Re: Some Good News and Some Bad News
- --- In firstname.lastname@example.org, Mitchell Timin <zenguyuno@y...> wrote:
> --- Colin Green <cgreen@d...> wrote:I'm not sure NEAT would solve this task with no problem. It
> > >>
> > From your description of the problem and knowledge
> > of what NEAT as
> > solved so far I would agree that NEAT is capable of
> > solving this with no
> > problem at all.
> Right. Your real target should be the six-card game,
> and larger. But you can do the 4-card game on the way
> there. It will be interesting to see how many neurons
> you wind up with. To play well the network must
> remember which two of the 4 cards were exposed by the
> first player, and whether each of those was a 1 or a
> Then it must use that data to turn over an unexposed
> card, and then pick the single exposed card that
> matches it.
seems that since NEAT can evolve topologies, it has a good
shot at finding a more efficient topology than a fully-recurrent
net. But how specific is that optimal structure? It really
would be quite interesting to see how NEAT would handle the problem.