i will train neural network off line with dataset ,the performance is very good, but when i use the neural network block in model Simulink don't give better results, the type of solver and the does affect in the results ? note that the model Simulink works in discrete time
Hi Rehouma Ridha
you leave a series of items open, which will make it hard to impossible to give you a detailed answer (type of problem, type of neural network, size of training data, size of validation data, etc. ... )
All together what you describe seems to be the classical situation, where the network (or any other machine learning system) has adapted to the training data, thus generating a low error during training, but performs bad on other data, having a bad generalization.
A possible cause might be, that the network is doing "overfitting", or "learning by heart".
A second possible cause, is that your training data is not representative for the application (the one you use in Simulink).
Here are some questions that might lead you to a solution:
Have you tried to use a smaller neural network? less hidden neurons, thus less weights, thus fewer parameters to learn by heart?
Have you divided your data into an explicit training-data-set and an explicit validation-data-set?
What happens to the performance of the network on the validation-data-set during training?
Have you thought about cross-validation?
I hope some of these thoughts will help you.
Feel free to ask more detailed questions, and/or get in direct contact.
Regards
Nils Goerke
- Badges
- Science topic
- Similar topics
- Chemistry
- Nanotechnology
- Nanostructures
- Thin Films