My summer research project.

Very nice, the project sounds interesting. I'm also using breve to simulate robots that I'm programming with different behaviors. Unfortunately for myself, I'm spending too much time with silly movement problems rather than my behaviors! This due to my unfamiliarity towards three dimensional movement, vectors, and rotation. I'm a decent enough programmer, but certain math skills of mine could definitally be improved.

Right now I'm working as a tech, and without much to do, I'm using my extra time to work on my project.

Hope your project continues going well!

~Moto

A lot of technical words for someone like me that was not born in the field, but it is still readable enough to get the main idea. ;-)

Since you've lauched this thread in order to stimulate discussion, well I'll try to jump in, if you don't mind. You must forgive my poor knowledge of the field from time to time (and also forgive me asking stupid or obvious questions). This having been said, let start.

So the idea, if I get the point, is to build a robot with a neural network that will only have a basic set of rules (instead of a "clever" well planned mind), and let GA work on this set of rules to get an "adapted" neural network that will fit the needs.

I've been reading a book on a-life (by Steven Levy), and I can't resist making the parallel between your project and some chapter of this book. The book is covering several aspect of a-life, it's a bit old, but still... So if you don't mind, I'll ask some questions...

One chapter of this book (Real Artificial Life) is talking about robots and about a guy named Brooks who "invented" the concept of subsumption. Basically, the way I see it, instead of trying to make robot smart from the beginning (AI way), they make them stupid. They told them, listen guy, you are a very dummy creature, you will do simple task. Like: if you find something that your sensor identifies as rock, you collect. If you have collected rock, you go back home. If you encounter something that looks like a beast, run as fast as you can and hide. Always prefer sun to darkness (so that your batery can keep being charged), and so on. But it turn out, that from outside these simple rules build up into an emergent behavior and gives the impression that the creature is "quite clever".

So here is the first "obvious" question, is your model based on subsumption (somehow)? You're talking about simple rules (basic assumption) that are used to define the basic behavior of the robot. So I guess that the idea is to carefully chose them to lead to an emmergent behavior...

There is also another interesting chapter (Artificial Flora, Fauna, and Ecologies) which talks about Holland Classifier concept. The way I see it, the classifier is an optional package to GA. It let creatures adapt their behavior (and learn) according to (from) the environment. What's interesting is that sometimes these learning tends to favorise creatures with a given characteristic (set of genes), which will then enforce this behavior from birth. The Baldwin "interpretation" of the Landmark idea, if I get the point.

So the second dummy question, does your model rely on Holland Classifier concept or something equivalent?

That's enough for now.

I hope you don't mind me asking you these questions... I have a last question: in the book I was reading, a-life is not so well seen among computer scientist, biologist and all those people. Is this still true?

So long!

My annual scholarly fiasco began this last month so I am slow to respond. My professor Jeff Horn was a graduate student of Goldberg and Goldberg was the first student to graduate (Ph.d) with a GA thesis as a student of Holland. So I guess you could say that I am one of Hollands many intellectual grandchildren. But to be more precise... Holland spent a couple of years in the psychology dept at U of M trying to understand the essential character of the brains type of computation (that is when he developed the classifier). I spent my first three years of undergraduate work doing the same thing (moonlighting in psychology). Holland's classifier is simple way to encode a decision surface in a symbolic markup (yes, no, do not care with single characters) however, reconciling a set of truths with a set of untruths is a threshold function that I don't believe he addressed with the encoding itself. Please correct me if I am wrong! It can be done easily enough with some jazz like: (((Y, Y, N), 6/10), N, N, Y, 7/10), meaning an affirmative response must agree with 6/10ths of the first condition (Y, Y, N) and the composite of that determination and N, N, Y at the next three positions must be at least 7/10ths true to qualify. But anywho, that is not what I am doing. I am using massive neural networks with meta-networks that can change the operation of the behavior generating neural network. I guess I think that the determination of partial truth is absolutely necessary for a robust autonomous agent and for that matter any creative or curious agent can be motivated by a determination of exactly how partial the classification is! Well to put it in a nutshell, a collection of connected Wiener filters (that is what an artificial neural network truly is), is more graceful under the ambiguity that results from the (nearly and effectively) continuous world we live in, than an LCS. I mean unless I am totally off LCS is rather discrete whereas artificial neural networks are continuous.
Correy Kowall