blondie24 (round robin) cig09 seminar
TRANSCRIPT
INTRODUCING A ROUND ROBIN
TOURNAMENT INTO BLONDIE24
Belal Al-Khateeb Graham Kendall [email protected] [email protected]
School of Computer Science (ASAP Group)
University of Nottingham
Outline
-Introduction
- Checkers
- Samuel’s Checkers Program
- Chinook
- Deep Blue
- Blondie24
- Blondie24-R
- Blondie24-RR
- Results and Discussion
- Conclusion
- Future Work
2
Checkers3
Opening Board of Checkers (Black moves first)
Checkers4
Black Forced to make Jump
move
Checkers5
Black Gets King
Samuel’s Checkers Program
- 1959, Arthur Samuel started to look at
Checkers
- The determination of weights through
self-play
- 39 Features
- Included look-ahead via mini-max (Alpha-
Beta)
- Defeated Robert Nealy
6
Chinook
- Produced by Jonathan Schaeffer in 1989.
- 40,000 openings.
- 8-piece endgame database in 1994.
- Won the 1989 Computer Olympiad.
- Chinook become the world champion. The
first automated game player to have
achieved this.
7
Deep Blue
- Developed by IBM in mid 1990s.
- An attempt to create a Chess program that
was capable of beating the world champion
at that time
- 30 processors with parallel search, could
evaluate up to 200 million chess positions
per second
- 8,000 different features
- The opening database in Deep Blue
consisted of 4,000 positions
8
Deep Blue
- The end game database of Deep Blue
consists of all positions with five or fewer
chess pieces on the board.
- Defeated Gary Kasparov in a six-game
match in 1997 to become the first computer
program to defeat a world Chess champion.
9
Blondie24
- Produced by Fogel in 1999-2000
- Neural network as an evaluation function.
- Values for input nodes
Red (Black) – positive
White – negative
Empty – zero
- Piece differential
- Subsections (sub-boards)
10
Blondie2411
Blondie24’s EANN Architecture
Blondie24
- Initial population of 30 neural networks
(players).
- Each neural network plays 5 games (as red)
against 5 randomly chosen players:-
+1 for a win
0 for a draw
-2 for a loss
-Best 15 players retained, the other 15 players
eliminated.
-Copy the best 15 players (replacing the worst
15) and mutate the weights.
12
Blondie24
- Repeat the process for 840 generations and
the best player after these generations is
retained.
- Played 165 games at zone.com.
- Rating: 2045.85 at that time
- In top 500 of over 120,000 players on
zone.com at that time.
- Better than 99.61% of registered players on
zone.com
End Product
13
Blondie2414
Blondie24 Performance after 165 games on
zone.com
Blondie24-R
- Has the same structure and architecture that
Fogel utilised in Blondie24.
- The only exception that the value of the King
is fixed to 2.
- The King is more valuable than an ordinary
piece, and this is a well-known, even to
novice players.
15
Blondie24-RR
- Eliminate the randomness in the evolutionary
phase of Blondie24-R.
- A league competition between all the 30
neural networks.
- All the neural networks play against each
other.
- The total number of matches per generation
will be 870 (30*29) rather than 150 (30*5).
- This increase (number of matches) will
decrease the number of generations (840
verses 145).
16
Results and Discussion17
Blondie24-R Blondie24-RR Online WinCheck3D SX checkers
Blondie24-R - Draw Win Lose (7-Piece) Lose (8-Piece)
Blondie24-RR Win - Win Lose (2-Pieces ) Lose (4-Pieces)
Results of Playing Against Selected Programs
Results and Discussion
- Blondie24-RR plays two matches (one as red
and one as white) against Blondie24-
R, Blondie24-RR.
- Wins as red against Blondie24-R.
- The result is draw when Blondie24-RR moves
second.
- Reflects a success for our hypothesis based
on the fact that both players are end
products.
18
Results and Discussion
- Blondie24-R and Blondie24-RR win against an
online program which can be considered as
another success.
- Plays against two programs (strong).
- For the first one Blondie24-RR lost with a two
piece difference, Blondie24-R lost with a seven
piece difference.
- Playing against the second program shows
that Blondie24-RR lost with a four piece
difference, while Blondie24-R lost with an eight
piece difference.
19
Conclusion
- The results show that Blondie24-RR is
performing better than Blondie24-R.
- Based on these results it would seem
appropriate to use the league structure,
instead of only choosing five random
opponents to play against during the
evolutionary phase.
20
Future Works
21
- Investigate if other changes are possible.
- Investigate using individual and social
learning methods in order to enhance the
ability of Blondie24-RR to overcome the
problem of being an end product.
References
22
1- Samuel, A. L., Some studies in machine learning using the game of checkers 1959,1967.
2- Fogel D. B., Blondie24 Playing at the Edge of AI, United States of America Academic Press, 2002.
3- Chellapilla K. and Fogel, D. B., Anaconda defeats hoyle 6-0: A case study competing an evolved
checkers program against commercially available software 2000.
4- Fogel D. B. and Chellapilla K., Verifying anaconda's expert rating by competing against Chinook:
experiments in co-evolving a neural checkers player.
5- Chellapilla K. and Fogel D.B., Evolution, Neural Networks, Games, and Intelligence,” 1999..
6- Chellapilla K. and Fogel D. B., Evolving an expert checkers playing program without using human
expertise.
7- Chellapilla K. and Fogel D. B., Evolving neural networks to play checkers without relying on
expert knowledge.1999.
Questions/Discussions
Thank You
23