EPSRC Network on Evolvability in Biology & Software Systems
Software Evolution and Evolutionary Computation Symposium Abstracts
University of Hertfordshire, Hatfield, U.K.
W. B. LANGDON
University College, London
London, WC1E 6BT, UK
Many common assumptions about the evolvability of both human produced and evolved programs turn out to be wrong. Even with modest sized programs, the fitness landscape does not appear to be rugged. Instead it appears to be riddled with neutral networks.
While it is very difficult to get rigorous results for every circumstance we have investigated a few genetic programming (GP) benchmark problems and have some proofs about the properties of fitness search spaces, which are applicable to many automatically generated programs [Foundations of Genetic Programming].
Potential practical implications for genetic programming and other forms of automatic programming will be discussed as well as for other open evolutionary architectures, including the utility of neutral pathways.