Yuting Yuan

We study the problem of predicting congestion risk in service systems, a factor associated with poor service experience, higher costs, and even medical risk (e.g., in ICUs). We define "high-risk states" in queuing models as system states that are likely to lead to a congested state in the near future, and strive to formulate simple, interpretable, rules for determining whether a given system state is high-risk. We employ methods from queueing theory, simulation, and machine learning (ML) to devise simple prediction rules, and demonstrate their effectiveness through extensive computational study, which includes a large scale ICU model validated using data. Our study suggests combining custom model-based interpretable features with linear models can accurately predict congestion in ICUs.

We study the problem of predicting in real-time customers' waiting time in partially visible queueing systems by a predictor who does not observe some of the customers that interact with the system. The problem arises due to technological limitations that prevent monitoring all customers.

We derive closed-form expressions for the prediction problem in a partially visible M/M/1 queue, and propose effective data-driven approaches to solve the prediction problem for more general queues. The effectiveness of our approach is evaluated using both simulated data and real data from a large outpatient hospital.

Omnichannel system is a common operation strategy which provides multiple ways for customers to experience the business. One feature of such system is the information heterogeneity across channels. For example, to customers, the queue length in the physical store is visible, while the queue length in the online ordering channel is invisible. We study customers’ decision making in an omnichannel system with partially observable queue (FCFS discipline or visible-class priority). We find that when market sizes under fully and partially observable systems are comparable, partially observable system generates higher throughput if customers are not scared away by invisibility. Surprisingly, even with less information, partially observable system can have higher social welfare when the invisible arrival rate is high. In that case, the customers are scared away and the system becomes less congested.

In operational planning problems, the advancement in technology has enabled organizations to collect real time data, learn the system mechanism, and take prompt actions accordingly. I identify three potential problems of this procedure: (1) the noise presented in data may cause biased predictions and hence suboptimal decisions; (2) it's nontrivial to predict system evolution under an action that never occurs in history (in other words, difficulty in counter-factual analysis); (3) in some context, practitioners are not likely to implement the recommended actions without understanding the reasons. I propose a new framework for sequential decision making that tackles these issues. The idea is to leverage the conceptual validity of operational models in data-driven learning and planning. My approach prescribes a data-driven policy that is regularized by a synthetic model, which is an optimal combination of several operational models. I am currently experimenting on synthetic data and real hospital data.