A technique for the optimization of electromagnetic annular phased arrays (APAs) for therapeutic hyperthermia has been developed and implemented. The controllable inputs are the amplitudes and phases of the driving signals of each element of the array. Magnetic resonance imaging (MRI) is used to estimate noninvasively the temperature distribution based on the temperature dependence of the proton resonance frequency (PRF). A parametric model of the dynamics that couple the control inputs to the resultant temperature elevations is developed based on physical considerations. The unknown parameters of this model are estimated during a pretreatment identification phase and can be continuously updated as new measurement data become available. Based on the parametric model, a controller automatically chooses optimal phases and amplitudes of the driving signals of the APA. An advantage of this approach to optimizing the APA is that no a priori information is required, eliminating the need for patient-specific computational modeling and optimization. Additionally, this approach represents a first step toward employing temperature feedback to make the optimization of the APA robust with respect to modeling errors and physiological changes. The ability of the controller to choose therapeutically beneficial driving amplitudes and phases is demonstrated via simulation. Experimental results are presented which demonstrate the ability of the controller to choose optimal phases for the APA using only information from magnetic resonance thermometry (MRT).