Last night I met with a couple of friends to discuss neural nets and our current AI project. We're about a week away from doing our first test, but that didn't stop us from outlining possible approaches to more difficult domains that we'll eventually be confronting.

One idea we'd strongly like to try is indirect genetic encoding. Right now we're using direct encoding. Here's the difference:

With direct encoding, a gene expresses one attribute. For example, one of the genes for our neural net may look like this:

[1->2,0.5]

Which would encode for a synapse (or connection) between neurons 1 and 2, with a connection weight of 0.5. The neural net genome may be thousands of such "genes" that our program will translate and build into a software neural net processor.

But with indirect encoding there is not a strict 1-to-1 relationship between genes and what they encode for. We could have a gene that expresses a neuron, and then another gene that basically copies that neuron a given number of times. So instead of encoding for just traits, you're also encoding for building rules.

Many organic structures are built this way. Your DNA doesn't encode for every single strand of hair on your head. It contains information on how to build a follicle, then basically repeats that information thousands of times over across your scalp.

And of course, large chunks of the brain are built this way, too.

We hope to develop, or possibly evolve, a logical grammar for indirectly encoding complex neural net structures.

Also, we're considering the idea of a neural structure with two main organs, one that sits atop the other. Human brains are like onions, with layers evolving around existing structures. At the base of our brain is the brain stem, a primitive structure found in many other vertebrates, that controls autonomic functions like heartbeat and breathing. The inner structures contain hardwired instinctual behavior related to basic animal functions (sex, aggression, etc.). The cortices have evolved on top of those existing structures, and they actually seem to be less complex structurally, and much more loosely wired and flexible.

We're thinking of employing this model for our game-playing AI, having a very set substructure that knows many of the unchangable foundations of its domain (like the rules, very basic strategic principles, etc.), with a very flexible superstructure that learns and adapts to changes it might encounter (like opponents with different strategic styles).

Anyway, all this stuff is interesting to think about, but we're still in the relative infancy of our project. To use the biological analogy, we want to build a brain, but haven't even built a ganglia yet. We're on the right track, though.