The training set is 350 and test data-set is 150. 100 or 1000 iterations? Is the training set large enough to go 1000 iterations and avoid over-fitting.
There is no optimal number of iterations. You should iterate until the error does not significantly decrease. To avoid overfitting, you should split the data set in training and test, and the error should be similar for both subsets.
The information given is insufficient for answering the question-especially since it isn't clear what overfitting here means and what's the figure of merit-that's what will decide: the residual error on the training set and the residual error on the test set are the typical quantities to monitor-assuming it's known that the network can learn the sets, of course.
Thanks Soutrik...And yes, a single hidden layer works. It outperformed ANN models with more than one hiden layer.
no specific iteration point, however not the difference of error rate during training and testing...........to determine if the model over-fits the situation.
Actually, I did not see a clear formula for setting the optimum iteration number, but you can figure out this issue by an iterative process by initializing the iteration number by a small number like 100 and then increase it linearly. This process will be repeated until the MSE of the test does not decrease and even may increase. The below link describes well:
https://www.quora.com/What-will-happen-if-I-train-my-neural-networks-with-too-much-iteration
- Badges
- Science topic