[ODE] Creature Evolution input

[skjold@cistron.nl]

> A purely vector based input would amount to around
> 200 inputs.
>
> Any thoughts on that?

If you mean to say that those 200 inputs are all used just to represent vectors (i.e. excluding any other sensory data you have for input), then it sounds like a lot. It's not necessarily a problem to have many inputs, but it really pays to carefully select sensory data representations that make the most sense to the task that your ANN is supposed to learn. That could be the primary cause for your ANN's to have difficulty learning.

To reiterate from my previous post, keep in mind in what way the ANN will control your creature. Vectors as input might do fine, I don't know. But I suspect that if your ANN outputs e.g. desired joint angles, then you only add unnecessary complexity to the problem if you input vectors rather than current joint angles, because then the ANN will first have to learn how to correlate those two radically different representations of the same information. Depending on the activation function and architecture you use, this correlation could even be impossible for the ANN to learn.

Additionally, if your creatures have complex physiologies, you might want to consider dividing the ANN into multiple smaller (hierarchically organized) ANN's that are each dedicated to controlling just one part of the creature (just like MIT's Attila or Karl Sims' block creatures). For example a simple hexapod could be controlled by seven little ANNs that are interconnected (one for each leg and one to coordinate the effort - the latter actually being optional).

Just some thoughts.
Greets,
Mark


> Thanks everybody :) - I'll combine the overall thought patterns and see what
> I'll come-up with...
> 
> the problem is a bit about representation: I simply feel that the ANN is
> very comfortable with a whole lot vectors, and that the evolution based
> purely on vectors would take very long. I've worked out a useful alternative
> representation - but it explodes the total required input to about 1500
> inputs (most of which would be zero... but still). A purely vector based
> input would amount to around 200 inputs.
> 
> Any thoughts on that?
> 
> _______________________________________________
> ODE mailing list
> ODE@q12.org
> http://q12.org/mailman/listinfo/ode
>