- author = "David Rushing Dewhurst",
- title = "Essays on modeling and analysis of dynamic sociotechnical systems",
- school = "The University of Vermont",
- year = "2020",
- address = "USA",
- month = may,
- keywords = "genetic algorithms, genetic programming, Specializing in Complex Systems and Data Science, ABM, shocklet transform, game theory",
-
URL = "
https://cdanfort.w3.uvm.edu/research/david-dewhurst-phd-dissertation.pdf",
- size = "249 pages",
-
abstract = "A sociotechnical system is a collection of humans and
algorithms that interact under the partial supervision
of a decentralized controller. These systems often
display intricate dynamics and can be characterised by
their unique emergent behavior. In this work, we
describe, analyze, and model aspects of three distinct
classes of sociotechnical systems: financial markets,
social media platforms, and elections. Though our work
is diverse in subject matter content, it is unified
though the study of evolution-and adaptation-driven
change in social systems and the development of methods
used to infer this change.
We first analyse evolutionary financial market microstructure dynamics in the context of an agent-based model (ABM). The ABM matching engine implements a frequent batch auction, a recently-developed type of price-discovery mechanism.We subject simple agents to evolutionary pressure using a variety of selection mechanisms, demonstrating that quantile-based selection mechanisms are associated with lower market-wide volatility. We then evolve deep neural networks in the ABM and demonstrate that elite individuals are profitable in back testing on real foreign exchange data, even though their fitness had never been evaluated on any real financial data during evolution.
We then turn to the extraction of multi-timescale functional signals from large panels of time series generated by sociotechnical systems. We introduce the discrete shocklet transform (DST) and associated similarity search algorithm, the shocklet transform and ranking (STAR) algorithm, to accomplish this task. We empirically demonstrate the STAR algorithm invariance to quantitative functional parameterisation and provide use case examples. The STAR algorithm compares favorably with Twitter anomaly detection algorithm on a feature extraction task. We close by using STAR to automatically construct a narrative time-line of societally-significant events using a panel of Twitter word usage time series.
Finally, we model strategic interactions between the foreign intelligence service (Red team) of a country that is attempting to interfere with an election occurring in another country, and the domestic intelligence service of the country in which the election is taking place (Blue team). We derive subgame-perfect Nash equilibrium strategies for both Red and Blue and demonstrate the emergence of arms race interference dynamics when either player has all-or-nothing attitudes about the result of the interference episode. We then confront our model with data from the 2016 USA presidential election contest, in which Russian military intelligence interfered. We demonstrate that our model captures the qualitative dynamics of this interference for most of the time under study.",
-
notes = "Is this GP?
Defense Date: February 14th, 2020
Dissertation Examination Committee: Peter Sheridan Dodds, Ph.D., Advisor: Christopher Danforth, Ph.D.
Safwan Wshah, Ph.D. Nicholas Allgaier, Ph.D.
Chair: Cynthia J. Forehand, Ph.D., Dean of Graduate College",
