Deep reinforcement learning for semi-active structural control using a variable inerter device

Structural control devices employing inerters have attracted considerable attention owing to their superior vibration mitigation performance. Variable inerter devices, which enable real-time adjustment of inertance and damping, offer further performance enhancement but also introduce a challenging nonlinear control problem. This paper proposes a deep reinforcement learning (DRL)-based control strategy for a tuned variable inertial mass electromagnetic transducer (TVIMET) installed in a five-story base-isolated building. The proposed controller is developed using recurrent proximal policy optimization (RPPO) with long short-term memory (LSTM) networks to account for both the partially observable nature of practical structural control and the history-dependent nonlinear behavior of the lead–rubber bearing (LRB) isolation system. Numerical simulations are conducted under a wide range of earthquake excitations, and the proposed method is compared with passive base isolation, a passive tuned viscous mass damper (TVMD), and a semi-active tuned inertial mass electromagnetic transducer (TIMET). The proposed TVIMET successfully adapts both the inertance and damping coefficient according to the changing dynamic characteristics of the structure. Unlike the passive TVMD, which is designed assuming post-yield structural behavior, the proposed controller remains effective even under earthquake excitations that do not induce significant yielding of the isolation layer. It also achieves the best overall vibration reduction, particularly in terms of the RMS structural responses. These results demonstrate the effectiveness of simultaneously adapting the inertance and damping through DRL for nonlinear structural vibration control.