Back to the program.
Analysis and Simulation of Intrademic Multilevel Selection
Tom Lenaerts, Ph.D
IRIDIA
Université Libre de Bruxelles
tlenaert@(nospam).vub.ac.be

Abstract

With the advent of theories on evolutionary transitions in biological complexity, interest in kinship, population structure and group selection has re-emerged. The relevance of these subjects is attributed to their constructive role in this newly emerging perspective on evolution: Before each transition, individual elements start forming closely collaborating groups, producing selection dynamics at a higher level. Especially in those cases where traits depend on the frequency of other organisms in the environment, this force can play an important role. Consequently the selfish gene perspective on evolution by natural selection shifts toward a perspective where groups play an important role. We investigate the selection dynamics produced by individuals within different groups and the higher-level dynamics between these groups. The resulting process is referred to as a multilevel selection process. Depending on the settings of the multilevel selection process, different dynamics will be observable. Through an analysis of these dynamics, we try to improve our understanding of the conditions under which certain equilibria are attained and how they differ from the outcome expected under standard selection dynamics. Furthermore, the mathematical and computational models will provide initial steps toward a validation of the role of multilevel selection in the context of a general model for evolutionary transitions. Besides the biological relevance of this study, multilevel selection was investigated because of the possible improvements it might introduce in context of learning and optimization by evolutionary algorithms. These possible improvements are envisioned in two area's. On one hand, a multilevel selection algorithm may improve the optimization capabilities of evolutionary algorithms for, for instance, those problems where the ``group-beneficial'' properties correspond to better solutions for particular problems (Pareto-optimal solutions). On the other hand, scalability of techniques like genetic algorithms toward structurally more complex solutions and larger problems is an important avenue of research. Evolutionary transitions and multilevel selection provide the natural metaphor to address these issues.Furthermore, the metaphor of transitions in complexity is not limited to only evolutionary machine learning methods. Other methods like reinforcement learning and learning automata might profit from certain ideas as well.

Keywords

multilevel selection, evolution