Architectures for Combining Discrete-event Simulation and Machine Learning

Andrew Greasley


A significant barrier to the combined use of simulation and machine learning (ML) is that practitioners in each area have differing backgrounds and use different tools. From a review of the literature this study presents five options for software architectures that combine simulation and machine learning. These architectures employ configurations of both simulation software and machine learning software and thus require skillsets in both areas. In order to further facilitate the combined use of these approaches this article presents a sixth option for a software architecture that uses a commercial off-the-shelf (COTS) DES software to implement both the simulation and machine learning algorithms. A study is presented of this approach that incorporates the use of a type of ML termed reinforcement learning (RL) which in this example determines an approximate best route for a robot in a factory moving from one physical location to another whilst avoiding fixed barriers. The study shows that the use of an object approach to modelling of the COTS DES Simio enables an ML capability to be embedded within the DES without the use of a programming language or specialist ML software.


Paper Citation