Deep Q-networks for an optimal placement of viscous dampers in 2D multi-story linear frames

This study proposes a reinforcement learning-based framework for the optimal placement of viscous dampers in multi-story frames, with a focus on enhancing structural performance under seismic excitations. A deep Q-Network (DQN) algorithm is developed to identify the most effective damper configurations across the structure, aiming to minimize peak interstory drift. In the literature, different solutions have been proposed mainly focusing on the characteristics of the structural system rather than on the variability of single seismic excitations. The proposed approach is designed to cope with the non-linear and highly variable nature of the cost function, which poses significant challenges to traditional optimization techniques. Therefore, the effectiveness of the DQN approach has been investigated by considering 85 seismic ground motions having various characteristics with the aim of defining the optimal arrangement of the same number of viscous dampers in a ten-story benchmark frame, with respect to the features of each record. To verify the effectiveness of the proposed algorithm, the results were, first, compared with those achievable through a simple uniform distribution strategy for damping, subsequently, tested to ensure their correspondence with the minimum points of the cost function by means of a sensitivity analysis. Finally, the DQN approach was further compared with three optimization methods: gradient descent, genetic algorithm and particle swarm optimization. The results confirm the potential of the reinforcement learning DQN algorithm as an effective optimization tool within performance-based seismic design.