Evolutionary optimization of weights of a neuro-fuzzy classifier and the effects on benchmark data and complex chemical data

The integral part of adaptive systems are learning algorithms that are often based on heuristics. Such algorithms are used for neural network, fuzzy system and neuro-fuzzy system training. The performance can be measured on benchmark data. In this contribution, we evaluate the performance of a neuro-fuzzy system with respect to the adapted weights. A comparison is given between the performance using the trained set of weights and two sets of optimized weights. Evolutionary algorithms can be used for optimizing the trained weights and for optimizing the weights directly without using the adapted weights as a basis. Additionally, all weights are set to one to measure the influence of the weight values. The results show that the weights have only a small impact on performance using benchmark data, but a high impact when using more complex, higher dimensional chemical data with skewed a priori probabilities.