PhD thesis, University of York, 2014

## Abstract

“Essentially, all models are wrong, but some are useful.” George E. P. BoxThis thesis explores the role of modelling and computational simulation, in relation to social systems, with specific focus on Schelling’s seminal models of segregation. It discusses the role of computational modelling and some techniques that can be used in the Social sciences. Computational modelling is potentially powerful; however, the complexity of the domains being modelled means caution is needed. Simulation of social interaction has consistently created debate in the Social sciences. From the differential equations of predator-prey models, to Agent Based Models of Schelling’s segregation, most models are dismissed as either too simplistic or unrealistic. In an attempt to counter these criticisms, more complex models have been developed. However, by increasing the complexity of the model, the underlying dynamics can be lost. Schelling’s models of segregation are a classic example, with much of the work building on his simple segregation model. The complexity of the models being developed are such that, real world implications are being inferred from the results. The Complex Systems Modelling and Simulation (CoSMoS) process has a proven track record in developing simulations of complex models. In a novel application, the CoSMoS process is applied to Schelling’s Bounded Neighbourhood Model. The process formalises Schelling’s Bounded Neighbourhood Model and, from the formalisation, develops a simulation. The simulation is validated against the results from Schelling’s model and then used to question the model. The questioning of the model is an attempt to examine the underlying dynamics of the segregation model. In this respect, two measures, static, regarding the final population mix (

M), and dynamic, counting the number of iterations (I), are used in the analysis of the results. In the initial experiment, the effect of ordered movement was tested by changing the movement, from ordered to random. A second experiment examined agents’ perfect knowledge of the system. By introducing a sample size (S), the agents’ knowledge of the system is reduced. The third experiment introduced a friction parameter (F), to examine the effect of ease of movement into and out of the neighbourhood. In the final experiment, Schelling’s model is recast as a network model. Although the recasting of the model is slightly unorthodox, it opens the model up to network analysis. This analysis allows the easy definition of a ‘social network’ that is overlaid on Schelling’s ‘neighbourhood network’. Two different networks are applied, Random and Small World, with the size of the network controlled bym. The results of the experiments showed, that Schelling’s model is remarkably robust. Whilst the adjustments to the model all contributed to changes in the output, the only significant difference occurred when the social network was added.

**Full thesis** 6.6MB PDF