The active magnetic bearing control through analytically designed linear PD regulator, with parallel nonlinear compensation represented by automatic approximator is described in this contribution. Coefficient (parameter) values come from actions of Continuous Action Reinforcement Learning Automata (CARLAs). Modified algorithm for automata implementation is used which continuously updates learning parameters according to former learning process. The goal of this on-line training is formulated as achievement of minimum mean square of control error. Described concept of control is proved by simulation study. It is shown that the significant improvement of whole system behavior can be achieved. and Obsahuje seznam literatury