Section: New Results

The devil in the details of evolvability

Participants: E Tannier, P Biller, V Liard, G Beslon

The theory of Evolvability consists in studying the evolution of living organisms as a computational learning process. It defines the possibilities of a population under Darwinian selection, to evolve in a certain direction, in a reasonable amount of time. While its robustness to certain parameters has been theoretically assessed, this theory has not been experimentally tested. We use a standard in silico experimental evolution tool to compare some predictions of the theory and the behavior of digital populations designed to resemble biological organisms. We obtain that the evolvability of monotone conjunctions under the uniform distribution of environmental conditions, presented as a major result of the theory, is not reproduced by the experiments. We show that this is due to different mutation algorithms, by a proof of exponential expectation time to target under theoretical conditions closer to the experiments. We examine into detail the choices of mutation algorithms. In the Evolvability theory it is any Turing machine, while much more restricted in the experimental design. This definition allows a wider range of conditions and in a certain way is conform to biological reality, where mutators evolve and can be selected. However it also allows, if it is misused, for the inclusion of oracles that are incompatible with the principles of a Darwinian evolution. Unfortunately these oracles are extensively used in the current evolvability proofs.