Game Design Tools for ML Data Generation in CPS

Abstract

The high complexity of Cyber-Physical Systems (CPS) necessitates novel approaches for system analysis, planning, anomaly detection and testing. Machine Learning (ML) methods are promising because of their ability to find underlying relations even in large, complex and conflicting data. While existing CPS produce large data sets, these might not cover the appropriate time frame, or the desired configuration. Therefore, the use of ML methods requires the use of simulation tools to generate the necessary data. There are numerous approaches to simulate CPS. However, they often have significant shortcomings regarding their expressiveness in regards to physical properties of system components, their scalability in the face of the ever-increasing complexity of CPS, their usability for simultaneous simulation of different aspects of CPS and interoperability between different simulation environments. Game and media creation tools have seen an impressive development in recent years with regards to their realistic representation of physical systems and simulation capabilities. These are already employed in some engineering challenges like training of algorithms for self-driving cars. They have huge potential for the application in simulation and analysis of CPS. In this work we provide an analysis of the shortcomings of currently used environments for modeling and simulation of CPS with regards to creating data for ML. We then analyze how currently existing limitations can be overcome by employing tools from game and media design, discussing possible use cases and applications of these tools. With this, we present a possible new direction of research which has the potential to improve modeling of CPS, especially with regards to their application for ML.