This is my 4th post today, and so far I've seen nobody but me post. Oh well,
I'll make it my goal to get this forum active again.
I've been fiddling around with designs for new creatures, and thinking about
how to design neural networks to get the desired behavior. I was thinking
about the problem people keep describing with the smell sensors, and how
best to approach the issue. Part of the difficulty, as I understand it, is
that the difference in the values of two sensors on a creature is too tiny
in proportion to the overall signal to be used effectively to direct
movement.
The easy fix for this aspect at least is to not use the smell values
themselves, but the difference between them.
Designing a network to do this is fairly simple, and here is the simplest
possible version in f0 code:
X[*:0.5][*:0.4][-2:1,-1:-1]
for the sake of simplicity I'm using hardwired neuron values, but the
process works just as well if they are any two neurons. The third neuron may
still have a very tiny value, but that can be corrected with a higher weight
when other neurons use it as an input. Of course, finding a single weight
that works at all distances from the energy source is a different matter.
It seems to me that you could get the most out of your training time if your
network is set up from the begining to provide it's central brain (which I'm
assuming for the purpose of this discussion you'll be evolving rather than
designing) with inputs containing the actual values it should utilize,
rather than hoping it will simultaneously evolve to correctly process the
sensory data and to use that processed data to make the correct decisions.
I think it would be a great help if there were a repository not of
creatures, but of neural-net functions, like my example above, or the
standard sinoid curve networks.
Intrest/reactions/comments? or even better, anyone have any neural net
functions of their own to throw out?
Will Thomas