文档介绍：GP-Gammon: Using ic Programming to
Evolve Backgammon Players
Yaniv Azaria and Moshe Sipper
Department puter Science, Ben-Gurion University, Israel
azariaya,sipper ***@,
{ }
Abstract. We apply ic programming to the evolution of strategies
for playing the game of backgammon. Pitted in a 1000-game tourna-
ment against a standard benchmark player—Pubeval—our best evolved
program wins 58% of the games, the highest verifiable result to date.
Moreover, several other evolved programs attain win percentages not far
behind the champion, evidencing the repeatability of our approach.
1 Introduction
The majority of learning software for backgammon is based on artificial neural
networks, which usually receive as input the board configuration and produce as
output the suggested next best move. The main problem lies with work’s
fixed topology: The designer must usually decide upon this a priori, whereupon
only the internal synaptic weights change. (Nowadays, one sometimes uses evo-
lutionary techniques to evolve the topology [1]).
The learning technique we have chosen to apply is ic Programming
(GP), by puter programs can be evolved [2]. A prime advantage of GP
over artificial works is the automatic development of structure, ., the
program’s “topology” need not be fixed in advance. In GP we start with an initial
set of general- and domain-specific features, and then let evolution determine
(evolve) the structure of the calculation (in our case, a backgammon-playing
strategy). In addition, GP readily affords the easy addition of control structures
such as conditional and loop statements, which may also evolve automatically.
This paper details the evolution of highly essful backgammon players via
ic programming. In the next section we present previous work on machine-
learning approaches to backgammon. In Section 3 we present our algorithm for
evolving backgammon-playing strategies using ic programming. Section 4
presents results, foll