Combining evolutionary algorithms and agent-based simulation for the development of urbanisation policies

Urban-planning authorities continually face the problem of optimising the allocation of green space over time in developing urban environments. To help in these decision-making processes, this thesis provides an empirical study of using evolutionary approaches to solve sequential decision making problems under uncertainty in stochastic environments. To achieve this goal, this work is underpinned by developing a theoretical framework based on the economic model of Alonso and the associated methodology for modelling spatial and temporal urban growth, in order to better understand the complexity inherent in this kind of system and to generate and improve relevant knowledge for the urban planning community. The model was hybridised with cellular automata and agent-based model and extended to encompass green space planning based on urban cost and satisfaction. Monte Carlo sampling techniques and the use of the urban model as a surrogate tool were the two main elements investigated and applied to overcome the noise and uncertainty derived from dealing with future trends and expectations. Once the evolutionary algorithms were equipped with these mechanisms, the problem under consideration was defined and characterised as a type of adaptive submodular. Afterwards, the performance of a non-adaptive evolutionary approach with a random search and a very smart greedy algorithm was compared and in which way the complexity that is linked with the configuration of the problem modifies the performance of both algorithms was analysed. Later on, the application of very distinct frameworks incorporating evolutionary algorithm approaches for this problem was explored: (i) an ‘offline’ approach, in which a candidate solution encodes a complete set of decisions, which is then evaluated by full simulation, and (ii) an ‘online’ approach which involves a sequential series of optimizations, each making only a single decision, and starting its simulations from the endpoint of the previous run