Fuzzy Classifier andtheEffects on BenchmarkData andComplex ChemicalData
Jürgen Paetz, J. W Goethe-University · 2005
Theintegral partofadaptive systems arelearning algorithms thatareoften basedonheuristics. Suchalgorithms areusedforneuralnetwork, fuzzy system andneuro-fuzzy systemtraining. The performance can be measured on benchmarkdata.In thiscontribution we evaluate the performance ofaneuro-fuzzy system withrespect totheadapted weights. A comparison isgiven between theperformance using thetrained setofweights andtwosetsofoptimized weights. Evolutionary algorithms canbeusedforoptimizing thetrained weights andforoptimizing theweights directly without using the adapted weights asabasis. Additionally, allweights aresettoone tomeasure theinfluence oftheweight values. Theresults show that theweights haveonly asmall impact onperformance using benchmark data, butahighimpact whenusing morecomplex, higher dimensional chemical datawithskeweda priori probabilities.