Virginia Tech leads DARPA project to develop algorithms and software tools for next generation social science
Making social science experiments rigorous, reproducible, and scalable to large populations is a major objective of the Defense Advanced Research Projects Agency (DARPA) Next Generation Social Science (NGS2) program.
To support this program, the agency has awarded Virginia Tech and four other university partners a potentially three-year cooperative agreement for up to $3 million to conduct research that will streamline modeling processes, experimental design, and methodology in the social sciences.
The Discovery Analytics Center and the Network Dynamics and Simulation Science Laboratory in the Biocomplexity Institute of Virginia Tech have joined with Stanford University, Cornell University, Johns Hopkins University, and Northwestern University to create a collaborative experimental effort that leverages a number of disciplines, including sociology, economics, network science, statistical physics, machine learning, and agent-based modeling.
Naren Ramakrishnan, the Thomas L. Phillips Professor of Engineering in the Department of Computer Science and director of the Discovery Analytics Center, and Madhav Marathe, professor of computer science and director of the Network Dynamics and Simulation Science Laboratory, will lead the project with primary research scientists Brian Goode and Chris Kuhlman, from the center and lab, respectively.
“The goal of this project is to scale up social science modeling and experimental methodology to potentially include thousands of people,” said Ramakrishnan. “We will focus on the difficult task of understanding the formation and dynamics of collective identity – the sense of self defined as belonging to a group rather than as an individual.”
Such groups can be rooted in a political affiliation, a public health issue, religious beliefs, educational status, or race and ethnicity, to cite just a few.
Collective identity occurs when a person’s alignment with a group places the needs/views of the group above his/her own. For example, one may participate in a protest to feel a greater sense of belonging to a group and the ideals they support.
“Designing controlled experiments in social sciences is never easy. But social behavior and its manifestation within a given context, the ability to tease out confounding effects, and causation versus correlation create even greater challenges when you are working on a massive scale,” said Marathe.
The research team will use a novel approach that combines large-scale agent-based causal modeling, data-driven model development, and rigorous statistical methods. They will develop algorithms and software based on a variety of theories and design experiments mimicking situations where an individual has to make a decision for personal versus group betterment.
The team has chosen to initiate its research on a version of the group ultimatum game, in which two teams must agree to a deal for a certain award distribution. Participants will undergo a series of exercises designed to strengthen the alignment of individual efforts to that of the team.
In an ultimatum game, for instance, Tom receives a sum of money (say $100) to share between himself and Jerry. Jerry can either accept or reject Tom’s proposed breakup of the money. If he rejects it, neither of them pocket any money. One would think that a possible outcome is for Tom to take $99 and to give Jerry $1 (which Jerry might accept since it is better than not getting any money at all).
“More often than not, when we do this kind of experiment with two people, we find that more equitable deals are encouraged. If the offer is too low, it may be rejected, even though it is still ‘free money,’” Goode said. “We would like to know what happens if you have groups of people – rather than individuals – interacting in this manner. Will groups cooperate? Will some groups collude to form partnerships against others?”
The group ultimatum game is only one experiment the researchers aim to perform. Results of this and similar experiments will shed light on collective identity formation, dynamics, and effects on experimental outcomes, Goode said.
Marathe said that plans to study language used on social media, such as Twitter and Reddit, in future experiments will allow the team to examine increased scales of collective identity effects – in particular, forming a collective identity for or against a cause or purpose.
However, focus is also on creating a broader interface and bringing accessibility to these types of experimental processes.
“Formal computer science methods can be very useful in social sciences,” Marathe said. “We aim to achieve reproducibility by formally specifying the experiments, conditions, and methods and creating a digital library for storing provenance information and sharing our results with the larger community so that other experts and stakeholders can review and critique the work – like the MATLAB of experimental social science.”
Others from Virginia Tech supporting this DARPA project are Patrick Butler, senior research associate, Brian Mayer, project manager, and Nathan Self, research associate, Discovery Analytics Center; Stephen Eubank, deputy director, Samarth Swarup, research assistant professor, and Anil Vullikanti, associate professor, Network Dynamics and Simulation Science Laboratory; Gizem Korkmaz, research assistant professor, and Chris Barrett, professor and executive director, Biocomplexity Institute of Virginia Tech.