Presented experiments show that usage ofevolutionary approach to feature - duction is justi?ed.Feature selection as well as construction gives goodresults. It is noticeable that attribute construction's best results assign higher classi?- tion accuracy than feature selection alone.That is why, carrying out selection before construction to decrease searchingspace isagoodsolution. Because of indeterministicbehavior of neuralnetworks,it was di?cultto - ducefeaturesetincaseofusingthemto evaluatecandidateresults.Forexample, aneuralnetworklearntverywellondatathatwasdescribedbyfullattributeset, but when thisset was decreased it had huge problems to do this duringrequired number ofepochs.That suggests that usingC4.5 ismuchmore preferred. Numerous experiments havebeen performed and observed.Analysis ofabove results allowsto put the hypothesisthat it is worth to use Construction module as the feature set reduction. But experiments show that Constructormodule does not work sowell whenitusesthe whole initial set offeatures - the search space istoo large.Soit is worth to use ?rstly Selectorand nextConstructor.
The second important issue isthatConstructor destructs the semanticmeaning of the features.New constructed features are notunderstandableforusers.In some real-liveproblems measuring offeature values isquite expensive, forsuch problems selector seems to be more suitable because itdiminishes a number of realfeatures.To constructonefeaturesa number ofreal(measured)featurescan be required. Obtainedresults haveencouragedus to extendour system,especiallythe c- structormodule.Weplan to developenlarged set offunctionsFwhich allowsto use the system with data containingdi?erenttype offeatures,not only nume- cal. Such system will be veri?ed usingagreater number ofbenchmark data sets as well as real data. Acknowledgments. This work ispartially ?nanced fromthe Ministryof S- ence and Higher Education Republic of Polandresources in 2008-2010 years as a Poland-Singapore joint research project 65/N-SINGAPORE/2007/0.