www.ailab.si

Download Report

Transcript www.ailab.si

CG 2006

Computer analysis of World Chess Champions

Matej Guid

and

Ivan Bratko

Introduction

Who was the best chess player of all time?

  Chess players of different eras never met across the chess board.

No well founded, objective answer.

Computers...

 Were so far mostly used as a tool for statistical analysis of players’ results.

High quality chess programs...

 Provide an opportunity of an objective comparisson.

Statistical analysis of results do NOT reflect:

 true strengths of the players,  quality of play.

I Wilhelm Steinitz, 1886 - 1894

Related work

Jeff Sonas, 2005:

  rating scheme, based on tournament results from 1840 to the present, ratings are calculated for each month separately, player’s activity is taken into account.

Disadvantages

  Playing level has risen dramatically in the recent decades.

The ratings in general reflect the players’ success in competition, but NOT directly their quality of play.

II Emanuel Lasker, 1894 -1921

Our approach

  computer analysis of individual moves played determine players’ quality of play regardless of the game score  the differences in players’ style were also taken into account  calm positional players vs aggresive tactical players  a method to assess the difficulty of positions was designed

Analysed games

  14 World Champions (classical version) from 1886 to 2004 analyses of the matches for the title of “World Chess Champion”  slightly adapted chess program Crafty has been used III Jose Raul Capablanca, 1921 -1927

The modified Crafty

   Instead of time limit, we limited search to fixed search depth. Backed-up evaluations from depth 2 to 12 were obtained for each move.

Quiescence search remained turned on to prevent horizont effects .

Advantages

  complex positions automatically get more computation time, the program could be run on computers of different computational powers.

Obtained data

    best move and its evaluation, second best move and its evaluation, move played and its evaluation, material state of each player.

IV Alexander Alekhine, 1927 -1935 and 1937 - 1946

Average error

  average difference between moves played and best evaluated moves basic criterion

Formula ∑|Best move evaluation – Move played evaluation| Number of moves

 “Best move” = Crafty’s decision resulting from 12 ply search

Constraints

  Evaluations started on move 12.

Positions, where both the move suggested and the move played were outside the interval [-2, 2], were discarded.

 Positional players are expected to commit less errors due to somewhat less complex positions, than tactical players.

V Max Euwe, 1935 - 1937

Average error

V Max Euwe, 1935 - 1937

Blunders

  Big mistakes can be quite reliably detected with a computer.

We label a move as a blunder when the numerical error exceeds 1.00.

VI Mikhail Botvinnik, 1948 - 1957, 1958 - 1960, and 1961 - 1963

Complexity of a position

Basic idea

 A given position is difficult, when different “best moves”, which considerably alter the evaluation of the root position, are discovered at different search depths.

Assumption

  This definition of complexity also applies to humans.

This assumption is in agreement with experimental results.

Formula ∑ |Best move evaluation – 2nd best move evaluation| best i ≠ best i - 1

VII Vasily Smyslov, 1957 - 1958

Complexity of a position

depth

2 3 4 5 6 7 8 9 10 11 12

1st

Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc1 Qc1 Qd4

eval

-0.09

+0.24

+0.08

+0.35

+0.07

+0.57

+0.72

+0.96

+1.30

+1.52

+4.46

2nd

Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc2 Qc2 Qc1

eval

-0.17

+0.16

+0.00

+0.30

+0.02

+0.55

+0.60

+0.87

+1.16

+1.26

+1.60

complexity = 0.00

Euwe-Alekhine, 16th World Championship 1935 VII Vasily Smyslov, 1957 - 1958

Complexity of a position

depth

2 3 4 5 6 7 8 9 10 11 12

1st

Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc1 Qc1 Qd4

eval

-0.09

+0.24

+0.08

+0.35

+0.07

+0.57

+0.72

+0.96

+1.30

+1.52

+4.46

2nd

Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc2 Qc2 Qc1

eval

-0.17

+0.16

+0.00

+0.30

+0.02

+0.55

+0.60

+0.87

+1.16

+1.26

+1.60

complexity = 0.00

Euwe-Alekhine, 16th World Championship 1935 VII Vasily Smyslov, 1957 - 1958

Complexity of a position

depth

2 3 4 5 6 7 8 9 10 11 12

1st

Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc1 Qc1 Qd4

eval

-0.09

+0.24

+0.08

+0.35

+0.07

+0.57

+0.72

+0.96

+1.30

+1.52

+4.46

2nd

Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc2 Qc2 Qc1

eval

-0.17

+0.16

+0.00

+0.30

+0.02

+0.55

+0.60

+0.87

+1.16

+1.26

+1.60

complexity = 0.00

Euwe-Alekhine, 16th World Championship 1935 VII Vasily Smyslov, 1957 - 1958

Complexity of a position

depth

2 3 4 5 6 7 8 9 10 11 12

1st

Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc1 Qc1 Qd4

eval

-0.09

+0.24

+0.08

+0.35

+0.07

+0.57

+0.72

+0.96

+1.30

+1.52

+4.46

2nd

Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc2 Qc2 Qc1

eval

-0.17

+0.16

+0.00

+0.30

+0.02

+0.55

+0.60

+0.87

+1.16

+1.26

+1.60

complexity = 0.00

Euwe-Alekhine, 16th World Championship 1935 VII Vasily Smyslov, 1957 - 1958

Complexity of a position

depth

2 3 4 5 6 7 8 9 10 11 12

1st

Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc1 Qc1 Qd4

eval

-0.09

+0.24

+0.08

+0.35

+0.07

+0.57

+0.72

+0.96

+1.30

+1.52

+4.46

2nd

Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc2 Qc2 Qc1

eval

-0.17

+0.16

+0.00

+0.30

+0.02

+0.55

+0.60

+0.87

+1.16

+1.26

+1.60

complexity = 0.00

Euwe-Alekhine, 16th World Championship 1935 VII Vasily Smyslov, 1957 - 1958

Complexity of a position

depth

2 3 4 5 6 7 8 9 10 11 12

1st

Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc1 Qc1 Qd4

eval

-0.09

+0.24

+0.08

+0.35

+0.07

+0.57

+0.72

+0.96

+1.30

+1.52

+4.46

2nd

Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc2 Qc2 Qc1

eval

-0.17

+0.16

+0.00

+0.30

+0.02

+0.55

+0.60

+0.87

+1.16

+1.26

+1.60

complexity = 0.00

Euwe-Alekhine, 16th World Championship 1935 VII Vasily Smyslov, 1957 - 1958

Complexity of a position

depth

2 3 4 5 6 7 8 9 10 11 12

1st

Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc1 Qc1 Qd4

eval

-0.09

+0.24

+0.08

+0.35

+0.07

+0.57

+0.72

+0.96

+1.30

+1.52

+4.46

2nd

Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc2 Qc2 Qc1

eval

-0.17

+0.16

+0.00

+0.30

+0.02

+0.55

+0.60

+0.87

+1.16

+1.26

+1.60

complexity = 0.00

Euwe-Alekhine, 16th World Championship 1935 VII Vasily Smyslov, 1957 - 1958

Complexity of a position

depth

2 3 4 5 6 7 8 9 10 11 12

1st

Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc1 Qc1 Qd4

eval

-0.09

+0.24

+0.08

+0.35

+0.07

+0.57

+0.72

+0.96

+1.30

+1.52

+4.46

2nd

Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc2 Qc2 Qc1

eval

-0.17

+0.16

+0.00

+0.30

+0.02

+0.55

+0.60

+0.87

+1.16

+1.26

+1.60

complexity = 0.00

Euwe-Alekhine, 16th World Championship 1935 VII Vasily Smyslov, 1957 - 1958

Complexity of a position

depth

2 3 4 5 6 7 8 9 10 11 12

1st

Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc1 Qc1 Qd4

eval

-0.09

+0.24

+0.08

+0.35

+0.07

+0.57

+0.72

+0.96

+1.30

+1.52

+4.46

2nd

Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc2 Qc2 Qc1

eval

-0.17

+0.16

+0.00

+0.30

+0.02

+0.55

+0.60

+0.87

+1.16

+1.26

+1.60

complexity = 0.00

Euwe-Alekhine, 16th World Championship 1935 VII Vasily Smyslov, 1957 - 1958

Complexity of a position

depth

2 3 4 5 6 7 8 9 10 11 12

1st

Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc1 Qc1 Qd4

eval

-0.09

+0.24

+0.08

+0.35

+0.07

+0.57

+0.72

+0.96

+1.30

+1.52

+4.46

2nd

Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc2 Qc2 Qc1

eval

-0.17

+0.16

+0.00

+0.30

+0.02

+0.55

+0.60

+0.87

+1.16

+1.26

+1.60

Euwe-Alekhine, 16th World Championship 1935

0.00 +

(1.30 – 1.16)

VII Vasily Smyslov, 1957 - 1958

Complexity of a position

depth

2 3 4 5 6 7 8 9 10 11 12

1st

Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc1 Qc1 Qd4

eval

-0.09

+0.24

+0.08

+0.35

+0.07

+0.57

+0.72

+0.96

+1.30

+1.52

+4.46

2nd

Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc2 Qc2 Qc1

eval

-0.17

+0.16

+0.00

+0.30

+0.02

+0.55

+0.60

+0.87

+1.16

+1.26

+1.60

complexity =

0.14

Euwe-Alekhine, 16th World Championship 1935 VII Vasily Smyslov, 1957 - 1958

Complexity of a position

depth

2 3 4 5 6 7 8 9 10 11 12

1st

Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc2 Qc1 Qc1 Qd4

eval

-0.09

+0.24

+0.08

+0.35

+0.07

+0.57

+0.72

+0.96

+1.30

+1.52

+4.46

2nd

Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc1 Qc2 Qc2 Qc1

eval

-0.17

+0.16

+0.00

+0.30

+0.02

+0.55

+0.60

+0.87

+1.16

+1.26

+1.60

Euwe-Alekhine, 16th World Championship 1935

3.00

+

(4.46 – 1.60)

VII Vasily Smyslov, 1957 - 1958

Complexity of a position

VII Vasily Smyslov, 1957 - 1958

Average error in equally complex positions

  How would players perform if they faced equally complex positions?

What would be their expected error if they were playing in another style ?

60 50 40 30 20 10 0 0,1 0,3 average 0,5 0,7 complexity Capablanca 0,9 Tal 1,1 VIII Mikhail Tal, 1960 - 1961

Percentage of best moves played

 It alone does NOT reveal true strength of a player.

IX Tigran Petrosian, 1963 - 1969

The difference in best move evaluations

X Boris Spassky, 1969 - 1972

Percentage of best moves played...

... and the difference in best move evaluations

XI Robert James Fischer, 1972 - 1975

Material

70 5 60 0 50 20 -15 10 31 31 41 41 51 51 61 61 71 71 81 81 91 XII Anatoly Karpov, 1975 - 1985

Credibility of Crafty as an analysis tool

  By limiting search depth we achieved automatic adaptation of time used to the complexity of a given position.

Occasional errors cancel out through statistical averaging (around 1.400

analyses were applied, altogether over 37.000 positions).

Using another program instead of Crafty...

  An open source program was required for the modification of the program.

Analyses of “Man against the machine” matches indicate that Crafty competently appreciates the strength of the strongest chess programs.

Deep Blue 0.0757

Deep Fritz 0.0617

Deep Junior 0.0865

FritzX3D Hydra 0.0904

0.0743

New York, 1997 Bahrain, 2002 New York, 2003 New York, 2003 London, 2005 Kasparov Kramnik Kasparov Kasparov Adams XIII Garry Kasparov, 1985 - 2000

Conclusion

 Slightly modified chess program Crafty was applied as tool for computer analysis aiming at an objective comparison of chess players of different eras.

 Several criteria for evaluation were designed:     average difference between moves played and best evaluated moves rate of blunders (big errors) expected error in equally complex positions rate of best moves played & difference in best moves evaluations  A method to assess the difficulty of positions was designed, in order to bring all players to a “common denominator”.

 The results might appear quite surprising. Overall, they can be nicely interpreted by a chess expert.

XIV Vladimir Kramnik, 2000 -

XIV Vladimir Kramnik, 2000 -