Machine Learning and AI with Simple Python and Matlab Scripts

3ANNs for Optimized Prediction

3.1 Introduction

ANNs can be effectively used for prediction of outcomes based on historical data if they are properly designed and effectively trained with sufficiently large numbers of samples. To be able train the weights of an ANN, a set of sample input and output data collected from the field or observed during an experiment are needed, as described in Chapter 2. Depending on the application, the number of samples typically ranges from thousands to hundreds of millions or even more. Input values of data samples are applied to the neurons of a hidden layer after being multiplied by the link weights. Following modification by summation and activation functions at each neuron, input signals propagate to the next hidden layer and finally to the ANN output layer. An error value is computed based on the difference between the real and calculated output values. These steps are called a forward propagation. Then, as part of the error minimization (or at least reduction) process, a back propagation is performed by computing partial derivatives of the error with respect to the link weights starting at output layer and working back to the input layer, as described in Chapter 2.

Consecutive forward and back propagation iterations (called epochs) over ANN layers reduces the error until either a satisfactorily small value is obtained, a limit for the number of epochs is reached or another stopping criterion is met. As mentioned earlier, for an ANN to be ...