next up previous
Next: Game of Hex Up: Higher order statistics in Previous: Higher order statistics in

Introduction

Historically, chess has been thought to be a good test-bed for artificial intelligence. In the last 10 years, though, computers have risen above human level. This is largely due to fast computers being able to evaluate a huge number of possibilities. One can argue, whether these programs show much intelligence after all.

There are still games that are similar to chess in some sence but have proved to be difficult for computer players. These games include Go, Hex, Y, and Havannah. The games have some common factors: First, the number of available moves at a time is large, which prevents the use of brute-force search to explore all the possibilities as the search tree grows exponentially with respect to depth. Second, there are lots of forcing moves which boost the so called horizon effect. If the program sees a limited number of moves ahead (search depth), there is often a way to postpone the interesting events beyound that horizon.

This paper introduces a method that avoids these two pitfalls. The search is performed by playing out games from the current situation to the end several times randomly. This is known as play-out analysis or Monte Carlo playing. The number of explored possibilities is constant with respect to search depth. Also, as the games are played out all the way to the end, there is no need to evaluate a game in progress, thus avoiding the horizon effect.



Subsections
next up previous
Next: Game of Hex Up: Higher order statistics in Previous: Higher order statistics in
Tapani Raiko 2006-09-01