neural network with constrained output

Hi all
I am a beginner of neural networks and I am trying to train a network with these characteristics:
14 input in the input layer
30 hidden nodes in the hidden layer
5 output in the output layer

I know from the nature of the problem that the sum of the 5 outputs must be 100 (or 1) and each output value belongs to [0,100] (or [0,1]).

I have trained a lot of NN playing with training parameters but I sometime obtain negative values (the sum of the output is always near 100).

How can I constrain the output to follow the physics? Change to logsig the transfer function in the hidden layer? Change the performance function adding a penalization term (how can I do it)?

This is the essence of my very simple script (matlab R2010b)

nninput=CCTInputs; %14x661
nntarget=CCTmicroTargets; %5x661
net=feedforwardnet(30);
net=init(net);
net.trainParam.max_fail=20;
[net,tr]=train(net,nninput,nntarget);
nnoutput = net(nninput)

I have also tryed logsig function in the hidden layer adding net.layers{1}.transferFcn='logsig';
and not changing anithing else, but the performance of the network in term of predictive capabilities have worsened.

Thank you in advance
bye!

Ale