Evolving urbanisation policies: Using a statistical model to accelerate optimisation over agent-based simulations

Agent-based systems are commonly used in the geographical land use sciences to model processes such as urban growth. In some cases, agents represent civic decision-makers, iteratively making decisions about the sale, purchase and development of patches of land. Based on simple assumptions, such systems are able broadly to model growth scenarios with plausible properties and patterns that can support decision-makers. However, the computational time complexity of simulations limits the use of such systems. Attractive possibilities, such as the optimisation of urban growth policies, tend to be unexplored since the time required to run many thousands of simulations is unacceptable. In this paper we address this situation by exploring an approach that makes use of a statistical model of the agent-based system's behaviour to inform a rapid approximation of the fitness function. This requires a limited number of prior simulations, and then allows the use of an evolutionary algorithm to optimise urban growth policies, where the quality of a policy is evaluated within a highly uncertain environment. The approach is tested on a typical urban growth simulation, in which the overall goal is to find policies that maximise the 'satisfaction' of the residents. We find that the model-driven approximation of the simulation is effective at leading the evolutionary algorithm towards policies that yield vastly better satisfaction levels than unoptimised policies.