An Initial Implementation Of A Multi-Agent Transport Simulator For South Africa
Transport demand planning in South Africa is a neglected field of study, using obsolete methods to model an extremely complex, dynamic system composed of an eclectic mix of First and Third World transport technologies, infrastructure and economic participants. We identify agent-based simulation as a viable modelling paradigm capable of capturing the effects emerging from the complex interactions within the South African transport system, and proceed to implement the Multi-Agent Transport Simulation Toolkit (MATSim) for South Africa's economically important Gauteng province. This report describes the procedure followed to transform household travel survey, census and Geographic Information System (GIS) data into an activity-based transport demand description, executed on network graphs derived from GIS shape files. We investigate the influence of network resolution on solution quality and simulation time, by preparing a full network representation and a small version, containing no street-level links. Then we compare the accuracy of our data-derived transport demand with a lower bound solution. Finally the simulation is tested for repeatability and convergence. Comparisons of simulated versus actual traffic counts on important road network links during the morning and afternoon rush hour peaks show a minimum mean relative error of less than 40%. Using the same metric, the small network differs from the full representation by a maximum of 2% during the morning peak hour, but the full network requires three times as much memory to execute, and takes 5.2 times longer to perform a single iteration. Our census- and travel survey-derived demand performs significantly better than uniformly distributed random pairings of home- and work locations, which we took to be analogous to a lower bound solution. The smallest difference in corresponding mean relative error between the two cases comes to more than 50%. We introduce a new counts ratio error metric that removes the bias present in traditional counts comparison error metrics. The new metric shows that the spread (standard deviation) of counts comparison values for the random demand is twice to three times as large as that of our reference case. The simulation proves highly repeatable for different seed values of the pseudo-random number generator. An extended simulation run reveals that full systematic relaxation requires 400 iterations. Departure time histograms show how agents 'learn' to gradually load the network while still complying with activity constraints. The initial implementation has already sparked further research. Current priorities are improving activity assignment, incorporating commercial traffic and public transport, and the development and implementation of the minibus taxi para-transit mode. Copyright