Hi Ken,

Thanks for the clarification.

As a follow-up, do you have any suggestions on how to handle mixed geometric and
non-geometric data?

For instance, if you're evolving a bot for a FPS, you may have some geometric
representation of the world. However, there may also be other relevant
information like health or ammo. It seems like there are only three solutions
right now:

1. Use regular NEAT and lose the regularity of the world.
2. Use HyperNEAT and force the health and ammo information to be represented as
extra dimensions even though they can be handled by a single input node.
3. Evolve a HyperNEAT agent separately then use its input(s) as secondary
input(s) to a regular NEAT controller.

Is there a fourth approach? Or is there generally a preference for one of the
above three approaches?

Wesley

Quoting Kenneth Stanley <kstanley@...>:

> Wesley, the main insight motivating r(x) is that it aligns the agent geometry
> with the team geometry. That is, the left side of the agent is on the same
> side of the agent as the left side of the team, which is how it usually would
> be in the real world. Nevertheless, you make a good point that the substrate
> could be 3D as an alternative. There is nothing wrong with doing it that way
> and it does have its own advantages. In fact, we have experimented with both
> approaches to laying out agents for multiagent learning (although only r(x)
> is in the GECCO paper), and both are valid.
>
> ken
>
> --- In neat@yahoogroups.com, Wesley Tansey <tansey@...> wrote:
> >
> > Hi everyone,
> >
> > I am working on a project that needs a substrate of more than 2
> > dimensions. Looking at David and Ken's GECCO'08 paper, I would have
> > expected a 3d substrate for the heterogeneous agent experiment. Is there
> > some insight I'm missing behind using the special r(x) function rather
> > than making a 3rd dimension?
> >
> > Wesley
> >
>
>
>