Tuning Differential Evolution For Artificial Neural Networks

The efficacy of an optimization method often depends on the choosing of a number of behavioural parameters. Research within this area has been focused on devising schemes for adapting the behavioural parameters during optimization, so as to alleviate the need for a practitioner to select the parameters manually. But these schemes usually introduce new behavioural parameters that must be tuned. This study takes a different approach in which finding behavioural parameters that yield good performance is considered an optimization problem in its own right and can therefore be attempted solved by an overlaid optimization method. In this work, variants of the general purpose optimization method known as Differential Evolution have their behavioural parameters tuned so as to work well in the optimization of an Artificial Neural Network. The results show that DE variants using so-called adaptive parameters do not have a general performance advantage as previously believed.