Automatic Ordering of Predicates by Metarules
Tristan Cazenave · 2001
I describe metarules which order predicates contained in first order logic rules. These metarules are applied to rules created by a learning Go program. After the ordering of the predicates, the rules are matched orders of magnitude faster than the original learned rules. Key words : Metaprogramming, Metareasoning, First Order Logic, Game of Go, Machine Learning. 1 Introduction One aspect of machine learning systems is the efficiency of the learned rules. This problem has been referenced as the utility problem [Minton 1988]. I have written a learning program which transforms its problem solving activity in the efficient matching of first order rules [Cazenave 1996]. A component of this system is the compilation of the learned rules. This paper describes how learned rules are compiled so as to match them orders of magnitude faster. This problem has already been adressed in papers such as [Ishida 1988]. My approach makes explicit use of metaprogramming, my system reasons on itself so a...