Estimating conditional probabilities

[Alex van der Spek]

Hi,

I need to estimate conditional probabilities, that is probabilities of
several types of behaviour of a machine only when that machine is running
(ON). The state of machine (ON or OFF) is determined first by a neural net.
I know how to do that.

What I could do is to first establish the ON/OFF state using a simple
backprop network with one logistic output function. Next when the state is
ON, pass the same input data to a second neural net with 1-of-C softmax
output units. Each output category corresponds to a particualr class of
machine behaviour. Whereas if the state is OFF, no further action is taken.
Thus the probability of each output category (machine behaviour type) is
conditional on the machine being ON or OFF. These machines are running 24/7
and most of the time they are ON.

This appears to be sub-optimal, especially since the number of inputs is
rather large (62 inputs).

Would anyone know how to devise a backprop NN to do both in one, i.e. both
the ON/OFF status as well as the C different types of behaviour.

Thanks in advance,
Alex van der Spek

[Greg Heath]

You can just add another softmax output with unipolar binary target
{0,1} that
indicates the machine being on (target = 0) or off (target = 1).

Hope this helps.

Greg

[Stephen Wolstenholme]

Why not use one NN with a boolean output for ON/OFF along with the
softmax outputs?

Steve

--
Neural Planner Software Ltd www.NPSL1.com

[Alex van der Spek]

Yes, this does help!

Very elegant and simple, just code ON=0 and OFF=1, instead of the other way
around!

Thanks much!
Alex

"Greg Heath" <he...@alumni.brown.edu> wrote in message
news:09af15e8-2213-4946...@l13g2000yqb.googlegroups.com...