COLLOQUIUM ON COMPUTATIONAL SOCIAL SCIENCE/DATA SCIENCES RESEARCH – Adding Narrative to ABMs and Identifying Which Ones Are Worth Reading – Brent Auble
COLLOQUIUM ON COMPUTATIONAL SOCIAL SCIENCE/DATA SCIENCES RESEARCH
Computational Social Science PhD Candidate
Department of Computational and Data Sciences
Adding Narrative to ABMs and Identifying Which Ones Are Worth Reading
Friday, March 29, 3:00 p.m.
Center for Social Complexity Suite, 3rd Floor Research Hall
All are welcome to attend.
One strength of agent-based models (ABM) is that they allow for collection of all details of the
characteristics and behavior over time of every individual agent and the environment, theoretically enabling
the analysis of micro-level interactions between individuals and within small groups. In practice, however, the
volume of raw data generated by each run of a model (thousands of which might be done to test a range of
parameters) makes it difficult to identify unusual agent behaviors and interactions, and analysis of models
typically ends up being done by aggregating data and reporting overall trends. One solution to the challenge of
falling back on aggregative analysis is to have the ABM itself generate narratives describing the behavior of
individual agents and the interactions of agents. I will argue that Humans tend to like stories and reading the
“life history” of an agent can be more easily understood than reviewing a series of numbers. Thus, a model that
generates narratives for each agent can allow for easier analysis of individual behavior. While narratives should
be easier to understand than purely numeric results, a model that generates potentially thousands of texts is not
likely to be more tractable than one that does not. A solution to this challenge is to identify those agents who
are most likely to be “interesting.” Interestingness is identified by selecting agents whose behavior is at the
extremes of expected values (e.g. the tails of a Gaussian distribution), and the most interesting agents are those
who are at the extremes of more than one variable. In addition, the model should select some agent narratives
whose behavior falls exactly where expected (e.g. the mean or median), in order to avoid biasing a researcher’s
analysis too greatly toward extreme results.
This talk will discuss techniques for generating narrative using examples from the Zero Intelligence Traders
ABM, Sugarscape in NetLogo, and more recent work on implementing narrative generation in the MASON
version of Sugarscape. In addition, it will discuss techniques for identifying the agents whose narratives are
most worth the time of a human to read.