Selected Technology Cluster Research Projects

The literature on assortment optimization problems mainly relies on the assumption that the uncertainties follow a specific probability distribution, leading to a neat analytical structure and hence a computationally efficient solution procedure. In this project, we study the robust setting in the sense that we do not have the exact knowledge of the distribution of the uncertainties—namely, the customers’ uncertain utility from each product. Specifically, we propose investigating two distributionally robust assortment formulations, robust assortment revenue optimization and robust assortment revenue satisficing. While the former uses a pre-specified ambiguity set to characterize the scope of probability distributions of customers’ utilities, the latter uses a target-driven approach to take all probability distributions into account.

As soon as the distribution deviates from a specific one, the typical closed-form solution structure is lost, which can cause computational intractability. However, in our preliminary study with the multinomial logit model and KL-divergence as the ambiguity measure, the optimal assortments exhibit a revenue-ordered property, i.e., products in the optimal assortment have higher individual revenues than those excluded. We aim to extend this robust assortment framework to more general settings (e.g., cardinality-constrained assortments) and develop efficient methods for obtaining optimal solutions. We will also compare the two robust assortment models theoretically and numerically to show, for instance, that the satisficing approach may yield solutions more robust with respect to a target revenue.

Specifically, we plan to identify a class of two-stage optimization problems featuring supermodularity and then develop efficient methods to solve them within the framework of distributionally robust optimization (DRO). Under certain distributional ambiguity sets (e.g., those characterized by supports, means, and mean absolute deviations), we hope to obtain explicit upper bounds on the expectation of supermodular functions and derive the worst-case distribution in closed-form. This can make the robust counterpart of such two-stage problems tractable. We will also explore necessary and sufficient conditions to determine whether a given two-stage problem possesses supermodularity. Unlike the usual approach in the literature (which often focuses on supermodularity in maximization), we investigate parametric minimization. We expect these conditions to hold in many operational problems—e.g., multi-product newsvendor, appointment scheduling, assemble-to-order systems—making them solvable via our approach, which otherwise might be too computationally complex.

Specifically, we extend from the Value at Risk (VaR). Although VaR has desirable properties such as Monotonicity and Positive Homogeneity, it cannot simultaneously guarantee performance across multiple probability levels. Also, due to Translation Invariance, VaR might ignore the target effect. We plan to develop a class of decision criteria that can encompass multiple quantile values and evaluate performance for the entire distribution’s tails. The new decision criteria will 1) satisfy normative properties like stochastic dominance, 2) allow specifying probability bounds over multiple thresholds, and 3) remain tractable within distributionally robust optimization. We aim to derive effective algorithms for both continuous and discrete decision variables. Preliminary work suggests that under certain conditions, solutions may be obtained via second-order cone programming or bisection search. Beyond the theoretical exploration, we will apply this framework to operational challenges such as shortest path problems and patient scheduling.

The system employs artificial intelligence (AI) techniques, including natural language processing (NLP), to understand email content, trigger appropriate workflows, and generate automatic reply messages. The result helps companies significantly reduce labor costs and improve accuracy in routine correspondence with customers.

This research leverages our recent work in the active RFID traceability technologies. Our system is developed with the capability to determine the indoor location of an active tag. This system was installed inside two wards (11A & 11C) of Prince of Wales Hospital for a 4-month formal pilot study with patient and medical staff participation to improve patient care.

Based on the user feedbacks collected, we find that time and space play an important role in shaping human behaviors, inferring human interactions, and influencing epidemic spread of infectious diseases in a healthcare institute. Our research aims at enhancing the capability of tracking the movement of people and tracing the infection of a hospital-acquired disease for each individual in order to monitor and control the epidemic spread of nosocomial infections. To achieve this goal, we plan to conduct the following research activities:

• network models of dynamic human behaviors for an enhanced spatial-temporal analysis to determine human interaction patterns and contacts in a relatively close community,
• epidemic models to track and trace nosocomial infections based on the time-varying contact networks to increase our understanding of transmission mechanisms in hospital,
• traceability analysis of tracking the mobility and infection for each individual to identify risk behaviors by linking both the network and epidemic models,
• effective and efficient control strategies against epidemic spread under different scenarios to take advantage of the network structure and disease dynamics.

This research is unique. First, since our focus is on the movement of people, our methods are very different from those used to track products in a supply chain. Further, we plan to conduct the traceability analysis of the dynamic interactions and infections of people at individual level for network epidemiology study in hospital. This differs from either the popular modeling of a static network or the traditional epidemiology study for a large region. Finally, many spatial-temporal analyses and epidemiology models are developed for outdoor applications while this work deals with indoor activities.

The Tree Guard Monitoring System will make use of various sensor technologies to continuously monitor tree statuses and detect for any abnormal activity being done to a tree, like sawing and hammering. In addition, the TGMS will utilize various wireless technologies to form a scalable wireless mesh network for transmitting various sensor data to a control server. The control server will further analyze the incoming sensor data and determine whether to issue a warning signal to corresponding management and operation parties. Moreover, historical data will be analyzed for deriving activity patterns to improve the detection accuracy as well as establish preventive or predictive measures to safe guard those trees.

The project will develop the hardware devices having sensing and wireless communication capabilities for being attached to trees. Each device will be capable of synthesizing various streams of sensor data to detect abnormal activities. The project will also develop the data communication protocols to support various communication transmissions between different components in the mesh network. Monitoring and analytics software will be developed to convert sensor information into warning signals.

The reason of our focus on uncertainty and risk is the increasing importance of robust planning and robust investment decisions in the modern telecommunication environment. Traditional telecommunication networks were dominated by public monopolies which provided only a few mature services with fairly predictable user demand. Unfortunately, this predictability is a feature of the past! Due to the fast evolution of modern high speed networks, deregulation, quick development of new services with shorter life cycle, unpredictable user response to introduction of new services, modern networks operate in the presence of substantial uncertainty. The emergence of a multitude of new actors, which assume new roles and engage in complex relations of competition and collaboration, further complicates the situation and compounds the risks. Traditional planning methods are not adequate for design of networks in this new environment due to their essential lack of capability to treat this instability and uncertainty. New planning methods are needed which would facilitate design of robust networks. These networks will be able, within reasonable bounds, to absorb unpredictable user reaction to new services, technological change and other uncertainties without consuming unreasonable amount of critical resources. Our aim is to develop planning tools for design of robust networks using the latest developments in optimization technology integrated with some ideas developed recently for evaluation of industrial projects. These tools will be useful both for industrial actors and for governmental regulatory bodies.