Inspired by Darwin's natural selection, genetic programming is an evolutionary computation technique which searches for computer programs best solving an optimization problem. The ability of GP to perform structural optimization at the same time of parameter optimization makes it uniquely suitable to solve more complex optimization problems, in which the structure of the solution is not known a priori. But, as GP is applied to increasingly difficult problems, the efficiency of the algorithm has been severely limited by bloating. Previous studies of bloating suggest that bloating can be resolved either directly by delaying the growth in depth and size, or indirectly by making GP to find optimal solutions faster. This thesis explores both options in order to improve the scalability and the capacity of GP algorithm. It tackles the former by firstly systematically analyzing the effect of bloating using a mathematical tool developed called activation rate. It then proposes depth difference hypothesis as a new cause of bloating and investigates depth constraint crossover as a new bloating control method, which is able to give very competitive control over bloating without affecting the exploration of fitter individuals. This thesis explores the second option by developing norm-referenced fitness function, which dynamically determines the individual's fitness based on not only how well it performs, but also the population's average performance as well. It is shown both theoretically and empirically that, norm-referenced fitness is able to significantly improve GP performance over the standard GP setup.
|Date of Award||1 Aug 2014|
- The University of Manchester
|Supervisor||Xiaojun Zeng (Supervisor)|