Research On Stochastic Dynamic System's Learning And Control

With the development of modern science and technology and production technology, control system become increasingly complex, and complex system modeling and control become one of hot researches. As in practice the systems have uncertainties, the control problem can not use a simple deterministic model to describe, and it need to combine controlling a system and learning a system uncertainty to an issue. The nature of control is that: on the one hand, the control signal can make the system output towards the desired goal (called control action); the other hand, the control signal can reduce the uncertainty of system parameters (called learning the parameter uncertainty). However, two hands are contradictory, and the former requires the control signal to change smoothly, while the latter wants to maintain certain amplitude of motivation, so control need trade-off.In this paper, linear, nonlinear stochastic system and multi-model nonlinear stochastic systems is studied.(1) For the unknown parameters with Gaussian white noise stochastic linear systems, using "utility function" it is presented a trade-off of learning and controlling of the control strategy. Controller. on the one hand, can control the system toward the desired output:on the other hand, can learn the unknown linear parameters. Simulation results show the validation of such approach.(2) With unknown parameters of the non-linear stochastic systems control problem, it is proposed a learning and controlling optimization algorithm. Controller can not only control the output to track the desired output, but also can using RBF networks online learn unknown parameters of nonlinear systems. Simulation results show superiority of the algorithm.(3) For a class of unknown parameters of nonlinear multi-model switching stochastic systems, it is proposed an algorithm, which use RBF network to learn nonlinear functions online, and Bayes posterior probability to estimate the model, according to the cost function obtain the control signal. By simulation, the algorithm can obtain a switching time of system exactly and can accurately track changes in the system.