I am using anfis for forecasting. The first problem is different runs ends up different results. fitness function is R^2 in test data. I have split my dataset in 70-30 range for train and test data.
For example one run results in 99% and another run results in 90%. even difference is in decimals. But the problem is the forecasting outputs for future data has massive difference. In a run with 99% result, output would be for instance about 7000. Accordingly another run with 99.99% result will output about 3000. How could I know which is the correct forecast. Also how to avoid instability in Anfis training results.
\
Dear Amirmahmood,
Here are links and attached files in topics.
- Business Intelligence and Performance Management: Theory, Systems ...
https://books.google.dz/books?isbn=1447148665
- Neuro-Adaptive Learning and ANFIS - MATLAB & Simulink - MathWorks
https://www.mathworks.com/.../neuro-adaptive-learning-and-anfis....
- Identification using ANFIS with intelligent hybrid stable learning ...
link.springer.com/article/10.1007/s00521-007-0168-9
Best regards
Some possible solutions are:
1. Train network until RMSE is stable (no of epoch).
2. Normalize your data if there is big difference in values.
3. Experiment with different membership functions.
4. No of membership function also effect output.
Dear zeeshan,
1. I will check this.
2. Dataset is normalized before training procedure.
3. as I am using fuzzy c-means gaussian is chosen which is experimentally suitable.
4. I have checked this . Still same
i want to train ANFIS with 12 input and 1 output
my dataset contain 3000 examples,the probleme is when i used genfis2 with different radii ,the error does not diminish and it remain 5.9.What can I do
- Badges
- Science topic