Poor Neural Network Performance

My Neural Network is giving me unsatisfactory results. The regression index comes out around 0.2ish all the time. I've tried initializing weights and changing the number of hidden neurons but they don't seem to work.

My project is about making a prediction program which predicts blood pressure from variables such as weight, age, etc. My input matrix is a 4x213 matrix of age, cholesterol level, sugar level and ECG value (I have normalized all the values to be between 0 and 1). My output matrix is a 1x213 matrix of blood pressure values. For the hidden neuron activation function, I've tried using logsin and tansin function, and for the output function I will have to use the purelin function since my output will be blood pressure values (typically between 60 to 150).

Below are some of the properties of my neural network:

- Network Type: Feed-forward backprop

- Training Function: TRAINLM

- Adaptive Learning Function: LEARNGDM

- Performance Function: MSEREG

- Epochs: 1000 - Goal: 0 (I also tried using 1e-06)

- Max-fail: 6 (I also tried other values such as 100 and 500)

This dataset has been used in other papers, so I don't think there is anything wrong with the dataset.

[Greg Heath]

Regression index = NMSE = mse(error)/var(target,1) ?

What do your 4 1-D plots of target vs input tell you?

What are NMSE for your 4 1-D NN models?

Greg

[In Kyu Sung]

These are the typical regression plots I get. Honestly, I do not understand what you mean by NMSE = mse(error)/var(target,1), since I am too new into this.

i am sure the plot is telling me about how my model predicts my input value. The ideal should be like the dotted line, however my regression plot does not seem anywhere near this.

Regards

[Greg Heath]

From the looks of your scatter plots, it seems that the data doesn't support an extremely accurate model. That is why I told you to look at the four separate 1-D models (y(x1), y(x2), ...)and plots to see how much you can get out of each separate input. It may or may not help your answer, but it should give you a better feel for why you are getting your current results.

Hope this helps,

Greg