Assuming that your membership function (MF) are gaussians and remembering that each gaussian has 3 modifiable parameters (mean, sigma and normalisation), I can infer that your ANFIS model has at least 17 MF * 3 param/MF = 51 modifiable parameters... But you have only 36 experimental data points, hence the warning message.

Don ´t overlook this warning message, it shows that you are over-fitting your data set, which you shouldn ´t do !

Please notice that I wrote "at least" as I didn't count modifiable parameters from input and output ANFIS units. And, of course, if your MF are not gaussians but another type will change the number of modifiable parameters per MF in the little computation above.

Thank you very much Mr. Leonard.

Can you please tell me how overfitting the data set will affect the final outcome?

in_fis = genfis3(Xin,Xout) ; output= evalfis(input,in_fis)

Thanks so much dear Léonard Studer for your nice answer 

let me answer Vishal John Mathai in this question.

overfitting means that your model is fitted to your training data so much so that have its best performance on them;but ,becasue of this fact(your model is focused only on training data),your final performance on testinjg data will become unpermissible and most of time bad.

Please try to use genfis3.

MATLAB support ANFIS by:Gengis1,Genfis2 and Genfis3

The different betwen this methods is rulling mthods.

Gengis1:

Generate Fuzzy Inference System structure from data using grid partition.

Genfis2

Generate Fuzzy Inference System structure from data using subtractive clustering

this two methos is supported by fuzzy logic toolbox

Genfis3

Generate Fuzzy Inference System structure from data using FCM clustering.

genfis3 generates a FIS using fuzzy c-means (FCM) clustering by extracting a set of rules that models the data behavior. 

maybe the number of rules in genfis3 become lower than 17 and I think so.

Please consider this point that genfis2 and genfis3 was introduced to decrease the number of rules(problem os Genfis1)

This option is not available in fuzzy logic toolbox and you should wite its program.It is easy and need to concentrate on it

use below codes:

Thank you Mr. Ali for that detailed explanation.

You are so welcome

all the best 

Overfitting: http://en.wikipedia.org/wiki/Overfitting

In case of overfitting, your final outcome will have poor perfomance on data your ANFIS system never saw. Your ANFIS model will be close-to-perfect on the very set of data you presented it during the learning phase. But new data presented to your ANFIS system after will be poorly modeled; it can be very bad and even dangerous for ANFIS systems used in real life objects.

Hi, In our analysis on ANFIS system for bio signal application, genfis 2 gives better result compared to genfis 3. In FCM, computation (no of rules) grows exponentially with the dimension of the problem but in SFC computation is proportional to problem size instead of the dimension. 

I have a doubt regarding the usage of data set.

For example, if my experimental design have 25 experiments and I have repeated those experiments twice (means I carried out each experiment twice), how should I use such a data for ANFIS?

Should I average out the readings (readings corresponding to the same experimental condition), which will result in 25 data sets or Can I use 25 + 25 = 50 data sets?

The modifiable parameters in ANFIS are of two types. Nonlinear parameters of the antecedent MF, and linear parameters of the consequent MF. The number of antecedent parameters depends on your selection of MF, number of MF per input, and number of inputs. The Gaussian MF has the minimum number of parameters (2 not 3, I believe). In your case the number of antecedent parameters would by 2*3*17 = 102, which is NOT aproblem at all. However, the number of consequent parameters depends on the partitioning of the fuzzy space, and the type of Sugeno model (constant or linear). In a 3-input ANFIS with 17 MF per input, you would have 17^3 = 4913 rules assuming grid partitioning. If you use constant model (zero order) you have one parameter per rule, which means 4913 parameters. This is OK. But, if you use linear model (first order polynomial), you will have 4 parameters per rule (3 inputs + a constant). This means 4*4913 = 19652 parameters. This is too much. Your computer will crash during training, particularly on a large training dataset. My advice to you, especially if you use linear Sugeno model, is to reduce the number of MF per input. It will help reduce the size of your modifiable parameters and avoid training crashes. In my 15-year experience of using ANFIS, I was never able to train a module as massive as yours. Good luck.

Thank you Mr. Awadallah for that explanation.

i m using genfid2 the given error occurred 

Attempted to access minX(2); index out of bounds because numel(minX)=1.

Error in subclust (line 130)

X(:,id) = (X(:,id) - minX(id)) / (maxX(id) - minX(id));

Error in genfis2 (line 83)

[centers,sigmas] = subclust([Xin Xout],radii,xBounds);

any one help me please. this is very necessary for me.

thanking you all

Once trained, i am nor sure which terms can be considered as configuration parameters

It can be refered to elements of the structure which are modified through the learning process