Miller, KeatonO'Briant, Colleen2024-08-072024-08-072024-08-07https://hdl.handle.net/1794/29825This dissertation aims to enhance transparency in AI systems by integrating methods from Machine Learning and Econometrics, specifically focusing on Dynamic Discrete Choice (DDC) models. In Chapter 2, I compare the Nested Fixed Point (NFXP) algorithm from Econometrics with Max-Margin Inverse Reinforcement Learning (IRL) methods from AI/ML, using Monte-Carlo experiments to demonstrate that preference shocks from Econometrics can resolve fundamental identification issues in IRL. The simulation results show that while Projection IRL is slightly less accurate than NFXP, IRL significantly reduces computational demands, requiring 20 times fewer dynamic programming problems to be solved. Chapter 3 investigates the practical applications of these methods by analyzing publicly available 2013 taxi data to compare IRL and NFXP in estimating payoffs for New York City taxi drivers during the morning commute. The analysis highlights that IRL’s flatter objective function has the problem of allowing a broader range of acceptable payoff functions, however its feature expectation matching technique provides valuable feedback on the smoothing parameter for kernel density estimation of the transition probability function. This chapter offers recommendations and identifies potential drawbacks of using IRL, thereby deepening our understanding of the real-world performance of the algorithm.In Chapter 4, the dissertation explores how small business owners may misattribute noise for profit signals using an instrumental variables approach and a rich dataset of product ordering decisions by Washington State marijuana dispensaries over the first three years of recreational marijuana legalization. The study examines whether entrepreneurs’ predictions about product profitability are influenced by exogenous weather shocks, assessing if owners with previous retail experience make more informed decisions, if attentiveness improves over time, and if living further from the dispensary increases the likelihood of conflating weather shocks with profitability signals.en-USAll Rights Reserved.Dynamic Discrete ChoiceInverse Reinforcement LearningElectronic Thesis or Dissertation