Understanding MLP for numeric prediction
by Bashyam, Navaneeth (CSI)

WEKA list,
I want to understand how MLP can be used for numeric prediction. I
started with a known function z=x+y; where x and y take values between
0.0 and 1.0 .
I create 5000 data points of different xi's, yi's and zi's.
I set up a MLP with 1 node forming the hidden layer. The simple MLP
looks like >-0-
Weka default MLP gives the following output.
Sigmoid Node:
_______________________
Threshold: 0.87709507261667
Weights for x : -0.893661993055382
Weights for y : -0.894017186952782
Linear Node:
_________________________
Weights : -4.91766646527907
Threshold: 2.47073476737811
How do I construct prediction model from the above information.
I tried the following. let us say training set is xt, yt and zp is
predicted value of z.
zp=
TANH((xt*(-0.893661993055382)+ yt *(-0.894017186952782)+
0.87709507261667))*(-4.91766646527907)
+ 2.47073476737811
The zp's do not match good with z= xt + yt. I am sure I am miising
something here. Because WEKA reports a very high correlation coefficient
for the above MLP run. I would appreciate if someone explains the
discrepancy.
Thanks,
Navaneeth
Navaneeth Bashyam
Senior Product Developer
Closure Systems International
Office: 765 364 7358
Fax : 765 364 5678
navaneeth.bashyam(a)csiclosures.com
The Mind once stretched by an idea, never regains its original
dimensions.