In an earlier publication, an Ant Colony Optimization — Artificial Neural Networks (ACO-ANN) based algorithm for feature subset selection was presented. The algorithm employed ANNs to evaluate subsets produced by ants. It is hypothesized that the performance of the algorithm depends on the generalization ability of the training algorithm used in the ANNs. This paper tests the performance of the algorithm for different types of neural network training techniques. The results obtained demonstrate that all the results from the studied training methods are competitive and the selection of an appropriate training algorithm depends on the customer requirements and his priorities such as desired accuracy or subset with least number of features or time of computation.