neural networks research group
areas
people
projects
demos
publications
software/data
Towards an Empirical Measure of Evolvability (2005)
Joseph Reisinger
,
Kenneth O. Stanley
,
Risto Miikkulainen
Genetic representations that do not employ a one-to-one mapping of genotype to phenotype are known as indirect encodings, and can be much more efficient than direct encodings for complex problems. Increasing a representation's capacity to facilitate effective search, i.e. its evolvability, has long been a goal of Evolutionary Computation. However, currently no benchmarks exist to measure evolvability. One reason is that it is difficult to decouple a representation's capacity to evolve under any fitness function, i.e. the latent evolvability, and its performance on a specific benchmark. Towards this goal, a method is proposed in this paper that measures the representation's ability to extract invariant properties from a changing fitness function. The test is applied to three distinct representations and it is able to distinguish all three. Ultimately, this test can serve as the foundation for performing controlled experiments determining what factors contribute to evolvability.
View:
PDF
Citation:
In
Genetic and Evolutionary Computation Conference {(GECCO2005)} Workshop Program
, 257-264, Washington, D.C., 2005. ACM Press.
Bibtex:
@InProceedings{reisinger:geccows05, title={Towards an Empirical Measure of Evolvability}, author={Joseph Reisinger and Kenneth O. Stanley and Risto Miikkulainen}, booktitle={Genetic and Evolutionary Computation Conference {(GECCO2005)} Workshop Program}, address={Washington, D.C.}, publisher={ACM Press}, pages={257-264}, url="http://nn.cs.utexas.edu/?reisinger:gecco05", year={2005} }
People
Risto Miikkulainen
Faculty
risto [at] cs utexas edu
Joseph Reisinger
Former Ph.D. Student
joeraii [at] cs utexas edu
Kenneth Stanley
Postdoctoral Alumni
kstanley [at] cs ucf edu
Projects
NEAT: Evolving Increasingly Complex Neural Network Topologies
2000 - 2011
Leveraging Evolvability in Search
2004 - 2007
Areas of Interest
Evolutionary Computation
Reinforcement Learning