Spatio-Temporal Game Theory & Real-Time Machine Learning for Advesarial Groups
The rise of extremist organizations such as ISIS and Jabhat al-Nusra represents a significant threat in both local and international settings. This threat may include one-off, large scale attacks, but more typically involves persistent day-to-day conflict. Such groups are often fragmented, heterogeneous and diffuse in their organization and their priorities. Individuals from different extremist organizations may cooperate for tactical reasons, but they may also engage in violent competition with one another, creating an ever shifting landscape of rivalries.
Extremist groups are surprisingly dynamic or adaptive, emerging and dissipating under directed pressure and/or in phase with broader social dynamics. Such organizations, of course, are able to project fear and actual violence at scales that grossly exceed their actual size. However, the atomized nature of the threat also means that both prediction of short-term risk and estimation of long-term trajectories for such groups is very challenging.
Integrated Methodological Approach
Threats posed by extremist organizations demand a robust response grounded in a rich qualitative understanding of how such organizations emerge and function as well as a quantitative basis for predicting their behavior. Specifically, our proposed research will reveal exactly how social, spatial and situational strategic factors underlie the violent action of extremist groups. The models and methods developed will be essential for predicting adversarial behavior and estimating long-term trajectories of adversarial groups.
Our approach is innovative in combining a diverse empirical record of culture, social organization and behavior, experimental field methods in ‘wild’ settings and rigorous mathematical and computational methods. Our highly integrated methodological approach includes three principal tasks: (1) ethnographic study of criminal street gangs; (2) experimental testing of gang decision making with mobile game theory; and (3) machine learning for predicting gang behaviors and trajectories, particularly in the context of algorithmic game theory.
Game Theory in the Wild
We propose to develop formal mathematical and computational methods to deal with the strategic decision making adversarial individuals and groups. We will deploy several smartphone-based mobile games to examine how cooperating and non-cooperating adversaries act in the face of pervasive imperfect information. We will develop and smartphone-based mobile platform for game theoretic experimentation in real-world settings. We will experimentally manipulate strategic match-making (e.g., pairing cooperators with non-cooperators), the spatial location of game play (e.g., inside or outside one’s own gang territory) and the temporal dynamics of game play (e.g., the speed with which a partner responds provides information about strategic orientation). Inputs from machine learning models will assist in building a rich substrate on which game theoretic models will build descriptions of human player behaviors.
We will use open-ended and semi-structured interviews to probe issues of within-group trust and its role in autonomy of action, mistrust and antagonistic interactions with rival gangs, the quality and quantity of information that gang members have about the actions of their own group and that of rivals, especially police as asymmetrical rivals, and how their willingness to use violence varies with both the surrounding audience and where in the socio-cultural landscape the opportunities for action arise. Ethnographic data will be incorporated into game theoretic and machine learning models of adversarial action.
Machine Learning Adversarial Behavior
We propose to develop novel machine learning methods focused on inferring societal-scale adversarial behavior given sparse, heterogeneous socio-cultural and environmental data. We will develop methods for training deep neural networks to discover, model, and disentangle the components of latent strategy phenotypes. Our machine learning approach will be able to infer contextual states from heterogeneous data, learn adversarial strategies from sequential activity data in games, and transfer adversarial strategies learned from game play to practical scenarios.
Our team brings to the project expertise in the ethnographic study of gangs (Jorja Leap, Co-PI), analysis of criminal event patterning and predictive policing (Jeff Brantingham, Co-PI), game theory for security problems (Milind Tambe, PI), and machine learning for latent variable inference and anomaly detection (Yan Liu, Co-PI).