[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

computer-go: relative strengths chess vs go



If Kramnik has a good position, he'll win from god.

This is the basic misunderstanding by bad players.

The best professional players can *finish* better positions no
matter how you play. 

There is a certain 'no point of return' point in gameplaying
where one can finish a position completely. 

Will a professional go player *ever* lose when he starts with 9
stones up and playing for his life?

At 02:38 PM 9/23/00 -0400, you wrote:
>Part 2 of Go-Devil.
>In my last  post, I talked about  rating curves in  Chess and tried to
>make a case that the very best human  players actually don't play very
>well compared to what would be possible by omniscient entities.

ratings in chess don't compare very well to GO. Apart from that
there are many ratings in chess too. I'm having
  - an international rating FIDE (at most a few games a year count for
    that rating, so it's highly inaccurate, only for professional players
    it's fairly accurate)
  - national rating
  - clubrating
  - blitzrating
  - a lot of different ratings at the internet at each server a different
    rating

For your perfect database vs best man experiment the rating system is
highly inaccurate anyway, as it doesn't take into account the draw
chance at all. It just reflects how you score (assuming you play
enough games).

In GO i guess draws are not very common, but in chess at the highest level
there are a few young players who simply hardly lose, on the other hand
it's relatively easy to get a draw against them. Very well known
top-gm is Peter Svidler. Great at blitz, but he has the 'name' to
draw a lot of games in tournaments.

In a post on the rating list one says 100 rating points are worth
a stone. 

The rating K-factor as defined for the international rating is 15 above 2000
and 10 above 2400. Under 2000 it's 30, but that's not interesting
anyway as you can only get an international rating when you're >= 2000
rated.

All japanese international rated chess players:

 7000383  Aoyama, Yoshiaki                   JPN  2232    9  01.07.60
 7000065  Gonda, Gentaro                     JPN  2191    6  06.02.50
 7000260  Ishii, Ichiro                      JPN  2056    6  04.09.59
 7000448  Kanbegawa, Minoru                  JPN  2210   12  01.06.62
 7000073  Matsumoto, Yasuji                  JPN  2205    0  20.11.33  i
 7000197  Matsuo, Tomohiko                   JPN  2186   20  26.06.26
 7000251  Nakamura, Ryuji                    JPN  2127   17  25.11.77
 7000219  Ramos, Domingo                  m  JPN  2309    0  09.10.60
 7000170  Sasaki, Katsumi                    JPN  2171    5  15.09.53
 7000022  Schmidt, Loren R                f  JPN  2340    0  01.02.54  i
 7000391  Shiomi, Ryo                        JPN  2044   18  19.06.75
 7000359  Sugimoto, Koichi                   JPN  2030    7  04.01.56
 7000308  Tabata, Minoru                     JPN  2130    4  01.01.55
 7000162  Takahashi, Takayoshi               JPN  2205    0  17.04.57  i
 7000332  Tanaka, Yuuki                      JPN  2126   18  20.08.77
 7000227  Watanabe, Akira                 f  JPN  2374    6  01.03.72

If mr. sugimoto with 2030 plays Kasparov who has:

 4100018  Kasparov, Gary                  g  RUS  2849   35  13.04.63

So if 100 points would be 1 stone, then Kasparov would play with 8 stones
behind against any 2049- rated person on the above list.

Kasparov is 9 dan professional, Kramnik too. 
Then big gap and a lot of 8 dans on the list.

Now a 2000+ rated person will make huge blunders most likely.
So 8 stones up for kasparov means kasparov will sweat like hell,
but he'll win.

 7000227  Watanabe, Akira                 f  JPN  2374    6  01.03.72

If kasparov plays with 4 or 5 stones less against Watanabe then 
Kasparov will get crucified, raped and completely annihilated, as
Watanabe won't make huge blunders. No way with 2374.

Most non-game playing persons are not able to imagine this very accurately.

The k factor 15 versus 10 says basically: it's 1.5 times harder to
win rating above 2400 as under 2400 rating.

Yet that would mean a stone difference of 4 x 1.5 = 6 stones.

How fast would kasparov lose the game then?

Under 2000 the K factor is 30, which means theoretical that it's 3 times
easier to win or lose rating under 2000. So if a 2000 rated plays
a 1400 rated, that would mean 6 stones difference, yet a 1400 rated
is what we call a 'weak club player'. He'll blunder away a queen and
will lose anyway. Yet if a 2200 plays a 2800 person then the equivalent
would be getting up 3 (k=30 ==> k=10)  x 6 (6x100 points) = 18
stones up, then the comparision is simply not true, as the 2200 rated
won't blunder away a queen.

So comparisions within the rating system are already impossible,
when you want to compare it with go and 'stones up'.

>I have some further thoughts on this taking it from a different angle.

>In the previous post I made reference to Kasparov's play and held this
>up  as an example of  the best human play.   However  there is kind of
>chess that is generally  acknowledged to be of  a much  higher quality
>than this,  and it's called   correspondence chess.  I have  talked to
>several master on this subject as well as one former world champion in
>correspondence chess  and the consensus is  that the level of  play of
>top  players   FAR  exceeds    that  of over    the  board   play.  In
>correspondence play,  each  player has  many hours  or  even  days  to
>respond to a single move, and uses this time to do serious analysis.

the same nonsense repeated. don't compare openingsnovelties from Kasparov
and Kramnik and their outstanding positional play with the average
correspondence game where moves are basically tactical not losing,
but after a move or 20 you get to a position where the advantage from
Kasparov/Kramnik and others is that big that they'll win anyway, and no
database will ever rescue you then.

DIFFERENT is it when you replace that database with top grandmaster
Shirov, though 8 dan
imho he takes into account the weaknesses of his opponent and gets to
positions where his opponents play weak.

This is described in JICCA as: "opponent modelling". Though the article
sucked, it reflects how Shirov plays very accurately.

Shirov managed to beat Kramnik in a match some years ago (when kramnik
played also real strong, but was a bit younger as he is now:
   4101588  Kramnik, Vladimir               g  RUS  2770   23  25.06.75

>As a thought experiment, if you could give a top correspondence player
>the   ability to  make  the same  high   quality moves  but  in normal
>tournament time controls, you  would  have a  player far  superior  to
>Kasparov in over  the board play.  I  have  never heard a  good player
>deny this conjecture and I have talked to several about it.
>
>When games of the top players are analyzed later, it is common to find
>flaws in the  play.  It is so common  in fact that  it is clear we are
>playing far from perfectly.  There is a still a very large gap between
>the current level of play of top humans and perfect play.
>
>Because of our limited perception, it is  difficult for us to see much
>farther than from our own vantage point.  To us, brilliant and perfect
>play is simply whatever the best of us  can achieve.  We are impressed
>with   our own   accomplishments.  But to   a  God,  what  we consider
>brilliant and perfect might very well be simple minded and naive.
>
>I would like to also point  out that player perfect  moves most of the
>time does  not  mean you  play good  Chess  (or Go)   in general.  The
>difference between a  great player and  an  excellent player might  be
>that the excellent  player plays  perfetly  90% of  the moves and  the
>Great player gets it  right 93% of  the time.   Over a long  series of
>games the great player will  amass a huge lead, especially considering
>that the bad moves of the weaker player will get punished more often.
>
>In a real sense, there is no such thing as a good or excellent move in
>2 player games  of perfect information like  Chess and Go.  There  are
>only bad moves.  A good move is a human label  applied to a move which
>is difficult to find or is  "pretty", a Go Devil  kind of move!  A bad
>move is  a  move that doesn't   maintain  the game  theoretic outcome.
>There is a maxim in chess that says the winner is the next to the last
>person to make a mistake!   

Again only someone who doesn't play game at high level can only
comment that there are only bad moves.

Basically in most positions there are 'default' moves as dictated
by historical insight, and outstanding moves. Just playing 'default' moves
is what you will see in most games. Outstanding moves not seldom take 15
minutes of insight from a strong players. 

One of my 'outstanding moves' i needed 54 minutes.

Usually people who are in 'trouble' go search to try to escape. 

Apart from all those discussions, there are not seldom positions where
several moves are game theoretical all leading to a win, but where the
human 'outstanding' move gives a very simple win, and the computer move
gives a tough position which you can still lose because it leads to
a complex position.

>Finally, I don't know how to apply this to Go.   I feel that we are at
>least a few hundred elo rating points from perfect chess, even as many
>as 1000 (which means a hypothtical player exists who would rarely ever
>lose to Kasparov, and yet there still exists a player who would rarely
>lose to this player!)

the whole rating point system is very relative. If there would be
a perfect playing player now who would win *all games* then
he can only get 750 points more as the players he wins from.

If those have on average 2600 he'll have 750+2600 = 3350 rating.

Yet if there are in a couple of tens of years time several players with
2800, so the average tournament this player joins in would be 2700 rated,
then he would be rated 3450.

Rating is completely defined by the players you meet!

Rating systems are like pyramids. The more players and the more games
get played the higher the pyramid, the higher the highest rated person.

Also important is K factor. At the internet one sees ratings of 3000+ very
common, but that's because the k factor is 32 there instead of 10 and
because of the many computers in the competitions who inflate the rating
system in advance.

A question which returns a lot in the chessworld is: "what was the
rating from Capablanca?" 

Now this man lived a century ago, so it's real hard to say what his
rating was, but it would be in advance lower as any of the worlds top 10
at the current list.

  1 Kasparov, Gary.............  g  RUS  2849
  2 Kramnik, Vladimir..........  g  RUS  2770
  3 Anand, Viswanathan.........  g  IND  2762
  4 Morozevich, Alexander......  g  RUS  2756
  5 Adams, Michael.............  g  ENG  2755
  6 Shirov, Alexei.............  g  ESP  2746
  7 Leko, Peter................  g  HUN  2743
  8 Ivanchuk, Vassily..........  g  UKR  2719
  9 Topalov, Veselin...........  g  BUL  2707
 10 Bareev, Evgeny.............  g  RUS  2702
    Krasenkow, Michal..........  g  POL  2702
 12 Karpov, Anatoly............  g  RUS  2699
 13 Kasimdzhanov, Rustam.......  g  UZB  2690
 14 Svidler, Peter.............  g  RUS  2689
 15 Gelfand, Boris.............  g  ISR  2681
 16 Smirin, Ilia...............  g  ISR  2677
    Short, Nigel D.............  g  ENG  2677
 18 Dreev, Alexey..............  g  RUS  2676
 19 Azmaiparashvili, Zurab.....  g  GEO  2673
 20 Ye, Jiangchuan.............  g  CHN  2670
    Rublevsky, Sergei..........  g  RUS  2670
 22 Xu, Jun....................  g  CHN  2668
    Almasi, Zoltan.............  g  HUN  2668
 24 Khalifman, Alexander.......  g  RUS  2667
    Gurevich, Mikhail..........  g  BEL  2667
 26 Movsesian, Sergei..........  g  CZE  2666
 27 Georgiev, Kiril............  g  BUL  2661
 28 Akopian, Vladimir..........  g  ARM  2660
 29 Nikolic, Predrag...........  g  BIH  2657
    Peng, Xiaomin..............  g  CHN  2657
    Tkachiev, Vladislav........  g  FRA  2657
 32 Polgar, Judit (GM).........  g  HUN  2656
 33 Beliavsky, Alexander G.....  g  SLO  2654
 34 Piket, Jeroen..............  g  NED  2649
 35 Lautier, Joel..............  g  FRA  2648
 36 Seirawan, Yasser...........  g  USA  2647

>Christoph gave the rough  rule of thumb  that  1 stone equals 100  ELO
>rating points.  If this is  approximately right, then  the idea of the
>best players  being  close to perfect  (within 1   or 2  stones) seems
>fairly ludicrous.  200 rating points (or two stones?) means that among
>equal players, the handicapped player would still win about 15% of the
>time although losing about 85% of the games.

>I think Go  is MUCH more profound and  difficult to master than Chess,
>and this by a really wide margin.  I have heard similar numbers quoted

this is completely untrue. There are more professional chessplayers
as there are professional go players.

So get as strong as kramnik you need to be able to win from all these
players.

If you play golf there is (if i remember well) a professional league of
ONLY 100 players. It's much easier to get #10 on that list as it is
to get #10 on the chess rating list.

Any achievement must be seen relative to the strength of other players.

There is no exception between chess and go there, only there are more
chessplayers on this earth, seeing the above rating list you'll see
that chess is getting bigger and bigger in china. 

Not so long ago not a single chinese player was on the top half of that
list. Now i count several real strong chinese players. Especially Ye
is amazing. About 4 years ago i was real dissappointed if i lost from
the #2 from china. Now the #2 from china would finish me completely with
a 100% score for him most likely. Also the strongest female chinese players
would completely kick my butt.

Chess is growing in strength everywhere, to get a top player is
REALLY hard. to get a top player in the USA is simply impossible,
in 99% of the USA one doesn't have even trainers that are adequate to
train you to that kind of level. So you won't get over 2600 (you need
over 2600 to get a reasonable living out of chess in west-europe, in 
the americas many are living from chess real well with way way less) 
in advance. 

>for  chess, where it's been  hypothesized that we  are only 200 or 300
>points away from perfection.  I even admit that it's hard to imagine a
>player  who could achieve  a score of  over 80% against  the great and
>wonderful Kasparov, but I think that kind of  thinking is what ties us
>in knots, we really shouldn't make our guesses by comparing to players
>we worship because they are too close to God in our minds!

Oh well as we saw kasparov is easy to beat, but try kramnik. He'll win
the next 50 years from computerprograms if they don't improve *bigtime*
in knowledge. 

Kramnik's style is easy to describe: positional, chanceless, 
seemingly a 'i'm in no hurry' style, but winning quick nevertheless.

Now the knowledge gap is quite huge, especially the relative tuning of
knowledge. When is pattern A more important and when is pattern B more
important?

Sure is that search next so many years won't solve it for us.

In this respect it's really interesting to see how quick go programs
will play strong real soon when they will get several plies brute force,
but then after that one will see in the 'newspaper' ads indicating
they get stronger where in fact there is a very slow progress. Yet
for commercial reasons it's not interesting to beat #1000 in the world
if someone else already beated #800.

this is the basic problem in computerchess. Apart from that there
appears to be another huge difference in chess from which i don't know
exactly to what extend it applies to go. That's the opening. 

In chess you start with a board full of pieces. Now you can play
quiet openings like Kramnik which simply don't give the opponent a chance
to get to a complex looking position. A simple choice wrong and you 
will already lose the game when having black.

Past few years chessprograms have definitely become stronger, but
if you study what really happened was a basically that their books
became stronger. Books made by humans. Based upon grandmasters.
Also endgame became a LOT stronger.

Now in go you already *start* with the endgame and you end up in
complex positions.

In those complex positions the branching factor is considerably smaller
as the start of the game, where in chess the same is happening now.

Slowly chessprograms get real *strong* in endgame. Search depth is not
the reason, but the improvement in knowledge. Only certain types of endgames
will be hard to play well for the next few years, but it certainly
isn't search depth related.

In contradiction. Old 'stupid' versions of my program searched up to 20
ply in endgame at the world championship 1999 (quad xeon 400 linux).

The current version at a dual 800PIII under NT, which is a lot faster
as that quad xeon (better speedup AND 10.8% faster as it's NT instead of
linux AND the PIII is like 12% faster as the PIIXeon) is not getting
near that 20 ply even. It's searching many plies less, despite that
nowadays i also have EGTB implemented.

Yet it plays *much* better. It's clear the knowledge is the big win here,
and that search depth is just not very interesting, as long as you see
the basic tricks.

Search combined with even better knowledge will make the endgame real
strong in future from chessprograms, though the many tens of thousands
of exceptions which a strong player knows or will realize, 
a program won't know.

As endgame in chess is very good comparable to opening in go,
the basic question is: how many 'exception' opening rules do strong
go players know by approximation?

For chess we have the approximation from De Groot, which is a very old
approximation. he thought masters would know about 100000 patterns.

Now personal i don't doubt this approximation at all, as for me in chess
everything is related to a lot of features at the board, from which i can't
start implementing it into my program. Even worse, it's much harder as 
just the 'pattern' idea, i think we had a discussion about that some time
ago here!

Yet the interesting thing is the big boost in endgame for many
chessprograms last few years. 

I wonder when that will happen in go too!

>Just for reference here is a simple table of expectancies that the ELO
>rating system predicts.  For instance the last entry indicates that if
>you are 400 points higher rated, you will win 97% of the games against
>your weaker opponent.  Most of  the experiments I mentioned earlier in

A perfect player needs about 750 points more as in the above example.

>computer chess, for those not familiar, were based on playing hundreds
>of  games against programs   of  various depths,  and calculating  ELO
>ratings   based on winning   percentages.  In the  experiments, it was
>shown  that  by searching 1  ply deeper  you  could  obtain a stronger
>program, one that could  expect to win about 80% or more of the time.

That's a table from 1984, when scientists like 
professor v/d Herik wrote in his book: "a 10 to 11 ply search will
probably make chessprograms stronger as any human"

At that time chessprograms not only lacked loads of knowledge,
but also search depth. 

I think your own program at 256 processors or so searched in
paderborn like 16 plies or something?

In draughts one can have the same discussion obviously.
There the branching factor is even smaller as in go or chess.

The branching factor is just 2. There is now way to win the game
in a tactical way against a draughtsprogram. It solves all tricks from
world champions within a few seconds at most.

Yet a national competition player can easily fool it and win from it!

I've done some big search experiments there, as one can easily search up to
30 plies there within a timespan of 24 hours.

At tournament level depths of about 12 to 14 plies are achieved,
which is quite comparable to search depths in chess.

Yet if i allowed the draughtsprogram to search to 30 ply in positions
where it made simple openings mistakes, then it didn't play any better
move at all. 

so the 1 ply deeper is 80% idea is completely invalid when you are
searching through the tactical barrier. Of course things are also
dependant upon how the programs play their game.

If a program choses, when it has to chose between 2 winning moves, the
move that leads to a complex tactical positions, then you ask for trouble
obviously. 

>    ELO diff    wins %
>    --------    ------
>        0        0.50
>       20        0.53
>       40        0.58
>       60        0.62
>       80        0.66
>      100        0.69
>      120        0.73
>      140        0.76
>      160        0.79
>      180        0.82
>      200        0.84
>      300        0.93
>      400        0.97
>
>