On the stability of example-driven learning systems: A case study in multirelational learning

Abstract

A popular way of dealing with the complexity of learning from examples is to proceed in an example-driven fashion. In the past, several researchers have shown that using an example-driven approach, it is possible to learn even structurally complex generalizations which would have been difficult to find using other multirelational learning (ILP) algorithms. On the other hand, it is also well known that the quality of the learning results in example-driven learning may depend on the ordering of the examples; however, such stability issues have received almost no attention. In this paper, we present empirical results in several multirelational application domains to show that instability actually affects the performance of a well-known example-driven ILP system. At the same time, we examine one possible solution to the instability problem, presenting an algorithm which relies on stochastically selected examples and parallel search. We show that our algorithm almost eliminates the instability of example-driven search with limited additional effort.