A review published in Smart Construction examines how reinforcement learning (RL) could help optimize steel structures by learning design strategies that may be reused across related structures. Focusing on section selection in steel moment-resisting and braced frames, the review compares RL-based approaches with traditional and surrogate-assisted methods and emphasizes that practical use will require robust code-compliance assurance, manageable computational cost, validated cross-structure generalization, and independent engineering verification.
Steel structures are assembled from standard sections, but choosing those sections is anything but routine. A lighter design can reduce material use and cost, yet every beam, column, and brace must still satisfy requirements for strength, stability, stiffness, drift, and detailing. Because member choices are discrete and structurally coupled, changing one section can redistribute forces across the system. Each candidate design may also require finite element analysis and code checks under multiple load cases, making the search space large and expensive to explore.
Metaheuristic methods such as genetic algorithms have long been used to search this space. Surrogate models can reduce the burden by estimating structural responses, feasibility, or design objectives before expensive analyses are run. Some can evaluate different structures within a defined design domain, but their primary role remains to accelerate evaluation and screening. RL places greater emphasis on learning how a design should be adjusted in a given state. A further question is whether the resulting policy can be reused across related structures.
To assess the promise and limits of that approach, researchers from Tongji University and the University of California, Berkeley, conducted a structured review of RL for steel structure optimization. The database searches yielded 4604 records, and 97 papers were ultimately included after screening, citation tracing, and additional manual searches. The review focuses on section optimization in steel moment-resisting and braced frames while also covering topology optimization and the optimization of detailing and device parameters.
Rather than treating all optimization studies as directly comparable, the review first organizes them by design variables, design spaces, objectives, and constraints. This distinction matters because choosing a section from a catalog, changing a load-carrying topology, and tuning a connection or damping device are different engineering tasks. Section optimization receives the greatest attention because it connects directly to common design workflows: once a structural layout is fixed, member or member-group sections are selected, analyzed, checked, and revised.
Within an RL framework, that iterative workflow becomes a sequence of decisions. The agent uses information about the current design and its performance to modify selected sections, then learns from the results of subsequent analysis and code checks. The review examines how this interaction is modeled, how structural information is represented, and how design decisions are formulated in single- and multi-agent settings.
One of the review's central findings is that the value of RL depends on how generalization is defined and tested. A policy trained for one structure may produce fast recommendations at inference time, but its training still requires repeated analyses and code checks. If a new policy must be trained for every design case, the training expense must be counted as part of the optimization cost for that case. Only when a policy performs reliably across a clearly defined family of structures can a larger upfront training investment translate into faster subsequent deployment.
That distinction also affects how the methods should be evaluated. Feasibility and material efficiency alone are not sufficient; results should be interpreted against appropriate baselines and in light of the computational effort required for both training and deployment. When a policy is intended to transfer across structures, its ability to generalize should be tested over clearly defined training and test domains that reflect meaningful differences in structural configuration, loading, and available design choices.
Code compliance is another critical issue. Although reward penalties and action limits can guide an agent toward feasible designs, they cannot guarantee that the resulting designs satisfy every applicable code provision. The review therefore treats RL as a design-support technology rather than a substitute for engineering verification. Before being used in practice, candidate designs must still be independently verified through high-fidelity finite element analysis and deterministic code checks, followed by professional review.
Future work should make RL training more efficient, improve decision-making in high-dimensional design spaces, and strengthen coordination among multiple agents in large structural systems. Learned policies also need to be tested across clearly defined structural families, including cases outside the ranges encountered during training. Surrogate and multi-fidelity models may help reduce analysis costs, provided that their errors near constraint boundaries are controlled. In practice, RL is most likely to operate within an auditable workflow that combines automated design exploration with high-fidelity verification and engineering judgment.
Overall, the review presents RL as a promising but still developing approach to steel structure optimization. Its engineering value will depend on whether learned policies can be reused reliably across related structures.
This paper was published in Smart Construction.
Gao Y, Lu Y, Du M, Wang W, Zhou G. Reinforcement learning for intelligent optimization of steel structures: a review. Smart Constr. 2026(3):0014, https://doi.org/10.55092/sc20260014.
DOI: 10.55092/sc20260014
Smart Construction
Literature review
Not applicable
Reinforcement learning for intelligent optimization of steel structures: a review
6-Aug-2026