I used NN as a predictor, i need to compansate the error of output and the time delay of the output.
What do you mean by "time delay of the output", processing time of the network?
Maybe he is using a recurrent NN, feeding the previous output to the new input.
Well for correcting the error you might need to change the values of learning rate, momentum or even add a few more neurons.
It depents... you have to try different values and keep the one that gives you the lowest error...
Making a NN behave ok its trial and error most of the times...
You have to check the testing and training error if possible at runtime, and when you see that test error is increasing while train error decreases that meen that your NN overtrains. So you have to decrease epochs, learning rate, monentum. Atherwise you will might have to increase one or some of these parameters.
You must check cross-validation 372-375 pp. Neural Networks for Pattern Recognition Christopher M. Bishop. Stop the training when you see test error is increasing.
There are several things that affect the error, the initial process (transformation), the selection of input, number of neurons, learning algorithm, learning rate and even activation function. We have to choose correct combination between input and number of neurons in hidden layer for the best NN. I agree with Mr Christos Kannas.
In my opinion, you should use historical inputs if you are using RNN. Only then output of the RNN is will match the target data.
Try including the current input in your input matrix.
nn input (k) = [input (k); input(k-1); input(k-2); ... ; output(k-1); output(k-2); ... ]
nn target(k) = [output(k)]
You can do it by several ways. Either by changing the internal structure like number of neurons and layers and epoch etc. or by using optimization algorithms. I am sure that the thesis of my student which was about the sensitivity analysis on ANN can be useful for you.
Article A Sensitivity Analysis on Application of Artificial Neural N...
- Badges
- Science topic