## A BP-Example with Values

### A BP-Example with Values

Hi,

I wrote a BP program and want to test it with the xor-problem. But it
seems, that the net needs much iterations for a good result.

Can s.o. please check, if my outputvalues are ok after 4000 iterations?!

I= Inputneuron
B= Bias
H= Hiddenneuron
O= Outputneuron

The xor-net with the inital weights:

I1 I2 B1
| \ / | ?
| \ / |0.4 ?/
0.1| \ / | ? /0.6
| \ /0.3 | ? /
| 0.2\ / | ?.5 /
| \ | /
| / \ ?| /
| / \? | /
| / ?\ | /
| / ? \ | /
| / ? \ |/
H1? H2 B2
\ 0.8/ ? 0.7\ / ? \ / ?.9
\ / ? \ / ? O?
Connectionweights:
I1 to H1= 0.1
I1 to H2= 0.2
I2 to H1= 0.3
I2 to H2= 0.4
B1 to H1= 0.5
B1 to H2= 0.6
H1 to O = 0.7
H2 to O = 0.8
B2 to O = 0.9

learnrate= 0.3
momentum = 0.9

Activationfunc: 1/ (1+ Exp(-1* PropagationResult))

After 4000 iterations:
I1 I2 O(must) O(is)
---+---+--------+------------------
0 0 0 0,0351244771951263
0 1 1 0,908663720173054
1 0 1 0,908485312804866
1 1 0 0,130843449323126

### A BP-Example with Values

OK, I'm pretty new to this group, and have only been working with
neural nets for a few months (in spare time), so forgive me if I am
misinformed about any of the following statements:

As far as I understand, XOR is not an easy problem, and often needs
a lot of iterations to achieve good results.

One thing I see there are some large initial weights. As far as I
remember, BP works best when the initial weights are small.
(I set mine to a random number with magnitude <= 0.25)

These results seem fine, the network is making the right decisions.

Geoff Holden
Computer Engineering Student
Memorial University of Newfoundland

--
Microsoft likes to buy the little guy.
The RIAA likes to sue the little guy.
SCO likes to sue the big guys to get bought by them