Genetic algorithms incorporating a pseudo-subspace method

GA performance in high-dimensional optimisation problems can be enhanced by the use of a 'pseudo subspace' technique. The method works by projecting the parameter space onto a lower dimensional subspace in the first stages of the optimisation process, in order to allow the GA search to discover the most promising area of the solution space. Subsequently, the dimensionality of the model is progressively increased until a predetermined limit is reached. Comparison between the pseudo-subspace procedure and a conventional GA, using two different GA implementations, shows the former to be more successful when applied to two geophysical problems characterised by different solution-space geometry and mathematics. This technique could be easily transferred to different image processing or pattern recognition problems where geometrical relationships between the parameters are maintained.