For the past two semesters, I’ve been doing some exploratory work marrying speciation with information theory in the framework of the Polyworld artificial life simulator. The simulation gives us a nice framework for mathematically “pure” evolutionary theory and exploration of neural complexity. We’ve applied clustering algorithms to the genetic information, revealing evidence of both sympatric and allopatric speciation events. The key algorithmic intuition is that genes which are highly selected for will conserve, while those which are not will descend to a random distribution (and thus high entropy), so each dimension (gene) can be weighted by its information certainty to alleviate the curse of dimensionality.
The work was accepted as a poster and extended abstract for the Genetic and Evolutionary Computing Conference (GECCO), and was accepted as a full paper for the European Conference on Artificial Life (ECAL). The full paper is substantially revised from the initial GECCO submission, and provides an introduction to several problems of biological, computational, and information theoretic importance. The visualizations, including several videos showing the cluster data, were especially fun to create, and I’m proud of the finished product.
There are still several more research directions from this work: the allopatric and sympatric effects have not been differentiated, only one environment was analyzed (consistent with past work on evolution of complexity), the clustering algorithm’s thresholds were not explored for hierarchical effects, alternate clustering algorithms were not explored (future open-source project for me: clusterlib), … Still, the present work is encapsuled, the source is in the Polyworld trunk, and it was accepted for publication.
Abstract, citation, and paper follow.
Complex artiﬁcial life simulations can yield substantially distinct populations of agents corresponding to different adaptations to a common environment or specialized adaptations to different environments. Here we show how a standard clustering algorithm applied to the artiﬁcial genomes of such agents can be used to discover and characterize these subpopulations. As gene changes propagate throughout the population, new subpopulations are produced, which show up as new clusters. Cluster centroids allow us to characterize these different subpopulations and identify their distinct adaptation mechanisms. We suggest these subpopulations may reasonably be thought of as species, even if the simulation software allows interbreeding between members of the different subpopulations, and provide evidence of both sympatric and allopatric speciation in the Polyworld artiﬁcial life system. Analyzing intra- and inter-cluster fecundity differences and offspring production rates suggests that speciation is being promoted by a combination of post-zygotic selection (lower ﬁtness of hybrid offspring) and pre-zygotic selection (assortative mating), which may be fostered by reinforcement (the Wallace effect).
Jaimie Murdock and Larry Yaeger. Identifying Species by Genetic Clustering. In Proceedings of the 2011 European Conference on Artificial Life. Paris, France, 2011. [paper]