Predicting Flus and Fender Benders
How agent-based simulation modeling from Research can improve our lives
What if scientists could examine school policies and predict which ones should be enforced to best slow the spread of the flu among students? How can a highway patrol force predict which types of penalties would most reduce local accidents? What if there was a way for us to try out different energy-saving behaviors, and then compare the predicted results, before we invested in actually doing them? How many lives, and how much money and time could be saved?
These are some of the questions that researcher Segev Wasserkrug and his IBM Research - Haifa team had in mind when developing models for agent-based simulation. Agent-based simulation modeling is a form of computational modeling based on studying the behaviors of agents—entities independent from their environments (people, in most cases). The modeling focuses on the interaction among agents, the agents’ likely responses to various factors, and the observation that past behavior is not always an accurate prediction of future performance. With traditional modeling methods, scientists study various groups within a population, based on history and on certain assumptions about behavior.
Modeling traffic trends
For traffic analysis, traditional modeling might involve studying the history of accidents for a particular stretch of road and applying those statistics to predict the number of accidents for the next year.
Certain trends would emerge over time—more accidents in the summer than in the winter months, or a ten percent increase in overall accidents each year, for example. That information could be somewhat useful in making decisions about when or how much to increase police presence each year; however, this type of modeling has its limitations. The model is based on the assumption that increasing police presence would result in drivers slowing down and driving more carefully, ultimately reducing the number of accidents.
Yet people do not always behave rationally, and if we don't have statistics showing specifically how increased police presence on that particular stretch of road affected the accident rate in past years, we can't predict that doing it now would have the desired effect. The more complex the factor we are interested in checking, the more limited we are by traditional modeling. If we want to compare types of police presence, such as marked or unmarked cars, or even types of penalties, like fines or license revocation, then a model based on past statistics isn't much help.
Agents to the rescue
Agent-based simulation modeling takes into account that not all agents behave the same way in response to the same factors, and that relationships among agents are more complex than can be demonstrated by traditional modeling. An agent-based simulation model divides the drivers into groups based on behavior (cautious drivers and aggressive drivers, for example), and then adds different types of interaction, such as horn honking or tailgating. A cautious driver caught in front of a tailgater or in front of a honking driver, might simply signal and move lanes, while a more aggressive driver might speed up even slam on the brakes. When many types of drivers along with many types of interactions are represented, an agent-based model can capture a realistic picture of the particular environment and population being studied.
At that point, outside factors, such as increased police presence, could be added into the mix. Rather than predicting how increased police presence would affect the environment (as in traditional models), an agent-based simulation model computes how each of the various groups will react to the change, based on their known qualities as a group, and then ultimately predicts an emergent behavior—how those reactions would influence the environment. Once the model is in place, it can be calibrated to changing conditions, and then rerun to predict the future reactions for each program in question.
US National Institutes of Health project known as MIDAS (Models of Infectious Disease Agent Study) agent-based modeling was used to help shape avian flu (H5N1) policy, by studying the ways in which diseases spread. The models were used to understand how people move around—children going to school, adults commuting to work, families traveling between cities, as well as how the different groups react to symptoms of a disease, often becoming less mobile as they become sicker. In addition to tracking the predicted spread of the disease, the models could check countermeasures as well. Airport closures, for example, may stop some people from traveling, but may encourage others to travel by train or bus. The same types of technologies will be essential in understanding the spread of other pandemics, such as swine flu (H1N1) and other emerging infectious diseases.In a
Agent-based simulation modes are already being used in the health, security, disaster-response, and energy-consumption fields. Because the models are graphic in nature, they can be easily read by policy-makers (who are not necessarily researchers) and the findings can be quickly interpreted and put into action. As the technology advances, the already widespread applications will continue to expand.