Thank you for the feedback, Jeff! Reading back, the article indeed glossed over that component rather quickly. I have made some adjustments and added more examples; I hope that clarifies the concept.

The artificial constraints are typically designed manually, reflecting aspects of the post-decision state arising after the decision. As such, the values \phi_f stem from some (linear) function (e.g., number of containers after shipping). The corresponding weights \theta_f are learned based on observations.