Essays on constrained decision-making and strategic interactions

Abstract

This dissertation examines decision-making under cognitive constraints and strategic interactions between agents in competitive environment. The first chapter establishes a micro labour market model that moves beyond traditional assumptions of perfect rationality by exploring the impact of limited attention on workers’ job search process and the resulting market outcome. I build on existing literature by allowing inattentive workers to have diverse priors and heterogeneous attention costs. The model reveals that labour market mismatch can stem from biases in workers’ default search strategies, and heterogeneous attention costs can contribute to greater variability in equilibrium outcomes. I also identify equilibrium multiplicity, which was not adequately addressed for in previous studies, and found that equilibria where workers adopt different application strategies may generate both higher market efficiency and lower monopsony power as compared to scenarios where workers use the same application strategies. The second chapter investigates how experience, serves as heuristic for future choices, shapes workers’ adaptive learning behaviour in job applications and affects coordination in the labour market. I analyse workers’ search behaviour when they do not observe wages and can only learn from feedback, as well as when wages are posted first, and workers’ choices depend on both the observed wages and their past experiences. I show that in presence of multiple equilibria, experience-based learning generally leads to more efficient outcome, where workers coordinate on applying with high probability to different firms, and this can be locally asymptotically stable. The final chapter explores strategic interactions among synthetic agents represented by various large language models (LLMs) in beauty contest games. By drawing parallels between experiments with LLM-based agents and human subjects, I suggest LLMs may be used as complements to human participants in future experiments and as simulation tools to mirror human-like complexities. I show that LLM-based agents exhibit reasoning depths ranging from level-0 to 1, which fall below that observed in human experiments, but these agents display similar convergence patterns toward Nash Equilibrium in repeated setting. I also demonstrate that environment with lower strategic uncertainty enhances convergence for LLM-based agents, and environments with mixed strategic types could accelerate convergence for all subjects.