ESSAYS ON INVENTORY MANAGEMENT, CAPACITY MANAGEMENT,
AND RESOURCE-SHARING SYSTEMS
by
Yang Bo
APPROVED BY SUPERVISORY COMMITTEE:
Milind Dawande, Co-Chair
Ganesh Janakiraman, Co-Chair
Alp Muharremoglu
Anyan Qi Copyright c 2017
Yang Bo
All rights reserved To my dear family and friends. ESSAYS ON INVENTORY MANAGEMENT, CAPACITY MANAGEMENT,
AND RESOURCE-SHARING SYSTEMS
by
YANG BO, BS
DISSERTATION
Presented to the Faculty of
The University of Texas at Dallas
in Partial Fulfillment
of the Requirements
for the Degree of
DOCTOR OF PHILOSOPHY IN
MANAGEMENT SCIENCE
THE UNIVERSITY OF TEXAS AT DALLAS
August 2017 ACKNOWLEDGMENTS
I am deeply thankful to my advisors, Drs. Milind Dawande and Ganesh Janakiraman, for patiently guiding me in my research endeavors, and for providing me with constructive advice and feedback. My frequent beneficial interactions with them have sharpened my research ability and have enabled me to become an independent researcher.
I owe much to my parents and my wife; their unconditional love and support continues to give me courage and make me stronger.
I would also like to express my gratitude to the following professors in my department, from whom I have benefited a lot: Alain Bensoussan, Metin Cakanyildirim, Dorothee Honhon, Elena Katok, Alp Muharremoglu, Shun-Chen Niu, Ozalp Ozer, Anyan Qi, Suresh P. Sethi, Kathryn Stecke, Serdar Simek, Shouqiang Wang, John J. Wiorkowski, and Shengqi Ye.
Last but not the least, I would like to thank my fellow PhD students: Ying Cao, Bahriye Cesaret, Jiayu Chen, Shaokuan Chen, Wei Chen, Ilhan Emre Ertan, Zhichao Feng, Blair Flicker, Shivam Gupta, Harish Guda, Bharadwaj Kadiyala, Ismail Kirci, Chungseung Lee, Jingyun Li, Ting Luo, Sandun Perera, Xi Shan, Sina Shokoohyar, Yulia Vorotyntseva, Xiao Zhang, and many others. Many thanks for making my stay at UTD joyful.
April 2017
v ESSAYS ON INVENTORY MANAGEMENT, CAPACITY MANAGEMENT,
AND RESOURCE-SHARING SYSTEMS
Yang Bo, PhD The University of Texas at Dallas, 2017
Supervising Professors: Milind Dawande, Co-Chair Ganesh Janakiraman, Co-Chair
This dissertation consists of three essays, each focusing on one of the following three impor- tant topics in operations management: inventory management, capacity management, and management of resource-sharing systems. These topics are each summarized below.
In the first essay, we investigate the following integrality question for inventory control on distribution systems: Given integral demands, does an integral optimal policy exist? We show that integrality holds under deterministic demand, but fails to hold under stochastic demand. In distribution systems with stochastic demand, we identify three factors that influence the gap between integral and real optimal policies: shipping cost variation across time, holding cost difference across stages, and economies of scale. We then obtain a tight worst-case bound for the gap, which captures the impact of all these factors.
In the second essay, we investigate the computational complexity of determining the capacity of a process, a fundamental concept in Operations Management. We show that it is hard to calculate process capacity exactly and, furthermore, also hard to efficiently approximate it to within a reasonable factor; e.g., within any constant factor. These results are based on a novel characterization, which we establish, of process capacity that relates it to the
vi fractional chromatic number of an associated graph. We also show that capacity can be efficiently computed for processes for which the collaboration graph is a perfect graph.
The third essay addresses an important problem in resource-sharing systems. We study the minimum-scrip rule in such a system: whenever a service request arises, among those who volunteer and are able to provide service, the one with the least number of scrips (also known as coupons) is selected to provide service. Under mild assumptions, we show that everybody in the system being always willing to provide service is a Nash Equilibrium under the minimum-scrip rule. This suggests that the minimum scrip rule can lead to a high level of social welfare.
vii TABLE OF CONTENTS
ACKNOWLEDGMENTS ...... v ABSTRACT ...... vi LIST OF FIGURES ...... xi CHAPTER 1 INTRODUCTION ...... 1 CHAPTER 2 ON INTEGRAL POLICIES IN DETERMINISTIC AND STOCHASTIC DISTRIBUTION SYSTEMS ...... 4 2.1 Introduction ...... 5 2.2 Preliminaries ...... 8 2.2.1 Network Representation and Notation ...... 8 2.2.2 Economies of Scale Structure ...... 10 2.3 Distribution Systems with Deterministic Demand ...... 10 2.4 Distribution Systems with Stochastic Demand ...... 13 2.4.1 Failure of Integrality When Assumption S Does Not Hold ...... 14 2.4.2 Failure of Integrality When Assumption H Does Not Hold ...... 16 2.4.3 Bounding the Integrality Gap ...... 17 2.5 Managerial Implications of Results ...... 22 2.5.1 Unpredictable Demand: Variation in Shipping Costs Across Time . . 22 2.5.2 Unpredictable Demand: Holding Cost Variation Across Stages . . . . 23 2.5.3 Unpredictable Demand: Economies of Scale ...... 24 2.5.4 Predictable Demand ...... 24 2.6 Proofs of Results ...... 24 2.6.1 Proof of Theorem 2.3.1 ...... 24 2.6.2 Proof of Proposition 2.4.1 ...... 26 2.6.3 Proof of Proposition 2.4.2 ...... 29 2.6.4 Proof of Theorem 2.4.1 ...... 31 CHAPTER 3 ON FINDING PROCESS CAPACITY ...... 40 3.1 Introduction ...... 40 3.1.1 A Brief Overview of the Related Literature ...... 43
viii 3.2 Preliminary Results ...... 44 3.2.1 Formal Definition of Process Capacity ...... 45 3.2.2 Sufficiency of Cyclic Policies and Redundance of Precedence and Non- Preemption Constraints ...... 47 3.2.3 Static Planning Problem for Collaboration ...... 50 3.3 Characterizing Process Rates for Subclasses of Cyclic Policies ...... 52 3.3.1 The Maximum Process Rate Over 1-Unit Cyclic Policies ...... 52 3.3.2 The Maximum Process Rate Over k-Unit Cyclic Policies ...... 54 3.3.3 The Maximum Process Rate Over Cyclic Policies (Process Capacity) 55 3.4 Hardness and Inapproximability Results ...... 57 3.4.1 Inapproximability of Determining Process Capacity ...... 57 3.5 Good News ...... 58 3.6 Extension: Relaxation of Assumption (3.2.14) ...... 61 3.7 Proofs of Some Results ...... 63 3.7.1 Proof of Lemma 3.2.1 ...... 63 3.7.2 Proof of Lemma 3.3.2 ...... 65 3.7.3 Proof of Theorem 3.4.1 ...... 66 3.7.4 Proof of Theorem 3.5.3 ...... 66 3.7.5 Proof of Lemma 3.6.2 ...... 71 CHAPTER 4 ANALYSIS OF SCRIP SYSTEMS ...... 76 4.1 Introduction ...... 76 4.2 Model ...... 78 4.3 Analysis and Results ...... 79 4.3.1 The Main Result ...... 80 4.3.2 Intermediate Results ...... 82 4.3.3 Proof of Theorem 4.3.1 ...... 83 4.4 Proofs of the Intermediate Results – Lemmas 4.3.2 – 4.3.4 ...... 84 4.4.1 Proof of Lemma 4.3.2 ...... 84 4.4.2 Proof of Lemma 4.3.3 ...... 86
ix 4.4.3 Proof of Lemma 4.3.4 ...... 87 APPENDIX PROOFS OF LEMMA 3.2.2 ...... 95 REFERENCES ...... 104 BIOGRAPHICAL SKETCH ...... 108 CURRICULUM VITAE
x LIST OF FIGURES
2.1 (a) A Distribution System and (b) An Assembly System...... 9 2.2 The Distribution Network in the Proofs of Propositions 2.4.1 and 2.4.2...... 15
A.1 Policy π1 in Example 1 ...... 97 A.2 Policyπ ˆ in Example 1 ...... 98
A.3 Policy π0 in Example 2 ...... 101 A.4 Policyπ ˇ in Example 2 ...... 102
xi CHAPTER 1
INTRODUCTION
This dissertation includes three essays. The first two essays focus on inventory and capacity management, and the third essay addresses an important problem in resource-sharing sys- tems. We now briefly describe the problems analyzed in these essays and summarize our contributions. In the first essay, we study the “integrality” question for dynamic optimization models of inventory control: Given integral initial inventory levels, capacity constraints, and demand realizations, does there exist an integral optimal policy? One practical implication of this question lies in whether or not full-truckload (FTL) shipping is optimal if customer demand is in integral number of truckloads. In this essay, we investigate the integrality question in single-product, multi-echelon distribution systems and show that integrality holds under deterministic demand, but fails to hold under stochastic demand. In distribution systems with stochastic demand, Less-Than-Truckload (LTL) shipping can be significantly cheaper than the cost of the optimal FTL shipping policy, even in the presence of economies of scale. For instance, this occurs in settings where shipping costs are expected to increase in the future and/or inventories are more expensive to hold upstream than downstream. In such situations, our results highlight the importance of strategically positioning inventory: LTL shipments can offer a more balanced allocation of inventory across the distribution network, leading to benefits that can exceed the savings from FTL shipments due to economies of scale. However, when the cost parameters are fairly constant across time and inventory holding costs are not significantly higher upstream than downstream, then the difference between the costs of optimal FTL and optimal LTL shipping is provably marginal. The second essay investigates the analytical tractability of process capacity. The concept of the capacity of a process and the associated managerial insights on the investment and management of capacity are of fundamental importance in Operations Management. Most
1 OM textbooks use the following simple approximation to determine process capacity: the
capacity of each resource is first calculated by examining that resource in isolation; process
capacity is then defined as the smallest (bottleneck) among the capacities of the resources.
In a recent paper, (Gurvich and Van Mieghem, 2015) show that this “bottleneck formula”
can be significantly inaccurate, and obtain a necessary and sufficient condition under which
it correctly determines process capacity. We provide further clarity on the intractability
of process capacity by showing that it is hard to calculate process capacity exactly and,
furthermore, also hard to efficiently approximate it to within a reasonable factor; e.g., within
any constant factor. These results are based on a novel characterization, which we establish,
of process capacity that relates it to the fractional chromatic number of an associated graph.
An important implication is that it is unlikely that we can replace the bottleneck formula
with a simple but close approximation of process capacity. From a practical viewpoint, our
analysis results in a natural hierarchy of subclasses of policies that require an increasing
amount of sophistication in implementation and management: While process capacity is the
maximum long-term process rate achievable over all feasible policies, we provide a precise
expression for the maximum process rate over policies in each subclass in this hierarchy, thus
highlighting the tradeoff between operational difficulty and the achievable process rate.
The third essay addresses an important problem in resource-sharing systems. In a recent
paper, (Johnson et al., 2014) use an infinitely-repeated game with discounting, among a
set of homogeneous players, to model a scrip system. In each period, a randomly-chosen
player requests service; all the other players have a choice of whether or not to volunteer
to provide service. Among the players who volunteer, the service provider is chosen using
the minimum-scrip rule: A player with the minimum number of scrips is chosen as the service provider. The authors study the always-trade strategy for a player, i.e., the strategy
of a player always willing to provide service, regardless of the distribution of scrips among
the players. A key result of their work is that, under the minimum-scrip rule, for any
2 number of players and for every discount factor close enough to one, there exists a Nash equilibrium in which each player plays the always-trade strategy. This result, however, is established under an assumption of severe punishment: If a player selected to provide service refuses to do so even once, then that player will be forever banned from participating in the system, thus losing all potential future benefit. (Johnson et al., 2014) explain that this assumption is unsatisfactory, particularly for large scrip systems, due to difficulties in detecting players who refuse to provide service and verifying the reason(s) for their refusal, and the consequent possibility of players being unfairly banned. Motivated by this concern, the authors suggest an important direction of future research – investigate whether or not the always-trade strategy is an equilibrium without the punishment assumption. In this essay, we address the question posed in (Johnson et al., 2014) for a generalization of their model. Our generalization allows for the possibility that players might genuinely not be available to provide service, say due to sickness. Without using the assumption of punishment, we show that when the number of players is large enough and the discount factor is close enough to one, there exists an -Nash equilibrium in which each player plays the always-trade strategy.
3 CHAPTER 2
ON INTEGRAL POLICIES IN DETERMINISTIC AND STOCHASTIC
DISTRIBUTION SYSTEMS
Authors – Yang Bo, Milind Dawande, Ganesh Janakiraman, and S. Thomas McCormick
Naveen Jindal School of Management, SM 30
The University of Texas at Dallas
800 West Campbell Road
Richardson, Texas 75080-3021
Portions of this chapter are reprinted by permission, Yang Bo, Milind Dawande, Ganesh Janakiraman, S. Thomas McCormick, “On Integral Policies in Deterministic and Stochastic Distribution Systems”, Operations Research, Published online in Articles in Advance 04 Apr 2017. Copyright (2017), the Institute for Operations Research and the Management Sciences, 5521 Research Park Drive, Suite 200, Catonsville, Maryland 21228 USA.
4 2.1 Introduction
In a recent paper, (Chen et al., 2014) pose the following fundamental “integrality” question
for dynamic optimization models of inventory control: When the starting inventory levels are
integer-valued and all demands are integer-valued random variables, for what kind of inven-
tory problems do there exist integral optimal ordering quantities, even if we allow them to be
real valued? For a single-product, multi-echelon assembly network with stochastic or deter-
ministic demand, (Chen et al., 2014) prove that there always exists an integral optimal policy.
Our main goal in this paper is to examine the integrality question for both deterministic-
and stochastic-demand single-product multi-echelon distribution networks, and contrast the
answers with those for assembly networks.
Why do the notion of integrality and our results matter? To answer this, consider the
following practical setting for an arbitrary distribution network:
Suppose that customer demand is in full truckloads (FTL) and that shipments to cus-
tomers from the retail locations are also required to be in full truckloads. We define a unit
to be a truckload of shipment. In any period, the ordering/shipping cost incurred in trans-
porting a quantity q from a node of the network to its immediate successor in the network is: