Neural-network control of building structures by a force-matching training scheme

A method to generate an efficient control law for a neural-network controller is presented to reduce the dynamic response of buildings exposed to earthquake-induced ground excitations. The proposed training scheme for the neural-network controller does not rely on the emulation of the structure to be controlled. The approach used for this work is based on a force-matching procedure, and it directly utilizes the dynamic data characterizing the structure response to generate an efficient training signal. The proposed controller has a feedback structure, utilizing a limited set of response quantities. A shear building actuated at its top by a tuned-mass damper is utilized to demonstrate the effectiveness of the controller. For training purposes, an ensemble of synthetically generated ground-motion time histories, with appropriate site spectrum characteristics, have been used. The performance of the trained controller is then evaluated for two different historic ground-acceleration records that do not belong to the training set of time histories. The numerical simulations show the control effectiveness of the proposed scheme with modest control requirements.