VII. Libneural, a software implementation of a Backpropagation Neural Network
OK, now you should understand the principle of this particular neural
network topology, and you want to play around or even you want to learn a
little more about it The solution for you is an already implemented and
stable library called libneural.
This C++ written library gives the powerful of a BPN neural network
without having to bother with implementing the training algorithms as
described in the previous section, actually you just have to do the
following steps
- Initialise your neural network (you need to specify the number of
inputs, outputs and hidden-layer nodes)
- Provide it with a set of example input values and desired output values
(to train the network), repeatedly cycling through your training data.
Inputs and outputs are one-dimensional arrays of floating point numbers,
which should be between zero and one. For best results, limit your desired
output to between 0.05 and 0.95, otherwise the network will not
converge.
- When training is finished, you can present a set of new input values and
it should give you the correct output.
- If it doesn't work, try changing the number of hidden nodes or the rate
of learning.
You can even save on disk all connections weight (also called the memory
of the network) to a file for later recovery (to avoid teaching your network
the same thing again).
