Robust Reinforcement Learning Decoupling Control Based on Integral Quadratic Constraints

Abstract. In order to keep stable in reinforcement learning process, a novel robust reinforcement learning decoupling control (RRLDC) based on integral quadratic constraints(IQC)is presented in this paper. It composes of a linear model to approximate the nonlinear plant, a state feedback K controller to generate the basic control law, and an adaptive critic unit to evaluate decoupling performance, which tunes an actor unit to compensate decoupling action and model uncertainty as well as system nonlinearity. By replacing nonlinear and time-varying aspects of a neural network and model uncertainty with IQC, the stability of the control loop is analyzed. As a result, the range of the adjusted parameters is found within which the stability is guaranteed, the control system performance is improved through learning and the algorithm convergence speed is accelerated. The proposed RRLDC is applied to gas collector pressure control of coke ovens. The simulation results show the proposed control strategy can not only obtain the good performance, but also avoid unstable behavior in learning process. It is an effective multivariable decoupling control method for a class of strong coupling systems such as the gas collector pressure control of coke ovens.the effectiveness of proposed control strategy for the collector gas pressure of coke ovens