Trained NN results

Hello,

I have this neural net:

private BasicNetwork createNetwork() {
this.network = new BasicNetwork();
this.network.addLayer(new BasicLayer(new ActivationSigmoid(), true, 10));
this.network.addLayer(new BasicLayer(new ActivationSigmoid(), true, 20));
this.network.addLayer(new BasicLayer(new ActivationSigmoid(), true, 1));
this.network.setLogic(new FeedforwardLogic());
this.network.getStructure().finalizeStructure();
this.network.reset();
return this.network;
}

and when I start to train with Backpropagation (or ResilientPropagation - results are same), it runs only for 5 loops(!) and error rate very quickly quick to 0.006 and then stop the training period (my max error rate is 2%). Is seems to be strange for me.

run:
8.10.2009 9:47:15 org.encog.neural.networks.training.propagation.Propagation iteration
INFO: Beginning propagation iteration
Epoch #1 Error:0.7027441258644403
8.10.2009 9:47:22 org.encog.neural.networks.training.propagation.Propagation iteration
INFO: Beginning propagation iteration
Epoch #2 Error:0.3750384946232957
8.10.2009 9:47:27 org.encog.neural.networks.training.propagation.Propagation iteration
INFO: Beginning propagation iteration
Epoch #3 Error:0.10244367070539317
8.10.2009 9:47:33 org.encog.neural.networks.training.propagation.Propagation iteration
INFO: Beginning propagation iteration
Epoch #4 Error:0.016045278537732646
8.10.2009 9:47:39 org.encog.neural.networks.training.propagation.Propagation iteration
INFO: Beginning propagation iteration
Epoch #5 Error:0.006197002799930156
BUILD SUCCESSFUL (total time: 32 seconds)

After it I save neural network into file on disk. When I restore NN and check its error with original training dataset, the error rate is same when training where finished. So neural network seems to be trained correctly(?):

run:
Error = 0.006100640877554247
BUILD SUCCESSFUL (total time: 2 seconds)

But, when I try to predict value from original dataset, generated output values (almost all) are totally different from original ideal values (differences are in thousands of percents). I can not explain these differents for myself when neural network's error rate is counted to 0.006. I think it is related to too quickly done learning period.

I can pass on more details and CSV file for neural dataset.

Thank you for any advices and ideas.