Analysing bioHEL using challenging boolean functions
Proceedings of the 12th annual conference companion on Genetic and evolutionary computation
| Hi-index | 0.00 |
Abstract: The number of samples needed to learn an instance of the representation class kDNF^s_n of Boolean formulas is predicted using some tolerance parameters by the PAC framework. When the learning machine is a simple genetic algorithm, the initial population is an issue. Using PAC-learning we derive the population size that has at least one individual at a given Hamming distance from the optimum. Then we show that the GA evolves solutions from initial populations rather far (Hamming distance) from the optimum.