Currently i am using the dataset (https://www.researchgate.net/publication/281449157_USDvINR_rate) to predict the future rate of INR against USD. The current training and testing mean square error are in the range of 0.0201 to 0.0235. The training set uses 5670 and test set uses 1170 random records from the dataset. What can be done to reduce the MSE further (approx 0.0001 to 0.0005)
For further info about the implementation refer to - https://www.researchgate.net/publication/280698216_FOREX_rate_prediction_using_a_Hybrid_System
Data USDvINR rate
Article FOREX rate prediction using a Hybrid System
Dear D'Lima,
You can try the following strategies:
- Reduce the number of input variables;
- Change the number of neurons in the hidden layer;
- Change the activation function of neurons in the hidden layer;
- Increase the number of training samples.
Regards,
Francisco R. Lima Jr
My current anfis system uses 2 inputs, 8 MFs per input, and it runs for 2000 epochs
I suggest increase the number of epochs and decrease the number of membership functions of the input variables.
number of epochs have to be increased. the initial selection of parameters in the forward and backward pass has to be fine tuned
Nathan,
You can do a lot of things. But, remember that lower error (especially on the training data) does not ALWAYS mean that your system performs better since ANFIS may become only a MEMORY remembering the training data behavior. In this case, ANFIS performance on the testing data may not be as much good. Keep in mind that you need to experiment on many factors till you obtain satisfactory results.
Here is what you can do.
1. Try different MF although they tend not to change the results much. If this is the case, I would stick to the Gaussian MF as it gives least number of adaptable parameters.
2. Try different number of MF per input. The rule of thumb is that the higher the number of MF per input, the better the results. However, you may find different behavior for your ANFIS depending on the nature of your training data.
3. Try different Sugeno models (linear vs. constant) although the linear model tends to give better results usually.
4. Increase the number of training epochs as long as the error decreases. Stop training once the error stagnates. To reduce the error further, you need to adjust the design of your ANFIS as more training may not help.
5. Change the number of data points and selection criteria in your training/testing datasets.
It is a tedious process of various experiments, but hopefully, one of these suggestions may help you. Good luck.
Mohamed
- Badges
- Science topic