A New Algorithm for the LQR Problem with Partially Unknown Dynamics

AbstractWe consider an LQR optimal control problem with partially unknown dynamics. We propose a new model-based online algorithm to obtain approximation of the dynamics and at same time during single simulation. The iterative is based on mixture Reinforcement Learning techniques. In particular, we use Gaussian distributions represent model uncertainty probabilistic updated each iteration using Bayesian regression formulas. On other hand, obtained in feedback form via Riccati differential equation. present some numerical tests showing that can efficiently bring system towards origin.KeywordsReinforcement learningLQR problemNumerical methods