Department of Pattern Recognition

(eng): The solution of repeated fixed-horizon trajectory optimization problems of processes that are either too difficult or too complex to be described by physicsbased models can pose formidable challenges. Very often, soft-computing methods - e.g. black-box modeling and evolutionary optimization - are used. These approaches are ineffective or even computationally intractable for searching high-dimensional parameter spaces. In this paper, a structured iterative process is described for addressing such problems: the starting point is a simple parameterization of the trajectory starting with a reduced number of parameters; after selection of values for these parameters so that this simpler problem is covered satisfactorily, a refinement procedure increases the number of parameters and the optimization is repeated. This continuous parameter refinement and optimization process can yield effective solutions after only a few iterations.