On Applying Multilayer Perceptron Learning Properties to (Pseudo) Random Number Generation and Evaluation