Computational Characteristics of Law Discovery Using Neural Networks

Lately the authors have proposed a new law discovery method called RF5 using neural networks; i.e., law-candidates (neural networks) are trained by using a second-order learning algorithm, and an information criterion selects the most suitable from law-candidates. Our previous experiments showed that RF5 worked well for relatively small problems. This paper evaluates how the method can be scaled up, and analyses how it is invariant for the normalization of input and output variables. Since the sizes of many real data are middle or large, the scalability of any law discovery method is highly important. Moreover since in most real data different variables have typical values which may differ significantly, the invariant nature for the normalization of variables is also important.