Go game AI "alphago"

[Estie]

They made a go program of professional strength, alphago, to the surprise and excitement of all the go world; today starts a 5 match challenge between the program and one of the worlds top players.

https://gaming.youtube.com/watch?v=vFr3K2DORc8

This is the live stream, but professional players doubt it will be a repeat of deep blue/Kasparov. Not yet, at least.


Edit: the first match ended: Alphago 1 - 0 Lee Sedol
Edit2: after 2nd match: Alphago 2 - 0 Lee Sedol
Edit3: after 3rd match: Alphago 3 - 0 Lee Sedol. The remaining 2 games will be played, but with this, Alphago has won the duel.

[Thraalbee]

Ouch. My bet was on Lee

[fph]

We will have true AI when...

• computers pass the Turing tests
• computers beat humans at chess
• my cellphone can understand human voice commands and answer my questions
• computer beat humans at Go
• ???
• Skynet conquers the world.

[Derakon]

We will have true AI when...

...we don't have to code a specialist system for every task we want a computer to perform. The way I see it, a "true AI" is really a learning AI, that can adapt to any task it needs to be able to perform. And people are working on that, but they're far behind the people making these expert systems that are very good at one thing.

[bio_hazard]

I've been reading through Nate Silver's book "The signal and the noise", and just finished the chapter on the Deep Blue vs Kasparov matches. I only knew the basic end result, so it was an interesting read.

I had kind of assumed that Deep Blue was programmed more or less from first principles. Actually a big part of it's programming was just checking a huge database of previous chess matches.

Also, Kasparov won the first game, but the book argues that the way he won it actually shook his confidence in such a way that he may have lost his mojo for the remaining matches. Deep Blue did a weird move before it conceded the first match, and Kasparov took it to suggest that the computer was able to see so far into future moves of the game to know that it's position was hopeless. Turns out it was a software glitch, and Deep Blue had fallen back on a "move randomly" strategy!

[JamesGoblin]

I've been reading through Nate Silver's book "The signal and the noise", and just finished the chapter on the Deep Blue vs Kasparov matches. I only knew the basic end result, so it was an interesting read.

I had kind of assumed that Deep Blue was programmed more or less from first principles. Actually a big part of it's programming was just checking a huge database of previous chess matches.

Also, Kasparov won the first game, but the book argues that the way he won it actually shook his confidence in such a way that he may have lost his mojo for the remaining matches. Deep Blue did a weird move before it conceded the first match, and Kasparov took it to suggest that the computer was able to see so far into future moves of the game to know that it's position was hopeless. Turns out it was a software glitch, and Deep Blue had fallen back on a "move randomly" strategy!

What actually happened is - Kasparov won the previous match with dry score (same as any he ever played against machines before that) and was mostly pushing for win even in drawn games of this one. Before the last game, match was tied (which Kasparov certainly didn't want to allow) so he played very aggressively for victory - with black! - and self-destructed.

Needless to say, journalists and IBM folks hurried to project the image of the world champion (which Kasparov wasn't back in the day, BTW - but that's different topic) being squashed like a bug by superior machine.

I believe that it was only 10 years later, or even more, that computers started playing better than best human - anyway, today's world champion won't let us test. He considers matches against computers "pointless". For example, six years later (2003) Kasparov tied a match against much stronger machine (Deep Junior).

PS the fact that he, AGAIN, wasn't world champion - is, again, different topic.

[Severian]

What actually happened is - Kasparov won the previous match with dry score (same as any he ever played against machines before that) and was mostly pushing for win even in drawn games of this one. Before the last game, match was tied (which Kasparov certainly didn't want to allow) so he played very aggressively for victory - with black! - and self-destructed.

Needless to say, journalists and IBM folks hurried to project the image of the world champion (which Kasparov wasn't back in the day, BTW - but that's different topic) being squashed like a bug by superior machine.

I believe that it was only 10 years later, or even more, that computers started playing better than best human - anyway, today's world champion won't let us test. He considers matches against computers "pointless". For example, six years later (2003) Kasparov tied a match against much stronger machine (Deep Junior).

PS the fact that he, AGAIN, wasn't world champion - is, again, different topic.

I thought Kasparov was still overall quite a bit stronger than Deep Blue when he lost the match. I'm not a very good Chess player (peaked at a bit over 2000 ELO (USCF) when I was 16, and stopped playing soon after) but.. among other things I later knew one of Deep Blue's seconds. My impression is that they were able to prepare for him and that he got caught by surprise. That match might have been won more by Deep Blue's "seconds" than by Deep Blue.

That said, Chess is pretty amenable to brute force, and computers have gotten a lot better than Deep Blue since just by getting faster than it was, with some tweaks to the evaluation functions. I'm not really surprised that computers are better than people at Chess now (of course people teamed with computers are strongest of all.)

Go is another animal, I think. The branching factor in Go is much higher than in Chess, and even beginners can be taught to read certain forced plays (like ladders) out to twenty moves, more or less reliably, where computer programs can't. Now you might think that you could just tell a program to read ladders out, but... what about intervening moves and double threats arising from them? Go is just not amenable to brute force in the same way Chess is.

I am worse at Go than I am at Chess- maybe 10 kyu KGS the last time I played. But when I started playing Go I could beat all the Go programs fairly easily- they were bad, when Chess programs were Grandmaster strength.

A few years ago there was a breakthrough in Computer Go that made Computer Go programs much stronger in one jump. I was actually familiar with the technique from my work, but I wouldn't have thought to apply it to Go, and I was (to put it mildly) surprised that it worked.

Someone thought to apply Monte Carlo techniques to Go and... I don't know how to put this simply and accurately, so I'll just put it simply and let you refine the accuracy on your own if you wish, but: basically they had the program play out a lot (like a _lot_) of almost random games from each prospective move and looked at which prospective move won the most.

That, with some special-casing, got computer programs to a level that beat me every time, soundly, even giving me a significant handicap, and could beat a pro given a four stone handicap (a four stone handicap is a huge advantage, but I would lose to the pro or even a decent amateur every time, badly, given that advantage.) It was a really surprising jump in strength, but I figured that was about that for Computer Go, for a while.

AlphaGo is yet another animal. It turns out that my landlord (also the friend who got me back into Angband, so I imagine he might eventually post here) is a very strong Go player- one of the best Caucasian North American amateurs (some weird modifiers there, but that's Go for you.) He thought Lee would crush AlphaGo. I did too. It turned out otherwise and....

I can, in retrospect, understand why Monte Carlo techniques worked for Go. I wouldn't have thought of that by myself, probably, but I understand it. I don't understand what AlphaGo is doing. All I know is that it's disconcertingly good at Go.

[Scraper]

I also play go and watched this series online. A couple of thoughts from memory.

1) All 5 games ended in resignation which is think is a sad loss for the go community. Most lower ranked players aren't good enough to really count a game properly until quite close to the end, so it's a shame the games weren't completed to the counting stage. I think the greater go community would have had more closure with an actual count.
Also, with resignations, people who know nothing about go are basically no closer to understanding a complete game which again is a shame since this was a huge stage for bringing go to a wider audience.(Despite being the 2nd most played game in the world).

2) In the last game, Alphago continued to play with a dead group despite even amateurs being able to read many moves ahead to know that this was a dead group.

3) According to the wiki page, the game that alphago lost was due to an error in how the program viewed move 78 "the hand of god" move. This means that the current version might beat Lee 5-0! eek!

[Estie]

I am not so sure about 1; the games were played according to chinese rules and the process of counting under that system might just add to the confusion of the uninitiated.

I belive alphago´s play in game 5 has been examined and understood; unlike in game 4, the ai didnt go crazy, but instead played correctly in a way that is very hard to understand even for professionals.

Here is an analysis where Fan Hui sat down with two chinese 9dan pros and, provided with information about the ai´s internal processes, went through the 5 games in addition to some alphago vs alphago games.

https://deepmind.com/research/alphago/alphago-games-english/

If nothing else, the comments make for a good reading by re-capturing the athmosphere of those historical days in Seoul.

[Scraper]

I am not so sure about 1; the games were played according to chinese rules and the process of counting under that system might just add to the confusion of the uninitiated.

Yeah, the Chinese counting would have been a bit confusing for those used to the Japanese way of counting, but I still think it would have given the spectators a more "complete" view of the game of go.

Thanks so much for posting that link. I knew there was analysis done after the games but hadn't managed to track it down.