A Preliminary Study of Regularization Framework for Constructing Task-Specific Simulators

Dilara Aykanat, Zeyu Zheng, Nian Si

Abstract: One approach to construct or calibrate simulators, when representative real data exist, is to ensure that the synthetic data generated by the simulated match the empirical distribution of the real data. However, such approach to construct simulators does not take into consideration where the constructed simulators will be used. For some applications, there are clear tasks (such as performance evaluation of different decisions) in users’ mind where the simulated data will serve as input to the tasks. In this work, we propose an approach to use the knowledge of these tasks to guide the construction of simulators, in addition to the distribution match of simulated data and real data by regularizing the objective function with a task related penalty. We conduct a preliminary numerical study of this approach to illustrate the effectiveness compared to not taking into consideration the specific tasks of the simulators.