Updating simulation model parameters using stochastic gradient descent

This paper presents a method to automatically improve simulation model accuracy by using a stochastic gradient descent algorithm. The proposed algorithm updates models' parameters based on data collected from the actual domain under investigation. Collected data and feedback are fed into the simulation model to get predictions. In a linear prediction model, this data, along with the predictions, would form a typical regression problem; however, the stochastic gradient descent algorithm was modified to update the simulation model parameters. A tunneling case study is presented here, and the results show that the proposed algorithm can decrease simulation error by more than 50%, even in the case of incomplete simulation models or a missing inter-relationship between elements. Besides improving initial models, this paper provides a new approach for achieving data-driven simulation models that are updated in real time based on feedback from the actual domain. • Proposing a data-driven approach to enhance simulation performance. • Using stochastic gradient descent algorithm to update simulation parameters. • Deriving mathematical model to calculate error rate from a simulation model. • Applying the proposed method in a real case study of a tunneling project. • Suggesting a new approach for fitting data to model a simulation.