Hello,
First all thanks for the explanation and the code showing was really helpful.
I had been working with Neuralnetworks for a while at Matlab Software and now i am working in projet to develop a Neural Network in a micro controller to give a product capability to learn and adapt to client use.
As the micro controller has lot of memory and processing limitations i am using a simple feedfoward NN, 1 Input, 1 Hidden Layer and 1 Output.
The learning process is going good and it is able to find to each pair I/O a MSE < Threshold. When i save the weights and the Bias associated and go for the use of the trained NN it all goes wrong. I am using a simple Input-Output Curve Fitting so the Output is simple the Summation from the Outputs from the Hidden Layers.
My main problem comes from the fact my input data goes from (0..48) simple times steps each one representing minutes. But my output must be for example 0 for a input but in the next step it will be 77 and it will stay in 77 for some time steps and then another big step from 77 to 150, so this step the Network isn't able to do it and it goes linear and starts incrementing in the good direction some steps before. May you suggest me a different output function or method? Or different way to represent my data as targets or inputs.
Smile | :) Wink | ;) Thumbs Up | :thumbsup:
Ferreira
Situation Example: This is my Targets during a specific time:
So what's happening it's that the Neural network isn't doing the "Steps" but being incremented proportional and it will end in no error, but i do want to have the steps so i may use that information in a Controller.
Hope you had understand my problem and may help me. Thank you.
