List of Speakers:
Mor Armony – New York University
Onno Boxma – Eindhoven University of Technology
Hong Chen – Shanghai Advanced Institute of Finance (SAIF)
Avigdor Gal – Technion — Israel Institute of Technology
Jim Dai – Cornell University
Junfei Huang – The Chinese University of Hong Kong (CUHK)
Haya Kaspi – Technion — Israel Institute of Technology
Ger Koole – Vrije Universiteit Amsterdam
Avi Mandelbaum – Technion — Israel Institute of Technology
Isaco Meilijson – Tel Aviv University
Petar Momcilovic – Texas A&M University
Kavita Ramanan – Brown University
Yaakov Ritov – Hebrew University
Haipeng Shen – The University of Hong Kong (HKU)
Sasha Stolyar – University of Illinois at Urbana-Champaign
Johan Van Leeuwaarden – Eindhoven University of Technology
Uri Yechiali – Tel Aviv University
Sergey Zeltyn – IBM Research Labs
Hanqin Zhang – National University of Singapore (NUS)
Abstracts:
Mor Armony. (New York University, Stern School of Business)
When to send a patient home? The effect of infection risk on hospital length-of-stay
Following treatment hematology patients face an increased risk of developing an infection. If a patient remains at the hospital their infection risk is higher than at home, but once an infection develops the mortality risk at the hospital is lower than at home due to quick access to appropriate treatment. We study the problem of dynamically determining discharge times for hematology patients subject to capacity constraints. We show that in an overloaded operating regime this dynamic problem reduces to a static problem with a simple solution of a two-class two-discharge thresholds form. We characterize the specific optimal solutions under empirically
driven time-to-infection distributions.
Based on joint work with Galit Yom-Tov.
Onno Boxma. (Eindhoven University of Technology, Department of Mathematics and Computer Science)
Delta probing policies for redundancy
We consider job dispatching in systems with N parallel servers. In redundancy-d policies, replicas of an arriving job are assigned to d out of N servers selected uniformly at random (without replacement) with the objective to reduce the delay. We introduce a quite general workload model, in which job sizes have some probability distribution while the speeds (slowdown factors) of the various servers for a given job are allowed to be inter-dependent and non-identically distributed. This allows not only for inherent speed differences among different servers, but also for affinity relations.
We further propose two novel redundancy policies, so-called delta-probe-d policies, where d probes of a fixed, small, size delta are created for each incoming job, and assigned to d servers selected uniformly at random. As soon as the first of these d probe tasks finishes, the actual job is assigned for execution — with the same speed — to the corresponding server and the other probe tasks are abandoned. We also consider a delta-probe-d policy in which the probes receive preemptive-resume priority over regular jobs. The aim of these policies is to retain the benefits of redundancy-d policies while accounting for systematic speed differences and mitigating the risks of running replicas of the full job simultaneously for long periods of time. Analytic and numerical results are presented to evaluate the effect of both probing policies on the job latency, and to illustrate the potential performance improvements.
Based on joint work with Sem Borst and Youri Raaijmakers.
Hong Chen. (Shanghai Advanced Institute of Finance)
The Path and Opportunities Toward Hydrogen Energy Development
Hydrogen has been considered by some the ultimate fuel and energy carrier. In this talk, I would like to address the following questions:
- Why is the hydrogen energy important?
- Have the “Four Miracles” (coined by Dr. Steven Chu, the Secretary of US DOE under President Obama) happened?
- What is the path toward the hydrogen energy development?
Moreover, I will introduce one breakthrough technology, Liquid Organic Hydrogen Carrier (LOHC), that would make happen the last two of the Four Miracles. This is a disruptive technology that will make the hydrogen become the main stream energy and make the earth really clean. As the main carrier is a liquid, the operational performance will be connected to modeling a fluid network (though it is not part of the talk).
Avigdor Gal. (Technion, Faculty of Industrial Engineering & Management)
From Queueing Theory to Queue Mining
In my talk I will unfold a story of a jointly supervised PhD student that taught Avishai and me how to integrate the results of statistical modeling with state-of-the-art machine learning towards a nice research contribution in the area of process mining.
Jim Dai. (Institute for Data and Decision Analytics (iDDA) at the Chinese University of Hong Kong, Shenzhen and School of Operations Research and Information Engineering Cornell University)
Hospital inpatient operations: modeling, analysis, and optimization
Prolonged waiting time for admission from the emergency department (ED) to inpatient wards is a leading cause of ED overcrowding and can negatively impact patient outcomes. In this talk, (a) I will review and compare two service time models that attempt to capture the time-of-day effect of this waiting time; they are the two-time-scale service time model proposed in Shi et al (2016) and the inspection-delay service time model proposed in Chan, Dong, and Green (2016). (b) I will demonstrate the role of Stein’s method to generate approximations for the waiting time performance that are accurate in a variety of parameter regimes. (c) I will discuss recent efforts in developing approximate dynamic programming algorithms to optimize the trade-off between waiting time and the fraction of patients who are assigned to non-primary beds.
Based on joint work with Pengyi Shi at Purdue University
Junfei Huang. (Chinese University of Hong Kong, Business School)
On the Taylor Expansion of Value Functions
We introduce a framework for approximate dynamic programming that we apply to discrete time chains on with countable action sets. Our approach is grounded in the approximation of the (controlled) chain’s generator by that of another Markov process. In simple terms, our approach stipulates applying a second-order Taylor expansion to the value function to replace the Bellman equation with one in continuous space and time where the transition matrix is reduced to its first and second moments. In some cases, the resulting equation (which we label TCP) can be interpreted as corresponding to a Brownian control problem. When tractable, the TCP serves as a useful modeling tool. More generally, the TCP is a starting point for approximation algorithms. We develop bounds on the optimality gap—the sub-optimality introduced by using the control produced by the “Taylored” equation. These bounds can be viewed as a conceptual underpinning, analytical rather than relying on weak convergence arguments, for the good performance of controls derived from Brownian approximations. We prove that, under suitable conditions and for suitably “large” initial states, (i) the optimality gap is smaller than a fraction of the optimal value, where is the discount factor, and (ii) the gap can be further expressed as the infinite horizon discounted value with a “lower-order” per period reward. Computationally, our framework leads to an “aggregation” approach with performance guarantees. While the guarantees are grounded in PDE theory, the practical use of this approach requires no knowledge of that theory.
Joint work with Anton Braverman and Itai Gurvich.
Haya Kaspi. (Technion – Israel Institute of Technology, Faculty of Industrial Engineering & Management)
From Multi Armed Bandits to Walsh spider Brownian Motion
This talk is based on joint work with Avi Mandelbaum back in the nineties and some discussions with Yannis Karatzas some twenty five years later. We start with Multi Armed Bandits with d arms, where each arm is governed by a linear Standard Brownian Motion and the reward rate of arm i at time t is equal to riBi(t) where Bi is the Brownian motion associated with arm i and ri ≥ 0. Applying to this bandit an index strategy and interpreting the times that we pull arm i as the times that the spider lives on ray Θi from the origin of ℜ2, we obtain a Walsh spider Brownian motion on the rays Θ1 ,….,Θd. Using the index strategy we can compute the probabilities (p1,…,pd) that, on its excursions from the origin, the spider will shoot off to the various rays.
Ger Koole. (Vrije Universteit Amsterdam, Department of Mathematics)
Optimal Contact Center Staffing and Scheduling with Machine Learning
We consider making weekly schedules in a multi-channel multi-skill call center.
The only known accurate solution method is simulation optimization, but this tends to be slow and therefore leading to highly suboptimal solutions. To speed up this process we approximate the performance by a predictive model based on a limited set of simulations and optimize using this approximation. We present numerical results from a real call center.
Based on joint work with Qingchen Wang and Siqiao Li.
Avi Mandelbaum. (Technion — Israel Institute of Technology)
On “Theomperical” (Operations) Research
or
On `Data-Based Network-Models’ of `Human-Centric Congestion-Prone Service-Systems’
Isaac Meilijson. (Tel Aviv University, School of Mathematical Sciences)
Pollaczek-Khintchine revisited: M/G/1 service and sojourn times under delayed report of task conclusions.
Reversing roles in the Pollaczek-Khintchine formula, sojourn times under FIFO M/G/1 determine service time distribution. Identifiability hinges on a novel formula for mean service time via Stieltjes-Laplace transform of sojourn times. The model is applied to service systems that may delay reporting task conclusions.
Petar Momcilovic. (Texas A&M University, Department of Industrial & Systems Engineering)
Data-driven appointment scheduling under uncertainty
We develop a novel, data-driven approach to deal with appointment sequencing and scheduling in a multi-server system, where both customer punctuality and service times are stochastic. Our model is calibrated using a data set of unprecedented resolution, gathered at a large-scale outpatient oncology practice. This data set combines real-time locations, electronic health records and appointments logs. Our approach yields tractable and scalable solutions that accommodate hundreds of jobs and servers. We demonstrate the performance of our algorithm by comparing it with existing state-of-the-art sequencing and scheduling algorithms.
Based on joint work with Avishai Mandelbaum and Nikos Trichakis.
Kavita Ramanan. (Brown University)
Rare Nash Equilibria in Games with Many Players
We study the asymptotic behavior of Nash equilibria in static games with a large number of agents. In particular, we establish law of large number limits and large deviation principles for the set of Nash equilibria and discuss applications to congestion games and the price of anarchy. Time permitting, we will also discuss related results in the case of differential games. This is joint work with Daniel Lacker.
Yaacov Ritov. (The Hebrew University of Jerusalem, Department of Statistics)
Learning the unutilized behavior of impatient patients.
We consider two separate subjects. In the first, we try to predict the future demand of customers (and patients) who leave the queue before declaring their demands. In the second, we consider patients who leave the queue at an undeclared time and compare them to patients that either declare their inpatient or were lucky to get service before their patient expired.
Haipeng Shen. (The University of Hong Kong, Faculty of Business and Economics)
A Statistician’s 18-year Journey in the Queueing Land with Avi
Back in 2000, Professor Avi Mandelbaum took the Bank Anonymous data to Wharton, which started my PhD dissertation under the supervision of Professor Larry Brown. This presentation reviews my 18-year journey in the queueing land together with Avi. I shall use examples from customer service call center workforce management and healthcare delivery systems, from a statistician’s perspective.
Alexander (Sasha) Stolyar. (University of Illinois, Industrial and Enterprise Systems Engineering)
A queueing system with on-demand servers
In the basic model randomly arriving customers are matched to servers, which are invited on-demand and arrive after random delays. One customer and one server are matched (for service) and leave the system. The basic objective is to minimize both customer and server waiting times. The model has several applications, including call/contact centers, inventory systems, telemedicine. We study two feedback-based adaptive server invitation algorithms, and derive fluid and diffusion limits (including steady-state limits) as the customer arrival rate becomes large. The results imply, in particular, that in the classical single-item inventory system with exponentially distributed lead times and order crossovers, it is possible to achieve potentially infinite improvement in the average inventory cost over the standard constant-base-stock algorithm. We also consider some extensions of the basic model, where servers may leave or remain in the system after each service completion.
Based on joint works with G.Pang (Pennstate), Q.Wang (UIUC), L.Nguyen (Lehigh Univ.)
Johan Van Leeuwaarden. (Eindhoven University of Technology, Department of Mathematics and Computer Science)
Asymptotically Optimal Load Balancing Topologies
Consider a system of N servers interconnected by some underlying graph topology G. Each incoming task is irrevocably assigned to whichever server has the smallest number of tasks among the one where it appears and its neighbors in G. This model arises in the context of load balancing in large-scale cloud networks and data centers, and has been extensively investigated in the case G is a clique (JSQ with parallel servers). For an arbitrary graph G, the lack of exchangeability complicates analysis, and queue lengths tend to be worse than for a clique. Using stochastic-process limits and coupling, we show when the queue length process on G is equivalent to that on a clique in the large-N limit, for subcritical (mean field) and critical (Halfin-Whitt or QED) regimes.
Based on joint work with Sem Borst and Debankur Mukherjee.
Uri Yechiali. (Tel Aviv University, School of Mathematical Sciences)
On Ticket Queues with Both Regular and Smart (Strategic) Customers
In his Master thesis, Avishai studied an M/G/1 queue with a single smart customer. While every other regular arrival joins the system unconditionally, the (only) smart customer is allowed to observe the queue and then choose between three options: (i) Join the system and stay until being served; (ii) Leave right away; or (iii) Wait until the next service completion, upon which a new decision – Join, Leave or Wait – is taken.
The current work studies a Markovian single-server ticket queue where, upon arrival, each customer can draw a numbered ticket, while the ticket number of the customer currently being served is displayed on a panel. The difference, D, between the above two numbers is called the virtual queue length. The population of customers is comprised of two types: ‘regular‘ (non-strategic), or ‘smart‘ (strategic). A regular customer draws a ticket regardless of the observed virtual queue length and stays in the system until being served. A smart customer, depending on the virtual queue length D, may either Join, Leave or go to ‘Orbit’ for a random duration time. If, upon return from orbit, a smart customer realizes that her/his turn passed, he/she balks. Otherwise, that customer assumes his/her original ticket-position in the queue.
We analyze this notorious system (only a single published paper is cited in Hassin’s (2016) book ‘Rational Queueing’), and calculate the optimal mean orbiting time of customers. Numerical examples are presented.
Joint work with G. Hanukov and S. Anily.
Sergey Zeltyn. (IBM Research – Haifa)
Reducing Costs and Improving Quality in Wastewater Treatment
Wastewater treatment is carried out through the interaction of complex biological, physical, and chemical processes. Typically, wastewater treatment plants operate in a conservative and inefficient risk-averse mode that makes it difficult to quantify the risks or truly minimize the costs. We developed an innovative operational control that applies descriptive, predictive, and prescriptive analytics to improve process efficiency and reduce costs. The descriptive analytics uses historical sensor data to build a simulation model and plant state estimator. The predictive analytics models the wastewater treatment process behavior using a transition probability matrix that we estimated. Finally, our prescriptive Constrained Markov Decision Process analytics offers recommendations for improved plant operation.
Based on joint work with A. Zadorojniy, S. Wasserkrug, and V. Lipets
Hanqin Zhang. (National University of Singapore, NUS Business School)
Impatience and Learning in Queues
Customers often abandon waiting in queues when they get impatient. Prior literature on Markovian queues shows that it is not rational for customers to quit “midway”: Customers should either quit immediately on arrival (balk) or wait till the completion of their service. We show how in-queue abandonment behavior can be rational in queues because of learning. We compare how rational Bayesian customers make abandonment decisions under different information disclosures. Our paper reveals interesting features in waiting behavior, showing that customers can be (rationally) more patient in slower shorter queues, than in faster longer queues. Using stochastic comparisons, we demonstrate that customers who anticipate a congested system can become more impatient. Finally, we show that Bayesian customers may exhibit a more conservative threshold joining behavior compared to myopic customers with the same priors.
Based on joint work with Senthil Veeraraghavan and Li Xiao.