Competitive Coevolution in Problem Design and Metaheuristical Parameter Tuning

Abstract

The Marine Research Institute of Iceland has over the last 20 years developed and used the program gadget for modeling of the marine ecosystem around Iceland. The estimation of parameters in this program requires constrained optimization in a continuous domain. In this thesis a coevolutionary algorithm approach is developed to tune the optimization parameters in gadget. The objective of the coevolutionary algorithm is to find optimization parameters that both make the optimization methods in gadget more robust against poorly chosen starting values and tries to reduce the computation time while maintaining convergence. This may also ease the tuning of optimization parameters for new users and may reveal other local optima in the likelihood, which may give hint of model misspecification. The algorithm is tested on functions that have similar characteristics as the log-likelihood functions in gadget and some results shown for the case of modeling haddock.