Many agent-based models require for their initialization a population which corresponds in its essential characteristics to the examined real population. This population should reflect the real distribution of agent attributes of interest, e.g., age, gender, marital status or income, as well as the social network between the agents. Since a disaggregated data set with all the required information is rarely available due to privacy concerns, a synthetic population needs to be created. Methods that assign realistic attribute values to agents are well studied in the literature. In contrast, the generation of a plausible social network has been less extensively researched yet, but several comprehensive adhoc models have been developed. The focus of this work is to introduce a reusable, generalized approach for the generation of a social network, that can be applied subsequently to the generation of synthetic individuals. The novel approach uses a symbolic regression, a technique from the field of genetic programming, to automatically train network generation rules based on a network sample, instead of having to define rules a priori. The rules are optimized until the statistical properties of the sample are approximated by a synthetically generated network. In addition, manually specified constraints can be taken into account to avoid implausible relationships.
Bemerkung: Die Masterarbeit wurde an der Nordakademie in Koooperation mit Prof. Jan Himmelspach angefertigt.