Deep Agent: A Framework for Information Spread and Evolution in Social Networks

Home / Projects / Deep Agent: A Framework for Information Spread and Evolution in Social Networks

Social SimulationOverview:
In project Deep Agent (DARPA Social Sim program, $6.2M, PI Garibay), we build a comprehensive, realistic and at-scale computational simulation of information spread and evolution in online social networks using a novel computational modeling paradigm: Deep Agent Framework (DAF).
The Deep Agent Framework unleashes the power of combining massively parallel computing, data analytics of large datasets and machine learning into assisting model designers to mix and match sub models in a semi-automated way, exploring, testing and validating not one but tens of thousands of models against not a single real world phenomenon but a large set of target behaviors. This process will aid model designers to continuously improve the best-so-far model as the problem challenges become more difficult. In fact, this process would also serve to automatically recombine and introduce variants to all the models produced by different SocialSim performers in order to obtain the overall best model. This framework will enable the creation of an accurate and at-scale simulation of information spread and evolution that can run on a typical off-the shelf commercial computer or small cluster.

The Deep Agent Framework posits the following: (1) The modeling of social dynamics can be accomplished by a network of computational agents endowed with deep neurocognitive capabilities via emotional, cognitive, and social modules. This is a synthesis of Agent_Zero (Epstein, 2014) and Homos Socialis (Gintis and Helbing, 2015) frameworks. (2) Instead of the creation of a single hand-designed model of information spread and evolution, we create a family of modular sub-components from which multiple plausible models can be systematically assembled, tested and validated. These subcomponents will be created from both leading social theory-driven models and data-driven models. (3) The use of machine learning techniques to aid expert model designers and social scientists in our team in the computer-aided exploration of tens of thousands of competing models of information spread and evolution. The search is guided by model accuracy, as measured by comparing model simulated outputs with real-world social dynamics data.

Principal Investigators:

  • Dr. Ivan Garibay
  • Dr. Mirsad Hadzikadic
  • Dr. Georgios Anagnostopoulos
  • Dr. Stephen Fiore
  • Dr. Gita Sukthankar
  • Dr. Alexander Mantzaris
  • Dr. Christina Bouwens
  • Dr. Joseph Whitmeyer
  • Dr. William Rand
  • Dr. Joshua Epstein

University of Central Florida
University of North Carolina Charlotte
Florida Institute of Technology
North Carolina State University
University of New York
University of Chicago

Defense Advanced Research Projects Agency

$6.2 Million DARPA Grant Awarded to Dr. Ivan Garibay
$12.5M Grant Will Help Predict Spread of Online Social Behavior