Supervisor: Schaefer Robert
Status: finished, master dissertation
Defense date: 2012-01-04
A Global Optimization Algorithm is defined as optimization algorithm that employs measures that prevent convergence to local optima and increase the probability of finding a global optimum. However, there are classes of problems in which instead of finding the global optimum we are interested in finding many local optimums whose basins of attraction are properly wide and deep. One way to achieve this goal is to perform several runs of a evolutionary algorithm and alter the fitness function in every subsequent runs of the algorithm in the way that prevents exploration of basins which were found in previous runs of the algorithm. Another cause for concern is how to interpolate fitness landscape in the area of basin of attraction for further fitness deterioration. Very often fitness function is computationally intensive and in such case it is unacceptable to perform classical interpolation of the fitness function. Making the assumption that clusters of population obtained after the single run of the algorithm are good estimators of basins of attraction, it would be better to exploit spatial characteristics of clusters and approximate basins of attraction by multidimensional Gaussain functions performing only a few fitness evaluation. Our goal is to create efficient method for fitness deterioration using the above schema and to analyze the relation between the deterioration accuracy and reproduction operators (mutation, crossover, etc.).