PUMA: Multi-agent Modelling of urban systems

It is increasingly recognised that land use change processes are the outcome of decisions made by individual actors, such as landowners, authorities, firms and households. As multi-agent models provide a natural framework for modelling urban processes on the level of individual actors, Utrecht University, Eindhoven University of Technology and RIVM are developing PUMA (Predicting Urbanisation with Multi-Agents), a fully fledged multi-agent system of urban processes. PUMA consists of various modules, representing the behaviours of specific actors. The land conversion module describes farmers', authorities', investors' and developers' decisions to sell or buy land and develop it into other uses. The households module describes households' housing careers in relation to life cycle events (marriage, child birth, aging, job change etc.). The firms module includes firms' demography and their related demand for production facilities leading to location choice processes. The daily activity pattern module describes the trips made and locations visited by individuals to carry out certain tasks. This module generates aggregated effects of individual behaviours (congestion, pollution, noise), affecting households' or firms' longer-term location decisions. The paper describes the model system architecture and the interactions between the modules. Particular attention is devoted to the households module that includes a behaviourally sophisticated model of households' process of awakening (deciding to actively search for another dwelling), search and acceptance of an offered dwelling. This model was calibrated on the Dutch Housing Preferences Survey. Based on the disaggregate housing search and acceptance model, the households module describes housing market dynamics and indicates the demand for new dwellings per region. The paper describes the model specification and calibration in detail. The households module was implemented and tested for the Northwing of the Dutch Randstad, including about 1.5 million households and 1.6 million dwellings. The paper describes the implementation and the first model results