Search All of the Math Forum:
Views expressed in these public forums are not endorsed by
NCTM or The Math Forum.



Re: neural network constrain?
Posted:
Jan 24, 2013 11:47 AM


"Apple" wrote in message <kdmcql$t3v$1@newscl01ah.mathworks.com>... > I try to have a time series nn with 2 outputs V1 and V2 and predict the next pair of values. the simplified medel can be explained as (I don't know how to upload the diagram): Current i1 and i3 in parallel, i1 comes into a R and i3 to a nonlinear load; then they merge into single point, then split into two: i2 goes to R and i4 goes to nonlinear load, then to ground. All current can't be described by equation and need NN is useful. > > As we can see, there are two constrains: > (1) V2>0 > (2) V1>V2 > > The NN model provide good training, but the outputs have V2<0 and V1<v2 values. > > Is it possible to set constrain to this time series NN?
I gather that your training inputs are i1 and 12, your training targets are V1 > V2 > 0 and you wish to predict V1 and V2 for s timesteps ahead.
Try narxnet with 'logsig' outputs for V2 and V1V2 that are normalized to the interval [0,1].
Hope this helps.
Greg
It seems to me that > > Apple



