Genetic algorithms are optimization programs similar to the linear programming models discussed in Chapter 9. Genetic algorithm software conducts random experiments with new solutions while keeping the "good" interim results. A example problem would be to find the best subset of 20 variables to predict the stock market. To create a genetic model, the 20 variables would be identified as "genes" that have at least 2 possible values. The software would then select genes and their values randomly in an attempt to maximize of minimize a performance or fitness function. The performance function would provide a value for the fitness of the specific genetic model. Genetic optimization software also includes operators to combine and mutate genes. This quantitative model is used to find patterns, like other data mining techniques.
Neural network tools are used to predict future information by learning patterns and then applying them to predict future relationships. According to Berry and Linoff (1997), neural networks are the most common type of data mining technique. Some people even think that using a neural network is the only type of data mining. Vendors make many claims for neural networks. One claim that is especially questionable is that neural networks can compensate for a lower quality of data. Neural networks attempt to learn patterns from data directly by repeatedly examining the data to identify relationships and build a model. They build models by trial and error. The network guesses a value that it compares to the actual number. If the guess is wrong, the model is adjusted. This process involves three iterative steps: predict, compare, and adjust. Neural networks are commonly used in a DSS to classify data and, as noted, to make predictions. Figure 10.2 shows that various inputs (from I1 to Ik) are transformed by a network of simple processors. The processors combine and weight the inputs and produce an output value.
Figure 10.2 Neural Network Example.