Linear Quadratic Gaussian using Kalman Network and Reinforcement Learning for Discrete-Time System

A problem arises when implementing the control scheme into a system, the limitation of the number of sensors used causes not all state information from the plant can be obtained. In addition, the data measurement process also often contains noise. Therefore, implementing the optimal control scheme directly becomes difficult. The status estimation method proposed in this study utilizes the use of an Artificial Neural Network (ANN) as a function approximation. On the other hand, the development of optimal control implementation has also experienced a fairly massive increase. Some of the existing optimal control solutions can only be carried out offline and depend on the dynamics' model of the system to be controlled. The combination of KalmanNet and the Value Iteration (VI) algorithm is used to adapt the role of the Linear Quadratic Gaussian (LQG) method in designing control procedures to produce high performance. Tests were carried out on a cart-pole system in a discrete-time domain. The use of the combination of these methods could make the closed loop system towards a value of 0 and the evolution of the estimator gain is smoother, namely with the norm of 2.33S4E-04.