This enables an appropriate training of both ordinary and structural parameters of the model. Note that the preference towards high entropy distributions (fewer assumptions) applies only within the admissible set of distributions P'"Y consistent with the constraints. |