7.6.2 Larger Example

In this example, a function with two outputs is approximated.

Read in the Neural Networks package.

In[1]:=

Generate data and look at the two outputs.

In[2]:=

An RBF network containing four neurons is chosen. You can modify the structure and repeat the example. You can also change it to an FF network.

Initialize an RBF network.

In[9]:=

Out[9]=

Different algorithms will be compared with respect to execution time and efficiency; that is, their RMSE rate of decrease. The Levenberg-Marquardt algorithm is tested first using the separable algorithm.

Train 15 iterations and find the time used.

In[10]:=

Out[11]=

Consider now the case where the separable algorithm is not used.

Train 15 iterations and find the time used.

In[12]:=

Out[13]=

Compare the obtained fit illustrated in the previous two plots. Normally, the separable algorithm is more efficient, showing a larger rate of error decrease per iteration using approximately the same time.

You can repeat the example with the other training algorithms, such as Gauss-Newton and s steepest descent, by changing the option Method.