Harnessing the complexity of natural evolution to enhance evolutionary algorithms

Evolutionary computation is a powerful optimization tool, and an invaluable test bed for population genetics. Evolutionary algorithms can become stuck on local optima, but can escape these traps by temporarily losing fitness in order to discover even higher fitness in a process called valley-crossing. Valley-crossing is fundamentally linked to the balance between the forces of selection and variation, and as such, controlling this balance is important for optimizing the efficiency of evolutionary algorithms. Nature, in contrast, is not actively optimized for performance, and yet nature seems to overcome many challenges that evolutionary algorithms do not. It is possible that nature benefits from a highly dynamic balance between selection and variation, and this constant flux helps natural populations avoid stagnation and overcome obstacles in the fitness landscape. Working with this hypothesis in mind, I investigate the nature of selection and how natural phenomena strengthen or weaken it.I find that selection strength can be thought of as the degree to which an evolving system is dissimilar to neutral drift. This perspective opens the door to accept all phenomena that affect the strength of selection as part of a unified theory of selection that treats selection strength as an emergent property. I present a new evolutionary dynamic---the free-for-all effect---that is the reduction of selection strength on organisms with higher-than-average fitness. Free-for-all can result in rapid evolutionary adaption that would otherwise seem impossible, and provides an elegant explanation for punctuated equilibrium. The discovery of free-for-all highlights the importance of spatial structure in evolving populations, and has led to the design of a new evolutionary search method called super explorers. Super explorers mimic the free-for-all effect, and improve evolutionary search, while placing full control into the hands of the algorithm designer.