I am using the Matlab Fuzzy Inference toolbox to build an adaptive neuro-fuzzy inference system. I first use genfis3 to generate a FIS system and do the clustering, then use anfis to further calibrate the FIS system. The number of clusters N and exponent for the partition matrix U are randomly generated with minimal value of 1 for both variables. The result of genfis3 is shown in fig 1, and the result of anfis is shown in fig 2. Then I use another class of data to validate it, but an error was occured:
"Error using evalfismex
Illegal parameters in fisGaussianMF() --> sigma = 0
Error in evalfis (line 84)
[output,IRR,ORR,ARR] = evalfismex(input, fis, numofpoints);"
Please advise me if anyone knows what's going on with this problem.
In a Gaussian function, a sigma value of zero in the denominator probably leads the overall result to be NaN. By the way, this is just a guess.
try the attached function .
add function to your Matlab path and set path for it.
do not run on 'auto' clustering mode, if your data is very large. it's very time consuming.
see the link for more information.
http://www.researchgate.net/publication/259912006_Application_of_Fuzzy_Modeling_and_Optimization_in_Enzymatic_Esterification_Process
Article Application of Fuzzy Modeling and Optimization in Enzymatic ...
Hi, in Matlab "NaN - Not a Number" generally appear when is not possible to represent the value, check the component of objective function.
Check the code…
Examples of operations that produce a Not-a-Number result are given at
https://nf.nci.org.au/facilities/software/Matlab/techdoc/ref/nan.html
In C-means (k-means) clustering method, it is possible that a cluster is empty (in some iteration of the method) and you are dividing by number of points in the cluster to get the centroid. Since, number of points in the cluster is zero, you are getting NaN output.
If the method uses FCM, the ranges for your parameters are wrong. First, the number of clusters should be > 1, not >=1. Second, the fuzzifier exponent should be >1. Taking 1 induces distances to be at the power 2/(1-1) which is a NaN.
- Badges
- Science topic