I used ANFIS to predict my experimental data. I used the MATLAB ANFIS toolbox.
The number of my data is few and I trained ANFIS by 24 data sets and tested with the 4 data set. The Gaussian functions are used to describe the membership degree of these inputs.
The number of membership functions were decided by the application of sub-clustering algorithm in which each data point belongs to a cluster at a degree specified by a membership grade.
The results show that there is a large gap between error in training and test data. [ RMSE and std for train data are 0.00037575 and 0.00038383, respectively, but for test data these are 0.35715 and 0.37864] . It seems that there is an overfitting in trained ANFIS. What is your idea to solve this problem?
The size of your data set is not the only factor affecting the results. The method you use to select your data is also important. Regarding the non-linearity of your problem, improper selection of training data such as clustered points adversely affect the training process. Your training points should present the linear or nonlinear behavior of the output function(s).
Your training data set is small. You have to check the data with annova for fit and desirability. If your model has error, that error will be propagated in your ANFIS model also. Pl check your data set
Hi, the most common causes of bad quality of a prediction model (in general) are:
- the data quality is not sufficient,
- the size of your data set is small,
- inadequate model type or model structure was chosen,
- the training, validation and test data sets are chosen not optimal.
So, you can either try to use other model structures or/and to try a new separating into training and testing data sets.
Dear Ali,
I would try to increase the size of the dataset, as well as checking whether it is representative enough for all the input parameter conditions. Try to randomly split the dataset intro training and verification set a few times and see whether the prediction error reduces.
Good luck!
Luiza
When that hapends, It is suggested to increase the number of hiden ellements (rules)
- Badges
- Science topic