Development of Methods How to Avoid the Overfitting-Effect within the GeLog-System

摘要

This article examines the methods how to avoid an over-fitting-effect within GeLog-systems. This effect can be observed in nearly all systems of inductive concept learning, if due to false classification of examples false, especially too specific theories, are learned. There are a number or procedures, how to counter the effects of the overfitting-effect or to avoid it. This article develops criteria for the selection of those procedures. In this context, the integrability into the GeLog-system, a system of genetic inductive logic programming, is of great importance. Finally, a filter procedure, based on the correlation heuristic, which is also used for top-down-pruning, is selected, as it promised the possible application to a relatively huge amount of problems. After that, the efficiency of the methods will be proven with the help of systematic experiments.