Good morning Aarti 

Yesterday we were running your data (in my AI class) and here we have the results:

As you may know, the anfis structure generated by matlab, uses clustering techniques or the initial generation of membership functions in the antecedent, and fixing these parameters uses a least squares methodology to obtain the consequent parameters. Nevertheless you should be aware of the nature of data by running first the clustering algorithms in order to determine a “good” set of parameters for training. I will lead you through several recommended steps:

1. Normalize the data without using the genfis/anfis commands, just normalize and take the data to the hypercube [0, 1] or  [-1, 1]

2. Select randomly a 60% of the data for training, 20% for test and 20% for validation VERY IMPORTANT. 

3. Run clustering algorithms over the whole set via subclust command in order to know the sensitivity of your data. 

You have just 524 patterns but 34 inputs and at least 3 outputs and some clustering algorithms such as subtractive provides 340 clusters with radii 0.5 or 31 clusters with radii 0.2 which means that your data is too sensitive to clustering algorithms. We like to run it several times looking for a nice parameter... but if you know too much the set you can semi-supervise the results.

4. Run fuzzy c means algorithm with the optimal number of clusters from subtractive and try to plot the 3-d projection of the centroids and clusters to verify if the results have any sense for you. I was just wondering that you work on power systems.

5. When you are satisfied with the number of clusters, use just the training set and run the genfis2 command and remember that you are going to build a MISO sugeno structure. Use the options menu to set a nice parameter for subtractive clustering!!!!

6. Also, try a genfis3 and remember to use the fuzzy c means right number of clusters as parameter and also try building a ‘mandami’ type of FIS (see the command line setup) because we find out a nice results with it. Also build a ’surgeon’ structure in order to compare.

7. Now we have 3 fis that reflects the training set with good accuracy, use evalfis command to evaluate antis results and see if it is ok with your research.

8. Try evaluating the test and validation set in order to check if the ANFIS can extrapolate the data

9. it is very important to realize that this structure is highly dependent of the number of patterns to properly converge to the right connections in the network. Try to use not too much clusters because you will be overtraining the net and will never reach the capability for extrapolation.

10. Finally you can train a 'sugeno’ type structure with anfis command but… it seems that genfis2 or  genfis3 can provide you a nice results. 

I will send the code to help you to compare your findings, but I really prefer that you try to replicate this exercise.

My kindest regards,

Lucia

Hi Aarti,

Before training the model normalize the output.

Hi

you should use normalized data, use the different trial and error to find the best number of membership function or you can check different FIS generation or different training algorithm to reach the best results for your problem

Dear Aliyeh and Isa,

I have tried normalizing the output as well as normalizing the data. Still I am not able to get the results. Would you like to suggest anything else?

 Dear Aarti

Today we will run your data in my AI class.

I will let you know.

Regards,

Lucia

Dear aarti

you must use ANFIS with genifs2 WITH NORMALIZED DATA.

if you have not the matlab code I can send you a matlab programm for ANFIS using GENFIS2

I am sure this is a good solution .

Good Luck

Good morning Aarti 

Yesterday we were running your data (in my AI class) and here we have the results:

As you may know, the anfis structure generated by matlab, uses clustering techniques or the initial generation of membership functions in the antecedent, and fixing these parameters uses a least squares methodology to obtain the consequent parameters. Nevertheless you should be aware of the nature of data by running first the clustering algorithms in order to determine a “good” set of parameters for training. I will lead you through several recommended steps:

1. Normalize the data without using the genfis/anfis commands, just normalize and take the data to the hypercube [0, 1] or  [-1, 1]

2. Select randomly a 60% of the data for training, 20% for test and 20% for validation VERY IMPORTANT. 

3. Run clustering algorithms over the whole set via subclust command in order to know the sensitivity of your data. 

You have just 524 patterns but 34 inputs and at least 3 outputs and some clustering algorithms such as subtractive provides 340 clusters with radii 0.5 or 31 clusters with radii 0.2 which means that your data is too sensitive to clustering algorithms. We like to run it several times looking for a nice parameter... but if you know too much the set you can semi-supervise the results.

4. Run fuzzy c means algorithm with the optimal number of clusters from subtractive and try to plot the 3-d projection of the centroids and clusters to verify if the results have any sense for you. I was just wondering that you work on power systems.

5. When you are satisfied with the number of clusters, use just the training set and run the genfis2 command and remember that you are going to build a MISO sugeno structure. Use the options menu to set a nice parameter for subtractive clustering!!!!

6. Also, try a genfis3 and remember to use the fuzzy c means right number of clusters as parameter and also try building a ‘mandami’ type of FIS (see the command line setup) because we find out a nice results with it. Also build a ’surgeon’ structure in order to compare.

7. Now we have 3 fis that reflects the training set with good accuracy, use evalfis command to evaluate antis results and see if it is ok with your research.

8. Try evaluating the test and validation set in order to check if the ANFIS can extrapolate the data

9. it is very important to realize that this structure is highly dependent of the number of patterns to properly converge to the right connections in the network. Try to use not too much clusters because you will be overtraining the net and will never reach the capability for extrapolation.

10. Finally you can train a 'sugeno’ type structure with anfis command but… it seems that genfis2 or  genfis3 can provide you a nice results. 

I will send the code to help you to compare your findings, but I really prefer that you try to replicate this exercise.

My kindest regards,

Lucia

Dear Aarti,

I've found in the literature that the size of training data should be 60 to 70 % of data sets. If only 3 samples are for testing purposes, fitting can be found by chance i.e. fitting is found for cases which do not represent given phenomenon. 

Secondly, a lot of inputs (34) plus multiple neurons in a hidden layer gives a lot of connections. Pls refers to the Kołomogorov theory and vc-dim criterion. They are shortened by practicers (of ANN) to the rule, that you should have 5 to 10 times more teaching data sets than connections between nuerons. 

Trimming paramiters of ANN I've got twice more with pretty good results. 

The next thing is number of inputs. I've checked which of them are strongly correlated with output (as a single predictors). I've left only these. Next (basing on litereture) I deleted these inputs which has been strongly correlated one to each other. I has much more improved the ratio (number of data sets to number of connections between neurons). 

So, only six from twelve input data has been left for calculation.

That's a part of my phd thesis. Unfortunately is written in Polish. After defence, I will create an article in English basing on that, but it will not be before beginning of 2018.

Best regards

Hubert

you can try substractive clustering instead of grid partitioning, thank you

Check the number of hiden neurons (Rules) , it is posible that they are too few