World #1 Go Player Ke Jie accepts Google Alpha Go Match..

[Laskos]

The only loss of AlphaGo one year ago against Lee sedol, analyzed by Leela. It is the best engine in seeing that something happened around move 80:



Now AlphaGo is probably 500+ ELO points stronger than AlphaGo of a year ago, no single hint of a loss.

[JJJ]

I hope Alphago will be trained to play Starcraft , the first one.

[Leto]

I hope Alphago will be trained to play Starcraft , the first one.

It is currently training on StarCraft 2. I'd say mastering StarCraft 2 would be a much tougher challenge than a game like Go where you only play one move at a time. If it masters StarCraft 2 I hope it tries its hands on a competitive card game such as Gwent.

[chetday]

Kai, thanks for posting that link to Leela. I've been wanting to learn how to play Go for several years and Leela appears to be just the ticket to get me off on the right foot.

[lkaufman]

mankind's last hope is played out in just a few hours

go Ke Jie!

This thing is probably 3 handicap stones stones stronger than Ke Jie. Cannot wait for it or something even remotely similar like Zen to appear for home PC.[/quote

Michael Redmond (9 dan Pro) thought two stones would be fair. Two stone handicap should produce about a 20 point win between even players, so if the final score of these even games would have been about 20 points plus for AlphaGo if AlphaGo were score-maximizing rather than win% maximizing, two stones should be even. That sounds about right to me, but who knows?

2 stones was based on "Master" online games in January. AlphaGo meanwhile must have improved, the advantage versus Ke Jie might be anywhere at 3-5 stones, or up to 1000 ELO points. It's very hard for me to guess, and I think very few know. It seems they are in different category, like pro versus amateur.

I yesterday discovered that Gian-Carlo Pascutto, author of Deep Sjeng, has made a top Go program, free, called Leela. The previous iteration is already no.8 on KGS, and the new one is *very* strong.

https://www.sjeng.org/leela.html

It is professional level program, the strongest available for public, free or commercial. I re-analyzed the three Ke Jia games, all were lost right after the opening, starting with moves 40-60:







They seem to indicate a huge strength difference. The third game went nasty for Ke Jie, AlphaGo was not content winning marginally, it thrashed Ke Jie.

Wow, already a pro level Go program for pc, and remarkably it's free! People were saying maybe we'd have a pro level Go engine for pc in a year or two after this match, but it already happened! I wonder if any of the techniques used in these go programs would help with chess?

[Eelco de Groot]

Leela has been around for a while, but it was a much weaker program. I can't really judge the strength as I hardly can play the game, but to be fair, all Go programs were much weaker then. I just checked, I have a free Leela Lite 0.2.9 from 2007 on my old Athlon computer with Windows XP. It had very nice graphics, but as far as I can see it seems to play only on a 9x9 board. I think it was much weaker than what Gian Carlo has now. From what Kai is saying GCP put a huge effort in it!

[Werewolf]

I hope Alphago will be trained to play Starcraft , the first one.

It is currently training on StarCraft 2. I'd say mastering StarCraft 2 would be a much tougher challenge than a game like Go where you only play one move at a time. If it masters StarCraft 2 I hope it tries its hands on a competitive card game such as Gwent.

As someone who plays SCII I'm looking forward to this. Is there any hint of public availability or if it will challenge the best players?

[jhellis3]

As someone who plays SCII I'm looking forward to this. Is there any hint of public availability or if it will challenge the best players?

If it is allowed to control every unit individually, it will likely crush the best players. The only hope for humans at that point is to prevent scouting and execute an attack with units which largely counter whatever the AI is building.

Once the AI gets vision though.... it is over. The power of perfectly tracking all of your opponents spent & remaining resources, unit composition, unit availability, and unit costs, and perfect timings on unit/upgrade availability means AI can always build (counter) perfectly and will always micro perfectly. Thus the game will essentially end up being a vision war with vision denial being a necessary but not sufficient condition for a human victory.

Gwent is a bit different in that optimal play will be much easier to establish, but there could be situations where, even with optimal play, one deck just beats another. Of course, the human would still have to play exceptionally well, but the best players should still be able to win games from time to time...

[Vladimir Xern]

As someone who plays SCII I'm looking forward to this. Is there any hint of public availability or if it will challenge the best players?

If it is allowed to control every unit individually, it will likely crush the best players. The only hope for humans at that point is to prevent scouting and execute an attack with units which largely counter whatever the AI is building.

Once the AI gets vision though.... it is over. The power of perfectly tracking all of your opponents spent & remaining resources, unit composition, unit availability, and unit costs, and perfect timings on unit/upgrade availability means AI can always build (counter) perfectly and will always micro perfectly. Thus the game will essentially end up being a vision war with vision denial being a necessary but not sufficient condition for a human victory.

While that may be a theoretical concern, do note that an API has been available for StarCraft: Brood War that accomplishes much the same thing for around 7 years now, receiving some academic attention. Despite very early on achieving superhuman feats of micromanagement (many videos on YouTube boasting 40K actions per minute), the level of these AIs still hovers around that of a very mediocre amateur player. At some point they may find a better way to bring that power to bear, but it hasn't happened yet.

Moreover, Google's first press release said that the idea ultimately was to eschew these hooks directly into the game's memory in favor of reading and evaluating the screen output. If it were confined to that and inputs through the StarCraft UI like a human, it would be much more crippled.

[IanO]

I yesterday discovered that Gian-Carlo Pascutto, author of Deep Sjeng, has made a top Go program, free, called Leela. The previous iteration is already no.8 on KGS, and the new one is *very* strong.

https://www.sjeng.org/leela.html

It is professional level program, the strongest available for public, free or commercial.

Wow, already a pro level Go program for pc, and remarkably it's free! People were saying maybe we'd have a pro level Go engine for pc in a year or two after this match, but it already happened! I wonder if any of the techniques used in these go programs would help with chess?

Thanks for letting us know about this, I thought Leela development had stopped.

Lest we forget, the original version of Leela was the first Go program for sale to incorporate Monte Carlo search innovations over a decade ago. I'm happy to see that GCP is still interested in implementing the latest Go algorithms for us, especially for free!

(P.S. I went over those recent computer vs. pro shogi games and was mystified at the strange openings, and that the pro resigned so early. I guess I need to find an annotated version of those games to understand.)