Dear sarvesh

series 1 is actual target and series 2 is for simulated value by network. The  left graph is for training set and right side graph is for test set. here I found error in test data set how can I minimize ??

Dear Lila,

I think that, you have some significant deviation between your training data sample and the testing data sample. That might be the reason for you to end up with a solution like that. That is also a problem with the generalization of your ANN.

Regards.

Dear Lila,

It appears that you are overfitting the training data set. This is a very common problem when training neural networks, but more in general in machine learning. How to prevent overfitting is a large area of investigation. I would recommend you looking more in detail in machine learning and neural network literature. Common methods are called regularisation, early stop of training, reducing the number degree of freedom of the model, i.e. number of parameters, increasing the size of the data-set. 

You have not posted something about the architecture of the network. it would be interesting to see more details both about architecture and about your setting in general. At first glace, it seems to be an overfitting issue. 

Cheers!!

Hi Lila,

You got expected results. If you correctly design ANN scheme you should alway get training set error minimal as possible. On other hand test set error would be higher compare to training set error. That's normal. Keep in mind when designing ANN model:

1. Overestimation

2. Underestimation

3. What criteria do you use compare the error - MSE, RMSE, R - cross validation coefficient.

4. Golden rule keep it simple most of the problems solved by one layer and few nodes.

There are tons of publications on the web how to train and test ANN models.  

Good luck,

Hi Lila,

I think the problem is in the selection of the training sample. Using very few samples may cause this problem. Samples should be enough for the ANN to be capable of right generalization.

Good luck