Bingnan (Luke) Lu

This is an academia-industry collaboration with a medical equipment company. Our goal is to develop an analytical tool that expedites the detection of product issues from emerging customer feedback by leveraging various data sources.

We have developed a Human-In-The-Loop (HITL) system that provides a framework to enhance machine learning models capable of identifying existing and emerging product issues with limited human inspection. In particular, we construct a multi-label classifier that can “understand” the customer’s textual feedback and apply methods in active learning to update the model with new information from selective product inspections. This adaptive learning process reduces the volume of training data by more than 80% and can identify new product issues without losing resolution to known ones. The company is in the process of integrating our HITL system with its existing workflow.

The recent global outbreak of the COVID-19 pandemic has impacted everyone’s life. Fast and efficient testing is key to identifying infected individuals and containing viral spread. In this project, we study the operation of a testing facility that pools individual samples and tests them together to minimize the waiting time for the test results.

We model this pooled testing process as a two-stage tandem queueing system with batch service and reentry. Using a matrix analytic method, we calculate the stationary probabilities of the number of customers in the system and find the optimal batch size that minimizes delay. We also provide an approximate analysis using the G/G/1 queueing model, which is highly accurate. It enables us to derive various monotonicity properties, providing a deeper understanding of how the prevalence rate and service rate affect the optimal batch size. Furthermore, we compare the delay optimization problem proposed in this study with the typical throughput optimization problem formulated in the literature. We demonstrate that using the latter for dynamic problems can lead to inappropriate results.

Dynamic matching is a core mechanism that empowers the operation of many on-demand services, such as ride-sharing transportation, food and goods delivery, and online matchmaking. In this project, we study a two-sided platform that matches units of demand and supply that arise over time with heterogeneous matching preferences and may abandon the system before being matched. A matching policy specifies which pairs of items to match for any given state of the system. We formulate the problem as a Markov decision process and seek to characterize the structure of an optimal matching policy. We develop and test conjectures as to what properties the optimal policy satisfies. Our goal is to discover useful structural properties to understand the features of an optimal matching policy and provide a basis for designing effective heuristics.

There are numerous service systems where the arrivals of customers are driven by scheduled appointments. Examples include arrivals to healthcare facilities, government agencies, restaurants, wellness centers, etc. In this project, we study a single-server queueing system that serves a finite number of scheduled customers who are not necessarily punctual and may arrive either earlier or later than their appointment times or may not show up at all. Our goal is to develop and analyze the optimal appointment schedule that increases the utilization of service resources and maintains a certain level of service reliability in terms of delay.

We develop both exact and approximate approaches to obtain the distribution of waiting time for each customer. The latter is much faster in computation with little cost on accuracy. We prove that the approximation provides an upper bound for the expected customer waiting time when non-punctuality is uniformly distributed. We also examine the impact of non-punctuality on system performance and prove that non-punctuality deteriorates waiting time performance regardless of its distribution. In addition, we illustrate how our approach can be used to support individualized appointment scheduling.

Tandem systems are common in many manufacturing and service industries. For instance, in a hospital, a tandem system diagnoses patients (jobs) through a sequence of medical tests. Since labor cost constitutes a substantial portion of operating cost in these environments, it is crucial to make effective use of the workforce of a tandem system to maximize its productivity.

In this project, we study a service system that requires a finite number of heterogeneous jobs to be processed by two stations in tandem. Each station serves at most one job at a time, and there is a finite buffer between the two stations. We consider two flexible servers that are cross-trained to work at both stations. The duration for a server to finish a job at a station is exponentially distributed with a rate that depends on the server, the station, and the job. Our goal is to identify an efficient policy to dynamically assign the servers to work on jobs at stations such that the expected makespan (duration to complete all the jobs) is minimized.

Given that servers must be non-idling to maximize system throughput, we first derive the expected makespan of a general non-idling service assignment policy. We then analyze three simple non-idling policies: the summation-myopic, the product-myopic, and the teamwork policies. We prove that (i) the product-myopic policy is optimal if the servers maintain the same service-rate ratio at each station for all the jobs, (ii) the teamwork policy is optimal if the servers maintain the same service-rate ratio at different stations for jobs that are sequenced near each other, and (iii) the summation-myopic policy is no worse than the teamwork policy. Our numerical study based on general service rates suggests that the summation-myopic policy can be better or worse than the product-myopic policy. We extend the model to {incorporate moving costs and service defects}, which are understudied in the literature. We derive the expected total moving cost and the expected number of perfect jobs (without service defects) under a general non-idling policy.