Training sample selection in learning control

Learning control from a human expert demonstration can he considered as building a mapping between system states and control inputs. The mapping precision of the controller relies primarily on training samples. However, due to system states often failing in Dense Region (neighbored region around system's equilibrium point) and seldom failing in Sparse Region(regions far from equilibrium point), the samples for training neural network controller are often unbalanced which leads to controller with different precisions in different regions. The trained controller will work well in Dense Region around the equilibrium point, but might deteriorate in Sparse Region. Thus the convergent region is relatively small, while in many control system we want the convergent region to be as large as possible. This paper proposes a novel solution, which resamples the original training samples to balance the sample sizes in different regions. The re-sampling approach adopted here is based on cluster sampling. Preliminary simulation results demonstrated the feasibility of this approach.