Improving controller performance for modeling two-person artificial wrestling by underestimating basins of attraction

Abstract

This study proposes a preprocessing method that overcomes the misprediction problems in the intelligent motion controllers (IMCs) of two-person wrestling models. In our previous study, the competitive problems were solved on a coupled inverted pendula model, in which an intelligent controller produced the desired final states by exerting impulsive forces on each pendulum. The final states were inferred from a look-up table (LUT) that stores the dynamic correspondences from the initial to the final states. In a numerical investigation, we demonstrated that the observed performance degradation resulted from mispredictions caused by the limited resolution of the LUT. We conjectured that these mispredictions occur near the basin boundaries of the desirable final states. In the present study, we first derive a simple algorithm that removes the basin boundary points from the LUT. We then show that the resulting controller largely reduces the number of mispredictions, but renders the controller overconservative. We hence propose an adaptive algorithm that removes only the boundary points causing the noted mispredictions. Our second algorithm successfully derives a less conservative controller and improves the overall control performance without increasing the resolution of the LUT, implying that the performance of such a competitive IMC can depend on tiny changes in the a priori knowledge of the system dynamics.