A Multiplatform Chemometric Approach to Modeling of Mosquito Repellents

■179

prior information related to the very nature of the problem. The modeling of Rindex of the

aforementioned set of natural and synthesized compounds based on BP, tPSA and ClogP

molecular descriptors (the same inputs as in MLR2 model) was carried out applying ANN

approach as well. The modeling was done applying multi-layer perceptron (MLP) feed-

forward neural networks with Broyden-Fletcher-Goldfarb-Shanno (BFGS) algorithm with

different hidden and activation functions. The main set of the compounds was divided into

three sets: Eq. 9.1 training set for the networks’ training, Eq.9.2 test set for determination

of generalization error and Eq.9.3 validation set for finding the best ANN architecture and

training parameters. The architectures of the obtained networks are presented in Table 9.1.

The networks’ architecture is described in the form of number of input variables –

number of hidden neurons – number of output variables. The results indicate a good qual-

ity of the obtained networks considering high correlation coefficients (R) of the training,

validation and test sets, as well as acceptable room mean square error (RMSE) values. The

predictive ability of the established ANNs was estimated based on the comparison of ex-

perimental and predicted values of Rindex and, as it can be seen on Figure 9.4, there is

quite good concurrence between the experimental data and data predicted by the estab-

lished ANNs. The best concurrence between the data was achieved by the network MLP

3-8-1. One of the main flaws of the ANN approach is that there is no the exact mathemati-

cal equation that describes the relationship between input and output variables. This is the

reason that some researchers consider neural networks to be a “black box”.

Another application of the non-linear modeling of mosquito repellency of acylpiperidines,

carried out by using the ANN modeling approach, was presented in the study by Katritzky

Figure 9.4: The comparison between the experimental (target) Rindex values and Rindex

values (output) predicted by the established ANNs.