Optimisation of a neural network model for calibration of voltammetric data

Neural networks are powerful tools for the calibration of multivariate analytical data, but the large number of network parameters make it difficult to obtain an optimum calibration model. Often, network parameters are guessed or chosen according to postulated ‘rules of thumb.’ In this paper, we perform a thorough network optimisation for the calibration of electrochemical data from tertiary aliphatic mixtures, with the dual aims of achieving the best possible calibration model, and identifying the most significant network parameters. Two optimisation methods are used: one involves testing networks over the available parameter space, while the other employs a genetic algorithm to perform a more focussed parameter search. The calibration accuracy achieved using the two methods is found to be similar, but the genetic algorithm is considerably more efficient. The number of network inputs and initial weights are found to have the greatest impact on network accuracy, while the number of training epochs and hidden layer neurons are seen to be much less important.