|
author |
David Nesterov-Rappoport
| title |
The Evolution of Trust: Understanding Prosocial Behavior in Multi-Agent Reinforcement Learning Systems
| abstract |
This thesis looks into what factors contribute to intelligent agents making the decision to
cooperate with one another in social dilemma-like interactions. Using concepts from
game theory, artificial intelligence, and biology, the work explores what considerations
push interacting agents towards prosocial or antisocial strategies. Cooperative behaviors
form the backbone of social organization, furthermore understanding their governing
mechanics is of the utmost importance. To achieve this, a custom piece of software is
developed to enable experimentation in the domain, a number of advanced machine
learning models are trained, and research from across different disciplines is synthesized
into a single perspective. At the core of the quantitative research lies the stag hunt family
of games, played by reinforcement learning agents which try to maximize their average
number of points earned. By observing their learning behavior in relationship to
configuration parameters, ideas from past research are validated, future avenues for
exploration are identified, and concrete principles about these systems are unearthed. On
the way there, the thesis summarizes the academic foundation for its methods and tools,
explains how they work, and elaborates on how they are to be coupled into a single
consistent system. Lastly, the implications of the research are related to the human
context and framed in concrete terms.
| school |
The College of Liberal Arts, Drew University
| degree |
B.A. (2022)
|
advisor |
Emily Hill
|
full text | DNesterov-Rappoport.pdf |
| |