Both GENFIS1 and GENFIS2 use a given training data set to generate an initial fuzzy inference system that can be fine-tuned via the ANFIS command. However, GENFIS1 and GENFIS2 differ in two aspect.

First, GENFIS1 produces grid partitioning of the input space (and thus is more likely to have the problem of the ``curse of dimensionality'', while GENFIS2 uses SUBCLUST (subtractive clustering) to produces scattering partition.

Secondly, GENFIS1 produce a fuzzy inference system where each rule has zero coefficients in its output equation, while GENFIS2 applies the backslash ("\") command in MATLAB to identify the coefficients. Therefore the fuzzy inference system generated by GENFIS1 always needs subsequent optimization by ANFIS command, while the one generated by GENFIS2 can sometimes have a good input-output mapping precision already.

http://www.cs.nthu.edu.tw/~jang/anfisfaq.htm

Both GENFIS1 and GENFIS2 use a given training data set to generate an initial fuzzy inference system that can be fine-tuned via the ANFIS command. However, GENFIS1 and GENFIS2 differ in two aspect.

First, GENFIS1 produces grid partitioning of the input space (and thus is more likely to have the problem of the ``curse of dimensionality'', while GENFIS2 uses SUBCLUST (subtractive clustering) to produces scattering partition.

Secondly, GENFIS1 produce a fuzzy inference system where each rule has zero coefficients in its output equation, while GENFIS2 applies the backslash ("\") command in MATLAB to identify the coefficients. Therefore the fuzzy inference system generated by GENFIS1 always needs subsequent optimization by ANFIS command, while the one generated by GENFIS2 can sometimes have a good input-output mapping precision already.

http://www.cs.nthu.edu.tw/~jang/anfisfaq.htm

Thank you for information and the ANFIS FAQ link. From the answer I conclude genfis2 deals better than genfis1 with more inputs and the `curse of dimensionality'.

Hello Omidji,

Please go through this , it may helpful to you,

CONSTRAINTS OF ANFIS

Anfis is much more complex than the fuzzy inference systems discussed so far, and is not available for all of the fuzzy inference system options. Specifically, anfis only supports Sugeno-type systems, and these must have the following properties:

•Be first or zeroth order Sugeno-type systems.

•Have a single output, obtained using weighted average defuzzification. All output membership functions must be the same type and either be linear or constant.

•Have no rule sharing. Different rules cannot share the same output membership function, namely the number of output membership functions must be equal to the number of rules.

•Have unity weight for each rule.

An error occurs if your FIS structure does not comply with these constraints. Moreover, anfis cannot accept all the customization options that basic fuzzy inference allows. That is, you cannot make your own membership functions and defuzzification functions; you must use the ones provided.

here is the syntax,

Genfis1

fismat = genfis1(data)

fismat = genfis1(data,numMFs,inmftype,outmftype)

Genfis2

fismat = genfis2(Xin,Xout,radii)

fismat = genfis2(Xin,Xout,radii,xBounds)

fismat = genfis2(Xin,Xout,radii,xBounds,options)

fismat = genfis2(Xin,Xout,radii,xBounds,options,user_centers)

Genfis1 produces a FIS structure based on a fixed number of membership functions.

This invokes the so-called curse of dimensionality, and causes an explosion of the number of rules when the number of inputs is moderately large (more than four or five).

http://www-rohan.sdsu.edu/doc/matlab/toolbox/fuzzy/fuzzyt39.html

Dear

Suresh and Behrouz. Thank you for your points.

Is there any limitation with the number of inputs?As you know when we have more than 5 or 6 inputs a single ANFIS using genfis1 fails and we must construct few anfis in a cascaded form which complicate the overall model. Is genfis2 outperfom genfis1 in such cases?

Dear Mahmoud,

In using both, the number of training data should be more than or equal to the number of parameters.

If there is less then 6 inputs and a large size of training data, GENFIS1 can be used and if there are more inputs, GENFIS2 should be used.

Thank your dear Behrouz. Is there any work which compared (speed, performance, ...) between these functions.

Hi,

ANFIS structure allows a good performance approximation to nonlinear systems.

In terms of performance and accuracy, not only GENFIS1 or GENFIS2 provides good results. Also GENFIS3 allows you to build an ANFIS structure with mamdani outputs, not suitable to train trough ANFIS command.

If you have more than 6 inputs, you should explore not only to use the GENFIS commands but also, try to set parameters in a smarter way.

This is: as professors said GENFIS1 and GENFIS2 uses grid partition and subclustering to define the membership functions in antecedent. GENFIS 3 uses fuzzy c means as mechanism to clusterize the inputs and allows you to define if the consequent is a mandani or sugeno output.

I recommend you for improving the results of the estimation or modeling, to explore the data set clusterization through the command subclust to find the number of optimal centers of the data and later uses the fcm command to find these centers. This way you can explore the data surface and build an idea about the results you should find through GENFIS.

Later, in GENFIS construction according to the command (GENFIS2 or GENFIS3), set the optional parameters with the best results obtained in the first step.

Sometimes, GENFIS build with this method, do not require ANFIS training.

In my experience, GENFIS2 behaves better when the input output estimation corresponds to time domain data set (MISO). GENFIS3 works better with other kind of problems.

hope my answer can help.

Thanks to all for this lively discussion.

I too have a question.

How can we design  ANFIS  in MATLAB if the number of outputs is more than one?

Hi all

see it

https://www.researchgate.net/post/What_software_can_I_use_for_ANFIS_calibration?cp=re68_x_p2&ch=reg&loginT=iHFhqqBI8l10AuW08kYuXke4huakYlbZ1vE4p9NLjf4%2A&pli=1#view=53ac5ca5d685cc1e458b46ab

Hi all,

This is an application of GENFIS3:

A novel approach for ANFIS modelling based on Grey system theory for thermal error compensation

Conference Paper A novel approach for ANFIS modelling based on Grey system th...

Thank you Sirs for your valuable inputs.

Is it possible to change the membership function type in genfis2?

The default function is Gauss.

If yes, how it can be done?

In new release of Matlab you can create initial FIS structure using three methods:

For genfis3 the input membership function type is 'gaussmf'. By default, the output membership function type is 'linear'. However, if you specify type as 'mamdani', then the output membership function type is 'gaussmf'. Link to ginfis3 document is attached:

http://www.mathworks.com/help/fuzzy/genfis3.html

Dear Mahmoud,

thanks to sharing this valuable material.

One of them genfis2 preprocess input data in clusters to generate the input partition

Input particion is developed by using the FCM approach

Other way to see this difference is to check both codes. They are avaliable in the toolboox

Difference betweene genfis1 and gnfis2:

{source website: http://www.cs.nthu.edu.tw/~jang/anfisfaq.htm#6)

Both GENFIS1 and GENFIS2 use a given training data set to generate an initial fuzzy inference system (represented by a FIS matrix) that can be fine-tuned via the ANFIS command. However, GENFIS1 and GENFIS2 differ in two aspect. First, GENFIS1 produces grid partitioning of the input space (and thus is more likely to have the problem of the ``curse of dimensionality'' described in question 5), while GENFIS2 uses SUBCLUST (subtractive clustering) to produces scattering partition. Secondly, GENFIS1 produce a fuzzy inference system where each rule has zero coefficients in its output equation, while GENFIS2 applies the backslash ("\") command in MATLAB to identify the coefficients. Therefore the fuzzy inference system generated by GENFIS1 always needs subsequent optimization by ANFIS command, while the one generated by GENFIS2 can sometimes have a good input-output mapping precision already.