[pluspy1337]

Very cool, but for those with no money you can use leela master by pangafu for more human-like moves in a bot. Just use it with sabaki or lizzie as you would another weight like 40 block or elf. The most recent weights from 15 block G series (which is what you should get) has 70% human weights if I understand correctly, and I'd say it is at least on par with LZ 157 if not stronger.

https://github.com/pangafu/LeelaMasterWeight

[And]

Crazy Stone Zero ON SALE

https://www.amazon.co.jp/アンバランス-最強の囲碁-Zero/dp/B07J9Z3R6C

[johnsmith]

Has anyone bought it and tried it, yet? Can new Crazy Zero precisely count difference in points?

[Glummie]

Would the new version be released on steam or something?