Lee Sedol vs Alpha go

[mhlepore]

After game 3, my advice to Lee Sedol is the same as before game 1: instead of fighting locally, he must, from the beginning, go for global, long-term interaction including good positional judgement...

Hi Robert,

I read this post and the other one you wrote giving similar advice. It intrigues me, but let me throw out a straw man for you to knock down.

1) Aren't you falling prey to the same flawed thinking that the Go professionals did after watching AlphaGo beat Fan Hui? Clearly AlphaGo had the ability to play at a higher level than it needed when it beat Fan Hui. Just because we didn't see it doesn't mean it wasn't there. In fact, you seem to say as much in your own post when you discuss Alpha Go losing points for a better win probability.

2) Assuming you are correct - that a person with brilliant positional judgement (Yoda in his prime?) - could give AlphaGo the best run for its money. Wouldn't that human success be temporary? That is, why couldn't AlphaGo be trained to adapt to this strategy as well?

Am interested in your thoughts.

[Cassandra]

Human beings do not own a measuring divice for "winning probability". But AlphaGo (as all other programs) does.
This makes it so difficult to "understand" / "judge" her moves.

Human players need to accept some kind of "indirect" measurement, like "difference in potential territory".

However, even Michael Redmond seems to have learned in the meantime that apparently "slack" / "slow" moves might be a valid indicator that AlphaGo securely assumes that she will win. Even if this behaviour can be seen as early as in midgame.

If a special ability / skill / technique is not necessary to apply, in order to increase the winning probability, AlphaGo will not show these.

However, if she is forced to use such ability / skill / technique, she will do so successfully.
Otherwise, she would have avoided to become caught in such a "dilemma" much earlier.

It is very, very hard to play an "adaptive" opponent.

[pookpooi]

Many internet comments in China want Lee Sedol to ko fight with AlphaGo (some even accuse DeepMind of bribing Lee Sedol not to play ko), now we know that AlphaGo respond to the ko fight perfectly, now what? I don't want the upper bound of AlphaGo's strength to be found by playing a handicapped game with pro, that'd be humiliating for pro...
Facebook's DarkForest give analysis of two first game and AlphaGo seems to be leading even in the opening...

[CnP]

If AlphaGo beats Lee 5-0 I think making the pro take a handicap is exactly what they should do. Could it beat him if Lee took a 4 stone handi? That would be very interesting to see.

[Simba]

If AlphaGo beats Lee 5-0 I think making the pro take a handicap is exactly what they should do. Could it beat him if Lee took a 4 stone handi? That would be very interesting to see.

4 stones is a huge handicap. It's impossible to be sure, but I think it's unlikely even perfect play would beat Lee on 4 stones.

I'd like to see a game with 2 stones, though. I think that's doable.

[Cassandra]

Facebook's DarkForest give analysis of two first game and AlphaGo seems to be leading even in the opening...

I have only CS2012 (and the much older MFGo12) as "measuring tool" at my sleeves.

My observations:

According to the assessement of CS, the winning probability in games between top-professionals seems to be in the range of (approximately) 45 % to 55 % over a quite long period of the games.

There are a lot of ups and downs, but the graph usually does not leave this region.
Until one of both players found the dicisive blow to win the game.


In AlphaGo's games, these graphs are (most of the time) always above 50 % (for her), usually gradually increasing over time.
In the second game against Lee Sedol, the value was lower than 50 % (until about move 40), but never below 49 %.
So her opponent was unable to get a decisive advantage in the opening.


Probably (just my assumption) AlphaGo reaches her "zone of comfort" at a time, when CS assesses her winning probability about 65 %.

Thereafter, AlphaGo's opponent will have to address a really "blind spot" of the program (if there were any), should he want the change the direction of that graph.


DeepMind will have better values, derived from AlphaGo's own calculations.

[RobertJasiek]

Cassandra, I have used probabilities since about 2 kyu whenever strategic ambiguity remains after application of all other theory.

mhlepore, from Alphago's strengths (reading, endgame) and mistakes against Fan Hui, I concluded my advice for Lee Sedol. The strengths already were convincing, so one must benefit from the earlier mistakes and, after improvement, relative weaknesses.

Like most, I have been greatly surprised by the degree of improvement. Therefore, of course I cannot be sure and it might be that also I am lost in Alphago's current planning level.

Yes, if some human players can still beat Alphago in its current version, this is temporary. Alphago's further improvement, if proceeding similarly, will soon make any human win impossible. With the possible exceptions I mentioned in the computer go mailing list, mirror go etc. Those still include complex kos, for which we cannot know whether Alphago handles them as well as the simple kos in game 3.

[Cassandra]

With the possible exceptions I mentioned in the computer go mailing list, mirror go etc. Those still include complex kos, for which we cannot know whether Alphago handles them as well as the simple kos in game 3.

Please do not forget that there is no programm yet that was able to understand Igo Hatsuyôron 120.

I would be very curious, whether AlphaGo assesses our amateurish solution to be correct.
Or finds an even better one.

[Bonobo]

I read on the web that The I_AM_A_SPAMMER_PLS_IGNORE GO System is going to challenge AlphaGo next ?

{URL Removed by admin}

LOLOLOL, this “I_AM_A_SPAMMER_PLS_IGNORE.” reminds me of “djhbrown” who used to spam somewhat similar stuff … highly arcane … mysterious … esoteric.

[Galation]

I really wish we got to see more live pro. games with English commentary. Watching these games has been so much fun, I wish it happened all the time! In fact, I think it would be even more thrilling if both players were emotional, mortal humans to whom we could relate in some way.

Me too!
Galation

[trout]

4th game;

Lee Sedol defeated AlphaGo by resign.

AlphaGo vs Lee Sedol 3:1

[Cassandra]

For game #4:

Here is what two electronical "seniors" thought about the winning probability of their youthful buddy.

Please be careful while interpreting the diagrams.
I suppose that "EQUAL" is ABOVE 50 %, maybe something about 55 %, at least.

DeepMind will have to work hard to stop such (temporarily) turning points in the gradiation of the graph as at approx. move 80 (apparently happening outside AlphaGo's "comfort zone") from becoming points-of-no-return.

CS2012.jpg (156.46 KiB) Viewed 10274 times

MFGo12.jpg (95.81 KiB) Viewed 10274 times

[Charlie]

Where can I get the SGF files for game 4?

(I'm trying Go4Go.net but they seem to require registration to download SGF files and I think their server is completely overwhelmed.)

Ok. I succeeded on Go4Go.net. The server responded.

[Uberdude]

After game 3, my advice to Lee Sedol is the same as before game 1: instead of fighting locally, he must, from the beginning, go for global, long-term interaction including good positional judgement...

Contrary to this advice, it was by putting pressure on AlphaGo with complicated tactical reading that Lee managed to elicit a mistake and win the 4th game.

[Charles Matthews]

After game 3, my advice to Lee Sedol is the same as before game 1: instead of fighting locally, he must, from the beginning, go for global, long-term interaction including good positional judgement...

Contrary to this advice, it was by putting pressure on AlphaGo with complicated tactical reading that Lee managed to elicit a mistake and win the 4th game.

Indeed, Lee is to be congratulated on digging himself out of the snowdrifts of a blizzard of well-intentioned advice.