I am trying to develop a neuro-fuzzy model for classifying two activity states from movement attributes.
I tried this earlier with Mamdani fuzzy where the rules look like as follows- "IF speed is less and change in acceleration is less and direction change is less THEN chance of being rest is high and chance of walking is less." and like this...there are some 12 rules.
As you can see this fuzzy rule has 3 antecedent variables and 2 consequent variables so we can call it a multiple-input multiple-output or MIMO structure.
I also developed another model in Neural network with these 3 inputs and 2 outputs in weka through multilayer perceptron. where I directly used the categorical values and the model worked perfect. so far so good.
Now the problem is I am trying to develop a neuro-fuzzy model (ANFIS in matlab) using the same three inputs and given outputs. However, as we know MATLAB ANFIS supports only one sugeno type zero order or 1st order output (class) for each record. Existing ANFIS allows mapping 'n' number of inputs to 1 output and that too with nominal or numerical value (no support for categorical data). If I use some nominal value for respective categorical data (say for rest 1 and walk 0) during learning phase the model is tuning the function parameters and on test data the output values are becoming too unreasonable (sometimes the values are floating, or beyond 0 ,1). Don't know how can I address this issue in Neuro-fuzzy. But it seems Matlab doesn't allow MIMO structure. So far all the neuro-fuzzy examples I found used a single output (defined on a continuous function and hence there is no problem with the float values). I want something like- "IF speed is less and change in acceleration is less and direction change is less THEN chance of being rest is 100 and chance of walking is 0." (see outputs are both constant and thus satisfy zero order Sugeno type) Is there any neuro-fuzzy tool which allows to model multiple outputs?
Hello,
I think that you are using toolbox.
You can solve the problem of 3 inputs and 2 outputs using command-line. I have also tried using it and it works.
Aarti
Hello
You can not do this with MATLAB
you should use CANFIS in this case
some other softwares are able to employed for your case
Cheers
Thanx Ali. Where can I get a software/API that supports CANFIS or is there already a dedicated software for this purpose? I searched on net but did not see any result for MIMO ANFIS except NEFCLASS and that doesn't run on recent version of Java.
@Aarti, can you refer me to a tutorial/code that you use for your MIMO structure? I am using genfis1 and genfis2 both and tried to develop 3 columns for input and 2 column matrix for output and fed them as training data but it is not working out. Are you sure you used Sugeno or Mamdani? Because in MATLAB, Mamdani fuzzy supports MIMO but Sugeno does not support MIMO structure- atleast that's what I have been experiencing.
Thnx once again to both of you Ali and Aarti.
Dear Rahul,
Check the help for anfis and you will get the code.
epoch_n = 20;
in_fis = genfis1([x y],5,'gbellmf');
out_fis = anfis([x y],in_fis,epoch_n);
a=evalfis(xt,out_fis);
Here, x,y are the training input and output respectively and xt is the test input.
And this forms the Mamdani Fuzzy Model.
Try it once and if you are still not getting the output please ask me.
Aarti
Dear Aarti,
Thnx for your reply. I have no issues with Mamdani fuzzy. The problem is with the neuro-fuzzy (ANFIS sugeno). Where the input vector has 3 elements and output vector has 2 elements. In your case (the example you provided) it has got only 1 element in the output vector and hence it is working fine. Actually Mamdani fuzzy works fine for more than one output. But I am trying to use ANFIS.
ANFIS works only with one output. Yo need as many ANFISes as outputs to model multioutput systems. As Ali pointed out, you could use CANFIS instead, but I think you have to program it yourself (no Matlab tool fot that).
However, if you only nedd to discriminate two classes (say rest/walk) and you obtain values floating or beyond (0,1), you can add a treshold at the output (and, eventually, a saturation function to 0/1: smaller values than 0.5 meaning "rest" (saturate to 0) and larger values meaning "walk" (saturate to 1).
On the other hand, why do you nedd a fuzzy system for that problem? Do you want to interpret the system after trainig? If not, use a neural network instead. If you feel that using gaussian-like functions works better for your problem you can use a RBFN insted of a MLP. In the end, an ANFIS is just a RBFN with some impossed restrictions.
One more thing.
When I said that you need as many ANFISes as outputs, I meant as many as output variables. In a classification problem you could consider that all classes are values of one variable (activity, in this case), so you can try to codify the activities using different numerical values of that variable (say different integer values between 0 and 10). Then, after training, you can round the output value of the ANFIS to the nearest integer.
@Rahul Deb you can do this in Matlab,
you have to just train for each parameter and save the trained ANFIS in .fis formate then simply cement the Matlab code with simple commands. If face some more difficulty email me on [email protected]
Thanks
Dear Rahul ,
Please Check this link.
Yours
http://www.mathworks.com/matlabcentral/fileexchange/36098-adaptive-neuro-fuzzy-inference-systems--anfis--library-for-simulink/content/Gradient%20Consistency%20Check/canfis_scatter_forward.m
Dear All,
In case you are still interested in this thread, you can see my latest publication where I successfully used ANFIS in a classification problem. I used MATLAB and JAVA to implement my model.
http://www.mdpi.com/2220-9964/5/11/207
Thanks to all
Yes, you have to implement it by yourself. Read my previous answer in this thread.