Agent-based modeling is a new and deeply powerful tool to understand the behavior of systems with many interacting parts.  The essence of the approach is that the salient behaviors of actors are encapsulated in software modules as agents who interact with the environment and with each other in a computer simulation.  The properties of the agents may be derived from empirical data or they may draw on the assumptions of the modeler or policy-maker.   An agent-based model can be used  to explore the consequences of policy decisions under diverse scenarios. Such models may lead to the development of more fundamental theories about the system in question, but their main advantage is that they make it possible to study many problems even when such theories are not available.

 

Most importantly for our purposes, agent-based models make it possible to study the future consequences of current choices in a wide variety of situations, and to explore how these consequences depend on other factors. Agent-based models have been successfully used  in a variety of contexts that were previously beyond quantitative study, ranging from models of the stock market to the outbreak of revolutions.  We are at the outset of a new era in which agent-based models will increasingly become used as the means to study a variety of policy issues.  These new technologies offer decision makers deeper understanding, greater precision, and the capacity to see the distribution of probable outcomes resulting from a course of action.

A simple example gives the flavor.  We are all familiar with traffic jams, particularly the irritating ones in which “nothing” has happened. One is slowed to a crawl or stop and start pace which may last for half a mile, and the end of which the traffic resumes normal speed. No accidents – on either side of the road, no partial obstructions to block traffic flow…nothing. Then what happened?

 

A simple agent-based model shows what happened. Model a car and its driver as an agent. Populate a model road with different densities of cars, traveling at different speeds, one speed for each case.  Now endow each driver with some simple rules: 1) If the car in front of me is slowing, then I too will slow – but my reaction time induces a delay in my hitting the brake. 2)  If I am far from the car in front of me, speed up a bit faster than the average velocity of the traffic on the road.

 

Here is what happens.  At low densities of cars, and slow mean speed, when the car in front of you slows, your reaction time is very fast compared to the time it takes to slow your car by using the brake. When the density is high, so you are close to the car in front of you and are traveling fast, the reaction time delay is a significant part of the total time it takes to apply the brakes and slow the car. Indeed, the delay is so great that you are forced to apply the brakes more firmly than did the car in front of you. So you slow down more abruptly and, thanks to reaction time again, tend to slow to a slower speed than did the car in front of you.  This dynamics leads to a backward propagating “wave” of cars that slow ever more abruptly until the stop and start traffic jam forms.

 

Now this verbal description can be captured in such an agent-based model, which rather delightfully exhibits the traffic jam.  Further the model allows one to explore the relation between reaction time, traffic density and mean speed and the onset of traffic jams at and beyond critical levels of these “control parameters”.  Ultimately, models along the lines described here are being used by traffic engineers to design roads and traffic rules that minimize traffic jams.  

 

You would think it would be easy to model this with differential equations, or some other standard mathematical formalism.   Surprisingly, for this simple system, and systems of much greater complexity, it is quite difficult to model the system using standard mathematical tools.

 

Thus, agent-based models are emerging as the methodology of choice to study the “collective emergent properties” of complex systems of many interacting parts.  At base, there is almost a new epistemology at work. The agent-based models allow us to delve into the detailed causal structure of the parts and their interactions, then study how the global system behaves.  AFTER we see the emergent collective behavior we may be able to derive some mathematical approximation to these collective behaviors, but this is a very non-trivial task.

 

The wider ramifications of agent-based models is that they can be used in a variety of contexts that were beyond quantitative study before, ranging from detailed interacting models of the stock market, to analysis of conditions leading to the outbreak of revolutions.  As such, we are at the outset of a new era in which agent based models will increasingly become used as the means to study a variety of policy issues that were formerly the province of verbal or war game approaches.  It is likely that these new technologies will bring a whole new rigor to the specification and precision of policy making, such that expected and unexpected consequences, dependence upon control parameters, and collective emergent behavior are all open to detailed simulation.