[combining island-style genetic algorithms with cultural evolutionary systems] - A New Kind of Science: The NKS Forum

A New Kind of Science: The NKS Forum

Pages:1


combining island-style genetic algorithms with cultural evolutionary systems

(Click here to view the original thread with full colors/images)


Posted by: inhaesio zha

Michael Bakkenson and I recently created some systems that combine island-style genetic algorithms with the cultural evolutionary systems defined and described in "the evolution of culture" posts to the "Pure NKS" part of this site.

(http://forum.wolframscience.com/sho...=&threadid=1415)

Island-style GAs mimic the isolation of species from each other in the way Australia or the Galapagos are separate for a long time from other land masses, giving species the chance to develop in little pockets, with limited exposure to other pockets...and they occasionally allow some type of exposure to occur between the evolutionary pockets.

One of the effects this can have is to postpone convergence of solutions--the Galapagos turtle might need to evolve for a very long time without competition from the Bengal tiger before it possibly arrived at a fitness that was greater than that of the Bengal tiger. Without partitioning of the solutions the tiger might kill off all the early turtles before they had a chance to become superior to the tiger. If the turtle evolves for some time on an island devoid of tigers, however, the outcome could be different.

To my understanding, in island-style GA systems, the programmer typically defines (either completely or to a large degree) the nature of the partitioning and the nature of the periodic interface of the partitions.

What Bakkenson and I built is a system that takes advantage of teoc systems' emergent partitioning, clumping, rejoining, and other cultural features; but that also evolves solutions to specific real-world problems.

If you look at the images of teoc systems I've posted elsewhere on this site, you will see that in general the systems evolve in ways where there emerge rules or guidelines which the organisms follow when arranging themselves: in some systems, there are regions of a certain color that seem to be generally unable to exist in the middle of regions of another color--these principles are not imposed by the programmer; they emerge from the behavior of the system itself. This clumping, isolation, merging, etc. is an organic, complex, emergent version of the island-style partitioning sometimes used with genetic algorithms.

We modified the original teoc systems such that an individual's ability to swap its position with a neighboring individual, and its ability to push some of its ideas/genome onto a neighboring individual, is determined by a comparison of the two individuals' fitnesses--fitnesses that are the result of applying an individual's genome to a real-world problem...in our initial case we used data about poisonous mushrooms. The real world problem was to classify a mushroom as either poisonous or edible by examining physical traits of the mushroom.

To calculate an individual's fitness, we took the teoc organism's genome and used that as the parameter to a mushroom classification function (details here are interesting in their own right, but left out, and really not important in this context). This makes it so that the way an individual moves and meets within the [cultural, geographic] environment is related to the individual's ability to tell whether a mushroom is poisonous: individuals who are better at solving this problem are more likely to travel and impress other with their ideas.

teoc's inherent cultural/geographic properties introduce partitioning into the GA-style search, but in a much more open-ended way than traditional island-style partitioning.

We haven't done methodical comparisons of quantitative properties of teoc-style GA partitioning versus those of traditional island-style partitioning. We've simply run some teoc-style GA systems and shown ourselves that they can find solutions to the mushroom classification problem that are at least as good as solutions found using the same types of genome interpretations but without any partitioning (classic GA style). In our small set of sample runs the teoc-style GA found solutions to that problem that were nominally better than the best solutions we have found by a similar, but unpartitioned, classic GA.

So: that's a weak claim...but (I think) an interesting idea to pursue.




Forum Sponsored by Wolfram Research

© 2004-2009 Wolfram Research, Inc. | Powered by vBulletin 2.3.0 © 2000-2002 Jelsoft Enterprises, Ltd. | Disclaimer
vB Easy Archive Final - Created by Xenon and modified/released by SkuZZy from the Job Openings