Large number of nodes in solution

[maxbennedich]

I'm wondering what the status is of this project; is it considered to be working as described in Ken Stanley's paper? The reason I'm asking is that when I run the "xor" experiment, the resulting network typically has around 20 hidden nodes, whereas in the paper it's stated that "NEAT very consistently uses 1 or 2 hidden nodes to build an XOR network".

Is this implementation based on / inspired by some other implementation, or implemented from scratch from the paper? Is there any other way to verify that it works as intended? Thanks!

Click to see an example of a network produced by the xor experiment

non-recurrent
24

1 4
5 -1.60176 2 1.9577 9 -0.0896985 2 0.668721 

1 0


2 8
0 -0.40958 1 -0.150364 3 -0.110997 4 1.60981 5 -1.48269 7 -1.8749 6 1.40913 18 -0.150364 

0 7
1 -0.0912658 4 -1.51859 0 -0.449885 5 -0.177615 6 -1.19628 14 1.27212 12 -1.86968 

0 4
0 -0.0975539 8 1.48903 7 -1.82679 12 1.1046 

0 4
7 1.29259 0 0.918139 8 -1.32898 22 0.918139 

0 3
5 -1.99141 15 -1.21041 14 -0.418487 

0 5
1 -0.673895 10 0.0527143 11 1.48281 17 -1.30425 0 -1.581 

0 3
0 0.998239 13 0.994958 1 -1.97802 

0 2
2 0.866348 7 -1.38062 

0 0


0 2
10 1.88048 17 1.12631 

0 1
1 0.64485 

0 2
10 1.02876 16 0.122552 

0 1
5 -0.67357 

0 0


0 2
10 1 11 1.73794 

0 0


0 2
1 0.985758 0 1.70312 

0 2
5 -1.51736 16 -0.815769 

0 1
0 0.486592 

0 1
4 1.47994 

0 1
0 1 

0 1
7 -0.535213 

[mcneb10]

I have also noticed this. I wrote a program that graphs the neural networks (I am working on a fork of this software cut down for microcontrollers), and I noticed most of the nodes (all except ~1/5) lead nowhere.

[mcneb10]

Here is the graph: link. I wasn't quite sure what I was doing when I made the program, and I still don't, so take this with a grain of salt. The graph/flow chart was made from "fit = 200". Something i noticed is that "fit = 152" and other were much simple (~3-5) nodes vs this one which has dozens.