Michael Kim

This dissertation studies how misinformation, bias, and strategic information provision affect learning and decision-making in operations and management contexts. These challenges commonly arise in healthcare, digital platforms, and revenue management, where decisions depend on complex information-sharing environments. To address them, we develop Bayesian control models to design information provision and robust learning strategies, examining the problems from complementary perspectives. The first chapter focuses on dynamic manipulation by a strategic firm that seeks to influence public beliefs through costly dissemination and distortion. We model the problem as a Bayesian dynamic program, where the firm balances immediate persuasion and future uncertainty. Using a variational representation, we derive a closed-form optimal policy that involves threshold-based dissemination and dynamic mean boosts to belief distributions, and characterize when and how manipulation is most effective over time.The second chapter extends this framework by incorporating social learning. Instead of a single aggregated learner, we model the public as a pair of partially-Bayesian social learners who update beliefs based on both private signals and each other’s opinions. We show that, although social learning may amplify manipulation locally, it guarantees asymptotic convergence to the truth. This highlights the long-run benefits of belief diversification innetworked environments, even under manipulation.The third chapter shifts to the learner’s perspective, introducing a hierarchical Bayesian network model where a decision-maker optimally acquires signals from multiple biased sources. We derive the optimal acquisition strategy under general earning objectives. We show that biased sources are complementary, and derive closed-form solutions for the optimal acquisition strategy, which diversifies across biased sources to mitigate bias under budget constraints. Simulation studies using healthcare nowcasting and demand forecasting data confirm this diversified learning approach’s advantages in efficiency and reliability.

View record

Decision making under uncertainty and limited information has been a critical challenge in operations. This thesis studies three applications and sheds light on how to make sequential decisions as data for learning are collected over time. In the first chapter, we consider the data-driven newsvendor problem where a manager makes inventory decisions sequentially and learns the unknown demand distribution based on observed samples of continuous demand (no truncation). We show that the widely-used sample average approximation approach is near-optimal. Moreover, we characterize how the best achievable performance depends on not only the time horizon but also the local flatness of the demand distribution.The second chapter considers a dynamic pricing and learning problem where a seller prices multiple products and learns from sales data about unknown demand. To avoid the classical problem of incomplete learning, we propose dithering policies under which prices are probabilistically selected in a neighborhood surrounding the myopic optimal price. We show that dithering policies achieve asymptotically optimal performance in three typical settings and their extensions with demand correlation, which demonstrates dithering as a unified approach to balance exploration and exploitation.The third chapter considers a sequential search over a group of similar alternatives to select the best one. The individual value of an alternative contains two components: an observable utility and an idiosyncratic value. Once an alternative is searched, the utility can be fully revealed, but the idiosyncratic value is unobservable and needs to be learned gradually by sampling. The utilities share an unknown population distribution, which captures the similarity across the alternatives and allows for knowledge transfer within the group. A novel feature of this problem is the combination of the individual and population levels of learning. We formulate the problem as a Bayesian dynamic program and show that the optimal policy can be found by comparing the mean estimates of the current alternative and the population. We also derive other structural properties to provide managerial insights and shed light on the two levels of learning.

View record