| dc.description.abstract | The Gene-pool Optimal Mixing Evolutionary Algorithm (GOMEA) is an optimization framework with a single tunable parameter: the population size. In this thesis we explore the use of parameter-less population schemes to remove this parameter. It allows users to skip parameter tuning and makes scalability analysis easier to perform for researchers. The Exponential Population Scheme (EPS) is tested, which simply restarts the population upon convergence with double the population size. According to traditional scalability analysis there is only minor constant overhead, but EPS inevitably throws away several populations worth of fitness evaluations. The Parameter-less GA would reduce this wasteful behavior using population racing, but it resulted in worse scalability instead. The population pyramid scheme is extracted from the novel P3 algorithm and is applied to EPS in order to reuse old populations. This makes EPS more efficient in the number of fitness evaluations and allows for a slower growth of the populations. The memory usage increases dramatically, but this is remedied by adding a selection procedure. As a result of these experiments, the run-time of P3 is significantly decreased by operating on populations (as EPS does) instead of single solutions. | |
