The optimization of steel structures encompasses a range of design tasks, including section and topology optimization, as well as the optimization of detailing and device parameters. For steel structures, section optimization is a representative task in which component sections are selected, adjusted, and verified through structural analysis and code-compliance checks. Traditional optimization algorithms have been widely used to identify feasible and economical designs, while surrogate-assisted methods have also been implemented to reduce the cost of repeated structural evaluations. However, these approaches are generally organized around case-specific searches rather than around the learning of reusable decision policies across related structural instances. Reinforcement learning (RL) provides acomplementary perspective by learning optimization strategies through interaction with structural environments. This review examines the use of RL for the intelligent optimization of steel structures, focusing on the section optimization of steel frames. First, the scope of the review is defined by organizing the main optimization tasks and forms of design variables in steel structures. Traditional optimization and surrogate-assisted methods are then reviewed as methodological foundations for single-case searches, constraint handling, and accelerated evaluation. Building on these foundations, RL-based studies are synthesized from the perspectives of structural-family formulation, environment-based policy learning, and generalization evaluation. The key barriers to developing code-compliant and engineering-oriented RL methods for steel-structure optimization are also discussed. This review provides a structured synthesis for understanding the role of RL in structural optimization and for developing more generalizable, verifiable, and engineering-oriented intelligent design methods.