Next: Evolution of Neural Up: Genetic algorithms Previous: Genetic algorithms

GA adaptation

The GA, in and of itself, continues to prove itself to be an extremely powerful adaptive method. John McInerney's dissertation investigated the role ``demes'' - partially segregated sub-populations - may have on evolutionary dynamics [1]. More recently we have also investigated the evolution of cellular automata, with particular interest in the ability of these representations to support global computations [37a] and self replication [37].

Neural networks, because of their powerful learning capabilities, remain the most attractive representation for the GA to manipulate; see below. However, we are also investigating evolution of other forms. Sorting networks are a well-analyzed class sort algorithms, with topological properties that have some similarities to NNets. We have successfully used the GA and a developmental grammar method to discover general sorting network solutions that are on a par with the best-known, analytically derived ones [36][12].