Improving scalability of large scale agent-based simulations

With the increase of our cities and urban environments being connected to the internet, we see a significant increase in the scale of these, so called, Internet of Things or Smart City environments. These environments are characterized by a large amount of heterogeneous devices and actors. This makes analyzing and testing of novel applications within these environment difficult. This is because these solutions can only be properly evaluated when deployed at the full scale of the environment, taking the full heterogeneity and complexity of the real environment into account. Deploying these solutions untested is too risky. Instead dfferent approaches are required. To cope with this complexity we see that modeling and simulation techniques can help. More specifcally, Agent Based Simulation (ABS) is a paradigm that is well-suited to simulate this type of environments, which consist out of many autonomous devices and actors. These techniques allow the behavior of these environments to be analyzed and evaluated. Furthermore, it allows for the creation of a virtual copy of the environment, which can be used to optimize certain behavior within this environment. In the last years we have seen examples of this in practice under the name of "Digital Twins". Such a Digital Twin, relies on real-time data originating from Internet-Of-Things devices, combined with simulated environments of its physical counterpart. Many cities utilize digital twins as a representation of the current state of their city. City related properties such as traffic behavior, noise and air pollution can be easily accessed. These virtual environments can even control certain parts of the environment, which could theoretically allow to dynamically optimize the aforementioned properties. But this is currently only possible in theory because modeling and simulation of such environments is strongly limited by its scaling capabilities. This is because they rely in most cases on centralized, monolithic simulation architectures, with statically defined computationally ineffcient models. In this thesis, we tackle the scalability capabilities of large-scale simulated environments. We start with an in-depth analysis of building and testing simulated large-scale Internet of Things environments. Based on this analysis we identify two opportunities. First, we look at optimizing the partitioning of distributed agent-based simulations. In the second we look at abstraction techniques that allow to switch abstraction levels of simulation regions. This enables us to balance between various levels of detail and levels of computational cost. Furthermore, we develop a generic methodology that allows dynamic optimization of simulation partitioning and model abstraction levels. We implemented and validated these techniques in a custom developed state-of-the-art traffic simulator, which we rigorously discuss throughout this thesis.

Publication year:2021

Keywords:Doctoral thesis

Accessibility:Open