Enhancing Realism in Simulation Through Deep Learning

Muhammad Shalihin Bin Othman, Gary Tan

Abstract: Modeling and simulation have been around for years and its application to study several different systems and processes have proven its practical importance. Various research has sought to optimize its performance and capabilities, but few address the issues of generating realistic inputs for simulating into the future. In this paper, some issues in the commonly used simulation flow were identified and deep learning was introduced to enhance realism by learning historical data progressively, so as to generate realistic inputs to a simulation model. We focus on improving the input generation phase and not the model of the system itself. To the best of our knowledge, this is the first work that realizes the possibility of integrating deep learning models directly into simulation models for general-purpose applications. Experiments showed that the proposed methods are able to achieve higher overall accuracy in generating input sequences as compared to current state-of-art.