Scalable Techniques for Behavioral Analysis and Forecasting

Loading...
Thumbnail Image

Files

Publication or External Link

Date

2011

Citation

DRUM DOI

Abstract

The ability to model, forecast, and analyze the behaviors of other

agents has applications in many diverse contexts. For example, behavioral models can be used in multi-player games to forecast an opponent's next move, in economics to forecast a merger decision by a CEO, or in international politics to predict the behavior of a rival state or group. Such models can facilitate formulation of effective mitigating responses and provide a foundation for decision-support technologies.

Behavioral modeling is a computationally challenging problem--real world data sets can contain on the order of 10^30,000 possible behaviors in any given situation. This work presents several scalable frameworks for modeling and forecasting agent behavior, particularly in the realm of international security dynamics. A probabilistic logic formalism for modeling and forecasting behavior is described, as well as distributed algorithms for efficient reasoning in this framework. To further cope with the scale of this problem, forecasting methods are also introduced that operate directly on time series data, rather than an intermediate behavioral model, to forecast actions and situations at some time in the future. Agent behavior can be adaptive, and in rare circumstances can deviate from the statistically "normal" past behavior. A system is also presented that can forecast when and how such behavioral changes will occur. These forecasting techniques, as well as any arbitrary time series forecasting approach, can be classified by a general axiomatic framework for forecasting in temporal databases.

The knowledge gained from behavioral models and forecasts can be employed by decision-makers to develop effective response policies. An efficient framework is provided for identifying the optimal changes to the state of the world to elicit desired behaviors from another agent, balancing cost with likelihood of success. These modeling and analysis tools have also been incorporated into a prototype decision-support system and used in several case studies of real-world international security situations.

Notes

Rights