Generalization Performance of Neural Nets versus Function Decomposition

Two learning algorithms, Neural Nets and Function Decomposition, are tested on a set of real-world problems of varying complexity and compared on two measures of generalization - a PAC-like measure called SAM and an average of the error over the learning curve, called ERR2. Function Decomposition generally performs better on both measures. Although it would be tempting to declare Function Decomposition the better algorithm, several considerations such as runtime and problem complexity distribution make such a conclusion tentative. 1 INTRODUCTION Inductive learning can be defined as the problem of filling in the missing values of a partially specified function based on the specified values [Michalski et al., 1983]. For theoretical purposes, an inductive learning algorithm is evaluated by how well it reconstructs a sampled total function, and how many samples it takes to do this to within a specified amount of error. Numerous techniques are used to solve the problem of inductive learnin..