Possible bug in ErrorCalculation class

Submitted by vancv786 on Tue, 06/23/2009 - 13:23

Using Encog in Java

Hello,

I recently downloaded encog-core-2.0.0-beta2. Very cool piece of software, and very well built! After playing around a little bit, I think I spotted one possible bug in ErrorCalculation class, in package org.encog.util. In method updateError(double[], double[]) on line 87, there's a line:

this.setSize += ideal.length;

This will add a full length of the training set in every iteration, in stead of only increasing setSize by one for every element checked. I think the line should be:

this.setSize ++;

Could you please check this if you already haven't caught this one?

Thanks for providing a great platform!
Vanja

Submitted by JPS on Wed, 10/07/2009 - 12:27.

Is there any update to this bug?

Submitted by jeffheaton on Wed, 10/07/2009 - 19:51.

It will be a very easy fix. I don't think it has a great deal of impact, but it does throw off the error calculation to some degree. It won't change the way that it is trained. It will be fixed on C# and Java sides shortly. I will reply when it is fixed.

Jeff

Submitted by JPS on Thu, 10/08/2009 - 04:00.

Thanks for the fix. I saw the checkin.

I was using the feed forward network with back-propagation with a multi layered network with one hidden layer. I was finding too low of error for what I would expect.

How can I become a contributor in case I find something more?

Submitted by jeffheaton on Thu, 10/08/2009 - 06:06.

Both fixes have been checked in.

As to contributing fixes, have you ever used SVN before? I would need to give you access to the Google code SVN source control system.