Computer Chess

1
Computer Chess
Deep Blue (Campbell, Hsu Hoane) Hydra
(Donninger) Shredder (Meyer-Kahlen) Junior (Ban
Bushinsky) Fritz (Friedel) Rybka (Rajlich) Hiarcs
(Uniacke)
They make a lot of counterintuitive, even
absurd-looking, moves that on close inspection
can turn out to be outrageously creative By
generating countless new ideas, they are
expanding the boundaries of chess, enabling top
players to study the game more deeply, play more
subtly. John Watson, 2005
Fritz 9
2
Checkers
Chinook (Schaeffer) http//games.cs.ualberta
.ca/chinook World Man-Machine Champion since
1994
Othello Logistello (Buro) Defeated human world
champion in 2002
3
Backgammon
TD-Gammon (Tesauro) Ranked among top 3 players
Go programs are amateurs
Goemate (Chen Zhixing) vs. Go (Michael
Reiss)
Wulu (Lei Xiuyu)
4
Tic-Tac-Toe game tree
5
minimax search
maximizer
3
original board state
player 1 moves
?
minimizer
new board states
3
2
player 2 moves
new board states
3 9 2 10
static evaluation
6
minimax search with alpha-beta pruning
maximizer
(5) 3
(8) 3
original board state
player 1 moves
(2) 3
minimizer
new board states
(7) 2
(4) 3
player 2 moves
new board states
static evaluation
3 9 2 10
(1)
(3)
(6)
?-? principle (Winston) If you have an idea
that is surely bad, do not take the time to see
how truly awful it is
7
Apply the minimax procedure with ?-? pruning
?-? pruning improves efficiency, but
doesnt change the result
maximizer
minimizer
maximizer
6 4 8 6 4 0 2
2
8
What allows computer chess programs to
compete with humans?
What distinguishes chess programs?
Processing power and speed?
Wheres the knowledge?
9
Static evaluation function
weighted combination of features of the board
position
S c1 material c2 pawn-structure c3
mobility c4 king-safety c5
center-control
e.g. material pawn (1), king (3), bishop (3.5),
rook (5), queen (9)
Too weak to predict ultimate success
10
Role of experience
Openings and endgames use databases of known
positions and moves
Move generator Careful ordering of
potential next moves increases efficiency by
increasing a-b pruning of the game tree
Typical chess program Time Move
generator 50 Static evaluation 40
Search control 10
11
Searching under time pressure helpful
heuristics
progressive deepening
search until quiescent singular extensions
These heuristics allow deeper searches where
needed