how i can find fittnes function for signal?

My project is EEG analysis using genetic algorithm , am writing a code to analysis my signal using matlab.. And almost to finish just I have problem with genetic algorithm in matlab because I couldn't found fitness function for my signal.. I have to option to use genetic algorithm : 1/ selection my features. 2/ adjust the weight of neural network to classify the signal or enter the genetic in neural network some how..

Answer 1

[Star Strider]

There are many resources on the Internet that discuss genetic algorithm EEG classification techniques. It is usually necessary to pre-process EEG signals before classifying them with any classifier (wavelet transforms are likely the best), but some papers have used short segments of unprocessed EEG signals. This paper ‘Research on Classification Method of Combining Support Vector Machine and Genetic Algorithm for Motor Imagery EEG’ (free on the ‘Net) may help you design your fitness function. Unfortunately, none of the resources I discovered go into significant detail on their methods. This paper seems to do better than most, but you can likely find papers more directly applicable to your project in your university library.

[Star Strider]

The expression of your fitness function depends on the data you want to use (segments of time-domain EEG, wavelet coefficients, Fourier transform coefficients, etc.) and some sort of classification metric. You already know the patterns of the learning set you want to classify and have labeled them appropriately as to the condition they represent. So for instance if a particular EEG record is Condition #1, your fitness function should return a value that indicates that it is in Condition #1. The distance (or difference) between that value and the correct value for that particular record is the fitness. The less the distance between the actual condition and the condition returned by the GA for all conditions and records, the greater the fitness for that particular individual. In the end, your GA should have established a set (matrix) of parameters, with one set for each condition you want to classify.

So I would start with a fitness function that would take a matrix with a size of (conditions x parameters) (your initial population of individuals) and matches the parameters in the data with the parameter elements of the matrix. Your fitness function should return the condition dimension of your matrix for every EEG record presented to it. The individual with the closest approximation of classification results to the actual classes is the most fit. This is one approach, there are likely others that would work as well.

Also consider using the patternsearch function. It may be more appropriate to your problem than GA, but that is your decision.

Classifying EEG conditions is not a trivial problem, and there are no easy solutions. I did it years ago using STFT preprocessing, principal components analysis, and a linear discriminant that slowed the institution mainframe to a crawl (and generated several less-than-supportive e-mails from the system administrators to me). I suggest you look for papers by Pfurtscheller (I believe I spelled his name correctly) and his co-authors. They were among the first to do this (independently and simultaneously with the group I was associated with at the same time), and are likely still the leaders in the field.

[Star Strider]

Yassmin Idriss’s ‘Answer’ (on 6 May 2014) moved here:

i had do preprocess by band pass filter and notch filter to remove EOG and EMG..

I was reed alot of papers about EEG and genetic algorithm but i dont know how to transform this in matlab ..

i hope if can do same work like this in paper .. what is the code for that

[Star Strider]

There is likely no generic code for EEG classification because each application is different. You will have to write your own. (For instance, I was classifying EEG activation patterns by laboratory task performed, and designed my signal preprocessing and classifier for that specific task.)