[ODE] Creature Evolution input

[Gene Ruebsamen]

Hello,

Been a while since I've had a chance to check out the ODE mailing list, and I'm
glad to see there is still alot of interest in evolving robots/virtual creatures.

I haven't really been keeping up with the latest techniques; however, for my
simulations, I determined that keeping the number of inputs to the ANN
reasonable is very important.  Reasonableness of inputs is hard to define, but
can be thought of as having two components.  

(1) The number of inputs into the ANN should be as few as possible without
resorting to combining multiple inputs into one. (ie. you dont want hundreds or
even thousands of inputs going into an ANN.  If the data on the inputs are all
very similar, it may make it difficult for the ANN to determine what inputs
should correspond with what outputs).

(2) Inputs that clearly represent one piece of data.
A touch sensor should represent a 1.0 for the body touching something, or a -1.0
if it is not.  (assuming you are using bipolar representation.. more on this
later).  Angle sensors should represent either the angle of a joint, or how many
degrees off center they are if you are trying to follow a target.  The important
thing is try not to combine multiple inputs into one just for the sake of saving
an input. 

My best results were, not surprisingly, when I used fairly simple creature
morphologies with fewer ANN inputs.  Typically, if you can find a way to
constrain one DOF of a joint w/o adversly affecting the movement of the
creature, then do it, as the ANN will have a much easier time learning how to
manipulate a hinge or a universal joint effectively, rather than a ball joint.

Furthermore, you certainly should consider using an ANN with some type of
feedback.  For my experiments, I used a modified Elman type network with a
bipolar data representation.  I found that a bipolar representation speeds
learning within the ANN.  

You can take a look at some of my results on the ODE community page under the
"evolving robots" link.  I hope to have my thesis & sources online soon.  Now
that I have a full time job, time is a precious commodity that I have too little
of =(.

Anyhow, good luck with your simulations, and please let us know how they turn out!

Gene Ruebsamen
http://www.erachampion.com/ai/

Quoting Henri Hakl <henri@cs.sun.ac.za>:

> Hi :)
> 
> I'm doing the ever popular evolving of AI control over an articulated model -
> the question is what are suitable inputs to the Neural Network? (For each
> body-segment input: position, relative position, velocity, force,
> orientation, other stuff...?)
> 
> thanks
>   Henri
> 

-------------------------------------------------
This mail sent through IMP: http://horde.org/imp/