Self-adaptive evolution of model structures and parameters / by Jason John Bobbin.

Abstract

This thesis proposes a new self-adaptive, symbiotic model evolution framework SASME. SASME is described generally and then applied to the specific task of evolving rule sets with explicit default hierarchies for learning problems. Experiments are conducted on the SASME framework to establish the efficiency of the self-adaptive mutation scheme when compared to apriori settings of the mutation rate. The self-adaptive scheme is found to perform optimally, removing the trial-and-error experimentation often required to get satisfactory performance from fixed mutation-rate evolutionary systems.