An intelligent agent-based method for model updating of structures with multiple uncertain parameters via deep reinforcement learning

Abstract

The paper presents a new method for performing model updating (MU) of structures characterized by uncertain multiple parameters, based on a deep reinforcement learning (DRL) approach. In particular, the idea behind the proposed approach, named DRUM-A (acronym of Deep Reinforcement learning for Updating Model Approach), consists of defining an intelligent agent that can support analysts in performing fine-tuning of the numerical model to update to match numerical and experimental data, by searching correct values of uncertain parameters with high accuracy. DRUM-A was idealized and organized in five consecutive steps. Given data of the monitoring campaign, the definition of the target variables of interest, and a numerical model of the structure in an externally accessible environment, the proposed approach defines a DRL-engine, in which a strategy for tuning the structural control parameters was defined. Contextually, training/test phases of the intelligent agent were performed, to derive the correct solution according to a statistical-based evaluation of results. The paper provides a detailed description of DRUM-A and subsequently reports the application on a dummy structure (and the related comparison with the output provided by a deterministic and a probabilistic approach), and a real-life structure, accounting for multiple uncertain parameters and variables in different scenarios. DRUM-A represents a paradigm shift toward current MU approaches, since it allows managing an increasing number of unknown variables without requiring strict engineering assumptions. In addition, DRUM-A allows performing an online MU, providing accurate estimates of uncertain parameters, with reasonable time and computational efforts.