How can i train a multilayer perceptron with complex data?

Dear all,

i am interested in using a neural network classifier (i.e. MLP) for equalization of a quadrature amplitude modulation (QAM) signal.

the MATLAB toolbox does not support complex data as far as i understand... is there a way around this? i guess that i could simply split the data into real and imaginary parts and use two classfiers? However, this does not seem to be the ideal solution in my mind.

If someone can provide somebody can confirm in either direction whether it is possible to use a training method with a complex valued dataset i would appreciate it. Otherwise i think it is time to write my own network...

many thanks!

[Greg Heath]

Why would you use a classifier for equalization?

Answer 1

[Greg Heath]

It can and has been done. Just use neural and complex in any search engine.

Although I tried it once or twice, I gave it up as a waste of time.

My solution was to double the number of inputs and outputs by using real and imaginary parts.

I am not familiar enough with the existing references to steer you in the right direction.

Beware and Good Luck.

Greg

[Greg Heath]

This answer is w.r.t. complex NNs; NOT to NNs for QAM equalization.

Answer 2

[Greg Heath]

Google:

neural qam adaptive equalization

~ 5000 hits.

Hope this helps.

Greg

Answer 3

[Paul]

Hi Greg,

thanks for your comments, interest and time.

to answer sequentially: viewing equalization as a classification problem is quite common. in fact the performance of classifiers such as MLPs is superior to "traditional" equalizers such as a DFE.

i think your second comment is the way i will proceed, thank you. i did not do this in the first place because it felt a bit like cheating as it is a simple modification, but ultimately i suppose it is simply a split-complex approach rather than fully complex which is not a problem.

thanks! indeed i understand the literature and theory, however it was not clear to me whether MATLAB could support the fully complex approach without having to write the whole code myself.

thanks once again for your all of your help.

best wishes, Paul