Automatic Generation of Search Engines

Abstract

A plethora of enhancements are available to be used together with the αβ search algorithm. There are so many, that their selection and implementation is a non-trivial task, even for the expert. Every domain has its specifics which affect the search tree Even seemingly minute changes to an evaluation function can have an impact on the characteristics of a search tree. In turn, different tree characteristics must be addressed by selecting different enhancements. This paper introduces Pilot, a system for automatically selecting enhancements for αβ search. Pilot generates its own test data and then uses a greedy search to explore the space of possible enhancements. Experiments with multiple domains show differing enhancement selections. Tournament results are presented for two games to demonstrate that automatically generated αβ search performs at least on a par with what is achievable by hand-crafted search engines, but with orders of magnitude less effort in its creation.