Computer Go

[John McKenna]

... Lee Sedol has been narrowing the gap in each game...

I somewhat exaggerated the smoothness of Lee Sedol's progress in the match with Alphago.

In game 1 he moved first but was behind from early on and resigned after 187 moves.

Game 2 was much closer and it took Alphago until move 211 to win.

In game 3 Lee held his own in the opening but was surprisingly outplayed in 176 moves -

https://gogameguru.com/alphago-shows-tr ... lee-sedol/

Game 4 saw Lee pull one back with a win in 180 moves.

[Christopher Kreuzer]

Thanks John. That is a good site for game analysis.

Game 4 (the human win) analysed here:

https://gogameguru.com/lee-sedol-defeat ... ck-game-4/

Seems that only really brilliant moves (the only ones that can win in a nearly lost position) can escape the notice of this sort of AI Monte Carlo-type gaming algorithm.

[MartinCarpenter]

That can happen to any sort of search, even human ones

Very interesting dicussion there thanks. I hadn't realised how horribly cruel the thing is. It doesn't try to optimise its final score, rather its probability of winning.

So if it gets a long way ahead AlphaGo can almost start messing about..... Must be a bit horrible to play against for a human.

[Roger de Coverly]

So if it gets a long way ahead AlphaGo can almost start messing about..... Must be a bit horrible to play against for a human.

Understanding how your opponent will play has long been a useful skill at top level chess, well pre-dating computer engines. I don't know to what extent databases of previous games are used by top Go players. The reports suggest that in game 4 at least, the human player had insights into his opponent's style and was able to counter it.

[John McKenna]

Yes, Lee Sedol is learning about Alphago as the match progresses.

The reverse - Alphago learning about Lee Sedol's style of play - seems not to be happening during the match. Though it was given access to his games prior to it -

Demis Hassabis has said that training and playing are two different processes, with training taking weeks of processing on a much larger computer, and play occurring via a smaller computer. For now AlphaGo hasn’t learned anything from the match.

Besides, it would be wrong to think that a couple of games with Lee Sedol would do much to alter its judgement. It has already chewed through the largest database of professional games that any computer program has ever ingested...

(European Champion Fan Hui, it's previous victim, didn't just play a 5 game match. He played extensively with Alphago and went from being in the top 600 Go players in the world to being in the top 300.)

[MJMcCready]

When is the next game? It's very interesting even to those of us with little understanding of the game.

[John McKenna]

As far as I can recall, no time to check now, the 5th and final game of the match is scheduled for 04:00 GMT on Tue!?

[John McKenna]

Yes, the final game of the match is as above and that means it should start in a little over 4 hrs.

See here for details and further relevant links -

https://gogameguru.com/alphago-lee-sedol-schedule/

[Clive Blackburn]

Live coverage of the fifth and final game here:-

https://www.youtube.com/watch?v=mzpW10DPHeQ

[Clive Blackburn]

Live coverage of the fifth and final game here:-

https://www.youtube.com/watch?v=mzpW10DPHeQ

AlphaGo won the final game, so 4-1 to the computer.

I am quite glad that humanity at least managed to salvage one point!

[MartinCarpenter]

Definitely The silicon still has nothing like the superiority that it has in chess.

Matter of time perhaps, although because so much of AlphaGo is locked into its neural net evaluation engines it really doesn't scale that well with added computing power. I guess having a lot more computer power would let them train (and use) bigger/more discriminating neural nets in the first place.

Lee Sodol's sum up in the final press conference seemed to be that he thinks humanity can match its pure Go skills, but the concentration/stability that the computer has is another matter. Not entirely unlike chess when they first got ahead of us of course, although by now they've blown way past us in terms of calculating power too.

Deepmind seemingly haven't decided what they're going to do with AlphaGo.

[NickFaulks]

The real test would have been for AlphaGo to speak for itself at the press conference.

[David Robertson]

The real test would have been for AlphaGo to speak for itself at the press conference.

That could be easily achieved. The real 'real test' is when it understands what it has done

[Christopher Kreuzer]

""After the match, South Korea's Go Association awarded AlphaGo the highest Go grandmaster rank – ninth dan. It was given in recognition of AlphaGo's "sincere efforts" to master Go's Taoist foundations and reach a level "close to the territory of divinity"""

http://www.straitstimes.com/asia/east-a ... go-ranking

FIDE haven't awarded any Grandmaster titles to computer programs, have they?

[Clive Blackburn]

FIDE haven't awarded any Grandmaster titles to computer programs, have they?

They may well have done, unknowingly