Memetic Algorithms (was Re: virus: RAND)

David McFadzean (david@lucifer.com)
Sun, 21 Mar 1999 13:57:27 -0700

At 01:19 PM 3/21/99 +0000, Robin Faichney wrote:

>I glanced at the ACE page, but can't find the time this
>deserves right now. Just one quick question, David: is
>anyone looking at agent learning-by-imitation yet? I
>mean, beyond the basic do-what-the-other-prisoner-did-
>last level.

You mean simulated memetics? Yes indeed, but I have to qualify that by saying it really depends on how you look at it.

Here's what is really happening: each agent in the simulation has its own digital signature that determines the agent's behavior, i.e. its strategy in the IPD games. The agents play against each other for a few rounds, then the most successful strategies are propagated to a new generation of agents while the less successful strategies are discarded[1].

If you prefer to look at the next generation of agents as the "children" of the previous generation, then the digital signatures are like genes and the process is called a genetic algorithm. This perspective is the standard one in the field.

If, on the other hand, you prefer to look at the next generation of agents as the same individuals as the previous generation, but learning from experience by mimicking their more successful neighbors, then the digital signatures are more like memes and the process could be called a memetic algorithm.

In either case drawing an analogy between the simulation and genes or memes is pretty tenuous so in some ways it is just a matter of taste or perspective. The reason I would tend to lean toward genetic algorithm to describe the process is because each new strategy is derived from combining the signatures of exactly two relatively successful agents[2], rather like parents in bio-life. If the algorithm was slightly altered to generate new strategies from more than two others I would tend to lean towards calling it a memetic algorithm instead.

[1] Technically the more successful strategies just have a higher probability of getting into the next generation.

[2] With some random mutations thrown in to ensure that no part of the fitness landscape is excluded a priori by the initial population.

--
David McFadzean                 david@lucifer.com
Memetic Engineer                http://www.lucifer.com/~david/
Church of Virus                 http://www.lucifer.com/virus/