You can download a toy database which will allow you to get started immediately and compare your results to those shown in the User Guide-Toy Problem.

To create the toy database, we used a function of three variables. We generated random points x,y,z and calculated the target, then added a random gaussian noise (very small).
Below are figures comparing the results obtained with Neuromat's Model Manager and NeuroSolutions's software. The latter does not provide a bayesian framework (and therefore no indication is given as to the uncertainty of fitting), in addition, it does not optimise the network architecture, nor does it build a committee of models.

The graphs show comparisons of the results obtained with Neuromat's Model Manager (MM), and NeuroSolutions (NS) using, in both cases, the default settings wherever possible. The correct value, as given by the function initially used to generate the data, are shown as (F).
Neurosolutions does not provide model comparison, so that the optimum number of hidden units (16) was taken from the Model Manager. For comparison, we have also shown results with a simpler model (NS-10) using Neurosolutions with 10 hidden units.

Note that in both cases, the models reached a virtually null test error when testing on cross-validation data. These data however, never show the effect of individual variables. While the model created with Model Manager has clearly grasped the correct individual effects of x, y and z, the other model has only recognised (partially) the impact of y but leads to severely erroneous predictions when y or z alone are varied.

All predictions are made within the range of variations of the database.