Home » Blog » ETech2008 » ETech2008: Paul Torens and modeling crowds

ETech2008: Paul Torens and modeling crowds

Dr. Paul M. Torrens from the Arizona State University School of Geosciences likes making people panic. Virtual people and virtual panic, that is. He studies crowd behavior, which involves putting virtual people into virtual disaster situations and seeing just what they do.

Crowds, he tells us, are pretty important to understanding social life, to understanding our “ambient social infrastructure.” They are instrumental in history’s most famous events, like storming the bastille. Crowds are complex adaptive systems – their swarming and flocking behavior helps us understand mob behavior, but also how people passing by retail storefronts pay attention to advertisements.

It’s hard to study this stuff in real life. “It’s very difficult to interview a rioter who’s throwing a molotov cocktail. “Excuse me, sir, on on a scale of 1-10, how upset are you right now?” So he builds models – simulations, used as alternative laboratories.

The state of the art in this research is coming from different fields: computer science, physics, movie special effects and urban design. The special effects designers want to find ways to create realistic crowds, composting computer-generated avatars against matte paintings. Physicists are interested in how crowds might serve as models for other adaptive systems, like particles in fluid flow. Urban designers want to know how line of sight moves people through space. And some scientists are interested in building “social force models” that model attraction and repulsion in crowd behavior.

Torrens wants to build scaleable, replicable toolkits to build these models. They’re agent-based – each character has its own motivations and acts autonomously, encountering other actors in the space. It allows modeling of meregence, path dependency and human factors like feedback and lock-in. This might be useful for builing AI models for crowd behavior in games, or for applications like designing better cities or buildings.

These models begin with motion capture data, filming people walking, running, or falling over. You can then use these models to introduce physics into the worlds – “You can try running rioters into steel walls, which is otherwise hard to get past the IRB board.” The motion capture models then get wrapped into textured envelopes, making them look realistic.

One major problem – the tools for building simulations are currently richer than the data sets we have. He’s capable of building very realistic looking and acting people, but modeling crowd behavior requires inferential statistics and other forms of intelligent estimation. Once he’s done this, he can map his virtual crowds into “geospatial exoskeletons” – real-world spaces modeled in code. We see one of these scenes in action: a highly realistic street scene, including young and old people, a few drunks, all intersecting in a crowded pedestrial boulevard.

This is cool, and looks great, but the hardcore stuff is the “extraordinary scenarios”. We watch a group of avatars attempt to find their way out of a crowded building. They’re very bad at it – they cluster in front of a doorway in a pair of wedges. The system works far better when a column is introduced off-center in front of the door. It’s counterintuitive, but it sends shock waves through the crowds to break up these patterns.

A much more complex situation attempts to figure out how police can control a crowd. A model includes aggrieved civilians, rebels trying to instigate rebellion and law enforcement, trying to keep the crowd from spinning out of control. He’s run base scenarios as well as ones that include tear gas, introducing panic in the rioters but allowing police to move unmolested. The idea is to learn from the models and figure out the signs of mob behavior before it breaks out. (This makes me a bit nervous, but not quite as nervous as his next project.) Another project models the city of Salt Lake City and tracks the path of three individuals based on their cellphone signals. Based on their path behavior, we can extrapolate where we might look at building CCTV surveillance tape to identify a set of kidnappers, for instance. (Or political dissidents. Hmmm.)

His work is extremely cool eye-candy with lots of fun implications behind it – I’m looking forward to learning more at Geosimulation.org.