© William W. Armstrong, 1999

We now use the estimate of the noise in our data to set the parameters of ALNBench in order to obtain a function with good generalization properties. We push the slider to the middle to increase the input tolerance to a value that corresponds roughly to about half the distance between training points. This way, we are smoothing over the gaps between training points.

An error on the training set significantly less than 0.36 would indicate overtraining. A little experimentation shows that by taking the output tolerance to be 0.36, we can get the ALN to have 0.3 error on the training set and 0.39 error on the test set. The trained function has 37 linear pieces active. The small number of pieces suggests the function will generalize well.

Recall that the set in our test set window is “estimate.txt”. We can do anything we like to try to get the error 0.39 on that set lower. Of course, we do not expect any method to reduce the error below 0.36.

Now we have to take the ALN function resulting from this process and test it against the data set we held back.

NEXT