Overfitting of Regression Plot for a Feedforward Neural Network

I am trying to make an estimation for the state-of-charge of a battery using the Feedforward Network. My inputs to the network are current, voltage and time. The network has 3 inputs (Voltage, Current and Time) with 1 Output (State-of-charge(SOC)). The network consists of 2 hidden layers with 10 neurons each and output layer with 1 neuron. Each Input has 54043 samples, so has the output.

I am training the network using the 'trainlm' and have set the performance goal to 1e-5. However, after training the network, the regression plot seems to have a R of 0.9999 in all the cases, which indicate to me that the network I am using is Overfitting the target.

Attched is the code and mat file containing the measurement data, any leads would be appreciated.

Below is the obtained regression plot:

Answer 1

[Greg Heath]

OVERFITTING IS NOT "THE" PROBLEM !!!

In general, the problem is

OVERTRAINING an overfit net.

My solution is simple: DO NOT OVERFIT !!!

Use no less training equations Ntrneq than the number of unknown weights Nw, i.e.

Ntrneq >= Nw

For a double hidden layer I-H1-H2-O = 3-10-10-1 net,

Ntrneq = Ntrn*O = 0.7 * 54043 * 1 = 37,830

Nw = ( I + 1 )*H1 + (H1+1)*H2 + (H2+1)*O

= 4*10 + 11*10 + 11*1

= 161

OBVIOUSLY, YOU HAVE SO MUCH DATA THAT YOU COULD HAVE EVEN USED MANY, MANY MORE HIDDEN NODES!

HOPE THIS HELPS.

Thank you for formally accepting my answer.

Greg

Greg

[Zeeshan Ahmad Khan]

Thank you so much Greg, your insight helped me solve the problem.