A Network Approach to Scaling Agent-Based Model Outputs

Abstraction of reality through agent-based modeling has enabled us to analyze and explore previously intractable problems. Yet, there is always a trade-off between the scope of the model and the constraints of time and computational resources. Various approaches have been developed to help overcome these constraints in large-scale agent-based models. This dissertation reviews the existing methods that help simulate agent-based models at large sizes and proposes a new methodology based on network analysis to predict the change in the model's outputs with the change in its size. Using the model outputs produced at a smaller scale, along with measurements of the changes in the underlying agent connections, this methodology is able to reproduce the outputs of a large-scale model without the need to run the model at full size. This approach is demonstrated using several well-known agent-based models, including models of segregation, disease spread, and traffic congestion. These examples demonstrate the ability of this methodology to predict model outputs at a large scale across changes to inputs, network topologies, and agent behavior. The use of the presented methodology can help address the growing complexity of agent-based models by providing researchers with a way to obtain large-scale model results using the streamlined process described in this dissertation. This dissertation further explores the nature of scaling in agent-based modeling, using insights gained from network analysis to provide guidance on scaling models to produce the most representative results.

Author

• Malikov, Maxim A

• Agent-based models
• Modeling
• Networks
• Scaling
• Simulations

Identifier

hdl:1920/15126

OAI identifier oai:identifier

oai:MARS:1920/15126