Experimental evaluation of a caching technique

Abstract

Inductive Logic Programming (ILP) is a Machine Learning technique that has been quite successful in knowledge discovery for relational domains. ILP systems implemented in Prolog challenge the limits of Prolog systems due to heavy usage of resources such as database accesses and memory usage, and to very long execution times. The major reason to implement ILP systems in Prolog is that the inference mechanism implemented by the Prolog engine is fundamental to most ILP learning algorithms. ILP systems can therefore benefit from the extensive performance improvement work that has taken place for Prolog. On the other hand, ILP is a non-classical Prolog application because it uses large sets of ground facts and requires storing a large search tree. One major criticism of ILP systems is that they often have long running times. A technique that tries to tackle this problem is coverage caching [?]. Coverage caching stores previous results in order to avoid recomputation. Naturally, this technique uses the Prolog internal database to store results. The question is: does coverage caching successfully reduce the ILP systems running time? To obtain an answer to this question we evaluated the impact of the coverage caching technique using the April [?] ILP system with the YAP Prolog system. To understand the results obtained we profiled Aprils execution and present initial results. The contribution of this paper is twofold: to an ILP researcher it provides an evaluation of the coverage caching technique implemented in Prolog using well known datasets; to a Prolog implementation researcher it shows the need of efficient internal database indexing mechanisms.