Question about convergence

Hello,

as stated in the book "Programming NN with encog2 in Java", until now I have been training my NNs until their error was less than 0.01 ( i.e., error < 1% ).

However, while I was doing some researching for my project, I have just run across a particularly complex NN which doesn't seem to converge to such a low error level.

I have thrown at it many different NN structures:

- one hidden layer : hidden neuron count = input neuron count
- one hidden layer: hidden neuron count = 2 X input neuron count
- two hidden layers with different hidden neuron counts
...etc.

The lowest error I've been able to get so far is 5,3% ( 0.053). And that was a good one: usually it won't go lower than 9% - 10%.

So, here's the question: a NN trained to around 8%, should be considered a "bad" NN? Does this mean that such a NN will simply do his work with a possible error of 8%, or are there other effects due to the NN not having been trained under 1%?

Finally, is this normal and/or acceptable? Should I continue trying to decrease the error level, or will I probably get nothing better than this?

thanks a lot,