pem: Adding noise model *worsens* fit?

I'm fitting a first order dynamic model using the following:

>> m0=pem(z,'P1','disturbancemodel','none')
>> m1=pem(z,'P1','disturbancemodel','arma1')

With no disturbance model, m0 gives a better fit than m1, as determined by both the System ID Toolbox's compare function and by computing the residual sum of squares. This doesn't seem to make sense to me, since adding parameters should be able to better fit the data.

Can anyone make sense of this? Could be it simply an artifact of the optmization algorithm?

Scott

Answer 1

[Rajiv Singh]

It could be. But to isolate the issue, set 'focus' to 'simulation' in the call to PEM command since by default you are minimizing 1-step ahead prediction error (which is difference from simulation error for the second model owing to the noise model). COMPARE, on the other hand, computes the simulation error by default.