Control Force Characteristics and Seismic Control Performance Produced by Deep Reinforcement Learning

Structural control plays a critical role in protecting civil structures from earthquakes and other external disturbances. Among various strategies, active control has been widely studied, which uses actuators to apply counteracting forces based on control algorithms. Instead of traditional control theories, recent advances in machine learning have motivated the exploration of deep reinforcement learning (DRL) as a new paradigm for active structural control. This study investigates the feasibility of DRL‐based seismic response mitigation, focusing on whether DRL can realize control force characteristics and response reductions consistent with the design intent of structural control engineers. In this research, the proximal policy optimization (PPO) algorithm is adopted as a representative DRL method suitable for continuous control tasks. The training environment incorporates domain randomization in ground motion generation using a Kanai–Tajimi filter, enabling the agent to adapt to diverse seismic excitations. To verify the effectiveness of the proposed approach, three numerical examples are examined, including single‐ and multistory structural models with one or two active bracing systems. Numerical simulation results demonstrate that the trained controllers achieved significant reductions in story displacements, interstory drifts, and accelerations, while generating force–displacement hysteresis loops that reflected the intended reward design. Depending on the reward formulation, the controllers also exhibited restoring‐force characteristics resembling negative stiffness, demonstrating the flexibility of DRL‐based approaches in capturing diverse structural behaviors. Furthermore, the controllers maintained robust performance against a wide range of previously unseen disturbances. These findings highlight DRL and PPO, in particular, as a promising framework for next‐generation active structural control under seismic loading.