This paper describes a novel data-based modeling and control method for dynamic systems. The model structure consists of locally linear, clustered principal component regression modules. This scheme is motivated by a novel neuro-cognitive theory; the goal in this paper is to assess the plausibility of the scheme. As an application example, a complex dynamical system, that is, a simulation model of a walking biped robot, is used. For the data collection, the robot is first controlled by simple separate PD controllers; it turns out that the clustered regression model is accurate enough to replace the original controller and to reproduce the example gait very well. The optimization of the behavior by updating the regression structure, however, appears to be quite difficult. It turns out that the robustness that is necessary for optimizing the model cannot be reached.