In the field of Genetic Programming, the phenomenon of bloat is a common cause of decreased performance in the search process, as it is typically accompanied by loss of diversity and premature convergence. As the solutions grow rapidly in size, the genetic operators like standard sub-tree swapping crossover and sub-tree mutation loose their original effect of exploitation and exploration, and further improvement of the current solutions is only seldom observable.
In this thesis we use Genetic Programming to evolve close combat strategies for agents in the game of Unreal Tournament. The work presented in Holm and Nielsen (2002) is used as the framework for this thesis.
We propose four different extensions to the basic Genetic Programming algorithm, in order to gain control over the growth in average size of the solutions, and to improve the genetic operators, with respect to maintenance of diversity.
In conclusion we find that a guidance of the genetic operators to be applied within effective code, shows good performance. A direct pressure that rewards unique solutions and punishes common solutions shows good performance and an insignificant degree of bloat.