|Institution:||University of Manchester|
|Keywords:||evolvability; selection for evolvability|
|Full text PDF:||http://www.manchester.ac.uk/escholar/uk-ac-man-scw:306938|
This thesis is about direct selection forevolvability in artificial evolutionary systems. The origin ofevolvabilitythe capacity for adaptive evolutionis of greatinterest to evolutionary biologists, who have proposed manyindirect selection mechanisms. In evolutionary computation andartificial life, these indirect selection mechanisms have beenco-opted in order to engineer the evolution of evolvability intoartificial evolution simulations. Very little work has been done ondirect selection, and so this thesis investigates the extent towhich we should select for evolvability. I show in a simpletheoretical model the existence of conditions in which selectionfor a weighted sum of fitness and evolvability achieves greaterlong-term fitness than selection for fitness alone. There are noconditions, within the model, in which it is beneficial to selectmore for evolvability than for fitness. Subsequent empirical workcompares episodic group selection for evolvability (EGS)analgorithm that selects for evolvability estimates calculated fromnoisy sampleswith an algorithm that selects for fitness alone onfour fitness functions taken from the literature. The long-termfitness achieved by EGS does not exceed that of selection forfitness alone in any region of the parameter space. However, thereare regions of the parameter space in which EGS achieves greaterlong-term evolvability. A modification of the algorithm, EGS-AR,which incorporates a recent best-arm identification algorithm,reliably outperforms EGS across the parameter space, in terms ofboth eventual fitness and eventual evolvability. The thesisconcludes that selection for estimated evolvability may be a viablestrategy for solving time-varying problems.Advisors/Committee Members: HANDL, JULIA JK, Handl, Julia, Knowles, Joshua.