Abstract
dc:description.abstractThe thesis focuses on networks and experimental economics. The first chapter theoretically studies network coordination games under adaptive learning agents. The second and third chapters conduct experiments on strategic interactions in networks and on network formation, respectively. The fourth chapter examines the behaviour of Generative Pre-trained Transformers (GPT) in classic game experiments. The first chapter theoretically examines how agents' behavioural traits influence long-term coordination outcomes in networks. Agents conduct experience-weighted attraction (EWA) learning, a model that captures various aspects of behavioural characteristics. I demonstrate that the range of possible long-term outcomes depends significantly on these behavioural traits, ranging from a unique outcome where all agents favour the risk-dominant option to a broader set of possibilities than that of Nash equilibria. Additionally, I show that the importance of each agent's initial preference to the network’s coordinated outcome can be summarised by the principal left eigenvector of a Jacobian matrix, which incorporates both the properties of the network structure and the distribution of individual behavioural traits. In the second chapter, coauthored with Syngjoo Choi, Sanjeev Goyal, and Frederic Moisan, we conduct an experiment to study games on networks involving strategic complements and strategic substitutes between connected individuals. Economic theory predicts that equilibrium actions are proportional to (Bonacich) centrality. Our data offer broad qualitative support for this prediction, but the relationship between actions and centrality is weaker than predicted by theory. We also identify effects of the strategic structure (strategic complements vs. strategic substitutes) and the complexity of networks on action levels. These patterns cannot be explained using standard behavioural models – such as other-regarding preferences, k-level reasoning, and quantal response equilibrium – but they are consistent with a model of behavioural attenuation. The third chapter, co-authored with Syngjoo Choi, Sanjeev Goyal, and Frederic Moisan, presents experimental evidence on games in which individuals decide their connections. These connection decisions give rise to networks. We consider two classes of situations: first, in which only the individuals who initiate the connection benefit (one-way flow), and second, in which both parties benefit regardless of who initiates the connection (two-way flow). Our experiments reveal that in the one-way flow model, subjects create sparse networks whose connectedness and efficiency decrease as group size increases. In contrast, in the two-way flow model, subjects create sparse, small-world networks whose connectedness and efficiency remain high in both small and large groups. In my fourth chapter, I conduct experiments to study the behaviour of Generative Pre-trained Transformers (GPT) in the repeated play of the ultimatum game and the prisoner’s dilemma. GPT is prompted to generate both its choices and reasoning. Results show that GPT's behaviour aligns with intuitive expectations and observed human behaviour in important aspects, such as offering non-trivial amounts, rejecting unfair offers, conditional cooperation, and reactive behaviour to game history. Prompting GPT with fairness concern or selfishness significantly impacts its choices. The reasoning statements generated by GPT help to explain the rationale behind its choices.
Degree
thesis:*- Name dc:type.qualificationname
- Doctor of Philosophy (PhD)
- Level dc:type.qualificationlevel
- Doctoral
- Grantor dc:publisher.institution
- University of Cambridge
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Guo, Fulin
- Advisor dc:contributor.advisor
-
- Goyal, Sanjeev
Subjects
dc:subject × 2Rights
dc:rightsIdentifiers
dc:identifier.*- DOI dc:identifier.doi
- https://doi.org/10.17863/CAM.119806
- OAI identifier oai:identifier
- oai:www.repository.cam.ac.uk:1810/386702