Encog regression with line output weight reliability. ( for error ponderation )

Question

I'm experimenting some tests with Encog and the AutoMPGRegression example :

https://github.com/encog/encog-java-examples/blob/master/src/main/java/org/encog/examples/guide/regression/AutoMPGRegression.java

My question is quite simple, is there a way to set weight reliability to the output of a data CSV line ?

In facts, some samples of data might be more reliable than others, so for not reliable lines data, we should tell to Encog that's is not so important if the neural network doesn't match for that output.

Extract of AutoMPGRegression sample csv data:

18.0   8   307.0      130.0      3504.      12.0   70  1    "chevrolet chevelle malibu"

15.0   8   350.0      165.0      3693.      11.5   70  1    "buick skylark 320"

18.0   8   318.0      150.0      3436.      11.0   70  1    "plymouth satellite"

I would like to tell to Encog that for the line 2, the output value "15" has a fiability of 50%, so the error should be ponderated for that line.

Thanks for help, trix

Answer 1

Good question! What you ask is actually correct classification of outliers: when some small part in teaching data is not good representation of whole training data. As usually neural networks are good in understanding that two exceptional rows are somehow outliers, and "understand" it without your help. If you have 1 - 5 % of outliers, neural network more or less will ignore them without your help. In some of my tests neural network was able to "ignore" even 10% of outliers. In other words neural network will consider outliers as noise. What you actually need is to be sure that your training set and validation set has approximately similar percentage of outliers.