Thread Subject: newrb prediction problem

Hi all,

I've been using different networks to predict data and the newlin and newff networks have both produced good results. I am tryin to use newrb / newrbe now to see how accurate these networks can be. The newrb network has learned the training data almost perfectly, however when i present new data for prediction, the prediction is always just a straight horizontal line (the data should look similar to a fuzzy sinwave).
I've used the following code:

P = input_data
T = target_ data

net = newrb(P,T, GOAL,SPREAD,MN,DF)
train_result = sim(net, P)

i then compared T to the train_result and they were practically the same. I then used new data, X and Y, which have a similar shape to P and T

result = sim(net,X)

but when i plotted result it was just a straight horizontal line which bears no resemblence to the data in T or Y.

I have tried varying the values of spread and MN but the prediction using X as an input remains a straight line. Have i left out an important line of code does anyone know?

Thanks a lot,
Stephen

On Oct 12, 11:08 am, "Stephen Dunne" <dunn...@tcd.ie> wrote:
> Hi all,
>
> I've been using different networks to predict data and the newlin and newff networks have both produced good results. I am tryin to use newrb / newrbe now to see how accurate these networks can be. The newrb network has learned the training data almost perfectly, however when i present new data for prediction, the prediction is always just a straight horizontal line (the data should look similar to a fuzzy sinwave).
> I've used the following code:
>
> P = input_data
> T = target_ data
>
> net = newrb(P,T, GOAL,SPREAD,MN,DF)
> train_result = sim(net, P)
>
> i then compared T to the train_result and they were practically the same. I then used new data, X and Y, which have a similar shape to P and T
>
> result = sim(net,X)
>
> but when i plotted result it was just a straight horizontal line which bears no resemblence to the data in T or Y.
>
> I have tried varying the values of spread and MN but the prediction using X as an input remains a straight line. Have i left out an important line of code does anyone know?

What are the dimensions of P and T?
Other info?

Greg

Greg <heath@alumni.brown.edu> wrote in message <bb83f380-ec29-42a0-be17-807c85226e7e@33g2000vbe.googlegroups.com>...
> On Oct 12, 11:08?am, "Stephen Dunne" <dunn...@tcd.ie> wrote:
> > Hi all,
> >
> > I've been using different networks to predict data and the newlin and newff networks have both produced good results. I am tryin to use newrb / newrbe now to see how accurate these networks can be. The newrb network has learned the training data almost perfectly, however when i present new data for prediction, the prediction is always just a straight horizontal line (the data should look similar to a fuzzy sinwave).
> > I've used the following code:
> >
> > P = input_data
> > T = target_ data
> >
> > net = newrb(P,T, GOAL,SPREAD,MN,DF)
> > train_result = sim(net, P)
> >
> > i then compared T to the train_result and they were practically the same. I then used new data, X and Y, which have a similar shape to P and T
> >
> > result = sim(net,X)
> >
> > but when i plotted result it was just a straight horizontal line which bears no resemblence to the data in T or Y.
> >
> > I have tried varying the values of spread and MN but the prediction using X as an input remains a straight line. Have i left out an important line of code does anyone know?
>
> What are the dimensions of P and T?
> Other info?
>
> Greg

Hi Greg,
 
P & T used for training:
P is a <5 x 264> double
T is a <1 x 264> double

X & Y used for predicting new data:
X is a <5 x 24> double
Y is a <1 x 24> double

The thing is I used these exact variables with the newff and newlin networks. Aswell as that I have now used this data with the newgrnn function (P, T, X and Y already created):

net = newgrnn(P, T, 100);
result = sim(net, X);

and it produces a very good prediction. Is there some inherent feature that means the 'newrb' network just won't produce a proper prediction?

Thanks again,
Stephen

On Oct 14, 6:50 am, "Stephen Dunne" <dunn...@tcd.ie> wrote:
> Greg <he...@alumni.brown.edu> wrote in message <bb83f380-ec29-42a0-be17-807c85226...@33g2000vbe.googlegroups.com>...
> > On Oct 12, 11:08?am, "StephenDunne" <dunn...@tcd.ie> wrote:
> > > Hi all,
>
> > > I've been using different networks to predict data and the newlin and newff networks have both produced good results. I am tryin to use newrb / newrbe now to see how accurate these networks can be. The newrb network has learned the training data almost perfectly, however when i present new data for prediction, the prediction is always just a straight horizontal line (the data should look similar to a fuzzy sinwave).
> > > I've used the following code:
>
> > > P = input_data
> > > T = target_ data
>
> > > net = newrb(P,T, GOAL,SPREAD,MN,DF)
> > > train_result = sim(net, P)
>
> > > i then compared T to the train_result and they were practically the same. I then used new data, X and Y, which have a similar shape to P and T
>
> > > result = sim(net,X)
>
> > > but when i plotted result it was just a straight horizontal line which bears no resemblence to the data in T or Y.
>
> > > I have tried varying the values of spread and MN but the prediction using X as an input remains a straight line. Have i left out an important line of code does anyone know?
>
> > What are the dimensions of P and T?
> > Other info?
>
> > Greg
>
> Hi Greg,
>
> P & T used for training:
> P is a <5 x 264> double
> T is a <1 x 264> double
>
> X & Y used for predicting new data:
> X is a <5 x 24> double
> Y is a <1 x 24> double
>
> The thing is I used these exact variables with the newff and newlin networks. Aswell as that I have now used this data with the newgrnn function (P, T, X and Y already created):
>
> net = newgrnn(P, T, 100);
> result = sim(net, X);
>
> and it produces a very good prediction. Is there some inherent feature that means the 'newrb' network just won't produce a proper prediction?
>
> Thanks again,Stephen- Hide quoted text -
>
> - Show quoted text -

I missed this when it was first posted. The most prevalent
reason for newrb failure is the creation of too many
cluster centers because

1. SSEgoal is too low
2. Spread is too low or too high

What values did you specify?
How many clusters were generated?
What was the final R^2 = 1-SSE/SSE00?

Hope this helps.

Greg