There is certainly a free R package able to do that, just google for it or go to the R repository site and try to find.
As I use the QNet2000 (Vesta Services, Inc.) I will comment only on its capabilities.
You may choose the transfer function between layers (among Sigmoid, Gaussian, Hyperbolic Tangent and Hyperbolic Secant) as well as the connections between individual neurons (connected to the previous layers neurons or breaking some of those individual connections).
It is a fairly easy to use software, fast and cheap (we bought it for $50 USD some years ago). It runs on Win 95/98/NT and on later Win versions in virtual machines. You can use (call the lib functions from) the libraries (DLL's) with other software.
The Tech support email is, according to the version we own, [email protected]. Drop them a message if you're interested in it.
That's it. Thank you Luis. Breaking some connections. That is the feature I'd like to test. I've read that if we have input factors from different areas (macroeconomics, contract and contractor in my case), the reasonable thing is to make grouped connections to the first hidden layer (not each to each neuron). Thank you.