ニューラルネットワークの解析的表現を用いたモデル予測制御

抄録

Neural Network (NN) has attracted increasing attention as a way to constitute a non-mathematical model of a complex robot's dynamics. So far, many successful methods and applications of NN-based robot control have been proposed in parallel with sophisticating control theory in robotics. One of the advanced control theories attracting robotics researchers, Model Predictive Control (MPC), which realizes an optimal feedback control for a nonlinear controlled object based on the prediction of the dynamics, has been intensively studied both theoretically and experimentally. To combine MPC and NN is thought of as a promising to integrate robotic motor learning, control theory, and reinforcement learning. In this paper, we extract a mathematical model from a NN trained to express a robot's dynamics, which has not been paid attention, in order to express the optimization problem of MPC using NN in the form of the Eular-Lagrange equation.