CoBrand-IORA_H-web-1
Search
Close this search box.

Book of Abstracts

DAY 1

Wednesday - 30 June 2021

Juan L. Senga

Ethically Complex Blood Supply Interventions During Pandemics

In this study, we examine the effectiveness of five ethically complex interventions in blood supply management during pandemics; namely, (1) reduction in donor regeneration time, (2) rationing of blood, (3) tapping non-traditional donors, (4) deferring demand, and (5) using paid donation. We employ a stochastic inventory model with regenerative capacity, followed by a discrete-event simulation study using pandemic and blood collection data. We are able to show the effects of the above interventions on fillrates during and after pandemics. We consider two main factors: donors' propensity to donate which is the change in blood donor behavior in the presence of emergencies, and maturity level of the blood bank. We find that pandemics lead to lower fill-rates in maturing blood banks, especially when propensity to donate is low. Ethically complex interventions increase the fill-rates for maturing blood banks with paid donation generating the largest improvement. In contrast, mature blood banks maintain high fill-rates during pandemics, and paid donation reduces the fill-rate. Multiple pandemic waves reduce the fill-rate even when donors have high propensity to donate, implying the severity of long pandemics. Lastly, some ethically complex interventions can set back maturing blood banks in the maturation process as an increase in efficiency can lull them into complacency.
Sun Qinghe

Data-Driven Optimization for Bunker Refueling

Bunker procurement decisions concerning operational hedging via speculative stockpiling over time and geographical location are million-dollar decisions. In this paper, we propose a nonparametric, data-driven approach that is consistent with the structural properties of the optimal liner refueling policies shown in the stochastic dynamic program. In particular, considering the fixed cyclic routing nature of liner shipping, we model the bunker refueling problem as a semi-Markov decision process (SMDP) with an objective of minimizing the long-run expected cost per unit of time along the sequence of ports visited under price uncertainty at each port of call. We show that the optimal policy follows a price-dependent capacity-adjusted buy-up-to policy. Specifically, the optimal buy-up-to target at each bunker location belongs to a set of finite targets that depends on the historical prices and real-time feature data. We provide a framework to learn the optimal policy via (i) empirical risk minimization with regularizations or (ii) policy gradient method with entropy regularized reward. Numerical experiments are performed on real data sets that show the attractive performance of the solutions.
Long Zhao

Not to Invest, It Should be An Option: A Better Pursuit of Maximum-Sharpe Portfolio

Research problem: Conventionally, portfolio optimization faces the inevitable challenge of estimation errors in the expected returns, which is further amplified by the prevalent assumption that one must invest at all times. However, we propose that not to invest should be an option. With this option, portfolio optimization can become more manageable: it only needs to address the cases when estimation errors are tolerable; when the estimation errors are large, one does not invest. Taking the maximum-Sharpe (MS) portfolio optimization, we want to demonstrate that implementing the no-invest option is doable and beneficial.

Key methodology and assumptions: We leverage a newly developed dimension reduction tool tailored to portfolio optimization. It resolves PCA's fundamental flaw in portfolio optimization: the investor prefers low risk while PCA ignores the low-risk PCs. The assumptions for this tool are 1. The eigenvalues of the true covariance matrix have a specific structure; 2. A particular requirement for the data generating process. We also use a simple robust optimization to build a portfolio when there are limited estimation errors. Moreover, spline coupled with dimension reduction is used to generate an evaluation system.

Major Results: Because the new dimension reduction tool could reliably measure the level of estimation errors, we propose a scenario-based investment strategy: If no portfolio has a statistically guaranteed positive Sharpe ratio, one should not invest. Otherwise, one invests. The minimum-variance (MinVar) portfolio, being free from estimation errors in expected returns, serves as the default portfolio unless its estimated Sharpe ratio is low, in which case we use robust optimization to combine the MinVar and MS portfolios. We design a new spline-based system to evaluate such strategies involving invest-or-not decisions. Empirical evidence shows that this simple strategy achieves outstanding performance.

Implications: The option to not invest is natural in real life: one might not want to choose a fund manager that forces him/her to invest no matter what happens in the market. This option could revive the strategies that only work well in some situations if one could reliably predict whether the situations are beneficial. Even better, different strategies might favor different scenarios. Thus, it is possible to alternate among several strategies to pursue an even higher benefit.

Nuno Antunes Ribeiro

Predictive and Prescriptive Analytics for Airport Slot Allocation

The International Air Transport Association (IATA) slot allocation process is the dominant demand management mechanism applied at the busiest airports worldwide. At these airports, airlines need to be assigned slots by a slot coordinator to schedule their flights. Slot allocation is driven by a set of rules and priorities specified in the IATA Worldwide Slot Guidelines (WSG). These rules introduce coupling constraints across the allocation of slots at multiple times of the day and on multiple days of the year, resulting in a highly complex combinatorial problem. While previous models have explored a wide range of approaches to account for the complexity of the problem, they commonly assume a flight-centric approach, which minimizes flight displacement metrics (i.e., the difference between the slot request time and the slot allocated time). Nevertheless, minimizing flight displacement does not necessarily translate into the most attractive itineraries for passengers. By focusing strictly on displacement, previous optimization models may eliminate important passenger connections or make existing connections longer. This paper develops an original passenger-centric approach to airport slot allocation that optimizes slot allocation decisions while maintaining the feasibility of important historical passenger connections. The proposed model optimizes slot assignments under existing guidelines, while minimizing displacement and the cost of accommodating passengers. We integrate two primary analytics models: a passenger flow prediction model and a slot allocation optimization model. We use historical data and machine learning models to predict passenger flows at Changi Airport. We then propose a new formulation to optimize slot allocation decisions accordingly. A heuristic solution method based on large-scale neighborhood search methods is also proposed to solve the full slot allocation problem at Changi airport using real data from slot requests. Computational results yield two main findings: (i) the passenger-centric model and solution approach can be solved in reasonable runtimes. (ii) the passenger-centric model can vastly increase the number of feasible connections, as compared to the baseline flight-centric model. This leads to large reductions in passenger displacement at comparatively small costs in terms of flight displacement.
Ling Qin

Bitcoin Mining and Electricity Consumption

We develop a dynamic equilibrium Bitcoin mining model with exit options, technology innovation, and exogenous entries to characterize miners’ optimal entry and exit strategy. Calibrated with real data, our model remarkably predicts the co-movements of computing power and electricity consumption of Bitcoin network. Moreover, our model can capture drops in computing power and electricity consumption, and incorporating exogenous entries significantly enhances our model's prediction power. Our calibration results indicate that substantial new miners keep entering the mining market.
Shouchang Chen

Managing Order-Holding Problems in Online Retailing Platforms

Problem definition: The booming of third-party logistics (3PL) changes the cost structure of an online retailer in the order fulfillment process. The online retailer pays a fixed amount of order arrangement fee to the 3PL to outsource the order fulfillment service for each service request. We study the problem on when an online retailer should send the service request. The trade-off is between the order arrangement fee and the holding cost.

Methodology: We model the order-holding problem as a Markov Decision Process (MDP). By reducing the MDP to a sequence of single-dimensional counterparts, we analytically characterize the optimal policy. We apply the consumer’s sequential choice model to calculate the transition probabilities, which captures the heterogeneity across different orders and admits a personalized order-holding policy.

Results: The optimal policy is characterized as a personalized threshold-type policy. The pending orders are held within a threshold but sent otherwise. If a new order arrives within the threshold, the holding time is reset and the threshold is updated according to the latest order. The threshold of an order depends on its features, including the product attributes and the involved customer’s demographic information, whose close form together with its piecewise linear approximation can be characterized. Extensive numerical tests based on the data set from the 2020 MSOM Data-Driven Research Challenge show that (1) The bound of gaps of piecewise linear approximation is as tiny as 1; (2) The proposed policy achieves a considerable cost reduction compared to two benchmarks in the literature, with an average 30.13% and 14.02% cost reduction for enterprise users in all instances compared with the first and second benchmark respectively.

Managerial Implications: The characterized optimal policy is easy to implement. The proposed methods characterize the order features and predict the best holding threshold for each newly arrived order.

Denis Tkachenko

Inflation and Output Growth Forecasting in a Data-Rich Environment Using Extreme Gradient Boosted Trees with Online Learning

Forecasting of key macroeconomic indicators plays a critical role in rational decision making for economic agents and institutions. Inflation forecasting is important as it anchors inflationary expectations for contracts set in nominal terms and guides Central Banks’ actions where optimal policy depends on optimal forecasts (Svensson, 2005). Industrial production growth forecasts are important as it proxies GDP growth which measures the state of the economy. Despite extensive effort devoted to this problem in the literature, it has been extremely hard to improve upon naïve benchmarks such as the random walk (RW) or autoregressive (AR) models (Atkeson & Ohanian, 2001; d’Agostino et al., 2006). Moreover, past studies that found models that significantly outperformed these benchmarks often did so only due to better performance during the more volatile pre-Great Moderation (GM) period and failed to do so afterwards (Rossi & Sekhposyan, 2010). The advent of big data and machine learning (ML) methods has yielded new avenues for developing predictive models in various contexts. With advances in the ML field and the increased availability of economic big data (Garboden, 2020), there has been much effort invested in examining whether macroeconomic forecasting can benefit from such advances, however, earlier attempts have often produced lackluster results due to limited experience of economic researchers with ML methods (Woloszko, 2020). Traditionally, in this setting the literature handles high-dimensional data via dimension reduction afforded by linear dynamic factor models (Stock and Watson, 2006). We explore a relatively underutilized in economic forecasting literature ML tool, eXtreme Gradient Boosting (XGB), that possesses strong regularization properties allowing for effective handling of high dimensional data while fitting flexible nonlinear functions for modelling the conditional mean. We introduce two modifications to increase its effectiveness on time series macro data: 1) adaptive trees (Woloszko, 2020); 2) a novel algorithm for online learning of boosting iterations. In a recent application to inflation forecasting by Medeiros et al. (2019), it was found that random forest with default values of tuning parameters produces the best forecasts among many commonly used ML methods such as variants of penalized regression, neural networks and boosted trees, including XGB with default settings. In this paper we show that our implementation of XGB significantly outperforms the baseline benchmarks and random forest, particularly at the highly empirically relevant forecast horizons of 12- and 24-months, as evidenced by the tests of superior predictive ability (Hansen, 2005) and Model Confidence Set (Hansen et al., 2011). Finally, contrary to common beliefs, ML methods need not be a black box, as advances in interpretable ML allow us to unravel the black box and discover important economic insights. Specifically, an analysis using SHAP values uncovers new stylized facts: 1) the predictive content of predictors across the board did not actually decline after the Great Moderation; 2) the reduced form inflation persistence was largely derived from inherited rather than intrinsic persistence.
ZHONG Zixin

Probabilistic Sequential Shrinking: A Best Arm Identification Algorithm for Stochastic Bandits with Corruptions

Motivated by the issues of fraudulent clicks in online recommendation systems and contaminated samples in medical trials, we consider a best arm identification (BAI) problem for stochastic bandits with adversarial corruptions. The goal is to identify the best arm with a fixed number of pulls (which are also known as time steps), in the presence of an adversary who can corrupt the stochastic outcomes of the arms. We design a novel randomized algorithm, Probabilistic Sequential Shrinking(u) (PSS(u)), which is agnostic to the amount of corruptions. We identify the amount of corruptions per step (CPS) to be a crucial parameter that characterizes the possibility of BAI. When CPS is below a threshold, PSS(u) identifies the best arm with probability tending to 1 as T →∞. Otherwise, the optimality gap of the identified item degrades gracefully with the CPS. We argue that such a bifurcation is necessary. In PSS(u), the parameter u serves to balance between the optimality gap and success probability. The injection of randomization is shown to be essential to mitigate the impact of corruptions. To demonstrate this, we design two attack strategies that are applicable to any algorithm. We apply one of them to a deterministic analogue of PSS(u) known as Successive Halving (SH) by Karnin et al. [2013]. The attack strategy results in a high failure probability of at least 0.5 for SH, but PSS(u) remains robust. In the absence of corruptions, PSS(2)’s performance guarantee matches SH’s. We show that when the CPS is sufficiently large, no algorithm can achieve a BAI probability tending to 1 as T →∞. Numerical experiments corroborate our theoretical findings.

DAY 2

Thursday - 01 July 2021

Jinglong Zhao

Design and Analysis of Switchback Experiments

Switchback experimental design is one of the most prevalent approaches in marketplace experiments. In switchback experiments, a firm sequentially exposes an experimental unit to a random treatment, measures its response, and repeats the procedure for several periods to estimate the treatment effect. Although practitioners have widely adopted this experimental design technique, the development of its theoretical properties and the derivation of optimal design procedures have been, to the best of our knowledge, elusive. In this paper, we address these limitations by establishing the necessary results to ensure that practitioners can apply this powerful class of experiments with minimal assumptions. Our main result is the derivation of the optimal design of switchback experiments under a range of different assumptions on the order of the carryover effect — that is, the length of time a treatment persists in impacting the outcome. We cast the experimental design problem as a minimax discrete optimization problem, whose objective is to minimize the variance of an unbiased estimator. We identify the worst-case adversarial strategy, establish structural results for the optimal design, and finally solve the problem via a continuous relaxation. The optimal design is in the form of splitting the time horizon into several equal length epochs (except for the first and the last epochs), such that one fair coin flip determines the assignments for all the time periods within each epoch. When practitioners believe there to be very little carryover effect, the optimal designs are almost the same, which adds an extra layer of robustness to our method. For switchback experiments conducted using the optimal design, we provide two approaches for performing inference after running the experiment. The first approach tests a sharp null hypothesis and provides exact randomization based p-values. The second approach uses a finite population central limit theorem to conduct conservative hypothesis tests and build confidence intervals. We further provide theoretical results for our inferential procedures when the order of the carryover effect is misspecified. We show that our estimation and inference still remain valid, with only a little more variance. For firms that possess the capability to run multiple switchback experiments, we also provide a data-driven strategy to identify the likely order of carryover effect. We conduct extensive simulations to study the empirical properties of our results and give empirical illustrations. We conclude the paper by providing some practical suggestions.
Manxi Wu

Efficient Carpooling and Toll Pricing for Autonomous Transportation

How can autonomous transportation technology be leveraged to incentivize riders to take carpooled trips and share network capacity in an efficient manner? We study this problem using a competitive market framework. In this market, the transportation authority sets toll prices on edges, and riders organize carpooled trips using driverless cars and split toll prices. Riders have heterogeneous preferences, with value of each trip depending on the travel time of chosen route and rider-specific parameters that capture their individual value of time and carpool disutilities. Market equilibrium is defined as an outcome in which no rider has an incentive to deviate from the organized trips or opt-out, and riders’ payments cover the toll prices plus the trip costs. Additionally, trip organization at equilibrium, if exists, ensures efficiency (i.e., maximum social welfare). However, such an equilibrium may not exist in general due to the integral nature of carpooled trips on general networks and riders’ incentive constraints. Thus, the question of equilibrium existence becomes central to the implementation of the autonomous carpooling market. This question cannot be fully addressed using classical competitive market analysis because both feasible trip organizations and rider valuations are constrained by the network topology. In this work, we develop a new approach to study the equilibrium existence and computation. Our approach draws ideas from network flow optimization and coalition games. We show that a market equilibrium exists when (i) network is series-parallel and (ii) riders have homogeneous levels of carpool disutility, although riders can still have heterogeneous trip values and values of time. In fact, these conditions are minimal in the sense that counterexamples can be easily constructed when (i) or (ii) are violated. Our result also provides guidance about the implementation of a market-based outcome. The condition(i) suggests that the transportation authority should avoid embedding a wheatstone network structure in designating a subset of routes for the autonomous carpooling service. On the other hand, condition(ii) implies that rider groups that can be differentiated due to different levels of carpool disutilities should be allocated differently sized autonomous cars. Furthermore, we identify a market equilibrium that is equivalent to the Vickery-Clark-Grove mechanism and propose an efficient algorithm to compute it. We show that when a neural platform collects riders’ reported preferences and organizes trips according to the computed equilibrium, all riders truthfully report their preferences and accept the organized trips. This equilibrium also achieves the highest rider utilities among all market equilibria, with minimum total toll prices collected. Thus, our construction allows centralized implementation of a competitive market equilibrium and serves as a benchmark for alternative market structures.
Ruihao Zhu

Calibrating Sales Forecast in a Pandemic Using Online Non-Parametric Regression Model

Motivated by our collaboration with Anheuser-Busch InBev (AB InBev), a consumer-packaged goods (CPG) company, we consider the problem of forecasting sales under the coronavirus disease 2019 (COVID-19) pandemic. Our approach combines online learning and pandemic modeling to develop a data-driven online non-parametric regression method that addresses the challenges of model misspecification and limited data. Specifically, the method takes the future COVID-19 case number estimates, which can be simulated via the SIR (i.e., Susceptible-Infectious-Removed) epidemic model, as an input, and outputs the level of calibration of the baseline sales forecast generated by AB InBev’s offline learning algorithm. To generate the calibration level, we focus on an online non-parametric regression setting, where our algorithm sequentially predicts the label (i.e., the level of calibration) of a random covariate (i.e., the current active case numbers) given past observations and the generative process (i.e., the SIR epidemic model) of future covariates. We evaluate the performance of our algorithm by regret, which is the difference between the squared l2-norm associated with labels generated by the algorithm and labels generated by an adversary and the squared l2-norm associated with labels generated by the best isotonic (non-decreasing) function in hindsight and the adversarial labels. We develop a computationally efficient algorithm that attains the minimax-optimal worst-case regret over all possible choices of the labels. We demonstrate the performances of our algorithm on both synthetic and AB InBev’s datasets of three different geographical regions. The AB InBev’s numerical experiments show that our method is capable of reducing the forecasting error in terms of WMAPE (i.e., weighted mean absolute percentage error) and MSE (i.e., mean squared error) by more than 50% for the company. Key words: sales forecast, online learning, non-parametric regression, COVID-19 pandemic
Rui Chen

Stochastic Sequential Assignment Problem with Batch Arrivals and Application to Truck Temporary Entry Permit Allocation

In this paper, we consider a stochastic sequential assignment problem with batch arrivals. The jobs arrive in sequence and batches with random values that are assumed to be independent and identically distributed. The decision maker has to assign the workers simultaneously and irrevocably to the jobs in the same batch immediately when they are arriving. Associated with each worker is a fixed success rate. The reward of each assignment is the product of the worker’s success rate and the job value assigned to this worker. The optimal policy for maximizing the total expected reward is characterized, and a recursive equation for computing the necessary constants for this optimal policy is proposed. Finally, we showcase our assignment policy in the context of container truck temporary entry permit allocation problem for port operators. Numerical examples, based on data from the Port of Singapore, demonstrate that the policy is able to provide high-quality solutions. Moreover, vessel arrival information may considerably improve the performance of the port operations and management. Thus, policy makers should explicitly recognize the impact of vessel arrival information when designing a truck appointment system.
Yu Long

Optimal Stockist Selection and Contract Design: Evidence from Supply Chain in Indian

Problem Definition: To develop supply chain, manufacturers need to contract with stockists to distribute goods, manage inventory, service retailers, and even make market, etc. In this project, we aim to investigate (i) how a firm selects stockist and design the incentive contract in an optimal way facing asymmetric information and opportunity cost of learning, and (ii) how a stockist operates optimally facing career concern and switching cost. We have special interest in studying these questions in emerging market where the level of mistrust is high, learning and contract switching are costly, and competition is intense.

Methodology: We derive a unified two-period model with one firm and two stockists, incorporating bandit selection, Bayesian learning, contract theory, and structural estimation. We solve the model explicitly and provide characterizations for the optimal contract design and optimal bandit selection. Using contract and sales data gathered from an Indian potato chips manufacture, we are able to calibrate the model through simulated method of moments and perform counterfactual analysis.

Results: Our model shows the optimal contract consists of incentives from: competition effect, career concerns, and compensation. Competition and career concern have quite opposite effects on both firm and stockists’ decision-making. Competition induces stockists to put more effort to win the contracts, and therefore allows the firm to offer a lower level of performance-based payoff. In contrast, career concern motivates stockists to make less effort to win the future contract since all surplus is distributed to the firm, and thus firm needs to offer a contract with higher performance-based payoff. Our calibration results demonstrate that there exist two types of stockists in the India supply chain data: one has higher mean ability but also higher variance compared to the other. Furthermore, the low-productive stockist faces higher switching cost and production cost, indicating that switching to another firm is difficult, and exerting effort is costly. From the numerical studies, we show firm’s stockist selection faces trade-offs between stockist’s mean ability and associated risks, such as idiosyncratic risk and learning risk.

Implications: Our analysis is helpful for manufactures to select stockists and design compensation contracts in supply chain. We also provide an analysis of data from India, and offer managerial insights of developing efficient supply chain in emerging markets.
Shih-Che Lo, Mr. Yi-Cheng Shih

Quantum-inspired Genetic Algorithm for Sustainable Logistics Management

Global warming and climate change is happening along with increasing amount of greenhouse gases emission. Having a sustainable logistics management can help global enterprise reduce greenhouse gases emission and improve our living environment. Reduce carbon emissions and green energy become important issues toward sustainability. Hence, we develop a hybrid algorithm combined with the modified k-means classification method as a clustering tool and the quantum-inspired genetic algorithm (GA) as an optimization model to solve the Pollution Routing Problems (PRPs) in this paper.The goal of this study is to develop a variant of GA along with quantum information processing concept to minimizing CO2 emission for the PRPs in sustainable logistics management. From the experimental results, we found out that quantum-inspired GA approach can provide relative good quality of solution to minimize CO2 emission for the PRPs in sustainable logistics management.
Shih-Fen Cheng

The Driver Guidance System for Taxi Drivers

With the emergence of ride-hailing services around the world, traditional taxi fleet operators are facing intense competitions. In this work, we discuss how efficient real-time demand prediction and large-scale optimization technology could help taxi drivers improve their efficiencies. Our research aims to address two major challenges commonly faced by taxi drivers (also ride-hailing drivers to some degree): on the demand side, although experienced drivers are usually proficient in identifying recurrent demand hotspots, they do not react well when demand patterns are disrupted; on the supply side, the lack of coordination among drivers implies that drivers are very likely to flock to a few locations with high demands, resulting in congestions and long roaming time. As a solution to the above two challenges, we propose the Driver Guidance System (DGS), which is designed to anticipate demand and perform global coordination at scale. To validate the effectiveness of our solution, we also performed extensive field trials involving real taxi drivers in both Singapore and Tokyo, demonstrating the sizable benefits for drivers following our guidance.
Karthyek Murthy

Efficient Simulation of Tail Risks in Prescriptive Analytics Models

Motivated by the increasing adoption of models which utilize contextual information in risk management and decision-making, this paper presents a novel Importance Sampling scheme for measuring distribution tails of objectives modelled with a combination of enabling tools such as feature-based decision rules, mixed integer linear programs, deep neural networks, etc. With the explosion in expressivity of models that are obtained by combining these tools, it is imperative that the risk management practice seeks to measure and manage the tail risks associated with their use. Conventional efficient simulation approaches however suffer from feasibility and scalability concerns in addressing this requirement due to their need to be tailored intricately to the underlying probability distribution and the objective.

This challenge is overcome in the proposed black-box scheme by automating the selection of an effective importance sampling distribution with a transformation that implicitly learns and replicates the concentration properties observed in less rare samples. This novel approach is guided by a large deviations principle that brings out the phenomenon of self-similarity of optimal importance sampling distributions.

The proposed sampler is the first to attain asymptotically optimal variance reduction across a spectrum of multivariate distributions despite being oblivious to the problem structure. While a variety of methods exist for searching optimal parameters within a chosen distribution family, to the best of our knowledge, this is the first paper to exhibit an automated approach for tackling the complementary and more challenging problem of selection of sampling distribution families with optimal variance reduction properties.

Due to its black-box nature, the proposed sampler is amenable for serving as an engine for obtaining variance reduction in various chance-constrained and CVaR-based optimization formulations. Besides guiding the sampler, the large deviations principle additionally results in new distribution tail asymptotics capable of yielding operational insights. The applicability is illustrated by considering shortest path problems, product distribution networks and portfolio credit risk models informed by neural networks as examples.
Zhou Minglong

The Ananlytics of Robust Satisficing

Satisficers, in contrast to maximizers, are content with attaining the reasonable target that they set for themselves. Such behavior is prevalent in decision makers especially when they are facing complex situations under risks and uncertainty (Simon 1955). While there is an abundance of prescriptive analytics tools for maximizers, it is not the case for satisficers. To fill this gap, we develop a new prescriptive analytics tool called robust satisficing that uses data to help a satisficer achieve her targeted expected reward or consumption as well as possible under ambiguous risks and prediction uncertainty. It builds upon the robustness optimization framework recently proposed by Long et al. (2021), and we extend it to incorporate aspects of predictive analytics. We adopt linear regression as the underlying predictive model and propose a new estimator uncertainty and residual ambiguity set to characterize the relations between the underlying regression coefficients, which is uncertain but non-stochastic, and the stochastic random variables representing residuals that have ambiguous distributions. The robust satisficing model is also useful in allocating resources for multiple satisficing agents to meet their expected reward targets. We present some useful robust satisficing models that can be solved efficiently, and provide tractable approximations to tackle adaptive linear optimization problems. The simulation studies for multi-product pricing and newsvendor problems elucidate the benefits of the robust satisficing framework in helping the firm attain the desired target expected profits, mitigate shortfalls, and limit target surplus, if desired. The robust satisficing model can also improve solutions over one that is obtained by solving a baseline empirical optimization model using estimated parameters. The improvement is also more pronounced when data availability is limited. Paradoxically, maximizers can also benefit from the analytics of robust satisficin.  
Ruijiu Mao

How to Apply UAVs in Collaboration with Existing Network

Package delivery is an essential part for customers’ online shopping experience. However, lengthy delivery time is prevalent in pure road freight services owing to poor road condition, traffic jams, under-staffing, or incompetently managed schedule. It is expected that UAV-based delivery can help to address this issue due to timeliness, robustness, and convenience. Nevertheless, the current largest obstacles for UAV-based cargo delivery application lies in the high cost and limited cruise range. Therefore, in this paper, we aim to design a land-air collaborated network to help courier firms improve service performance, and guarantee equity with minimum cost.

Note that the current land-carriage is conducted with hierarchical structure consisting of online retailer warehouses (origin), at most two layers of sorting centers, and customer locations (destination), we intend to determine the number and location of UAVs and their bases. For every origin and destination (OD) pair, we find the route and transportation mode, i.e., truck or UAV, in each segment.

Specifically, we formulate an optimization problem that minimizes the cost of UAV system investment to achieve a target in terms of the CVaR of delivery time. Here we model the waiting time of transporting with UAV using an M/D/c queueing system, in which the service time distribution stems from the heterogeneous travel time between the UAV base and spreading discharging locations (including sorting centers and customer locations). To deal with the nonlinearity of the waiting time in the queue, we resort to incorporating discretization and power function approximation.

 By conducting a case study on the road delivery performance data of JD Logistics in Shaanxi province, China, we develop a land-air collaborated network which decreases system average of worst 5% delivery time by 14.3% with 8 UAVs and 4 bases. We also discuss the following insights into the impacts of infrastructure design, UAV cruise range limitation, and number of delivery hierarchical layers, etc. That is,

1. Deploying a few UAV and bases can achieve improvement in the target. For example, under the 5% CVaR target (the mean waiting time of the worst 5% delivery), the land-air collaborated system can shorten the time by 5.9% with 3 UAV and 1 base, by 12.2% with 6 UAV and 2 bases, by 14.3% with 8 UAV and 4 bases.

2. The firm should prioritize to enable UAV between the sorting centers involved in the routes of high-volume OD pairs.

3. Double the cruise range can save 70% of the cost under the same 5% CVaR target of 41 hours.

4. Adding an additional transit at the sorting centers in the delivery can save 53.3% of the cost under the same 5% CVaR target of 41 hours.
Sundara Natarajan Panchanatham

Be the Match. Optimizing Capacity Allocation for Allogeneic Stem Cell Transplantation

Background: Every year, thousands of people in the US are diagnosed with hematological (blood-related) diseases such as leukemia, lymphoma, severe aplastic anemia etc. The treatment of such diseases requires transplantation of genetically compatible hematopoietic stem cells (HSCs) extracted from the bone marrow (BM) of live donors or the umbilical cord blood (CB) of babies. For the transplant to take place successfully, the patient in need of the stem cell must be genetically compatible with the source and the compatibility is determined by similarity in the genetic codes. Put specifically, the compatibility often referred to as a “match", is the number of places the genetic codes are the same for a patient and the BM donor/CB cells in a string of genetic codes of length six (called phenotype). If all the six places in the phenotype matches between the BM donor/CBU cells and the recipient, this is referred to as a 6/6 match. If there is a mismatch in one place, it is a 5/6 match and so on. Every year approximately 13,000 individuals require BM donations (outside of the family members) or compatible CB cells held in storage for treatment.

Problem Definition: The two sources of HSCs differ along different dimensions. Most importantly, for a successful BM transplant, a perfect 6/6 match is required while for CB transplant the matches as low as 4/6 can still lead to a successful transplant. In addition to their matching requirements, these two sources of HSCs also differ in their sourcing costs, availability, and, importantly, medical effectiveness (where BM donations performs slightly better than cord blood cells). To facilitate the search for HSCs, BM registries (BMR) that collect the details of potential donors and public CB banks (CBB) store units of CB (CBU) donated by the public. A program setup by the National Marrow Donor Program (NMDP) called “Be the match" acts as interface that connects the patients with BMR/CBB. With over 10 million unique phenotypes in the US population and limited inventory relative to this variety, then the appropriate sizing of these two institutions (BMR and CBB) represents an important and challenging task from a public policy viewpoint. Most related works in the relevant literature, develop separate static models for estimating the sizes of BMR and CBB but consider only each institution in isolation and not in unison.

Methodology: We develop a simulation-based joint (BM + CB) modelling approach to estimate the temporal variation in matching probabilities before incorporating the associated regression parameters into a mathematical model closely matching the research context. Results are contrasted against a simplified mathematical model, highlighting the importance of the dynamic setup.

Results: Inventories of 17.5 million registered BM donors and 335 thousand CB units are estimated as optimal for the U.S. population under reasonable assumptions. Expanding capacity to these levels would satisfy 33% of the currently unmet demand, increasing the transplantation rate to 98.7% and delivering $97 million of extra social surplus annually.
Arjun Kodagehalli Ramachandra

Conic Robustness Optimization

A robustness optimization framework to deal with uncertain optimization problems has recently been proposed by Long et al. (2020) ,which, for the same computational effort has the advantage of greater resistance to uncertainty as compared to the corresponding robust optimization model. This work extends the robustness optimization framework to a class of two-stage adaptive conic uncertain problems which encompasses a wide range of constraints including bi-convex functions, and as a result, generalizes similar problems considered in recent papers. Constraints convex in both the decision variables and the uncertain parameters cannot be handled by classical robust optimization techniques. We overcome this challenge by consecutively dualizing over the wait-and-see and uncertain parameters, thereby absorbing the conicity of the original problem into a new uncertainty set while the polyhedral structure of the original uncertainty support appears in the linear constraints of the resulting reformulation, which under assumptions of complete recourse is an equivalent two-stage robustness optimization problem. We further show that when approximated in linear decision rules, our reformulation always allows a feasible solution. For the special case of a non-negative orthant cone, we prove that despite being simpler, the linear decision rule approximation of the dual reformulation provides a closer approximation of the original problem as compared to the non-linear decision rule approximation of the original problem itself. Finally, we demonstrate the improved performance of our dual robustness reformulation over the classical robust model with three numerical examples for three different cones: a growth-optimal portfolio example for a single quadratic constraint with ellipsoidal uncertainty where an exact reformulation exists, a geometric programming formulation for multiple two-term log-sum-exp constraints with a box uncertainty set, and lastly a two-stage network lot-sizing example for linear constraints with a budgeted uncertainty set . For the latter two examples, we approximate the robust and robustness formulations using linear decision rules which are feasible as long as the original uncertainty support is polyhedral. Additionally, for the lot-sizing example, our dual formulation shows a remarkable improvement in computational speed concurring with earlier known results.

DAY 3

Friday - 02 July 2021

Nicole Bertani

Spatiotemporal Modeling with General and Geographical Covariates. Insights on Crime in Philadelphia

We explore whether crime depends on the urban geography (i.e. the individual components of the city, such as parks, hospitals, or churches). To do this, we extend Gaussian Process modeling to high-dimensional count data and devise a general procedure to relate any spatiotemporal phenomenon with important characteristics of its environment. The procedure encodes raw spatial objects (e.g the point location of a restaurant or the perimeter of a park) into geographical covariates. Geographical and general covariates (e.g. socioeconomic indicators such as income or population density) are then used to estimate how, how far, and how strongly they relate to crime. We consider the city of Philadelphia as a case study. Our results show that, when controlling for the urban geography, socioeconomic indicators are weak predictors. In particular, demographic information emerges as irrelevant to predict crime. In addition, the results indicate possible interventions to improve safety in the roughest parts of Philadelphia. Finally, we apply the model to make policy evaluation and assess the predicted effect on crime of the ongoing urban redevelopment in Sharswood-Blumberg. This study only relies on freely accessible noncommercial data, making it replicable for other municipalities.
Yangfan Liang

Model Mis-Specification and Algorithmic Bias

Concerns about algorithmic bias arise as we increasingly use machine learning algorithms to support high-stake decision making. Various metrics have been proposed to measure algorithmic bias, and several methods have been proposed to reduce or eliminate it. However, we still do not have a complete understanding of why machine learning algorithms lead to biases. It is obvious that algorithms may produce uneven predictions for individuals in different demographic groups when the sensitive demographic attributes (e.g. race or gender) are provided as input features. However, previous studies have documented that machine learning models would generate biased predictions even when the sensitive attributes are not included in the model for a wide range of algorithms and applicants. A common belief is that such biases are caused by input features being correlated with sensitive attributes (for example, distribution of education levels are different across demographic groups), known as redundant encoding. In this work, we show that redundant encoding alone does not lead to biased predictions. When a model is correctly specified, it will produce unbiased predictions even when the input features have different distributions in different groups. When the model is mis-specified, redundant encoding leads to biased predictions. We focus on one common type of model misspecification: omitted variable problem, which happens when certain relevant variables, i.e. variables that appear in the true data generating process, are omitted in the algorithms. Since the true data generating process is usually unknown and the data we observe are limited, the omitted variable problem is ubiquitous and inevitable in most of the cases.

We analytically show how such bias arises in the case of linear regression. Our analysis suggests that omitting relevant variables would lead to different prediction error distributions for different groups. In particular, the mean prediction errors would have different signs when there are only two groups (e.g. a protected group and a regular group), which suggests that the predictions for one group are systematically overestimated while those for the other group are systematically underestimated. We quantify such mean errors by the group sizes, and variance/covariance of relevant variables. We partially extend our analytical results to probit regression and logistic regression, and perform numerical studies on synthetic and real data to confirm the theoretical findings.

 
Gaurav Kankanhalli

Corporate Hiring under Covid-19

Big data on job vacancy postings reveal multiple facets of the impact of Covid-19 on corporate hiring. Firms have disproportionately cut on hiring for high-skill jobs (within-firm downskilling). Financially-constrained firms scale back their hiring (particularly high-skill) the most, as do firms with workforces more adaptable to “working-from-home.” Applying machine learning to job ad texts, we show that firms have skewed their hiring towards operationally-core positions. New positions take longer to fill and display greater flexibility regarding schedules, tasks, and requirements. Financing constraints amplify pandemic-induced changes to the nature of positions that firms seek to fill, with constrained firms’ new hires witnessing larger adjustments to jobs and employment arrangements.
Guodong Lyu

Adaptive Shortest-Path Network Interdiction

We study adaptive shortest-path network interdiction (SPNI) problem, where the interdictor sequentially allocates limited resources to interdict a subset of arcs in a network so as to maximize the length of the shortest-path traversed by a malicious evader. In contrast to the class of static interdiction policies, which is configured a priori and implemented throughout the entire game without any changes, the interdiction decisions are sequentially updated after gaining more information of the evader under the adaptive interdiction policy. In this work, we explore the value of adaptability to the network interdiction problem in the following two scenarios:

Interdiction with the Knowledge of Actual Destination Information. We first consider the case when the interdictor has access to the actual destination information of the evader and can observe her real-time traveling path at each and every stage. We formulate the interdiction problem under the framework of follower-leader (zero-sum) game, and propose a class of adaptive interdiction policies by sequentially re-solving a branch of static SPNI problems. Leveraging on the structure of the interdiction game, we theoretically demonstrate that the adaptive interdiction policy outperforms the optimal static interdiction benchmark, regardless of the traveling strategies implemented by the evader.

Interdiction without the Knowledge of Actual Destination Information. We extend our algorithmic framework to address the case when the interdictor does not know the evader's actual destination. We reformulate the interdiction model based on linear duality theory, and revise the adaptive interdiction policy. Interestingly, we show the superior performance of this revised adaptive policy, which can also outperform the optimal static benchmark consistently. Furthermore, the sequential interdiction manner allows the interdictor to utilize appropriate destination inference models to improve the interdiction performance.

Encouraging computational results are reported on the Chicago sketch road network with 933 nodes and 2950 arcs. We propose an early-termination parallel algorithm to boost the computational efficiency. The numerical experiments validate the superior performance of our proposed adaptive interdiction policy in comparison with the optimal static policy and state-of-the-art benchmarks. Interestingly, we observe that our adaptive policy could strategically induce the evader into a trap so that all the arcs emanating out of the trap node can be interdicted more effectively. This phenomenon illustrates the operational insight of our adaptive policy.
Natarajan Karthik Balkrishnan

Tight Probability Bounds with Pairwise Independence

While some useful probability bounds for the sum of n pairwise independent Bernoulli random variables exceeding an integer k have been proposed in the literature, none of these bounds are tight in general. In this paper, we provide several results towards finding tight probability bounds for sums of Bernoulli random variables where the information is specified with univariate probabilities P(c ̃_i=1)=p_i for all i bivariate probabilities P(c ̃_i=1,c ̃_j=1)=p_i p_j for all i≠j.

Firstly, when k=1, the tightest upper bound on the probability of the union of n pairwiseindependent events is provided in closed form for any input univariate marginal probability vector p∈〖[0,1]〗^n. To prove the result, we show that a positively correlated Bernoulli random vector c ̃ with the univariate marginal probability vector p∈〖[0,1]〗^n and transformed bivariate probabilities p_i p_j/p where p∈[max_i p_i,1], always exists. The feasibility problem is of independent interest initself, since even testing for feasibility of Bernoulli random vectors with a given covariance matrixstructure is known to be a NP-complete problem. Building on this, we show that the ratio of Boole’s union bound and the tight pairwise independent bound is upper bounded by 4/3 and thebound is attained. In contrast from prior work, the ratio of Boole’s union bound and the probabilitywith mutually independent random variables is known to be upper bounded by e/(e-1).

Secondly, for k≥2 and any input univariate marginal probability vector p∈〖[0,1]〗^n, new upperbounds are derived exploiting ordering of probabilities. Numerical examples are provided to illustrate when the bounds provide significant improvement over existing bounds. Lastly, while theexisting and new bounds are not always tight, we provide special instances when they are shownto be tight.We believe these results and some of the extensions have implications on the price of correlationanalysis which compares stochastic optimization and distributionally robust optimization problems.
Chaoyu Zhang

Capacitated SIR Model with an Application to COVID-19

As the COVID-19 pandemic rages worldwide, all countries are working diligently to find effective ways to control this epidemic. Researchers have studied various mathematical models to understand the spread of the virus and its transmission pattern. Among them, the classical SIR model and its variants, such as SEIR, are arguably the most popular ones. However, insufficient testing capacity has not been appropriately accounted for in those classical models. At the time this research was conducted in the summer of 2020, there was a growing consensus that testing is a critical step to contain the outbreak, and in many countries, the testing capacity is still below the desired level which can successfully mitigate the spread of the virus.

We build on the classical SIR model and incorporate the constraint of testing capacity. In particular, we consider a parsimonious compartmental model with four simplified states: S (susceptible), A (infected without symptoms and not yet confirmed), I (infected with symptoms and not yet confirmed), and C (confirmed). The parameters associated with the progress among states include the infection rate of symptomatic population , the infection rate of asymptomatic population , the testing capacity K, the effectiveness of contact tracing  and the proportion of testing people without symptoms which is the sum of S and A (resp., people with symptoms which is I),  (resp., ). Here, in the early of a pandemic,  may reflect the degree of panic run and  can be interpreted as the reciprocal of incubation and testing turnaround time. To the best of our knowledge, this study is one of the first to thoroughly examine the impact of insufficient testing capacity theoretically, which leads to a set of critical policy implications. We study one critical measure S2, the total number of infectionsin period 2. We show that the following first- and second-order properties with respect to K, , ,  ,  and  hold for this measure.

First, as expected, increasing the testing capacity can reduce the total number of infections over time. Having said that, we show that only when the testing capacity is large enough, the marginal value of having more testing capacity can be significant. This implies the critical importance of having a large number of testing kits, or alternatively, using the group testing technique to significantly “expand” testing capacity. Second, given that the testing capacity may not be elastically adjusted, it is beneficial to limit the testing of people showing no symptoms, these people can be virus-free or asymptomatic COVID-19 infections. The limited capacity imposes a negative externality on the society: the higher the degree of testing virus-free people and asymptomatic COVID-19 infections is, the less testing kits are available for those who are truly in need, leaving more infected cases undetected and leading to more widespread infections. Such a squeeze comes from the testing resources “wasted” on virus-free people unintentionally, resulting in less testing resources available for symptomatic infected people. Third, it is well known that when we have more information on the infected population, such as their close contacts and places they have been, the number of infections can be reduced. And the marginal value of gaining more information on the infected people is increasing in controlling COVID-19. Finally, we find that reducing panic run and increasing testing capacities are complementary in helping suppress the total number of infected cases. Meanwhile, we show that decreasing testing turnaround time and increasing the testing capacity are also complementary. These results suggest that the increasing limited amount of testing capacity becomes more effective when paired up with reducing panic run or shortening the testing turnaround time.

Videos Sessions

Yuchao Dong

Q and A : 30 June 2021 - 01:00 - 01:20 PM SGT

Learning Equilibrium Mean-Variance Strategy

We study a dynamic mean-variance portfolio optimization problem under the reinforcement learning framework, where an entropy regularizer is introduced to induce exploration. Due to the time-inconsistency involved in a mean-variance criterion, we aim to learn an equilibrium strategy. Under an incomplete market setting, we obtain a semi analytical, exploratory, equilibrium mean-variance strategy that turns out to follow a Gaussian distribution. We then focus on a Gaussian mean return model and propose an algorithm to find the equilibrium strategy using the reinforcement learning technique. Thanks to a thoroughly designed policy iteration procedure in our algorithm, we can prove our algorithm's convergence under mild conditions, despite that dynamic programming principle and the usual policy improvement theorem fail to hold for an equilibrium solution. Numerical experiments are given to demonstrate our algorithm.
Xiaomin Liu

Q and A : 30 June 2021 - 01:00 - 01:20 PM SGT

Accurate Information Better? Sharing Commitment and Capacity Allocation under Supply Competition

The COVID-19 pandemic hit global supply chains and revealed the importance of supply chain resiliency. Firms are strengthening their supply chain resilience by using multi-sourcing strategy, increasing capacity buffers, improving supply chain transparency, etc. In this paper, we investigate the implication of demand information sharing in the supply chain where the manufacturer sources one emerging component from two capacity-constrained suppliers of different costs. Two suppliers compete for both the manufacturer’s component order in the emerging market and another component demand in a mature market. Two common rules are examined for the manufacturer to allocate her order: fixed allocation rule, with which the ratio that each supplier obtains is fixed, and proportional allocation rule, where the proportion depends on the reserved capacities of two supplies for the manufacturer. We model the problem as a Stackelberg game and characterize the firms’ equilibrium decisions. Our results show that under fixed allocation rule, the manufacturer always has no incentive of sharing. Otherwise, for proportional allocation rule, the manufacturer shares the information under moderate mature market potential with conditions of the information: if the accuracy of information is low, the manufacturer will share the information with both suppliers voluntarily; if the accuracy of information is moderate, the suppliers should buy the information from the manufacturer. Interestingly, the voluntary sharing of the manufacturer always hurts the supplier of low cost. This result is in contrast to the conventional wisdom that the supplier always benefits from the manufacturer’s voluntary sharing of accurate information. The reason is that the sharing behavior induces the low-cost supplier to allocate more capacity for the emerging market, and therefore less capacity is left for the mature market, which generates a competitive disadvantage in the mature market. An important managerial insight here is that since the information sharing increases the supply chain profit, the sharing agreement can be implemented through a positive transfer from the supplier of high cost to the low one, achieving a Pareto improvement for all parties. Moreover, we investigate the impact of information sharing on the choice of allocation rule. When the mature market potential is either low or high, no sharing happens and the preference of the manufacturer depends on the ratio of fixed allocation rule and the mature market demand. When the mature market potential is moderate: if the information accuracy is high, the manufacturer chooses no sharing and selects proportional allocation rule under the case that a high ratio is allocated to the supplier of high cost for fixed rule; if the information accuracy is low, information will be shared under proportional allocation rule and the sharing behavior pushes the manufacturer more likely to select proportional allocation rule compared with no information sharing.
Pei-Ling Yeh
Q and A : 30 June 2021 - 01:00 - 01:20 PM SGT

Implementation of Strengthened brainwave signals

Internet of Medicine (IoM) refers to the integration of nuclear magnetic resonance (NMR), electroencephalography (EEG), mixed Reality (MR), and Da Vinci through the technology of cloud Internet. Modern medical AI technology based on big-data of surgery and patients. For the treatment of tumor mutations, chronic diseases, minimally invasive medicine, etc., genetic mutation diseases and degenerative symptoms have made great progress from quantitative changes to qualitative changes. It may also promotes the integration of clinical medicine and 5G medical network to artificial intelligence and sensory motors. In the coming era, the existing non-invasive head-mounted brain wave device captures the signals of subconscious. Thus, the proposed research is forcused on strengthened the discrete brainwaves through the K-means clustering and deep learning algorithm and the research experience of ryodoraku theory derived by University of Tokyo, Japan. Based on the clinical experience of meridians acupuncture in the Emperor’s bible written by researchers from traditional Chinese medicine, the magnetic field of brain waves and pulse waves circulates the internal organs. Acupuncture treatment of epilepsy, depression and convulsions, etc. can effectively integrate brain waves and ryodoraku. It also serves as the basis for clinical testing of internal organs and their bio-energy as well as the human-machine interface of brain wave and pulse field. Therefore, the detection of traditional Chinese medicine can be effectively improved. Finally, The concept of the sensory motors, the brainwave, and the pulse information are physiologically integrated. The prospective goal of exploring prevention and psychiatric clinical research can be achieved in the coming decades.
Jingying Ding
Q and A : 30 June 2021 - 01:00 - 01:20 PM SGT

Feature-Based Nonparametric Inventory Control with Censored Demand

We study stochastic periodic-review inventory systems with lost sales, where the decision maker has no access to the true demand distribution a priori and can only observe historical sales data (referred to as censored demand) and feature information about the demand. In an inventory system, excess demand is unobservable due to inventory constraints, and sales data alone cannot fully recover the true demand. Meanwhile, feature information about the demand is abundant to assist inventory decisions. We incorporate features for inventory systems with censored demand. We propose two feature-based nonparametric inventory algorithms called the feature-based adaptive inventory algorithm and the dynamic shrinkage algorithm. Both algorithms are based on the stochastic gradient descent method. We measure the performance of the proposed algorithms through the average expected regret in finite periods: that is, the difference between the cost of our algorithms and that of a clairvoyant optimal policy with access to information, which is acting optimally. We show that the average expected cost incurred under both algorithms converges to the clairvoyant optimal cost at the rate of 𝑂 (1/ √T). The feature-based adaptive inventory algorithm results in high volatility in the stochastic gradients, which hampers the initial performance of regret. The dynamic shrinkage algorithm uses a shrinkage parameter to adjust the gradients, which significantly improves the initial performance. The idea of dynamic shrinkage for the stochastic gradient descent method builds on a fundamental insight known as the bias variance trade-off. Our research shows the importance of incorporating the bias-variance in a dynamic environment for inventory systems with feature information.
Zhao Jingyi
Q and A : 30 June 2021 - 01:00 - 01:20 PM SGT

Dynamic Programming-Based Split Algorithm for Solving the Multi-Trip Time-Dependent Multi-Vehicle Routing Problem

We consider a multi-vehicle multi-trip time-dependent vehicle routing problem (MVMT-TD-VRP). Our formulation aims to serve all requests with the given resources and to minimize the total travel cost. The multi-vehicle and time-dependence conditions, which are more realistic in practice, relax several restrictive constraints in the classical VRP---namely that there is a single vehicle and that its speed is constant. However, these conditions result in a VRP, which is already known to be an NP-hard problem, that is even more difficult to solve. We thus propose a mathematical formulation of MV-MT-TD-VRP and develop a dynamic programming-based algorithm (TD-Split) to find the allocation of vehicles, their routes, and the determination of each trip's departure time for given sequence of nodes of each vehicle. Moreover, we develop a memetic algorithm with customized mutation operators to generate giant tours for TD-Split. We test TD-Split on real-world data from a Singapore food and beverage company, compare the performance with results obtained using an off-the-shelf solver. Lastly, TD-Split is adapted to solve the classic capacity vehicle routing problem (CVRP).
Panca Jodiawan
Q and A : 01 July 2021 - 01:00 - 01:20 PM SGT

The Two-Echelon Vehicle Routing Problem with Relay Points and Occasional Drivers: Formulation and An Adaptive Large Neighborhood Search

This research introduces a new variant of two-echelon routing problems called the Two-Echelon Vehicle Routing Problem with Relay Points and Occasional Drivers (2E-VRPRP-OD). In the 2E-VRP-RP-OD, a two-echelon distribution network is addressed. A depot is located at the first echelon network, and all customers are located at the second echelon network. To serve the customers, the freights first need to be delivered from the depot to a set of intermediate locations called satellites by a set of first-echelon trucks. Two types of vehicles exist in the second-echelon network, i.e., city freighters and occasional drivers (ODs). The ODs are available on-road crowds who will perform deliveries in exchange for a predetermined amount of compensation. The employed ODs can collect packages not only at satellites but also at relay points. The utilized relay points need to be served by city freighters before ODs can collect the assigned packages. The key decisions involve (1) the assignment of freights to available satellites in the first echelon network, (2) the assignment of freights to available city freighters and ODs, and (3) the routes of first-echelon trucks, city freighters, and ODss. The objective of 2E-VRP-RP-OD is to minimize the total distribution costs consisting of the distribution cost of first-echelon trucks, city freighters, and ODs. We formulate this problem into a Mixed Integer Non-Linear Programming model solvable by an off-shelf exact solver, Gurobi, and develop an Adaptive Large Neighborhood Search (ALNS) to solve a set of newly generated 2E-VRP-RP-OD instances. Based on the conducted experiments, the proposed ALNS effectively solve the 2E-VRP-RP-OD instances by outperforming the results obtained by Gurobi in terms of both solution quality and computational time. We also show that considering occasional drivers to perform deliveries leads to lower operational costs than fully utilized city freighters in the second echelon networks. Moreover, sensitivity analyses regarding to characteristics of occasional drivers are presented as interesting managerial insights from solving the 2E-VRP-RP-OD
Yu-Ting Chiu
Q and A : 01 July 2021 - 01:00 - 01:20 PM SGT

Replenishment cycle based on SKU classification in Robotic Mobile Fulfillment System

In recent years there are already many types of research about the Robot Mobile Fulfillment System (RMFS) to support the development of smart warehouses in the e-commerce business. However, adjusting the layout and number of Automatic Guided Vehicles (AGVs) to fit a certain order pattern can be costly in the real world. Therefore, this study aims to build a simulation platform to easily adjust the parameters of RMFS in the smart warehouse. Previous researches use queuing theory and probability to simulate the traveling time of AGVs, but it involves a complicated queuing network that ignores details in the real-world system. RMFS are very complicated dynamic systems, with different settings that may affect the whole system throughput dramatically due to non-identical system bottlenecks.


This research focuses on fitting the real world by simulating the actual process of RMFS. AGVs bring shelves (pods) to the picking station, workers in the picking station pick items to fulfill the orders, then AGVs bring pods back to the storage area or bring them to the replenishment station to replenish Stock Keeping Units (SKUs) on the pod. The process also takes the replenishment part into consideration, while others mostly consider the picking process only.


The following could be modified in the simulation system: warehouse layout, number of AGVs, order pattern, SKU with its ABC classification, number of picking stations and replenishment station, and picking or replenish efficiency. The simulation could also implement different strategies for each process, including SKU to pod assignment, order to pod assignment, robot to pod assignment, robot motion control, and queuing of AGV strategies to the picking station.


This research focuses on the replenishment and SKU assignment part. In the replenishment part, it will decide which pod needs to be replenished based on the inventory level of the pod. Lower inventory level means the pod can’t fulfill the order in the higher service level, vice versa. On the other hand, SKU assignment is one of the important factors in traditional or modern warehouse management. Additionally, the timing of replenishment and AGV assigning is also challenging. The differences between these are the number of SKUs in ecommerce is too big, which is difficult to see the confidences or correlations between SKUs. Although this challenge exists, SKUs can be classified and stored based on the classification. The two parts have many benefits to optimize the e-commerce warehouse such as improving throughput, reducing robot travel distance, and reducing the number of robots. This study emphasizes the importance of building real-world simulation aims to help decision-making in complicated RMFS.
Nguyen Thi Anh Tuyet
Q and A : 01 July 2021 - 01:00 - 01:20 PM SGT

Economic Feasibility of Floating Offshore Wind Systems considering Technology, Depth of Sea Level and Distance to Shore

In recent decades, advantageous achievements of the technology have promoted the expansion of floating offshore wind energy to be one of the most promising renewable energy sources. Due to location characteristic and harsh weather conditions of the floating offshore wind farms, the operation and maintenance are hard–to–reach and the economic feasibility is highly uncertain. A huge number of researches have emerged recently to deal with the uncertainty of the economic feasibility of floating offshore wind energy. However, the researches considering the effects of various influential parameters remain small and limited, especially thosethat consider the impacts of different floating offshore turbine technologies, depth of sea level and distance to shore. Thisstudy aims to analyse comprehensively the economic feasibility and Levelised Cost of Energy (LCOE) of floating offshore wind energy considering the impacts of different floating offshore turbine technologies, depth of sea level and distance to shore. The results indicate that economic feasibility of floating offshore wind energy are strongly dependent on depth of sea level and distance from shore. The results also show that the current LCOE for a large floating offshore wind farm is still high cost with the range of 153-276 US$/MWh. Therefore, the government should provide appropriate subsidies to promote the development of floating offshore wind energy.
Mengyu Ji
Q and A : 01 July 2021 - 01:00 - 01:20 PM SGT

Automated Taxi Queue Management at High-Demand Venues

In major cities around the globe, taxi-like services have become increasingly dominant in providing point-to-point transportation services. This is especially true for locations that generate large amount of demands in a spiky manner and are relatively isolated, such as airports, stadiums, and exhibition centers. At these locations, it is common to see hundreds of people requesting for rides simultaneously, and this usually results in significant waiting time for passengers and congestions for drivers.


A common approach adopted by most venue operators to address such issue is to broadcast potential needs for taxis (or ride-hailing cars) ahead of or during the demand peaks. However, most implementations of such approach are ad hoc in nature and are rarely coordinated across the planning horizon. As a result, a balance in passenger demand and vehicle supply for rides is rarely achieved.
In this paper, we propose an automated taxi queue management framework that combines demand-side predictive analytics with supply-side policy optimization to come up with recommendations on the number of drivers to contact during different time periods, in anticipation of current and predicted future demand levels. Although the problem of taxi queue management at highdemand locations is general in nature, and can be applied to many different types of venues, each venue type brings different set of challenges, which can be driven by passenger demand patterns, venue design, and geographical locations.


To concretely demonstrate the value of combining both predictive analytics and control policy optimization, we choose to focus on the setting of an airport, more specifically, the Changi Airport in Singapore. Before the COVID-19 pandemic, Changi Airport was consistently ranked as one of the best and busiest airports in the world. For an airport that handles more than 68 million passenger traffic (in 2019), it is extremely challenging to maintain a high level passenger experience.


And the optimization of the taxi queue operations will be important in achieving this objective.
In solving the real-world use case at the airport, we aim to make the following contributions:
- We first propose a highly effective passenger demand prediction system that is based on the real-time flight arrival information. By monitoring cumulative passenger arrivals, and control for factors such as flight’s departure cities, we demonstrate that a simple linear regression model can accurately predict the number of passengers joining taxi queues.
- We then propose an optimal control strategy based on a Markov Decision Process to model the decisions of notifying individual taxis that are at different distances away from the airport. By using a real-world dataset, we demonstrate that an accurate passenger demand prediction system is crucial to the effectiveness of taxi queue management. In our numerical studies based on the real-world data, we observe that our proposed approach of optimal control with demand predictions outperforms the same control strategy, yet with Poisson demand assumption, by 43%. Against the status quo, which has no external control, we could reduce the gap by 23%. These results demonstrate that our proposed methodology has strong real world potential.
Jiayi Chen
Q and A : 01 July 2021 - 01:00 - 01:20 PM SGT

System Dynamics Analysis on the effectiveness of the vaccination and social mobilization Policy for COVID-19 in United States

The COVID-19 outbreak has brought along significant public health and economic challenges worldwide. Various health policies and non-pharmaceutical interventions have been adopted by different countries to control the spread of the virus. To shed light on the impact of the vaccination and social mobilization policies during this health and economic crisis, this paper proposes a system dynamics analysis on the effectiveness of these two policies to pandemic containment and economy in the United States. Based on the simulation of different policy scenarios, the findings are expected to help decisions and mitigation of the ramifications throughout this pandemic and beyond.

The Institute of Operations Research and Analytics (IORA) is a part of NUS’s Smart Nation Research Cluster.

Innovation 4.0, 3 Research Link, #04-01, Singapore 117602

IORA

Resources