I want to argue with you a little.

In my view, at first you must tell about your problem in detail, it would be better as a mathematical model. You never tell about your problem but consider Genetic Programming as a solution approach to solve your problem. I want to ask: Why do you want to use Genetic Programming ? Why you don't consider any other method but Genetic Programming ? In my opinion, Genetic Programming can not be competitive for combinatorial models of high dimensions. I think, use of other methods can be more successfull.

Dear Amer Bakhach, The articles accessible via the links provided below may help you.

http://sep.stanford.edu/sep/gabriel/Papers/micro_GAs.pdf

http://link.springer.com/chapter/10.1007%2F978-3-642-37189-9_5

dear Gennady;

i wish you can refer to more specifics methods/models.

i think my research match exactly with Genetic Programming technique: searching for best non-linear function and i don't really know how much is the impact of each variable (ranging from 0 to 1) and most importantly i am not concern with the function itself as much as i am interested in the associated statistical measures (e.g. MSE, variance) of the resulted function;

Regards;

Dear Afaq,

thank you for your concern;

i start to read the first references; and according to the abstract i hope this will be useful. however i am not sure how much the second reference is related to my research i need to read it all before deciding;

thank you again; and i will keep you updated;

Regards;

Dear Amer Bakhach,

I thank you for your reply to my comment. I observed for a long time: 99,9% of researchers on ResearchGate ask their questions without any models. I want to cite Nobel winner, physicist, academician L.D. Landau: "Looking through a paper, I only read problem description and author's results. As it is necessary to solve this problem I know better than he". I quoted this phrase to express my thought as clear as I can. I'm interested to know only about problem description without any solution approaches. I want to view the mathematical model only. Please, give any links to read about your model in detail.

Thank you in advance

Dear Dr. Fedulov

i really want to thank you for your concern.

for details about the mathematical model please refer to:

Granger, C. W. J. (1969). "Investigating Causal Relations by Econometric Models and Cross-spectral Methods".Econometrica 37 (3): 424–438. doi:10.2307/1912791. JSTOR 1912791

best Regards;

Dear Amer,

as a first question, I would ask you if the all variables (50) are necessary or not? have you carried out any analsysi to select the best feature? if no, I would try to select the most relevant feature as the first step. If yes, then we have to find a way. could you tell me please what are the rows (samples) number of yur matrix (with 50 columns)? GP wuld be abale to map the relation between input-output set through its nature for non-stationary data, but at first we should know what is the final scope of your work? once we knew, there are alot of programms, e.g. with MATLAb, Discipolus, GeneXpro, Applet, Treenet, etc...

Looking forward to hearing from you.

_Best

Dear Jalal,

i am so grateful for your concern;

your comment inspire me a couple of ideas;

in fact i didn't try yet to find "relevant feature"; in fact i depend on Discipulus: it can measure the importance of each variable.

anyway, attached is sample of my data;

my goal is to use the resulted equation (i.e. tree) by GP for forecasting issue.

kindest regards;

Amer.

Dear Amer,

Thank you very much or your information. As I could see in the attached file, there are 3 variables (N, C and F) as well as different lag patterns of the output (E) to be used as inputs for your model. Therefore, we have 4 independent-one dependent variables to be mapped through GP. I would ask you, here, what was the reason for selecting those time lags for N, F, C and E. Did you perform any ACF/PACF analysis (for best E time lags) or CCF (for selecting N, F and C time lags)? If no, any reviewer/reader would ask you whay you have selected those time lags?If you are going to make these results as a peer-reviewed paper or a research project, this question would appeared, of course. Therefore, kindly let me know if you have carriedout such analysis or not.

Neverthless, the reason for selecting N, C and F (among other possible variables which could affect your model) should be clarified.

Finally, although GP (Discipulus) can give weights, but at first, we have to select the most informative variables, then ask GP to proceed with its logical way.

I look forward to hearing from you.

Kind Regards,

Jalal.

Dear Jalal,

thank you for interest;

in the matter of fact, i didn't apply any of ACF nor PACF nor CCF;

this is the pure data (no preprocessing have been applied)

what i am trying to do is applying the theory of Granger-Causality (check the link below) using GP;

however, i am also trying to overcome the limit of lag-length in Granger theory (m=q) because it was proven in other article that it's not true in many real world cases.

check this link for Granger theory:http://www.jstor.org/discover/10.2307/1912791?uid=3738432&uid=2&uid=4&sid=21103177588121

finally, i want to say that i will try ACF and CCF during the next 48 hours;

i am really thankful for your help;

best regards;

Amer.

Dear Amer,

you're welcome.

OK. at the moment, I see no need for any pre-processing (normalization, etc) on your data and I guess, it will not be necessary, but I am not sure yet. Let's see what will be the outcome of ACF/PACF as well as CCF for your input-output set. Then, we might be more specified. Please use the raw data from your analysis.

Kind Regards,

Dear Jalal;

sorry for the delay;

kindly find attached the results: ACF, PACF, CCF;

however,

i think that a main problem in those measures is that: all of them can detected only "linear" dependency or correlation.

any advice?

best regards;

Amer.

Dear Amer,

You're welcome. Thanks for your efforts. I agree with you that the dependencies are linear, so it is difficult to make a decision. in this case, one alternative might be the application of PCAs...principles component analysis...Please find a gap and go through PCA to see what will be resulted. If doesn't work, we shall make the final decision based on GP. I will discuss this more in depth, after PCA...good luck

Kind Regards,

So did u use genexpro tool n pl send me the setup actually i have downloaded but not running

hello Gupta

i am using R for statistical computing and Discipulus as Genetic Prog. tool;

by the way, i was reading the attached paper;

and id didn't understand how they extend the search space for GP?? based on what statistical metrics??