Please use this identifier to cite or link to this item:
|A Machine Learning Approach to Guiding Altruistic Strategic Behavior
|Operations Research and Financial Engineering
|Cursory scans of social media forums, public platforms, and discussion boards show that our societies are currently facing the negative consequences of major global issues concerning both the social (e.g. racism, prejudice, etc.) and natural (e.g. climate change, environmental destruction, etc.) sciences. Though these problems manifest in very different ways, there is a common, recurrent structure that underpins these phenomena that hearkens back to the more general problems of game theory. This thesis seeks to use this game theoretic analysis to propose and establish the relevant root causes and identify promising solutions. The approach is two-fold. First, we consider the generic normal-form game, detached from real-world context. Specifically, we start with the premise that selfishness at the individual level in seeking maximum payoff per player results in systems with non-optimal social payoffs as seen, for example, in the classic Prisoner’s Dilemma scenario. We then propose a convolutional neural network borrowing from existing architectures to capture the complexities of individual human strategic decision-making, incorporating considerations for multi-agent (i.e. more than two people at the same time) interactions, repeat interactions, imperfect information, altruistic tendencies, and collusion. Using this architecture, we seek to create a holistic view of broader social dynamics in order to devise inputs and learning functions that might be able to encourage a natural or otherwise self-sustaining trend towards altruistic individual behavior (i.e. individuals compromising their own payoffs) within a population of purely selfish agents. Second, we use our analysis as a lens to view social phenomena. We apply these findings to analysis of real-world payoff data to seek an understanding of (a) the degree and structure of incentives or policies required to promote social change, (b) population thresholds and long-term behavior for potential altruistic behavior trends, and (c) the influence of performative activism and other additional considerations. The hope is that the conclusions from our research will offer some insight into how to approach human strategic decision making in a quantified way, what the relevant components to a machine learning architecture modeling society might be, and how the architecture could be used as a tool to progress towards a future with less prejudice and suffering.
|Type of Material:
|Princeton University Senior Theses
|Appears in Collections:
|Operations Research and Financial Engineering, 2000-2023
Files in This Item:
|Request a copy
Items in Dataspace are protected by copyright, with all rights reserved, unless otherwise indicated.