.oO(I wonder if I'm too partisan to answer this question)

I agree with cazz about the dangers of not using the tool specified. But if you want to use perl:

I really struggled using AI::jNeural & AI::NeuralNet::Backprop, which is why I ended up writing AI::NNFlex. AI::Perceptron only supports simple Hebbian learning as far as I know, so you wouldn't be able to simulate non linearly separable problems with that package.

I'd suggest using either AI::NNFlex, AI::NNEasy or AI::NeuralNet::Simple.

AI::NeuralNet::Simple and AI::NNEasy both use inline C for speed. AI::NNFlex is pure perl, so it uses various mathematical neural net tricks to speed things up (momentum learning, fahlman constant, non linear error functions etc).

Which to use really comes down to your requirements - you'll get more granularity of control from AI::NNFlex, but you might find AI::NeuralNet::Simple or AI::NNEasy a bit easier to use.