In this talk, we will discuss the formulation of subgrid-scale (SGS) models and wall models for large-eddy simulation as a control strategy using reinforcement learning. We train SGS models and wall models in turbulent zero-pressure gradient flat-plate boundary layers and smooth body separation over a Gaussian bump.  The non-deterministic control-strategy model allows for physically more realistic results that are not arbitrarily correlated with the input variables. We show that these models are successful as a predictive tool, even outside the training parameters. We discuss best practices in terms of state selection for model generalization and future steps.