| This thesis presents an implementation of supervised learning performed with a stochastic neural net, as applied to a classification of business firms based on financial statements and ratios. The general Boltzmann Machine Algorithm is presented in detail, as well as the specific version designed for this application. We describe in detail several classifications that we attempted to reproduce, and some of the different possible ways of encoding the data issued from the income statements and the balance sheet. The results are shown to depend on the chosen encoding. We also confirm that the neural net behaves better when it is trained on the most difficult cases. |
