Evolution strategies with subjective selection of the offspring can be applied to optimization problems, when no objective evaluation of the offspring is possible. For this class of optimization problems, strategies with insensitivity towards incorrect selection decisions are necessary. With subjective selection, incorrect decisions are unavoidable and have the effect of noise being overlaid onto the fitness value. Three examples of different evolution strategies (ES) are presented which are resistant against wrong selection decisions. The convergence of an ES with manually controlled stepsize for the adaptation of a mixture of colored liquid components towards a target color is shown. This first example can be regarded as a real world experiment. In two computer experiments an ES with automatically adapted stepsize has been applied for the first time to this type of optimization problems. The color and shape of a rectangle will be adapted towards a target rectangle, and the adaptation of the shape of a polygon towards a symmetric dodecagon will be shown. With these two test problems the convergence of different strategy variants with and without recombination is examined.
展开▼
