Go has fallen to computer domination?

[bob]

Three of the team members have now said they are 'quietly confident' for the Lee Sedol match, which makes me think they must have something now they didn't use for the previous match. I don't think they would say that if the ELOs are to come from bigger hardware only. Re-training the networks requires more games, takes a valuable time which could be used for other things, and even then may not bring much benefit since the 57% prediction rate they have now could be close to optimal.

In those monte carlo random type approaches, going from say 1000 cores to say 1 million cores doesn't bring that much playing strength wise.

Elo isn't getting used in Go, last time i checked.

Realize very few people play go at a stronger level, yet don't underestimate the strength of those 9-dan players.

Question is: is it another kasparov scandal, as deep blue 2 typically was rated about todays 2300 elo practical. Some FM today who doesn't blunder too much totally would bust that 1997 machine and program, even if you'd equip it with a modern book and its book was pretty bad for those years. See how much effort kasparov did do to lose that 2nd match versus deep blue.

All kind of bizarre moves i would never play.

And you can lookup my rating.

So the question is what is the intention of alphabet with alphago?
Is it a real honest match or another kasparov scandal?

My suggestion is to completely ignore the match against the 2 dan guy. That's elo 1900 or something. A guy who makes lots of small mistakes and no one ever confronted that guy with those mistakes (they don't have very strong programs right now). Even if that would be closer to FM level than 1900 level, his only job was to take care he lost all games and quickly hand over the job to a 9-dan guy as it's google man. Alphabet is huge company, if not largest in the world practical right now, they already look better by just playing a 9-dan of course, so that's what you do.


BY THE WAY - i'm very amazed they called that 2 dan player "european champion". Some years ago i played here a 4-dan guy (he of course toasted me in a manner world hadn't seen before you could toast someone that bad). And he spoke Dutch... ...yet odds are he sits in Japan already for a few years now...

Was deep blue lucky to beat Kasparov? Most likely. But it would NOT be crushed by a 2300 player today. That thing was extremely strong, it played far too many games to build that 2600+ Elo performance rating it held to win the Fredkin prize stage II (and III). Are programs today better? Of course. 20 years later. At the time they were untouchable vs computers and almost all humans. If they had continued, they would be untouchable today, still.

[jdart]

At least in chess, just not making any large mistakes and exploiting any of the opponent's large mistakes is worth a lot of rating points. GMs blunder all the time. They mis-remember openings. They overlook shallow combinations. They exchange into losing endgames. They fail to find key moves that are in tablebases. True, they do it less often than weaker players but it is hard for any human to keep concentration and memory working well enough over a game to not make any serious mistake.

Then there's been a huge increase in search depth due to better hardware and algorithms over the past 20 years. So now computers are just tactical monsters and they find things no GM is going to find OTB.

Go is much more difficult. In Go the large branching factor and the difficulty in building a good evaluation are serious obstacles, but recent work is showing considerable progress in tackling those issues.

[diep]

At least in chess, just not making any large mistakes and exploiting any of the opponent's large mistakes is worth a lot of rating points. GMs blunder all the time. They mis-remember openings. They overlook shallow combinations. They exchange into losing endgames. They fail to find key moves that are in tablebases. True, they do it less often than weaker players but it is hard for any human to keep concentration and memory working well enough over a game to not make any serious mistake.

Then there's been a huge increase in search depth due to better hardware and algorithms over the past 20 years. So now computers are just tactical monsters and they find things no GM is going to find OTB.

Go is much more difficult. In Go the large branching factor and the difficulty in building a good evaluation are serious obstacles, but recent work is showing considerable progress in tackling those issues.

Kasparov liked the money too much. Greedier than Kasparov in those days they haven't been born.

Look at the last tournament played. Wijk aan zee. What was the prize money?

10k euro first place. number 14 spot something like 1000 euro.

In reality of course Carlsen doesn't get out of his bed until you pay well over 100k euro 'start fee'.

If you'd pay tennis or soccer or basketball players in the same way chessplayers get paid - namely lumpsum - then not a single soccer player would ever be sweatted.

Look at what i wrote about that Kramnik match i negotiated for to play.

back end 90s, lots of GMs were total jealous on how Kasparov could sell games and still be number 1 in the world.

I remember how i did make a beginners mistake in a simultaneous exhibition not too long ago. It was dozens of kids and just 30 boards. I had to run fast to give all kids a chance to play at least 1 game. there were so many more than 30 kids who wanted a shot at it.

I ran so fast. After 2 hours of running board after board (it's 5 seconds a board roughly for longer periods of time) suddenly the head master came to me: "we have to soon stop as next event is going to take place soon".

I quickly finished. I suddenly realized very ashamed: "i didn't give a single kid a draw".

There was elo 1900 kids also joining in that simultaneously exhibition. Just so you know. I'm just in the 23xx.

In the 90s and start 21th century, playing a computerchess exhibition for all those GMs was nothing more than a simultaneous exhibition. Couldn't lose rating with it - just give a good show.

Somewhere a year or 10 ago, with some seconds from Topalov i checked Diep's world championship book. They toasted it like it didn't exist that book. The GMs with the black colors.

Main variations they had analyzed further till move 30. whereas Diep was out of book in such lines at move 14 to 16 or similar.

Only approaching move 30 suddenly scores would drop to -2.x or lower. Really 2 pawns down in evaluation.

That's why the world top is to scary always to play a mainline. Go is a different game. like from those 361 moves you can make first move, already 350 you can throw away.
You can hard prune 300.

That's not possible in chess.

[diep]

Three of the team members have now said they are 'quietly confident' for the Lee Sedol match, which makes me think they must have something now they didn't use for the previous match. I don't think they would say that if the ELOs are to come from bigger hardware only. Re-training the networks requires more games, takes a valuable time which could be used for other things, and even then may not bring much benefit since the 57% prediction rate they have now could be close to optimal.

In those monte carlo random type approaches, going from say 1000 cores to say 1 million cores doesn't bring that much playing strength wise.

Elo isn't getting used in Go, last time i checked.

Realize very few people play go at a stronger level, yet don't underestimate the strength of those 9-dan players.

Question is: is it another kasparov scandal, as deep blue 2 typically was rated about todays 2300 elo practical. Some FM today who doesn't blunder too much totally would bust that 1997 machine and program, even if you'd equip it with a modern book and its book was pretty bad for those years. See how much effort kasparov did do to lose that 2nd match versus deep blue.

All kind of bizarre moves i would never play.

And you can lookup my rating.

So the question is what is the intention of alphabet with alphago?
Is it a real honest match or another kasparov scandal?

My suggestion is to completely ignore the match against the 2 dan guy. That's elo 1900 or something. A guy who makes lots of small mistakes and no one ever confronted that guy with those mistakes (they don't have very strong programs right now). Even if that would be closer to FM level than 1900 level, his only job was to take care he lost all games and quickly hand over the job to a 9-dan guy as it's google man. Alphabet is huge company, if not largest in the world practical right now, they already look better by just playing a 9-dan of course, so that's what you do.


BY THE WAY - i'm very amazed they called that 2 dan player "european champion". Some years ago i played here a 4-dan guy (he of course toasted me in a manner world hadn't seen before you could toast someone that bad). And he spoke Dutch... ...yet odds are he sits in Japan already for a few years now...

Was deep blue lucky to beat Kasparov? Most likely. But it would NOT be crushed by a 2300 player today. That thing was extremely strong, it played far too many games to build that 2600+ Elo performance rating it held to win the Fredkin prize stage II (and III). Are programs today better? Of course. 20 years later. At the time they were untouchable vs computers and almost all humans. If they had continued, they would be untouchable today, still.

the games were not such high level, maybe 2200 at most Bob. If Kasparov was winning bigtime, he'd give away a pawn just to take care deep blue got back in the game. If he was having a crushing winningly strategic position he'd go moves like Ng5-h3 with white just to avoid playing Be3 giving white a blindfolded winning position.

And game 6 is a line that's also in a book i have. Yes, co authored by Kasparov...

Bob, Deep Blue wasn't lucky to beat Kasparov - Kasparov loved the money too much - and contractual he had to lose anyway (and i have that from an insider who was helping out Kasparov at that time).

Deep Blue by 1997 was so total outdated that they had to end it simply. It got 10-12 plies as you can see from the logs. Back in 1997 that was a good depth. 2 years later world champs 1999 that was not so great. IBM was clever enough to realize THAT.

Someone not in the know of all this also could have smelt it.

Directly after the match Kasparov praises Deep Blue bla bla bla.
then 1 hour later. Kasparov learned he wouldn't get a rematch. He's completely upset and starts spitting out nonsense. Fire comes out of his mouth. He slowly starts to realize what happened.

As a world champion he ALWAYS got another match. He naively hadn't thought about that until 1 hour after the match.

So the question i have is whether this will happen with Alphago as well.

A professional player lives from MONEY. And giants like IBM and Google have that in abundance.

[diep]

John, Seriously, if go would've been a world wide game like chess is and chess just would be popular in japan and korea - then in 1997 kasparov would have played go and lost from IBM deep blue go and we would be busy with computerchess right now.

It's about how much scientific effort you put in a game.

[Uri Blass]

Three of the team members have now said they are 'quietly confident' for the Lee Sedol match, which makes me think they must have something now they didn't use for the previous match. I don't think they would say that if the ELOs are to come from bigger hardware only. Re-training the networks requires more games, takes a valuable time which could be used for other things, and even then may not bring much benefit since the 57% prediction rate they have now could be close to optimal.

In those monte carlo random type approaches, going from say 1000 cores to say 1 million cores doesn't bring that much playing strength wise.

Elo isn't getting used in Go, last time i checked.

Realize very few people play go at a stronger level, yet don't underestimate the strength of those 9-dan players.

Question is: is it another kasparov scandal, as deep blue 2 typically was rated about todays 2300 elo practical. Some FM today who doesn't blunder too much totally would bust that 1997 machine and program, even if you'd equip it with a modern book and its book was pretty bad for those years. See how much effort kasparov did do to lose that 2nd match versus deep blue.

All kind of bizarre moves i would never play.

And you can lookup my rating.

So the question is what is the intention of alphabet with alphago?
Is it a real honest match or another kasparov scandal?

My suggestion is to completely ignore the match against the 2 dan guy. That's elo 1900 or something. A guy who makes lots of small mistakes and no one ever confronted that guy with those mistakes (they don't have very strong programs right now). Even if that would be closer to FM level than 1900 level, his only job was to take care he lost all games and quickly hand over the job to a 9-dan guy as it's google man. Alphabet is huge company, if not largest in the world practical right now, they already look better by just playing a 9-dan of course, so that's what you do.


BY THE WAY - i'm very amazed they called that 2 dan player "european champion". Some years ago i played here a 4-dan guy (he of course toasted me in a manner world hadn't seen before you could toast someone that bad). And he spoke Dutch... ...yet odds are he sits in Japan already for a few years now...

Was deep blue lucky to beat Kasparov? Most likely. But it would NOT be crushed by a 2300 player today. That thing was extremely strong, it played far too many games to build that 2600+ Elo performance rating it held to win the Fredkin prize stage II (and III). Are programs today better? Of course. 20 years later. At the time they were untouchable vs computers and almost all humans. If they had continued, they would be untouchable today, still.

<snipped>
the games were not such high level, maybe 2200 at most Bob

The facts are that deep thought some years earlier performed a lot better than 2200 in tournaments and not against a single human but against many humans so they did not lose on purpose.

It is not logical that deep blue2 was weaker than deep thought.

It is clear that deep blue2 was clearly weaker than the programs of today and I can believe it is possible that it was at level of 2700 and weaker than kasparov but I do not believe it was at level of 2200 or 2300.

[diep]

Analyze the games Uri.

I know you are a weak player OTB especially strategical yet with your correspondence skills where your computer performs pretty well at ICCF, it should be not a big deal to analyze it.

2200 at most yeah. Realize also they had a very bad book. Just some pgn games. No nothing manual.

Kasparov *couldn't* play any mainline - he'd be +10 out of book otherwise every game.

[duncan]

What's this Go thing?

google has alphago and how realistic is this facebook program darkforest?

Who are the programmers of alphago?

What i do understand is that you quickly declare a program to be strong enough to play one of the 9 dan players. What i do understand from my own go program experiments is that already 20 years ago i outsearched silly any existing go program - they just didn't know how to search. Of course the life and death analysis i hardly did do.

What i do understand is that in 2005 i talked to set up a match chessprogram versus Kramnik and very quickly the negotiation was about that some sort of appointed lumpsum kramnik wanted 100% pay out in case he lost or drew the match and 20% of the lumpsum had to get paid when software would lose the match.

Now i tried to reverse that 20% payout - yet that was non-negotiable.

In that sense it's peanuts for alphabet to buy a 9 dan player - especially if the guy says: "heh my 9 dan status i got from playing games against a human, some sort of exhibition match i don't care about". Usually that caring starts *after* they figure out there won't be a rematch.

How honest will this match be?

What i do understand is the computer shock. I had it myself as well. No go player is used to play a strong computer, like chessplayers are by now. Result in chess is that the top players are tactical way stronger nowadays in their games. Sure they make mistakes yet they blunder less pieces away.

If you do a simple blundercheck over games from start 90s, that's hundreds of elopoints weaker in world top versus todays world top - thanks to the computer.

In go they didn't go through that yet.

What i do understand is that the best chessplayers come from all over the world. If we select the best player from a couple of billions of players then you end up with way stronger players of course than if you select it from a 100 million japanese and some dozens of millions of koreans - the best chinese players play chess nowadays - that's for sure

What i do understand is that go is easy to search in a dubious manner. In chess you still need a way to explore even the most weird line as it's all about capturing his royal highness - whereas in go you can safely forward prune.

What i do understand from some stronger go players is that the average deep combination is 30 plies in stronger go (ladders not counted) whereas in stronger chess that's 10-12 plies (checks not counted).

What i do understand is that in opening as a FM, nearly IM though (got 2 IM norm results and would be pretty easily to get IM in fact if i'd play a tournament instead of just competition where you sometimes of course are in bad shape), that even todays software doesn't have a clue and that in go it actually seems tougher yet if you analyze it, that might not be the case.

You start with empty board. So that's similar to an endgame you start with. Now the real problem with go is that the branching factor is huge - yet other than that it should be way easier than the opening in chess from knowledge viewpoint seen.

If we look objectively then chessplayers still are impressive in making choices in opening/middlegame. Yet far endgame - i refer basically to endgames with less pawns or a few passed pawns - once a very weak area of chessprograms, they are really total superior there nowadays over humankind. And i say that as someone who is GM level in endgame (though not in my last game in Belgium league where i didn't win the endgame yet drew it - bit busy building 3d printer prototype for sales).

By the way Bob - a neural network only optimizes the heuristical PARAMETERS. It doesn't generate *knowledge*. We know from earlier attempts in computer go with good parameter tuning that this was very freakingly effective.

Yet that was playing go programs that already prune in the ROOT. Please realize that hard reality in go.

If you have something that randomly searches versus something that forward prunes hard in the root - then that explains the huge improvement of computer go software.

What i do realize is that because the board is larger yo usimply need a specific number of nodes a second. Some hundreds of millions preferably, to profit more from algorithmic improvements. You need to get that 30 plies.

What i do not know is how well it would scale strengthwise in go.

We'll see in March whether it's the kasparov/kramnik scenario. Yet i really wonder. Is facebook also serious or is this already a run race with some big pile of dollars changing hands and alphabet has something to brag about?

did you once say that you had a new way to program go to get to the top, but it will need 3 years of work.?

[bob]

At least in chess, just not making any large mistakes and exploiting any of the opponent's large mistakes is worth a lot of rating points. GMs blunder all the time. They mis-remember openings. They overlook shallow combinations. They exchange into losing endgames. They fail to find key moves that are in tablebases. True, they do it less often than weaker players but it is hard for any human to keep concentration and memory working well enough over a game to not make any serious mistake.

Then there's been a huge increase in search depth due to better hardware and algorithms over the past 20 years. So now computers are just tactical monsters and they find things no GM is going to find OTB.

Go is much more difficult. In Go the large branching factor and the difficulty in building a good evaluation are serious obstacles, but recent work is showing considerable progress in tackling those issues.

Kasparov liked the money too much. Greedier than Kasparov in those days they haven't been born.

Look at the last tournament played. Wijk aan zee. What was the prize money?

10k euro first place. number 14 spot something like 1000 euro.

In reality of course Carlsen doesn't get out of his bed until you pay well over 100k euro 'start fee'.

If you'd pay tennis or soccer or basketball players in the same way chessplayers get paid - namely lumpsum - then not a single soccer player would ever be sweatted.

Look at what i wrote about that Kramnik match i negotiated for to play.

back end 90s, lots of GMs were total jealous on how Kasparov could sell games and still be number 1 in the world.

I remember how i did make a beginners mistake in a simultaneous exhibition not too long ago. It was dozens of kids and just 30 boards. I had to run fast to give all kids a chance to play at least 1 game. there were so many more than 30 kids who wanted a shot at it.

I ran so fast. After 2 hours of running board after board (it's 5 seconds a board roughly for longer periods of time) suddenly the head master came to me: "we have to soon stop as next event is going to take place soon".

I quickly finished. I suddenly realized very ashamed: "i didn't give a single kid a draw".

There was elo 1900 kids also joining in that simultaneously exhibition. Just so you know. I'm just in the 23xx.

In the 90s and start 21th century, playing a computerchess exhibition for all those GMs was nothing more than a simultaneous exhibition. Couldn't lose rating with it - just give a good show.

Somewhere a year or 10 ago, with some seconds from Topalov i checked Diep's world championship book. They toasted it like it didn't exist that book. The GMs with the black colors.

Main variations they had analyzed further till move 30. whereas Diep was out of book in such lines at move 14 to 16 or similar.

Only approaching move 30 suddenly scores would drop to -2.x or lower. Really 2 pawns down in evaluation.

That's why the world top is to scary always to play a mainline. Go is a different game. like from those 361 moves you can make first move, already 350 you can throw away.
You can hard prune 300.

That's not possible in chess.

I don't buy this for Kasparov / Deep Blue. Kasparov won in 1996. IBM re-designed the hardware and set up a re-match in 1997. Had he won, they would have been back. Throwing the match would have ended all hope of another payout... He simply screwed up in the last game, but was outplayed in a couple of others as well.

[bob]

Three of the team members have now said they are 'quietly confident' for the Lee Sedol match, which makes me think they must have something now they didn't use for the previous match. I don't think they would say that if the ELOs are to come from bigger hardware only. Re-training the networks requires more games, takes a valuable time which could be used for other things, and even then may not bring much benefit since the 57% prediction rate they have now could be close to optimal.

In those monte carlo random type approaches, going from say 1000 cores to say 1 million cores doesn't bring that much playing strength wise.

Elo isn't getting used in Go, last time i checked.

Realize very few people play go at a stronger level, yet don't underestimate the strength of those 9-dan players.

Question is: is it another kasparov scandal, as deep blue 2 typically was rated about todays 2300 elo practical. Some FM today who doesn't blunder too much totally would bust that 1997 machine and program, even if you'd equip it with a modern book and its book was pretty bad for those years. See how much effort kasparov did do to lose that 2nd match versus deep blue.

All kind of bizarre moves i would never play.

And you can lookup my rating.

So the question is what is the intention of alphabet with alphago?
Is it a real honest match or another kasparov scandal?

My suggestion is to completely ignore the match against the 2 dan guy. That's elo 1900 or something. A guy who makes lots of small mistakes and no one ever confronted that guy with those mistakes (they don't have very strong programs right now). Even if that would be closer to FM level than 1900 level, his only job was to take care he lost all games and quickly hand over the job to a 9-dan guy as it's google man. Alphabet is huge company, if not largest in the world practical right now, they already look better by just playing a 9-dan of course, so that's what you do.


BY THE WAY - i'm very amazed they called that 2 dan player "european champion". Some years ago i played here a 4-dan guy (he of course toasted me in a manner world hadn't seen before you could toast someone that bad). And he spoke Dutch... ...yet odds are he sits in Japan already for a few years now...

Was deep blue lucky to beat Kasparov? Most likely. But it would NOT be crushed by a 2300 player today. That thing was extremely strong, it played far too many games to build that 2600+ Elo performance rating it held to win the Fredkin prize stage II (and III). Are programs today better? Of course. 20 years later. At the time they were untouchable vs computers and almost all humans. If they had continued, they would be untouchable today, still.

the games were not such high level, maybe 2200 at most Bob. If Kasparov was winning bigtime, he'd give away a pawn just to take care deep blue got back in the game. If he was having a crushing winningly strategic position he'd go moves like Ng5-h3 with white just to avoid playing Be3 giving white a blindfolded winning position.

And game 6 is a line that's also in a book i have. Yes, co authored by Kasparov...

Bob, Deep Blue wasn't lucky to beat Kasparov - Kasparov loved the money too much - and contractual he had to lose anyway (and i have that from an insider who was helping out Kasparov at that time).

Deep Blue by 1997 was so total outdated that they had to end it simply. It got 10-12 plies as you can see from the logs. Back in 1997 that was a good depth. 2 years later world champs 1999 that was not so great. IBM was clever enough to realize THAT.

Someone not in the know of all this also could have smelt it.

Directly after the match Kasparov praises Deep Blue bla bla bla.
then 1 hour later. Kasparov learned he wouldn't get a rematch. He's completely upset and starts spitting out nonsense. Fire comes out of his mouth. He slowly starts to realize what happened.

As a world champion he ALWAYS got another match. He naively hadn't thought about that until 1 hour after the match.

So the question i have is whether this will happen with Alphago as well.

A professional player lives from MONEY. And giants like IBM and Google have that in abundance.

You do realize that they were looking ahead? Hsu has already mentioned that the hardware had null-move search built in, but it was not used as there was not enough time for testing. There was nothing going on in 1997 that they could not replicate in their hardware. Today they would certainly be using all current search tricks, and where we are seeing depth=30, they would be hitting way beyond 40. Where I can hit 100M-200M nodes per second, they would be hitting 100B - 200B nodes per second. They would still be untouchable.

The ASICs they used would be considered heirlooms/antiques today.