Title: Decision-based model selection

Abstract

In data-driven optimization problems, simple mathematical models that discard important factors may sometimes be preferred to more realistic models. This may occur if the parameters of the simple model are easier to estimate than the parameters of the complex model, or if the optimization problem corresponding to the simple model can be solved exactly whereas the optimization problem corresponding to the `realistic model' is intractable. This trade-off between three sources of errors (modeling, estimation, and optimization errors) is encountered in many stochastic optimization problems. 
The question we address is: how can one determine if it is better to use a simplified model, rather than a more realistic model? In other words: given a particular optimization problem and a data set at hand, how do we know whether the model-misspecification error of a simple model is dominated by estimation and optimization errors of more realistic models? 

Location: KdVI meeting room, Science Park 107, room F3.20