CONCLUSION

As shown in the results, GP seems to be a powerful technique for extracting knowledge from different systems. In this chapter, it has been applied to two systems, a database and an ANN, using a well-known problem, the iris flower data. We achieved good results with the additional advantage of having as results presented as mathematical expressions which show the relationships among the parameters.

When we used GP to extract knowledge from databases, we saw two different points of view, using one expression or three different expressions to make the classification.

In the first attempt, the one-tree classifier, we show how we can adapt GP to produce decision rules with the desired shape, and, therefore, how we can obtain high-level explicit knowledge about the system. In this part, we can also see that it is better to do a pre-processing of the data to improve the performance of GP. This is so because we gave all of the parameters the same rank; so, the system finds it easier to work with all variables and constants in the same rank (i.e., it does not find problems in combining constants and values with much different values). We can conclude that, with a minimum analysis of the data, we can improve the process in two ways: in the final success and in the time needed to obtain it.

The second attempt, the three-tree classifier, gave additional knowledge. With the construction of three different Boolean expressions, one for each class, we obtained an additional knowledge. Now we can detect errors made by the system.

GP, then, is shown to be a suitable technique for extracting knowledge from databases, not only in classification problems. Its ability to adapt to many different environments (the user selects which operator is needed to be included in the sets) allows for the extraction of mathematical relations, decision rules, etc.

But we also used GP to extract knowledge from a more complicated system, an ANN. The extraction of the knowledge contained in it has made it possible to understand the network. This rule extraction process for ANNs can be used on any network, therefore, making possible their use in many other application areas where the ability to explain the reasoning processes is important.