how to improve performance of a neural network model

I used neural network toolbox to predict some data. I have both inputs (23*3043) and outputs (6*3043). I have tried every train function and reset every parameter of the toolbox, but the performance was bad. After weeks' trial, I think there are no problems about toolbox itself and my data.

Usually my model did iterations just for 10 or 20 times and stopped because of validation checks reaches to maximum (I raised it to 15 but it did not work). I thought maybe it is because of local optimization. Does anyone know how to add functions in the neural network codes (generated by neural network toolbox) so that i can get global optimization? I used trainlm as train function and mse as performance function.

What I mean is that how to add functions in the neural network codes so that they can be used with trainlm together to get global optimization of mse? If it works, I think I will get better performance of the nn model.

The following is my performance result. They are not good obviously although R is high.

Answer 1

[Greg Heath]

From the plots it seems that trend is captured and you probably can't do much better.

In order to test that hypothesis try using all of the data for training and increase the number of hidden nodes until performance stabilizes.

Also, from the plots, it seems that there may be some merit in using 3 separate nets and combining the results.

Hope this helps.

Thank you for formally accepting my answer

Greg

[jiabo he]

Thank you so much for your reply. I did not cancel test data because I think it is important. I cancelled validation data and use 80% of data as training and 20% as testing. The result is a little better but still not very accurate. I also tried to use 3 separate nets but the results are not good. Thank you again for your help.