This reminds me of Tierra and Core Wars.

Now core wars was fun! Not so much AI as the ultimate in coding... but a lot of fun!

Timkin

whats core wars?

http://mcraeclan.com/Graeme/CoreWars.htm

Check out ''classifier systems'' - they are (a kind of) programs that re-write themselves, or evolve.

It''s kindof a mix between rulebased systems (rules call eachother in an iterative framework), genetic algorithms (rules are formulated like GA-chromosomes) and neural networks (successful rules, and their ''supporters'' are strengthened)

Wish I had time to implement one of those myself

-----------------------
Always down - Never out

If you mean to suggest that a classifier system could be used to evolve bitstrings (that can be interpreted as machine code) then I don''t see the benefit over just using a genetic algorithm. The classifier system is better used for its implementation of an evolvable finite state machine that incorporates reinforcement learning. Evolving code does not require the code to be expressed in condition:action pairs, just bitstrings to be interpreted. Hence, a GA would be a better candidate for evolutionary search.

If you had something else in mind, please explain it further.

Cheers,

Timkin

I was thinking that the (evolvable) rules in the classifier system would Be the program, itself - not construct one.

The bitstring representation is just a formalism that allows the rules to be treated as GA chromosomes, I guess.

-I kindof like the idea that the generated ''program'' is a FSM in the form of a production system. That should be easier to evaluate (and handle in general) than one represented in machine code.
-Less open-ended, probably - depending on how you''d connect it to the ''outside'' machine, but it would feel a lot safer.

I''ve glanced on ''genetic programming'' stuff, but I don''t know if I''d dare doing something like that - my programs crash often enough as it is

Does anybody know any good ''classifier systems for dummies''-site?! - tell me!

-----------------------
Always down - Never out

Dr Hugo DeGarris at Utah State University is doing some research into making an artificial "brain" using fpgas. He basically use this machine to run genetic algorithms, keeping the good ones and merging them and running their offspring against each other until he has the best one.

Another idea is, in a class I took we wrote (using lisp) a program that took our earlier programs (dfs and heuristic searching) and removed all the recursion by turning it into a sequence of huge loops and pattern matching. The speed up was incredible

Dr Hugo DeGarris at Utah State University is doing some research into making an artificial "brain" using fpgas. He basically use this machine to run genetic algorithms, keeping the good ones and merging them and running their offspring against each other until he has the best one.

Another idea is, in a class I took we wrote (using lisp) a program that took our earlier programs (dfs and heuristic searching) and removed all the recursion by turning it into a sequence of huge loops and pattern matching. The speed up was incredible

David Goldberg''s book, "Genetic Algorithms in Search, Optimization and Machine Learning", has a couple of chapters on classifier systems. You can also obtain the original John Holland papers online (Citeseer has them I think), plus there are a few other interesting papers you can easily obtain with a little searching. Timkin should have some good references as I believe some of his research is based around classifier systems.




ai-junkie.com