Life In 19x19

I think it's mostly clear that AlphaGo seems to play more conservatively when it thinks it's significantly ahead and perhaps aggressively when behind.

The paper says that it was trained on KGS games, but it used a weighting dependent on the strength of the players. I suggested that they ought to use GoGoD games too, but that was subsequent to the match.

I don't see any reason for Lee to vary from what he would do against another middle-ranking pro. Game 1 will probably be, for him, a question of trying to establish how strong AlphaGo is. The alpha-male strategy of trying to win convincingly at the start of the match is, amusingly, something AlphaGo is immune to.

mumps wrote:I think it's mostly clear that AlphaGo seems to play more conservatively when it thinks it's significantly ahead and perhaps aggressively when behind.

The most interesting thing, apart from the scoreline, will be to see AlphaGo put through its paces. I like the way it plays: generally lucid and with a stable feel even in heavy fighting.

I'm not yet quite ready to articulate where the style comes from. I suspect it is "thematic", a step up from "shapeful". In other words it may have learned that certain types of things "in the offing" should affect the way you play.

David Ward went round Korean clubs: and he reported to me that the local strong amateurs whom he played adopted a steady style, looking to punish errors, rather than be provocative/disruptive. This is also pro common sense for the "stranger in town" opponent, I suppose.

So, all told, there may not be the playing to the gallery, at least at the start.

One question I haven't yet found an answer to is when Alpha started training against itself, i.e. How much time did it take to achieve it's current level (I haven't read the paper yet).

mumps wrote:I think it's mostly clear that AlphaGo seems to play more conservatively when it thinks it's significantly ahead and perhaps aggressively when behind.

That observation is less useful than it might appear. Consider the following substitution

I think it's mostly clear that "a strong go player" seems to play more conservatively when "he/she" thinks "he/she is" significantly ahead and perhaps aggressively when behind. We could even replace "go" with "any game where the result is either win or lose and margin considered irrelevant".

In other words, simply the correct way to play. A consequence of AlphaGo being a strong player.

Uberdude wrote: $1m is pocket change.

This is admittedly unrelated, but it reminds me of the story of the NHL glow-puck back in the 90s.

Engineer: “I tracked things a lot harder than a hockey puck for the military...but you couldn’t afford it.”

Media exec: “Just how much would it cost?”

Engineer: “It would take two years to develop and cost about $2 million,” is the reply.

Media Exec: “You don’t understand the economics of sports. Write a memo.”

[A few days later]

Other Media Exec: “David says you can track and highlight a hockey puck...and it’d only cost $2 million...That is now your highest priority. If anybody asks you about the money, tell them to call me.”

Uberdude wrote: $1m is pocket change.

I also noticed that Google recently overtook Apple as the most valuable company in the world by market capitalization. I can't help but wonder if part of that rise in share price* was thanks to the DeepMind announcement, which I'm sure went down well with investors. Spending $1m for $billions of stock growth is a great return.

* though most of the overtaking of Apple was due to Apple going down.

Match will be on "9, 10, 12, 13, 15 March - will be livestreamed on YouTube." from https://twitter.com/demishassabis/statu ... 7870008321

emeraldemon wrote:Goratings only has 11 games from Fan Hui (2 wins 9 losses) and the last game was December 2014. The rating algorithm probably has considerable uncertainty about his true rating. Also the 5-0 isn't a particularly large sample size either. It's still possible the odds for AlphaGo could be <1%. And of course we could also argue the program will continue to improve and the odds are much higher than 5%.

It now has 5 wins and 9 losses, and rates him at 3013. If the four wins from the European professional championship get added, who knows how high he could go ?

What if the match ends in a tie? Will there be a tie-break game?

Reportedly, Chinese rules will be used, which should mean that no long cycles other than sending-two-returning-one are prohibited.

The match involving a computer, it should follow the superko rule. I guess that all repetitions are prohibited.

The gap between the bottom professionals and the top ones is very wide. Cho Chi Kun (born in 1956) became one of Kitani's pupils in
1962 when he moved to Japan. He reached professional 1 dan in 1968 and 9 dan in 1981. He won his first title in 1976, The Oza,
but he didn't start to take the big titles until 1980 (Meijin), 1981 (Honinbo) and 1983 (Kisei).

So it took him 6 years to get from strong amateur to pro 1 dan (1962 to 1968) and about 12 more years to start winning the big titles
(1968 to 1980). AlphaGo has done the "easy" bit - getting to a low ranked professional, the hard work is getting from there to where
Lee Sedol is - winning the big tournaments. It took Cho twice as long to do that as it took him to reach 1 dan pro.

When I was in Japan in 1970 I went on the Nihon Kiin staff outing to Expo 70 in Osaka, on the coach I sat next to Murakami Bunsho. He
had already been Amateur Honinbo and had beaten several professionals at both two stones and in even games. He had beaten Sakata in
an even game in 1969. He said that he would need two stones to be confident of beating Sakata, when he played him he was aware of how big
the gap was between him and the top professionals. I would estimate Murakami's playing strength as that of a low ranked pro, similar
to Fan Hui, probably stronger. <There are some 82 of his games on GoGod>

So fwiw I think the gap from Fan Hui and AlphaGo to Lee Sedol is really big. I think AlphaGo might win one game, no stress, no fatigue and
the ability to read out the really difficult fights. Lee Sedol will win the other four games - he can push the game into a really difficult
ko fights and he has the experience of the big occasions.

So 4-1 to Lee Sedol. I decided to join Life in 19x19 today - so this is my first post!

John

I really hope AlphaGo plays its first move on Tengen. Bonus points for following up with opposing 4-4 and 3-3, but that is really too much to hope for.

John Tilley wrote:The gap between the bottom professionals and the top ones is very wide.

Couldn't you have said the same about Kasparov and ordinary GM's back in the day of Deep Blue?

John Tilley wrote:So it took him 6 years to get from strong amateur to pro 1 dan (1962 to 1968) and about 12 more years to start winning the big titles(1968 to 1980). AlphaGo has done the "easy" bit - getting to a low ranked professional, the hard work is getting from there to where Lee Sedol is

Interesting POV, but IMO maybe a bit too human-centric? You can't compare the learning curve between humans and machines - humans slow down drastically once they reach around 7d because that is near the limit of their brain capacity, but machine capacity limits are very different. It's like comparing the graphs of f(x)=X^0.5 and f(x)=100X^0.5.

OTOH, though, one thing that might make your argument valid is the fact that AlphaGo is learning through human-produced games. So the ceiling will be set similar to that of humans, at least learning material-wise. But imitating human-produced games is not the only method it uses for study, in fact the very reason AlphaGo is so strong is because it applies many different learning networks(i.e. learning how to predict moves, learning how to evaluate position, learning how to pick the best move etc.) at the same time.

From what I understand AlphaGo's knowledge of pro games are mainly used for reading ahead(predicting moves) and selecting candidate moves. Since those two compose a bulk of one's playing skill, I think it's likely that AlphaGo's skill will be stuck at 'human level' for quite some time, until its experience piles up enough to see the flaws in those man-made textbooks. So maybe your guess will be right after all, at least for a while.

John Tilley wrote:I decided to join Life in 19x19 today - so this is my first post!

Nice to see a fellow newcomer who was brought in by AlphaGo. It sure generates interest, doesn't it?

CutFirstThinkLater wrote:From what I understand AlphaGo's knowledge of pro games are mainly used for reading ahead(predicting moves) and selecting candidate moves. Since those two compose a bulk of one's playing skill, I think it's likely that AlphaGo's skill will be stuck at 'human level' for quite some time, until its experience piles up enough to see the flaws in those man-made textbooks. So maybe your guess will be right after all, at least for a while.

Don't you think that this effect ("seeing" some flaws) might have happened once already, with the textbooks written by strong amateur players (KGS 6d - 9d) ?