Our objective was to design a pedestrian simulator that would provide reliable information to those responsible for making important decisions about the design and operating procedures of major public buildings. For that we needed it to accurately represent the way pedestrians behave.
We needed to know how they behave as commuters, tourists, spectators, shoppers, etc. We needed our model of pedestrian behavior to be generic, ie: to describe the behavior of all types of pedestrians (men, women, age, nationality, tourists, commuters, etc.) in all kinds of places & circumstances.
Measurements existed but they were mostly limited in scope, inadequate and/or inaccurate. Many measurements were anecdotal, ie: they were mostly measurements of particular people in particular places. So we needed to do it ourselves: to systematically observe and measure the behavior of a large quantity and variety of people in all the types of circumstance that we wished to simulate.
We needed to be able to simulate large crowds (eg: for the Olympics), so the model had to be accurate but simple enough to be scaled up. To achieve the variety of different behaviours we needed to simulate each pedestrian as an individual agent (with its own specific preferences, perception & patterns of behaviour) and that meant implementing our model in a multi agent system in which each agent (virtual pedestrian) would have its own (artificial) intelligence. That wasn’t simple.
Once we had built our multi agent system, the last (but not least) of the challenges was:
- to calibrate the behavior of the individual agents in a wide variety of circumstances against individual data to ensure accurate behavior in all the circumstances, and
- to validate the simulated crowd behavior (resulting from the interactions of the agents) against real crowd behavior (in identical circumstances).