Hi Afaq,
The best stopping criteria is application dependent. For instance:
- If your application cares about the speed of the training process, you might want to choose 'epochs: Maximum Number of Training Epochs (Iterations)' as your stopping criteria and set it to a low number.
- If the application cares about the accuracy of the training, 'max_fail: Maximum Number of Validation Increases' would be a good choice.
- 'min_grad: Minimum Gradient Magnitude and goal: Minimum Performance Value' can be reasonable choices if the efficiency of the algorithm is of importance.
Finally, if you are not sure what is the best stopping criteria for your application, you can simply try them all and compare their performances and then choose the one that best fits your application.
I hope the above information helps.
Ghada
