Evolvability and Evolution of Hierarchy

In the context of developmental mechanics, ecological dynamics, and inheritance, evolution (as well as other natural processes) can be treated as an unsupervised learning system without explicit learning rules or the benefit of negative feedback. These conditions limit the types of structures it can learn, and previous work from the group focussing on symmetric correlation learning. This project looks at asymmetric systems, and is primarily an investigation into the evolution of hierarchy in natural systems, and the implications of such hierarchy for the apparent evolvability of complex organisms and ecological systems. Transferring methods and observations from machine learning, we are working toward a broader understanding of high-level learning in natural systems through the adaptation of evolutionary processes, in an attempt to account for the evolvability and robustness observed in modern organisms which is not adequately explained by a naïve process of incremental natural selection. Current work is focussing on hierarchical modules within idealised gene regulation networks (a classic developmental model) evolving in modular environments, probing two (by no means comprehensive nor mutually exclusive) general hypotheses for why such structures might evolve without assuming that such evolvable structures provide an inherent fitness advantage:
- Such structures evolve as a consequence of being evolvable: there is explicit selection for more evolvable types during transition periods which filters for evolvable types
- Such structures evolve as a consequence of feedback processes within the developmental processes which provide local fitness gradients toward evolvable structures