[ODE] Creature Evolution input

[Ian Macinnes]

Hi Henri,

Like (I think) quite a few people on this list, I evolve agent physical
and controller morphologies using ODE for movement and beacon following.
I've managed to evolve agents using a wide variety of movements made
from blocks connected by powered hinge joints, and the only inputs to
the network I use is the angle of the joints at a each timestep, which
adds a helpful negative feedback for the pattern generation - that's if
I'm evolving for movement only.

If you're just trying to get movement, you can get away with no real
inputs at all, and just use a sine wave input and allow the neural
network to modulate that!

For beacon following, I use two types of sensor input. I use directional
beacons. For these, the input to the neuron is the angle between a 
vector normal to some genetically determined position on a block, and a
vector pointing to the the beacon, this works fine.

The other type is a more realistic camera sensor, for which I use OpenGL
and off-screen rendering, but this is very slow. I've evolved an agent
that can visually locate and approach a block of a certain colour. This 
is a sensor with a directional field of view, and the neuron input is
the percentage that a certain colour fills it. I'm trying to work out a
faster way of doing this though!

So you can get interesting behaviour with very simple sensors!

Actually, I've be interested to find out what other people have done re:
evolved morphologies/controllers with ODE...?


Ian


On Wed, 12 Mar 2003, Henri Hakl wrote:

> 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
> 

---------------------------------------------------------
Ian Macinnes         01273 872952 / ianma@cogs.susx.ac.uk

Centre for Computational Neuroscience and Robotics (CCNR)
School of Biological Sciences
University of Sussex
Brighton, U.K.
BN1 9QG
---------------------------------------------------------