Robotics Reference
In-Depth Information
This event was not a match for the World Championship. My
program did defeat the current World Champion, but it was an
exhibition match not involving the title. Further, a match to 7
points is not considered very conclusive in Backgammon. A good
intermediate player would probably have a 1/3 chance of winning
such a match against a world-class player, whereas in a match to 25
points his chances would be considerably diminished. At the time
of the match the bookmakers in Monte Carlo were quoting odds
of 3 to 2 if you wanted to bet on the machine, and 1 to 2 if you
wanted to bet on Vila. Thus the bookmakers apparently thought
the program to be very slightly better than a good intermediate
player. . .
...Further, the conditions of play may have worked somewhat
againstVilatakingtheprogramseriouslyatfirst...encasedinsidea
Backgammon playing robot
...therobotgaveasemi-comicimpressionduringtheevent,rolling
around controlled by a remote radio link and occasionally speaking
and bumping into things.
...It should also be pointed out that BKG9.8had somewhat the
better of the dice rolling. However, its dice were being rolled by a
human assistant, not by itself. [7]
Neurogammon
For about a decade after Berliner's success, little was heard from the
world of computer Backgammon. But then, from IBM, came a program
called Neurogammon, developed by Gerald Tesauro. Neural networks
had become the Artificial Intelligence community's learning method of
choice,
43
so it was hardly surprising when Tesauro and others started
applying neural networks to game playing.
Neurogammon was taught only the starting position for the game and
a few evaluation features that incorporated the most important Backgam-
mon concepts employed by expert human players. The program consid-
ered the game to be divided into six phases and employed a different
neural network for each phase. For example, the networks responsible
for making moves throughout most phases of the game were trained on
a set of positions from 400 games in which Tesauro, himself a strong
Backgammon player, took both sides. These networks were instructed
43
See the section “Artificial Neural Networks” in Chapter 6 for a simplified explanation of how
they work.
