International Series in

Operations Research & Management Science

Tadeusz Sawik

Supply Chain
Disruption
Management Using
Stochastic Mixed
Integer Programming
International Series in Operations Research
& Management Science

Volume 256

Series Editor
Camille C. Price
Stephen F. Austin State University, TX, USA

Associate Series Editor

Joe Zhu
Worcester Polytechnic Institute, MA, USA

Founding Series Editor

Frederick S. Hillier
Stanford University, CA, USA
More information about this series at http://www.springer.com/series/6161
Tadeusz Sawik

Supply Chain Disruption

Management Using
Stochastic Mixed Integer
Programming

123
Tadeusz Sawik
Department of Operations Research
AGH University of Science and Technology
Kraków
Poland

ISSN 0884-8289 ISSN 2214-7934 (electronic)

International Series in Operations Research & Management Science
ISBN 978-3-319-58822-3 ISBN 978-3-319-58823-0 (eBook)
DOI 10.1007/978-3-319-58823-0
Library of Congress Control Number: 2017941460

© Springer International Publishing AG 2018

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made. The publisher remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.

Printed on acid-free paper

This Springer imprint is published by Springer Nature

The registered company is Springer International Publishing AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
To Bartek,
Siasia,
Kappa,
and Toranoko with love
and to the Memory of my Parents
Preface

Scope

This book deals with stochastic combinatorial optimization problems in supply

chain disruption management, with a particular focus on management of disrupted
flows in customer-driven supply chains. The problems are modeled using a
scenario-based stochastic mixed integer programming to address risk-neutral,
risk-averse, and mean-risk decision-making in the presence of supply chain dis-
ruption risks. One of the main objectives of this book is to present a computa-
tionally efficient portfolio approach to integrated decision-making in global supply
chains under disruption risks, where the portfolio is defined as the allocation of
demand for parts and finished products, respectively, among suppliers and pro-
duction facilities. The allocation of demand for parts among suppliers is defined as a
supply portfolio, whereas the allocation of demand for products among production
facilities (e.g., assembly plants) is defined as a demand or capacity portfolio. Unlike
most of reported research on the supply chain risk management which mainly
focuses on the risk mitigation decisions taken prior to a disruption, the proposed
portfolio approach combines decisions made before, during, and after the disrup-
tion. When a disruption occurs, the primary portfolios determined prior to a dis-
ruption are replaced by recovery portfolios. The selection of portfolios will be
combined with management of disrupted material flows, i.e., supply, production,
and distribution scheduling under disruption risks. This book demonstrates that the
developed portfolio approach leads to a well-structured decision-making and
computationally efficient mathematical formulations, in particular, to stochastic
mixed integer programs with a strong LP relaxation. Moreover,
• integrated versus hierarchical decision-making is compared depending on the
available information on disruptive events;
• a multi-objective decision-making is analyzed to trade off between: cost versus
service level objective functions, fairness versus non-equitability of objective
functions, average versus worst-case performance of a supply chain, etc.;

vii
viii Preface

• a multi-period decision-making is modeled to capture dynamics of disruption

and recovery processes, i.e., static versus dynamic portfolios, scheduling of
supply, production, and distribution operations to control disrupted flows under
time-varying conditions such as demands and capacities;
• a multi-level disruption scenarios are modeled to capture partially disrupted
flows, partially fulfilled orders, partially recovered facilities, and partially
available capacity.
This book also addresses the issue of fundamental understanding of average-case
and worst-case performance of a global supply chain in the presence of flow dis-
ruption risks as well as understanding of the recovery mechanisms.
A straightforward computational approach used in this book is to solve the
deterministic equivalent mixed integer program of a two-stage stochastic mixed
integer program with recourse, which allows for a direct application of commer-
cially available software for mixed integer programming. In the computational
experiments reported throughout this book, an advanced algebraic modeling lan-
guage AMPL (see, Fourer et al. 2003) and the CPLEX, Gurobi, and XPRESS
solvers have been applied.

Content

This book is divided into an introductory Chap. 1, where an overview of supply

chain disruption modeling and management is provided, and the five main parts.
Part I addresses selection of a supply portfolio, Part II, integrated selection of
supply portfolio and scheduling, Part III, integrated, equitably efficient selection of
supply portfolio and scheduling, Part IV, integrated selection of primary and
recovery supply (and demand) portfolios and scheduling, and finally, Part V
addresses disruption management of information flows in supply chains.
Part I (Chaps. 2–4) introduces the portfolio approach for supplier selection and
order quantity allocation in the presence of supply chain disruption risks, i.e., for
determining a supply portfolio. The proposed portfolio approach allows the two
popular in financial engineering percentile measures of risk, Value-at-Risk
(VaR) and Conditional Value-at-Risk (CVaR), to be applied for managing the
risk of supply disruptions. For a finite number of scenarios, CVaR allows the
evaluation of worst-case costs (or worst-case service level) and shaping of the
resulting cost (service level) distribution through optimal supplier selection and
order quantity allocation decisions, i.e., the selection of optimal supply portfolio.
Part I is comprised of these chapters:
• Chapter 2, Selection of Static Supply Portfolio. This chapter deals with selection
of a static supply portfolio under disruption risks, i.e., for determining a
single-period allocation of demand for parts among selected suppliers to mini-
mize expected or expected worst-case cost or maximize expected or expected
worst-case service level.
Preface ix

• Chapter 3, Selection of Dynamic Supply Portfolio. In this chapter, the static

portfolio approach and the stochastic mixed integer programming formulations
presented in Chap. 2 are enhanced for a multi-period supplier selection and
order quantity allocation in the presence of both the low-probability and
high-impact supply chain disruption risks and the high-probability and
low-impact supply chain delay risks. The suppliers are subject to local delivery
delay risks, and both local and regional delivery disruption risks. In the delivery
scenario analysis, both types of the supply chain risks are simultaneously
considered.
• Chapter 4, Selection of Resilient Supply Portfolio. In this chapter, the portfolio
approach and SMIP (Stochastic Mixed Integer Programming) models presented
in Chap. 2 are enhanced for the combined selection and protection of part
suppliers and order quantity allocation in a supply chain with disruption risks.
The protection decisions include the selection of suppliers to be protected
against disruptions and the allocation of emergency inventory of parts to be
prepositioned at the protected suppliers so as to maintain uninterrupted supplies
in case of natural or man-made disruptive events.
Part II of this book concerns with integrated selection of supply portfolio and
scheduling. The medium- to short-term decisions of the supplier selection and order
quantity allocation, driven by the time-varying customer demand, are made along
with scheduling of customer orders execution and distribution. The advantage of
such a joint decision-making is especially evident in the presence of supply chain
disruption risks. Part II has two chapters:
• Chapter 5, Integrated Selection of Supply Portfolio and Scheduling of
Production. This chapter proposes a SMIP approach to integrated supplier
selection and customer order scheduling in the presence of supply chain dis-
ruption risks, for a single, dual, or multiple sourcing strategy. The suppliers are
assumed to be located in two or more disjoint geographic regions: in the pro-
ducer’s region (domestic suppliers) and outside the producer’s region (foreign
suppliers). The supplies are subject to independent random local disruptions that
are uniquely associated with a particular supplier and to random regional dis-
ruptions that may result in disruption of all suppliers in the same geographic
region simultaneously.
• Chapter 6, Integrated Selection of Supply Portfolio and Scheduling of
Production and Distribution. The purpose of this chapter is to study the inte-
grated decision-making to simultaneously select suppliers of parts, allocate
order quantity, and schedule production and delivery of finished products to
customers in a supply chain under disruption risks. In addition to supplier
selection, order quantity allocation, and scheduling of customer orders, distri-
bution of finished products to customers is simultaneously considered with
different shipping methods to optimize the trade-off between cost and service
level. The three different shipping methods will be modeled and compared for
the distribution of products: batch shipping with a single shipment of different
x Preface

customer orders, batch shipping with multiple shipments of different customer

orders, and individual shipping of each customer order immediately after its
completion.
Part III addresses equitably efficient selection of supply portfolio and scheduling.
A fair optimization of an average performance of a supply chain with respect to
equally important conflicting objective functions and a fair mean-risk optimization
of average-case and worst-case performance are considered in the presence of
supply chain disruption risks. The conflicting and equally important objective
functions are expected values of cost and customer service level and the corre-
sponding expected and expected worst-case values, respectively. The fairness and
the mean-risk fairness reflect the decision maker's common requirement to maintain
an equally good performance of a supply chain with respect to equally important
objectives and under varying operating conditions. Part III has two chapters:
• Chapter 7, A Fair Decision-Making under Disruption Risks. In this chapter, the
two risk-neutral conflicting criteria—expected cost and expected service level—
are fairly optimized to achieve an equitably efficient supply portfolio and
production schedule in the presence of supply chain disruption risks. In order to
obtain an equitably efficient solution, the ordered weighted averaging (OWA)
aggregation of the two conflicting objective functions is applied. The equitably
efficient solutions obtained for the ordered weighted averaging aggregation of the
two conflicting objective functions will be compared with non-dominated solu-
tions obtained using the weighted-sum aggregation approach.
• Chapter 8, A Robust Decision-Making under Disruption Risks. In this chapter,
we look for an equitably efficient solution with respect to both average-case and
worst-case performance measures of a supply chain. Such an equitably efficient
average-case and worst-case solution or equivalently equitably efficient
risk-neutral and risk-averse solution will be called a fair mean-risk solution. The
solution will equitably focus on the two objective functions: the expected value
(average-case performance measure) and the expected worst-case value
(worst-case performance measure), i.e., Conditional Value-at-Risk of the
selected optimality criterion, cost, or service level. The fair mean-risk
decision-making aims at equalizing the distance to optimality both under
business-as-usual and under worst-case conditions, which reflects a common
requirement to maintain an equally good performance of a supply chain under
varying operating conditions. Therefore, the mean-risk fairness, i.e., the equi-
tably efficient performance of a supply chain in the average case as well as in the
worst case, in this chapter will be called robustness.
Part IV focuses on selection of primary and recovery portfolios and scheduling.
The selection of primary suppliers and order quantity allocation to mitigate the
impact of disruption risks is combined with selection of recovery suppliers and
assembly plants to optimize the recovery processes. The two decision-making
approaches will be considered and compared: an integrated approach with some
information about the future potential disruption scenarios available ahead of time
Preface xi

and a hierarchical approach with no such information available. In the integrated

approach, which may account for all potential disruption scenarios, the primary
portfolios that will hedge against all potential disruptive events will be determined
along with the recovery portfolios for each scenario. In the hierarchical approach,
first the primary portfolios are determined, and then, when the primary portfolios
are impacted by a disruptive event, the recovery portfolios are selected to optimize
the process of recovery from the disruption. Both the integrated and the hierarchical
decision-making account for time and cost of mitigation and recovery processes and
aim at optimizing cost and service level as the two equally important, conflicting
objective functions. Part IV has two chapters:
• Chapter 9, Selection of Primary and Recovery Supply Portfolios and
Scheduling. In this chapter, the portfolio approach presented in the previous
chapters for the selection of primary suppliers and order quantity allocation to
mitigate the impact of disruption risks is enhanced also for the recovery process,
i.e., for the selection of both primary and recovery suppliers and order quantity
allocation to mitigate the impact of disruption risks and optimize the recovery
process. Unlike most of reported research on the supply chain risk management
which focuses on the risk mitigation decisions taken prior to a disruption, this
chapter combines decisions made before, during, and after the disruption.
• Chapter 10, Selection of Primary and Recovery Supply and Demand Portfolios
and Scheduling. In this chapter, the portfolio approach proposed in Chap. 9 for
the selection of primary and recovery suppliers and order quantity allocation to
mitigate the impact of disruption risks is enhanced also for the recovery process
of the firm’s assembly plants for finished products. Unlike most of reported
research on supply chain disruption management, a disruptive event is assumed
to impact both a primary supplier of parts and the buyer’s firm primary assembly
plant. Then, the firm may choose alternate (recovery) suppliers and move pro-
duction to alternate (recovery) plants along with transshipment of parts from the
impacted primary plant to the recovery plants. The resulting allocation of
unfulfilled demand for parts among recovery suppliers and unfulfilled demand
for products among recovery assembly plants determines recovery supply
portfolio and recovery demand portfolio, respectively.
Part V deals with disruption management of information flows in supply chains
caused by cybersecurity incidents. The supply portfolio approach applied to miti-
gate the impact of supply disruptions has been modified to select countermeasure
portfolio to mitigate the impact of information flow disruptions. Part V has one
chapter:
• Chapter 11, Selection of Cybersecurity Safeguards Portfolio. This chapter deals
with the selection of countermeasure portfolio in cybersecurity planning to
prevent or mitigate the impact of information flow disruptions on a supply chain.
A scenario-based bi-objective SMIP approach with CVaR as a risk measure is
proposed for the decision-making. Given a set of potential threats and a set of
available countermeasures, the decision maker needs to decide which
xii Preface

countermeasure to implement under limited budget to minimize potential losses

from successful cyberattacks. The selection of countermeasures is based on their
effectiveness of blocking different threats, implementation costs, and probability
of potential attack scenarios. The bi-objective trade-off model provides the
decision maker with a simple tool for balancing expected and worst-case losses
and for shaping of the resulting cost distribution through the selection of optimal
subset of countermeasures for implementation, i.e., the selection of optimal
countermeasure portfolio.
Parts I–IV and the chapters within each part are arranged in the order recom-
mended for reading, while Part V with Chap. 11 can be read independently of the
other chapters. Each chapter ends with the end-of-chapter problems to help the
reader a self-check of material comprehension and to encourage for a further
self-study.
This book can be considered a companion as well as a follower of my previous
book on scheduling in supply chains using mixed integer programming (Sawik
2011a), where deterministic MIP approaches were developed for integrated
scheduling in customer-driven supply chains, in particular, in the electronics supply
chains. The reader interested in knowing more about stochastic programming is
referred to the monographs by Birge and Louveaux (2011) or Kall and Mayer (2011).
For a general introduction to mixed integer programming models and techniques, the
reader is referred to the application-oriented book by Chen et al. (2010) or to the
seminal work in the field by Nemhauser and Wolsey (1999). The fundamentals of
supply chain theory are well presented by Snyder and Shen (2011), and for an
engineering-oriented general reference work on supply chains, the reader is referred
to the book by Dolgui and Proth (2010). Finally, some books cover supply chain risk
management in general, e.g., Kouvelis et al. (2011), and some of these emphasize
supply chain disruption management, e.g., Gurnani et al. (2012).

Audience

This book is addressed to practitioners and researchers on supply chain risk man-
agement and disruption management, and to students in management, industrial
engineering, operations research, applied mathematics, computer science and the
like at masters and Ph.D. levels. It is not necessary to have a detailed knowledge of
stochastic programming and integer programming in order to go through this book.
The knowledge required corresponds to the level of an introductory course in
operations research and supply chain management for engineering, management,
and economics students.

Kraków, Poland Tadeusz Sawik

April 2017
Acknowledgements

This book is based on the results of my research on supply chain scheduling and
disruption management by stochastic mixed integer programming over the last
decade. I wish to acknowledge many anonymous reviewers for their comments and
suggestions on my submissions to different international journals, in particular,
Omega: The International Journal of Management Science, International Journal
of Production Research, Computers & Operations Research and Decision Support
Systems.
The material presented in this book is in part based on the results of different
research projects on supply chain optimization and risk management, funded by
Polish Ministry of Science and Higher Education and by NCN. This project has
been partially supported by NCN research grant #DEC-2013/11/B/ST8/04458.
Thanks are also due to the AGH University of Science & Technology for its support
of research (grants #11.11.200.273 and #11.11.200.324) on supply chain risk
management over the last five years.
Finally, I wish to express my thanks to Springer Editor, Camille Price, who
invited me to prepare this monograph for International Series in Operations
Research & Management Science.

xiii
Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Overview of Supply Chain Disruption Management . . . . . . . . . 1
1.2 Value-at-Risk Versus Conditional Value-at-Risk . . . . . . . . . . . . 3
1.3 Local, Regional and Global Disruptions . . . . . . . . . . . . . . . . . . . 6
1.4 Two-level Versus Multi-level Disruptions . . . . . . . . . . . . . . . . . 8
1.5 Risk-Neutral, Risk-Averse and Mean-Risk Decision-Making . . . 9

Part I Selection of Supply Portfolio

2 Selection of Static Supply Portfolio . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3 Models for Risk-Neutral Decision-Making . . . . . . . . . . . . . . . . . 18
2.4 Models for Risk-Averse Decision-Making . . . . . . . . . . . . . . . . . 22
2.5 Models for Mean-Risk Decision-Making . . . . . . . . . . . . . . . . . . 24
2.6 Computational Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.6.1 Single-Region Sourcing . . . . . . . . . . . . . . . . . . . . . . . . 26
2.6.2 Multi-region Sourcing . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.7 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3 Selection of Dynamic Supply Portfolio . . . . . . . . . . . . . . . . . . . . . . . 43
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.2 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.3 Models for Risk-Neutral Decision-Making . . . . . . . . . . . . . . . . . 46
3.3.1 Dynamic Supply Portfolio for All-or-Nothing
Delivery Scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.4 Models for Risk-Averse Decision-Making . . . . . . . . . . . . . . . . . 51
3.5 Models for Mean-Risk Decision-Making . . . . . . . . . . . . . . . . . . 53
3.6 Computational Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.6.1 Worst-case Scenarios with Longest Delays
and Disruptions . . . . . . . . . . . . . . . . . . . . . . . . ....... 60

xv
xvi Contents

3.7 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4 Selection of Resilient Supply Portfolio. . . . . . . . . . . . . . . . . . . . . . . . 69
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.2 Problem Description: Single-Level Protection . . . . . . . . . . . . . . 70
4.3 Resilient Supply Portfolio with Single-Level Protection . . . . . . . 73
4.3.1 Model for Risk-Neutral Decision-Making . . . . . . . . . . . 74
4.3.2 Model for Risk-Averse Decision-Making . . . . . . . . . . . 77
4.3.3 Model for Mean-Risk Decision-Making . . . . . . . . . . . . 78
4.4 Protection Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.5 Resilient Supply Portfolio with Multi-Level Protection . . . . . . . 80
4.6 Computational Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.6.1 Single-Level Protection . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.6.2 Multi-level Protection . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.7 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99

Part II Integrated Selection of Supply Portfolio and Scheduling

5 Integrated Selection of Supply Portfolio and Scheduling
of Production . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.2 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.3 Models for Risk-Neutral Decision-Making . . . . . . . . . . . . . . . . . 108
5.3.1 Minimization of Cost . . . . . . . . . . . . . . . . . . . . . . . . . . 108
5.3.2 Maximization of Customer Service Level . . . . . . . . . . . 113
5.4 Models for Risk-Averse Decision-Making . . . . . . . . . . . . . . . . . 114
5.4.1 Minimization of Cost . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.4.2 Maximization of Customer Service Level . . . . . . . . . . . 116
5.5 Model for Mixed Mean-Risk Decision-Making . . . . . . . . . . . . . 118
5.6 Computational Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
5.6.1 Single Versus Dual Sourcing . . . . . . . . . . . . . . . . . . . . 120
5.6.2 Single Versus Multiple Sourcing . . . . . . . . . . . . . . . . . 129
5.6.3 Order Versus Demand Fulfillment Rate . . . . . . . . . . . . 137
5.6.4 Conditional Service-At Risk Versus Expected Cost . . . 141
5.7 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
6 Integrated Selection of Supply Portfolio and Scheduling
of Production and Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
6.2 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
6.3 Models for Risk-Neutral Decision-Making . . . . . . . . . . . . . . . . . 153
6.3.1 Scheduling with Single Batch Shipping . . . . . . . . . . . . 154
Contents xvii

6.3.2 Scheduling with Individual and Immediate

Shipping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ....... 158
6.3.3 Scheduling with Multiple Batch Shipping
and Limited Transportation Capacity . . . . . . . . . . . . . . 159
6.4 Models for Risk-Averse Decision-Making . . . . . . . . . . . . . . . . . 160
6.5 Expected Value Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
6.6 Computational Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
6.6.1 Risk-Neutral Decision-Making . . . . . . . . . . . . . . . . . . . 167
6.6.2 Risk-Averse Decision-Making . . . . . . . . . . . . . . . . . . . 178
6.6.3 Expected Value - Based Decision-Making . . . . . . . . . . 180
6.7 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184

Part III Equitably Efficient Selection of Supply Portfolio

and Scheduling
7 A Fair Decision-Making Under Disruption Risks . . . . . . . . . . . . . .. 189
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 189
7.2 The Lexicographic Minimax Optimization and Equitable
Aggregation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 190
7.3 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 193
7.4 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 195
7.4.1 Minimum and Maximum Values of the Objective
Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
7.4.2 Model Limitations and Possible Enhancements . . . . . . . 199
7.5 Computational Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
7.6 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
8 A Robust Decision-Making Under Disruption Risks . . . . . . . . . . . . 211
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
8.2 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
8.3 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
8.3.1 Equitable Minimization of Average
and Worst-Case Costs . . . . . . . . . . . . . . . . . . . . . . . . .. 215
8.3.2 Equitable Maximization of Average
and Worst-Case Service Level . . . . . . . . . . . . . . . . . .. 217
8.3.3 Minimum and Maximum Values of the Objective
Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
8.4 Computational Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
8.5 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
xviii Contents

Part IV Selection of Primary and Recovery Portfolios

and Scheduling
9 Selection of Primary and Recovery Supply Portfolios
and Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
9.2 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
9.3 Models for Risk-Neutral Decision-Making . . . . . . . . . . . . . . . . . 243
9.3.1 Integrated Selection of Primary and Recovery
Supply Portfolios . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 243
9.3.2 Multiple Part Types and Product Types . . . . . . . . . . .. 248
9.3.3 Hierarchical Selection of Primary and Recovery
Supply Portfolios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
9.4 Models for Risk-Averse Decision-Making . . . . . . . . . . . . . . . . . 252
9.5 Computational Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
9.5.1 Risk-Neutral Decision-Making . . . . . . . . . . . . . . . . . . . 256
9.5.2 Risk-Averse Decision-Making . . . . . . . . . . . . . . . . . . . 266
9.6 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
10 Selection of Primary and Recovery Supply and Demand
Portfolios and Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
10.2 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
10.3 Models for Risk-Neutral Decision-Making . . . . . . . . . . . . . . . . . 276
10.3.1 Integrated Selection of Primary and Recovery
Supply and Demand Portfolios . . . . . . . . . . . . . . . . . .. 276
10.3.2 Multiple Part Types and Product Types . . . . . . . . . . .. 281
10.3.3 Hierarchical Selection of Primary and Recovery
Supply and Demand Portfolios . . . . . . . . . . . . . . . . . . . 284
10.4 Models for Risk-Averse Decision-Making . . . . . . . . . . . . . . . . . 290
10.4.1 Integrated Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
10.4.2 Hierarchical Approach . . . . . . . . . . . . . . . . . . . . . . . . . 291
10.5 Computational Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292
10.5.1 Risk-Neutral Decision-Making . . . . . . . . . . . . . . . . . . . 294
10.5.2 Risk-Averse Decision-Making . . . . . . . . . . . . . . . . . . . 304
10.6 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311

Part V Information Flow Disruption Management

11 Selection of Cybersecurity Safequards Portfolio . . . . . . . . . . . . . . . . 315
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
11.2 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316
Contents xix

11.3 Models for Risk-Neutral Decision-Making . . . . . . . . . . . . . . . . . 318

11.3.1 Nonlinear Model for Risk-Neutral Decision-Making . . . 318
11.3.2 Linear Model for Risk-Neutral Decision-Making . . . . . 320
11.4 Model for Risk-Averse Decision-Making . . . . . . . . . . . . . . . . . . 323
11.5 Models for Mean-Risk Decision-Making . . . . . . . . . . . . . . . . . . 324
11.6 Computational Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
11.7 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
Acronyms

ABC_E Model for risk-neutral decision-making

ABC_CV Model for risk-averse decision-making
ABC_ECV Model for mean-risk decision-making
CVaR Conditional Value-at-Risk
CVaRc Conditional Cost-at-Risk
CVaRsl Conditional Service-at-Risk
Ec Expected cost
Esl Expected service level
MIP Mixed Integer Programming
SMIP Stochastic Mixed Integer Programming
VaR Value-at-Risk
VaRc Cost-at-Risk
VaRsl Service-at-Risk

xxi
List of Figures

Fig. 1.1 A supply chain network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 2

Fig. 1.2 Distribution of cost: VaRc versus CVaRc . . . . . . . . . . . . . . . . .. 6
Fig. 1.3 Distribution of service level: VaRsl versus CVaRsl . . . . . . . . . .. 6
Fig. 2.1 Risk-neutral supply portfolio for model SP_E(c):
10 suppliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 29
Fig. 2.2 Risk-averse supply portfolios and cost distributions
for model SP_CV(c): 10 suppliers . . . . . . . . . . . . . . . . . . . . . .. 30
Fig. 2.3 Risk-averse supply portfolios for different local disruption
probabilities for model SP_CV(c): 10 suppliers . . . . . . . . . . . .. 33
Fig. 2.4 Pareto front for mean-risk model SP_ECV(c): 10 suppliers,
a ¼ 0:9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 34
Fig. 2.5 Suppliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 35
Fig. 2.6 Risk-averse supply portfolios: a model SP_CV(c), b model
SP_CV(sl) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 37
Fig. 3.1 Suppliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 54
Fig. 3.2 Risk-neutral dynamic supply portfolio: a model DSP_E(c),
b model DSP_E(sl) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 57
Fig. 3.3 Expected risk-neutral dynamic supply portfolio: a model
DSP_E(c), b model DSP_E(sl) . . . . . . . . . . . . . . . . . . . . . . . . .. 58
Fig. 3.4 Risk-averse dynamic supply portfolio for model DSP_CV(c):
a ¼ 0:9 and a ¼ 0:99 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 61
Fig. 3.5 Risk-averse dynamic supply portfolio for model
DSP_CV(sl): a ¼ 0:9 and a ¼ 0:99 . . . . . . . . . . . . . . . . . . . . .. 62
Fig. 3.6 Risk-averse dynamic supply portfolios for model
DSP_CV(c) and LDN scenarios: a a ¼ 0:9, b a ¼ 0:99 . . . . . .. 64
Fig. 3.7 Probability mass function for model DSP_CV(c): a ¼ 0:9
and a ¼ 0:99. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 65
Fig. 3.8 Probability mass function for model DSP_CV(c)
and LDN scenarios: a ¼ 0:9 and a ¼ 0:99 . . . . . . . . . . . . . . . .. 65

xxiii
xxiv List of Figures

Fig. 4.1 Optimal supply portfolio for model RSP_E: a p ¼ 0:001,

pi 2 ½0; 0:06; i 2 I, b p ¼ 0:01, pi 2 ½0:06; 0:15; i 2 I
(P - protected supplier) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 87
Fig. 4.2 Optimal supply portfolio for model RSP_CV: a p ¼ 0:001,
pi 2 ½0; 0:06; i 2 I, b p ¼ 0:01, pi 2 ½0:06; 0:15; i 2 I . . . . . .. 89
Fig. 4.3 Pareto front for model RSP_ECV: a ¼ 0:99, p ¼ 0:001,
pi 2 ½0; 0:06; i 2 I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 91
Fig. 4.4 Optimal supply portfolio and probability mass function
for model RSP_ECV: p ¼ 0:01, pi 2 ½0:06; 0:15; i 2 I,
a ¼ 0:99, k ¼ 0:99 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 92
Fig. 4.5 Suppliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 94
Fig. 4.6 Optimal supply portfolio for model RSP(mlp)_CV:
a p ¼ 0:001, pi 2 ½0; 0:06; i 2 I, b p ¼ 0:01,
pi 2 ½0:06; 0:15; i 2 I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 96
Fig. 5.1 A three-echelon supply chain . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
Fig. 5.2 Suppliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
Fig. 5.3 Distribution of cost per product for risk-averse supply
portfolio with a ¼ 0:99 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
Fig. 5.4 Distribution of customer service level for risk-averse
supply portfolio with a ¼ 0:99 . . . . . . . . . . . . . . . . . . . . . . . . . . 126
Fig. 5.5 Expected production schedule for risk-averse supply
portfolio with a ¼ 0:99 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
Fig. 5.6 Suppliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
Fig. 5.7 Distribution of cost per product for risk-averse solution
with a ¼ 0:99 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
Fig. 5.8 Distribution of customer service level for risk-averse
solution with a ¼ 0:99 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
Fig. 5.9 Expected cumulative production for risk-averse solution
with a ¼ 0:9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
Fig. 5.10 Distribution of demand fulfillment rate: a model
SPSm_CV(sl) for demand fulfillment rate, b model
SPSm_CV(sl) for order fulfillment rate . . . . . . . . . . . . . . . . . . . . 140
Fig. 5.11 Expected and expected worst-case cumulative production
for a ¼ 0:9 and k ¼ 0:5: a model SPSm_E(c)CV(sl)
for demand fulfillment rate, b model SPSm_E(c)CV(sl)
for order fulfillment rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
Fig. 6.1 A multi-echelon supply chain . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
Fig. 6.2 Suppliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
Fig. 6.3 Expected solution values: a single batch shipping,
b individual shipping, c multiple batch shipping . . . . . . . . . . . . 170
Fig. 6.4 Nondominated supply portfolios: a single batch shipping,
b individual shipping, c multiple batch shipping . . . . . . . . . . . . 171
Fig. 6.5 Expected supplies, production and deliveries for maximum
service level (k ¼ 0): single batch shipping . . . . . . . . . . . . . . . . 172
List of Figures xxv

Fig. 6.6 Expected supplies, production and deliveries for k ¼ 0:5:

single batch shipping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
Fig. 6.7 Expected supplies, production and deliveries for minimum
cost (k ¼ 1): single batch shipping . . . . . . . . . . . . . . . . . . . . . . . 174
Fig. 6.8 Expected supplies, production and deliveries for k ¼ 0:5:
individual shipping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
Fig. 6.9 Expected supplies, production and deliveries for k ¼ 0:5:
multiple batch shipping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
Fig. 6.10 Supplies, production and deliveries for the expected value
problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
Fig. 7.1 Expected production schedules: a order fulfillment rate;
b demand fulfillment rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
Fig. 7.2 Expected production schedules for diversified customer
orders: a order fulfillment rate; b demand fulfillment rate . . . . . 206
Fig. 7.3 Nondominated supply portfolios: diversified customer
orders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
Fig. 7.4 Pareto front for model WSPS_E: diversified customer
orders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
Fig. 8.1 Distribution of cost per product
for a ¼ 0.75 and a ¼ 0.99 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
Fig. 8.2 Distribution of service level for a ¼ 0.75 and a ¼ 0.99 . . . . . . 227
Fig. 8.3 Expected production schedules for a ¼ 0:99 . . . . . . . . . . . . . . . . 228
Fig. 8.4 Expected worst-case production schedules for maximum
service level, a ¼ 0:99 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
Fig. 8.5 Expected worst-case production schedules for minimum
cost, a ¼ 0:99: a hj ¼ 2daj maxi2I ðoi Þe,
b hj ¼ 4daj maxi2I ðoi Þe . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
Fig. 9.1 Hierarchical selection of supply portfolios . . . . . . . . . . . . . . . . . 250
Fig. 9.2 Suppliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
Fig. 10.1 Hierarchical selection of supply and demand portfolios . . . . . . . 285
Fig. 11.1 Probability mass function for risk-neutral model CP_E
with B ¼ $150; 000: Expected loss ¼ $63,842 . . . . . . . . . . . . . . 328
Fig. 11.2 Probability mass function for risk-neutral model CP_EB:
Required budget ¼ $28,000, Expected cost ¼ $132,545 . . . . . . 329
Fig. 11.3 Probability mass function for risk-averse model CP_CV
with B ¼ $150; 000: a ¼ 0:99, CVaR ¼ $921,449,
VaR ¼ $414,500, Expected cost ¼ $92,045 . . . . . . . . . . . . . . . . 330
Fig. 11.4 Probability mass function for risk-averse model CP_CVB:
a ¼ 0:99, CVaR ¼ $652,214, VaR ¼ $59,000, Required
budget ¼ $258,000, Expected cost ¼ $31,332 . . . . . . . . . . . . . . 331
Fig. 11.5 Pareto front for mean-risk model CP_EBCV: a ¼ 0:9 . . . . . . . . 332
List of Tables

Table 2.1 Notation: static supply portfolio . . . . . . . . . . . . . . . . . . . . . .. 17

Table 2.2 Variables: static supply portfolio . . . . . . . . . . . . . . . . . . . . .. 19
Table 2.3 Risk-neutral solutions for model SP_E(c): single-region
sourcing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 27
Table 2.4 Risk-averse solutions: single-region sourcing . . . . . . . . . . . .. 28
Table 2.5 Probability of cost per part for optimal risk-averse
supply portfolios: 10 suppliers . . . . . . . . . . . . . . . . . . . . . . .. 31
Table 2.6 Solutions results for model SP_CV(c) with binary
assignment variables vij 2 f0; 1g: 10 suppliers . . . . . . . . . . . . 32
Table 2.7 Risk-averse solutions: multi-region sourcing . . . . . . . . . . . . . 36
Table 3.1 Notation: dynamic supply portfolio . . . . . . . . . . . . . . . . . . . . 44
Table 3.2 Variables: dynamic supply portfolio . . . . . . . . . . . . . . . . . . . . 47
Table 3.3 Risk-neutral solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
Table 3.4 Risk-averse solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
Table 3.5 Risk-neutral solutions for LDN scenarios . . . . . . . . . . . . . . . . 63
Table 3.6 Risk-averse solutions for LDN scenarios:
model DSP_CV(c) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 63
Table 4.1 Notation: single-level protection . . . . . . . . . . . . . . . . . . . . . .. 71
Table 4.2 Variables: resilient supply portfolio with single-level
protection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
Table 4.3 Notation: multi-level protection . . . . . . . . . . . . . . . . . . . . . . . 81
Table 4.4 Solution results for model RSP_E . . . . . . . . . . . . . . . . . . . . . 86
Table 4.5 Solutions results for model RSP_CV . . . . . . . . . . . . . . . . . . . 88
Table 4.6 Probability of cost per part for optimal supply portfolios:
p ¼ 0:001, pi 2 ½0; 0:06; i 2 I . . . . . . . . . . . . . . . . . . . . . . .. 90
Table 4.7 Nondominated solutions for mean-risk model RSP_ECV:
a ¼ 0:99, p ¼ 0:001, pi 2 ½0; 0:06; i 2 I . . . . . . . . . . . . . . . . 90
Table 4.8 Solution results for model RSP(mlp)_E . . . . . . . . . . . . . . . . . 94
Table 4.9 Solutions results for model RSP(mlp)_CV . . . . . . . . . . . . . . 95
Table 5.1 Notation: selection of supply portfolio and scheduling . . . . . . 106

xxvii
xxviii List of Tables

Table 5.2 Variables: selection of supply portfolio and scheduling . . . . . 109

Table 5.3 Risk-neutral solutions: single versus dual sourcing . . . . . . . . . 122
Table 5.4 Risk-averse minimization of cost: single versus dual
sourcing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
Table 5.5 Risk-averse maximization of service level: single versus
dual sourcing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
Table 5.6 Risk-averse supply portfolios for suppliers offering
similar prices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
Table 5.7 Risk-averse solutions for reduced penalty costs: single
versus dual sourcing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
Table 5.8 Risk-averse minimization of cost: single versus multiple
sourcing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
Table 5.9 Risk-averse maximization of service level: single versus
multiple sourcing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
Table 5.10 Single versus multiple sourcing solution . . . . . . . . . . . . . . . . 136
Table 5.11 Order versus demand fulfillment rate: risk-averse multiple
sourcing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
Table 5.12 Nondominated solutions for mean-risk model SPSm_E(c)
CV(sl): a ¼ 0:9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
Table 6.1 Notation: supply chain scheduling . . . . . . . . . . . . . . . . . . . . . 151
Table 6.2 Variables: supply chain scheduling . . . . . . . . . . . . . . . . . . . . . 154
Table 6.3 Notation: expected value problem . . . . . . . . . . . . . . . . . . . . . 164
Table 6.4 Risk-neutral solutions: single batch shipping . . . . . . . . . . . . . 167
Table 6.5 Risk-neutral solutions: individual shipping . . . . . . . . . . . . . . . 168
Table 6.6 Risk-neutral solutions: multiple batch shipping . . . . . . . . . . . 169
Table 6.7 Risk-averse solutions for model SCS1_CV . . . . . . . . . . . . . . 179
Table 6.8 Solution results for the expected value model ESCS1 . . . . . . 180
Table 7.1 Notation: selection of equitably efficient supply portfolio
and scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
Table 7.2 Variables: selection of equitably efficient supply
portfolio and scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
Table 7.3 Solution results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
Table 7.4 Solution results: diversified customer orders . . . . . . . . . . . . . . 205
Table 7.5 Nondominated solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
Table 7.6 Nondominated solutions: diversified customer orders . . . . . . . 207
Table 8.1 Notation: selection of robust supply portfolio
and scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
Table 8.2 Variables: selection of robust supply portfolio
and scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
Table 8.3 Risk-neutral solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
Table 8.4 Risk-averse solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
Table 8.5 Robust solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
List of Tables xxix

Table 8.6 Minimum and maximum values of objective functions . . . . . 225

Table 8.7 Solution results for a double unit penalty cost,
hj ¼ 4daj maxi2I ðoi Þe . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
Table 8.8 Mean-risk solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
Table 9.1 Notation: selection of primary and recovery supply
portfolios and scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
Table 9.2 Variables: selection of primary and recovery supply
portfolios and scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
Table 9.3 Multi-level disruption scenarios:
k1s ; k2s ; k3s ; k4s ; s ¼ 1; . . .; 256 . . . . . . . . . . . . . . . . . . . . . . . . 257
Table 9.4 Solution results for a common disruption start time:
C=D ¼ 0:47 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
Table 9.5 Solution results for a common disruption start time:
C=D ¼ 0:93 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
Table 9.6 Solution results for a common disruption start time:
C=D ¼ 1:4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
Table 9.7 Optimal recovery supply portfolios for a common
disruption start time, C=D ¼ 1:4, g ¼ 1; 10; 102 ; 103 ; 104 :
integrated approach/hierarchical approachy . . . . . . . . . . . . . . . 261
Table 9.8 Solution results for different disruption start times:
C=D ¼ 0:93 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
Table 9.9 Worst-case scenarios: C=D ¼ 0:93; g ¼ 100 . . . . . . . . . . . . . . 264
Table 9.10 Risk-averse solutions for a common disruption
start time: C=D ¼ 0:93 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
Table 9.11 Risk-averse solutions for different disruption start times:
C=D ¼ 0:93 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
Table 10.1 Notation: selection of primary and recovery supply
and demand portfolios and scheduling . . . . . . . . . . . . . . . . . . 273
Table 10.2 Variables: selection of primary and recovery supply
and demand portfolios and scheduling . . . . . . . . . . . . . . . . . . 276
Table 10.3 Notation and variables: multiple part types and product
types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282
Table 10.4 Supply and demand portfolios . . . . . . . . . . . . . . . . . . . . . . . . 296
Table 10.5 Supply and demand recovery portfolios,
s
ðV11 þ V12s
; V21
s
þ V22 s
; V31
s
þ V32 s
; V41s
þ V42 s
; zs1 ; zs2 Þ
and associated cost and service, for selected supply
and production disruption scenarios s 2 S: integrated
approach, g ¼ 100 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
Table 10.6 Supply and demand recovery portfolios,
s
ðV11 þ V12s
; V21
s
þ V22 s
; V31
s
þ V32 s
; V41s
þ V42 s
; zs1 ; zs2 Þ
and associated cost and service, for selected supply
and production disruption scenarios s 2 S: hierarchical
approach, g ¼ 100 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
xxx List of Tables

Table 10.7 Worst-case disruption scenarios: g ¼ 100, integrated

approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
Table 10.8 Selected worst-case disruption scenarios with respect
to cost and service level: g ¼ 100, hierarchical approach . . . . 302
Table 10.9 Risk-averse solutions for model DSupport_CV(c),
g ¼ 100 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
Table 10.10 Risk-averse solutions for model DSupport_CV(sl) . . . . . . . . 306
Table 10.11 Worst-case scenarios with respect to cost and service
level: model DSupport_CV(c), g ¼ 100 . . . . . . . . . . . . . . . . 307
Table 10.12 Risk-averse solutions for model RDSupport_CV(c),
g ¼ 100 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
Table 10.13 Risk-averse solutions for model RDSupport_CV(sl) . . . . . . . 309
Table 11.1 Notation: selection of cybersecurity safeguards portfolio . . . . 317
Table 11.2 Variables: selection of cybersecurity safeguards portfolio . . . . 319
Table 11.3 Solution results for risk-neutral models CP_E
and CP_EB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
Table 11.4 Solutions results for risk-averse model CP_CV . . . . . . . . . . . 327
Table 11.5 Solutions results for risk-averse model CP_CVB . . . . . . . . . . 327
Table 11.6 Nondominated solutions for mean-risk model CP_ECV:
a ¼ 0:9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329
Table 11.7 Examples of mixed integer program size and CPU time . . . . 332
Chapter 1
Introduction

1.1 Overview of Supply Chain Disruption Management

A typical global supply chain in modern industry is a network of multiple part

suppliers at different geographical locations and multiple production plants and dis-
tribution centers, where supplied parts are assembled into finished products and
next distributed to customers. Figure 1.1 shows a schematic diagram of a supply
chain network, where each vertical level (suppliers, producers, distribution centers,
customers) is called an echelon, and the arcs represent material flows. In order to
achieve a high customer service level at a low cost, a variety of complex, intercon-
nected decision-making problems should be solved. The decision-making problems
are strictly associated with the control and optimization of material flows (as well
as financial and information flows) in the network, in particular, optimization of dis-
rupted flows. Different types of material flows (e.g., flows of parts from suppliers to
producers, flows of semi-finished products at producers, flows of finished products
from producers to distribution centers and from the distribution centers to customers)
should be coordinated in an efficient manner. In global supply chains, the control
and optimization of material flows are accomplished by scheduling of manufactur-
ing and supplies of parts, scheduling of production and customer orders for finished
products, and scheduling of deliveries to customers. All those scheduling decisions
should be efficiently coordinated to fulfill customer demand, especially in the pres-
ence of supply chain disruption risks. The schedule of customer orders immediately
depends on the schedule of parts supplies, which in turn depends on supplier selec-
tion and order quantity allocation, that is, on supply portfolio. On the other hand, the
schedule of customer orders implicitly defines the schedule of deliveries of finished
products to customers, which in turn depends on customer order allocation among
assembly plants, that is, on demand portfolio. In view of the recent trend of out-
sourcing and globalization, coordinated decision-making, e.g., selection of primary
and recovery part suppliers and allocation of order quantities, selection of primary
and recovery assembly plants and allocation of customer demand, and scheduling of

© Springer International Publishing AG 2018 1

T. Sawik, Supply Chain Disruption Management Using Stochastic Mixed
Integer Programming, International Series in Operations Research
& Management Science 256, DOI 10.1007/978-3-319-58823-0_1
2 1 Introduction

customer orders in the assembly plants may significantly improve performance of a

multi-echelon supply chain under disruption risks. However, most work on coordi-
nated supply chain scheduling focuses on coordinating non disrupted flows of supply
and demand over a supply chain network to minimize the inventory, transportation
and shortage costs. The research on quantitative approaches to coordinated schedul-
ing of disrupted flows in supply chains has not been sufficiently reported in the
literature and is mostly limited to separate investigation of supply, production or
distribution stage.

Fig. 1.1 A supply chain network

In modern global supply chains, disruption risk management has become a vital
part of supply chain management strategy. Material flows in supply chains can be dis-
rupted by unexpected natural or man-made disasters such as earthquakes, fires, floods,
hurricanes or equipment breakdowns, labor strikes, economic crisis, bankruptcy or
by a deliberate sabotage or terrorist attack. The low-probability and high-impact flow
disruptions and the resulting losses may threaten financial state of firms. For example,
the Taiwan earthquake of September 1999 created huge losses for many electronics
companies supplied with components by Taiwanese manufacturers, e.g., Apple lost
many customer orders due to supply shortage of DRAM chips (Sheffi 2005). The
Philips microchip plant fire of March 2000 in New Mexico resulted in 400 million
Euros in lost sales by a major cell phone producer, Ericsson (Norrman and Jansson
2004). The disruptions in the automotive and electronics supply chains that occurred
in 2011 after the Great East Japan Tohoku earthquake on March 11 and then the Thai-
land flooding in October, resulted in huge losses of major automakers and electronics
manufacturers, e.g., Haraguchi and Lall (2015), Park et al. (2013). Similar effects
were observed after the recent Kyushu earthquake in April 2016, e.g., Marszewska
(2016). Toyota supply chain has been again severely disrupted when two plants of
Aisin Seiki, a key supplier of car body and engine components were destroyed, and
Renesas, a key supplier of semiconductors for Toyota engines, had to halt production
in Kumamoto plant.
1.1 Overview of Supply Chain Disruption Management 3

In order to minimize losses caused by the shortage of material supplies, customer

companies (firms) apply different disruption management strategies. For example,
the firms may participate in supplier’s recovery process after disruption to reduce
recovery time. When the Tohoku earthquake and tsunami disrupted Toyota sup-
ply chain and, in particular, supply chain of automotive semiconductors, Toyota
supported recovery of its suppliers, Fujimoto and Park (2013). The automotive
semiconductors were manufactured by Renesas Electronics, who shares over 44%
of world-wide automobile microcontroller units, and its main plant in Naka was
severely damaged by the earthquake. The shipping of automotive semiconductors
was expected to be stopped for eight months. In order to shorten the expected recovery
time, Toyota and other Japanese automotive, electronics and semiconductor equip-
ment manufacturers sent to Naka over 2500 engineers to support plant recovery. As
a result, the recovery time to start shipping of automobile microcontroller units was
shortened from eight to five months, Matsuo (2015). Another example of helping by
customer companies in supplier’s recovery process was the case of Riken Corpora-
tion, the largest supplier of piston rings to all Japanese automobile manufacturers.
In July 2007, Riken plant in city of Kashiwazaki was hit by a strong earthquake
and severely damaged. Immediately after the shutdown, the Japanese automakers
coordinated by Toyota, sent over 650 people including many equipment engineers
to help its recovery and as a result the stoppage of piston rings production was short-
ened to two weeks only, Whitney et al. (2014). A similar action has been applied
to help recovery of Aisin Seiki plants after the Kyushu earthquake in April 2016,
(Marszewska 2016).
The above real-world examples illustrate a well-known disruption management
strategy of helping a primary supplier recover more quickly. However, when a pri-
mary supplier is hit by disruption, the customer company may choose either to sup-
port recovery of disrupted primary supplier, rely on a preselected backup supplier or
select an alternate (recovery) supplier, non-disrupted or disrupted less severely than
the primary supplier. In a similar way, a firm whose primary plant is hit by disruption
may either stop production until recovery process is finished or move production to
alternate (recovery) plants along with transshipment of parts to the recovery plants.
The complex decision-making that involves various characteristics of a supply chain
should be supported by optimization models to minimize cost (or maximize profit)
and maximize service level as typical objective functions. The objective of this book
is to present such optimization models and to stimulate further research on funda-
mental understanding of various mitigation and recovery policies in the presence of
flow disruption risks in global supply chains.

1.2 Value-at-Risk Versus Conditional Value-at-Risk

A common tool for supply chain optimization under disruption risks is stochastic
programming, in particular stochastic mixed integer programming (SMIP), e.g.,
Heckmann et al. (2015). SMIP is an exact mathematical modeling approach that
4 1 Introduction

allows for the inclusion of uncertainty by probabilistic scenarios of disruptive events

and for finding the optimal solutions with respect to multiple objective functions. In
this book the stochastic combinatorial decision-making problems will be formulated
as multi-period, multi-objective stochastic mixed integer programs with expected or
expected worst-case performance measures and trade-offs between various objective
functions. In order to mitigate the impact of supply chain disruptions, the two pop-
ular in financial engineering percentile measures of risk, Value-at-Risk (VaR) and
Conditional Value-at-Risk (CVaR), will be applied for the decision-making, e.g.,
Sawik (2011, 2012a, b, c, 2016). This section briefly defines and compares VaR and
CVaR, e.g., Sarykalin et al. (2008).
Let FW (u) = Pr ob{W ≤ u} be the cumulative distribution function of a random
variable W representing cost. The VaR (Value-at-Risk) of W with the confidence
level α ∈ [0, 1) is defined as a lower α-percentile of the random variable W

V a Rα (W ) = min{u : FW (u) ≥ α}.

VaR represents the maximum cost associated with a specified confidence level of
outcomes (i.e., the likelihood that a given portfolio’s costs will not exceed the amount
defined as VaR). However, VaR does not account for properties of the cost distribution
beyond the confidence level and hence does not explain the magnitude of the cost
when the VaR limit is exceeded. Moreover, VaR is not a coherent measure of risk
since it fails to hold the sub-additivity property ( f (x + y) ≤ f (x) + f (y) where
f (.) is the risk measure). VaR of a portfolio can be higher than the sum of VaRs of
the individual assets in the portfolio.
On the other hand, CVaR focuses on the tail of the cost distribution, that is,
on outcomes with the highest cost. Assuming that FW (u) is a continuous distrib-
ution function, the CVaR of W with the confidence level α ∈ [0, 1), CVaRα (W ),
equals the expectation of W subject to W ≥ V a Rα (W ). However, in the general
case CVaRα (W ) is not equal to an average of outcomes greater than V a Rα (W ) and
is defined as the mean of the generalized α-tail distribution
 ∞
C V a Rα (W ) = ud FWα (u),
−∞

where

0 if u < V a Rα (W )
FWα (u) = FW (u)−α
1−α
if u ≥ V a Rα (W ).

CVaRα (W ) can be considered as a generalization of the expected value, when α = 0

they are equivalent. On the other hand, the higher is the confidence level α, the closer
are values of V a Rα (W ) and CVaRα (W ).
Alternatively, CVaRα (W ) can be defined as the weighted average of V a Rα (W )
and the conditional expectation of W subject to W > V a Rα (W ).
1.2 Value-at-Risk Versus Conditional Value-at-Risk 5

When the distribution function has a vertical jump at V a Rα (W ) (the probability

interval of confidence level α with the same VaR), a probability “atom” is said to be
present at V a Rα (W ). For example, when the distribution is modeled by scenarios,
the probability measure is concentrated in finitely many points and the corresponding
distribution function is a step function (constant between jumps) with jumps at those
points. Since CVaRα (W ) is obtained by averaging a fractional number of scenarios,
the V a Rα (W ) atom can be split. When FW (V a Rα (W )) > α, then probability 1 −
FW (V a Rα (W )) of the cost interval [V a Rα (W ), ∞) is smaller than 1 − α.
Note that if FW (V a Rα (W )) = 1, so that V a Rα (W ) is the highest cost that may
occur, then CVaRα (W ) = V a Rα (W ).
Summarizing the above definitions (from now on, VaR and CVaR will be denoted
without subscript α, and with superscripts c and sl to denote cost and service level,
respectively):
• Cost-at-Risk (VaRc ) at a 100α% confidence level is the targeted cost such that for
100α% of the scenarios, the outcome will not exceed VaRc . In other words, VaRc
is a decision variable based on the α-percentile of costs, i.e., in 100(1 − α)% of
the scenarios, the outcome may exceed VaRc .
• Conditional Cost-at-Risk (CVaRc ) at a 100α% confidence level is the expected cost
in the worst 100(1 − α)% of the cases. In other words, we allow 100(1 − α)% of
the outcomes to exceed VaRc , and the mean value of these outcomes is represented
by CVaRc .
In other words, VaRc is the acceptable cost level above which we want to minimize
the number of outcomes and CVaRc considers those portfolio outcomes, where costs
exceed VaRc (see, Fig. 1.2).
Generally, confidence level α indicates the level of conservatism that a decision-
maker is willing to adopt. As α approaches one, the range of acceptable worst-cases
becomes narrower in the corresponding optimization problem. Figure 1.2 clarifies
the concept of CVaR and demonstrates that CVaR is the conditional expected value
exceeding the VaR.
Similar definitions of VaR and CVaR for service level are given below.
• Service-at-Risk (VaRsl ) at a 100α% confidence level is the targeted service level
such that for 100α% of the scenarios, the outcome will not be below VaRsl . In
other words, VaRsl is a decision variable based on the α-percentile of service level,
i.e., in 100(1 − α)% of the scenarios, the outcome may be below VaRsl .
• Conditional Service-at-Risk (CVaRsl ) at a 100α% confidence level is the expected
service level in the worst 100(1 − α)% of the cases. In other words, we allow
100(1 − α)% of the outcomes to be below VaRsl , and the mean value of these
outcomes is represented by CVaRsl .
In other words, VaRsl is the acceptable service level below which we want to maxi-
mize the number of outcomes and CVaRsl considers those portfolio outcomes, where
service levels do not exceed VaRsl (see, Fig. 1.3).
6 1 Introduction

Since VaR and CVaR measure different parts of the cost (service level) distribution,
VaR may be better for optimizing portfolios when good models for tails are not
available, otherwise CVaR may be preferred.

Fig. 1.2 Distribution of cost: VaRc versus CVaRc

Fig. 1.3 Distribution of service level: VaRsl versus CVaRsl

1.3 Local, Regional and Global Disruptions

Assume that the supply chain consists of I interconnected facilities (nodes in this
network) that are located in R disjoint geographic regions. A facility in a supply
chain network may be a supplier of raw material or components, a manufacturer or an
assembly plant, a distribution center, or a retailer. Denote by I r ⊆ I = {1,
 . . . , I } the
subset of supply chain facilities in region r ∈ R = {1, . . . , R}, where r ∈R I r = I .
The supply chain facilities are subject to random independent local disruptions that
1.3 Local, Regional and Global Disruptions 7

are uniquely associated with a particular facility, which may arise from equipment
breakdowns, local labour strikes, fires, etc. Denote by pi the local disruption probabil-
ity for object i ∈ I , i.e., the state of facility i is “on” (non-disrupted) with probability
(1 − pi ), or is “off” (disrupted) with probability pi .
In addition to independent local disruptions of each facility individually, there
are also potential regional disruptive events that may result in correlated regional
disruption of all facilities in the same geographic region, and global disruptive super
events that may simultaneously impact all the facilities, i.e., the entire supply chain.
For example, such regional disruptive events may include, flood, earthquake, regional
strike in a transportation sector, whereas global disaster super events may include an
economic crisis, widespread labor strike in a transportation sector, etc.
Let pr and p ∗ be the probability of correlated regional disruption, simultaneously
of all facilities i ∈ I r in region r ∈ R, and correlated global disruption, simulta-
neously of all facilities i ∈ I , respectively. The global disruptive super event, the
regional disruptive events in each region and the local disruptive events are assumed
to be independent events. Thus, the disruption probability, πi , of every facility
i ∈ I r , r ∈ R is

πi = p ∗ + (1 − p ∗ ) pr + (1 − p ∗ )(1 − pr ) pi ; i ∈ I r , r ∈ R. (1.1)

Denote by Ps the probability that disruption scenario s is realized, where each

scenario s ∈ S is comprised of a unique subset Is ⊂ I of facilities that are in state
“on” (non-disrupted), and S = {1, . . . , S} is the index set of all disruption scenar-
ios. Each scenario s ∈ S can be represented by a 0–1 vector λs = {λ1s , . . . , λ I s },
where λis = 0 denotes disruption of facility i ∈ I under scenario s ∈ S, and λis = 1
denotes a normal, non-disruptive state of facility i ∈ I under scenario s ∈ S. Given
the number of facilities I , the total number
 of scenarios in which at least one facility
I I
is non-disrupted is given by i=1 = 2 I − 1. Including the scenario in which
i
all facilities are disrupted, there are a total of S = 2 I potential disruption scenarios.
For each scenario s ∈ S, the facilities i ∈ I \ Is , can be disrupted either by a local,
regional or global disaster event.
The probability Ps of each disruption scenario s ∈ S is derived as follows. First,
the probability Psr of realizing disruption scenario s for suppliers  in I r is deter-
mined. If there are non-disrupted suppliers in region r , i.e., I r Is = ∅, then Psr is
the product of regional non-disruption probability,
 (1 − pr ), local probabilities of
non-disrupted suppliers, (1 − pi ), i ∈ I r Is , and local probabilities of disrupted
suppliers,
 pi , i ∈ I r \ Is . Otherwise, i.e., if all suppliers in region r are disrupted,
I r Is = ∅, then either the entire region is disrupted with probability, pr , or the
region is non-disrupted with probability, (1 − pr ), and every supplier i ∈ I r is locally
disrupted with probability, pi . Thus, the probability Psr is
 
(1 − pr ) i∈I r  Is (1 − pi ) i∈I r \Is pi if I r Is = ∅
Psr =  (1.2)
pr + (1 − pr ) i∈I r pi if I r Is = ∅.
8 1 Introduction

The probability Ps for each disruption scenario s ∈ S with the subset Is of non-
disrupted facilities, and with all possible combinations of different disruptive events
considered, is

(1 − p ∗ ) r ∈R Psr if Is = ∅
Ps = ∗ ∗ (1.3)
p + (1 − p ) r ∈R Psr if Is = ∅,

If global and regional disruption probabilities are negligible, i.e., p ∗ = 0 and pr =

0, r ∈ R, the probability Ps of disruption scenario s in the presence of independent
local disruptive events only, reduces to

Ps = (1 − pi ) · pi . (1.4)
i∈Is i ∈I
/ s

1.4 Two-level Versus Multi-level Disruptions

In this section the scenarios with the two-level, all-or-nothing disruptive events con-
sidered in Sect. 1.3 are enhanced for the multi-level (partial) disruptive events. In
contrast to yield uncertainty (e.g., defective products) that occurs, for instance, when
the quantity of supply delivered is a random variable, typically modeled as either a
random additive or multiplicative quantity, the multi-level disruptions are modeled
as events of different level (e.g., partial capacity available) which occur randomly
and may have a random length, e.g., Schmitt and Singh (2012).
Assume that each facility i ∈ I is subject to random independent local disruptions
of different levels, l ∈ L i = {0, . . . , L i }, where the disruption level refers to the
available fraction of full capacity of a facility (e.g., a partial fulfillment of an order
by a supplier, a partial fulfillment of a customer order by a producer, etc.).
Level l = 0 represents complete shutdown of a facility, i.e., no capacity available,
(e.g., no order delivery) while level l = L i represents normal conditions with full
capacity available (e.g., full order delivery). The fraction of full capacity of facility
i available under disruption level l is described by γil
⎧
⎨= 0 if l = 0
γil ∈ (0, 1) if l = 1, . . . , L i − 1 (1.5)
⎩
=1 if l = L i .

Denote by S = {1, . . . , S} the index set of all disruption scenarios, and by Ps the
probability of disruption scenario s ∈ S. Each scenario s ∈ S can be represented by
an integer-valued vector λs = {λ1s , . . . , λ I s }, where λis ∈ L i is the disruption level
of facility i ∈ I under scenario s ∈ S. When all potential disruption scenarios are
considered, then S = i∈I (L i + 1).
Assume that for each scenario s ∈ S, each facility can be disrupted either by a
multi-level local disruptive event or by a two-level regional disruptive event. The
1.4 Two-level Versus Multi-level Disruptions 9

probability Ps for disruption scenario s ∈ S with the subset Is of non-shutdown

facilities is given by (1.3). Now, the probability, Psr , of realizing of disruption scenario
s for facilities in I r is (e.g., Sawik 2015b)
 
(1 − pr ) i∈I r
( pi,λis ) if I r  Is = ∅
Psr = (1.6)
pr + (1 − pr ) i∈I r pi0 if I r Is = ∅,

where pi,λis is the probability of occurrence the disruption at level l = λis at facility i.

1.5 Risk-Neutral, Risk-Averse and Mean-Risk

Decision-Making

In this book we consider three types of decision-making policies.

1. Risk-neutral, that is based on expected value optimization approach, e.g., expected
cost minimization or expected service level maximization approach. The risk-
neutral policy focuses on an average performance of a supply chain.
2. Risk-averse wherein, rather that optimizing the expected value of an objective
function, the decision-maker uses a Conditional Value-at-Risk (CVaR) approach
to measure and quantify risk and to define what comprises a worst-case scenario.
The CVaR methodology allows the decision-maker to evaluate worst-case values
of an objective function, to specify to what extent worst-case scenarios should
be avoided and to shape distribution of the resulting objective function values,
associated with such a policy. The risk-averse policy focuses on worst-case per-
formance of a supply chain.
3. Mean-risk wherein the decision-maker seeks for a best trade-off between expected
value and CVaR of an objective function. The mean-risk policy focuses on both
the average and the worst-case performance of a supply chain, simultaneously.
In this book we utilize stochastic mixed integer programming approach, that leads
to a two-stage optimization problem. The decisions that are made ahead of time are
considered the first stage decisions and are represented by the first stage decision
variables.
Typical first stage decision variables are
• binary selection variables, such as supplier selection variables, supplier protection
variables,
• fractional allocation variables, such as order quantity allocation variables, emer-
gency inventory pre-positioning variables.
In the risk-averse decision-making, VaR (cost-at-risk, service-at-risk, etc.) can also
be interpreted as first stage variables.
10 1 Introduction

The second stage decision variables represent decisions that are made after the
realization of the random events (e.g., supply disruptions) is known. The second
stage variables are dependent on the realized random event.
Typical second stage decision variables are
• binary selection variables, such as recovery supplier selection variables, recovery
plant selection variables,
• time-indexed binary assignment variables, such as production scheduling vari-
ables, distribution scheduling variables,
• fractional allocation variables, such as recovery order quantity allocation variables,
recovery demand allocation variables, emergency inventory usage variables, trans-
shipment variables,
• continuous variables, such as tail cost, tail service level.
Risk-Neutral Decision-Making
The stochastic formulation of the risk-neutral decision-making problem aimed at
loss (cost) minimization can be written as

min c T x + E[Q(x, ξ s )], (1.7)

x∈X

where c T x + E[Q(x, ξ s )] is the total cost function of the first-stage problem

and

Q(x, ξ s ) = min {(qs )T ys } (1.8)

y ∈Y
s s

is the optimal solution of the second-stage problem corresponding to the first

stage decision variables x and the realization of the random data ξ s for scenario
s ∈ S, denoted by ξ s = (qs , Y s ).
E[Q(x, ξ s )] is the expected “cost” taken with respect to random scenario
s ∈ S.

The objective function Q(x, ξ s ) of the second-stage problem (1.8), also known
as the recourse (cost) function, is a random variable.
Here x and ys are the vectors of first stage and second stage decision variables,
where the first stage decisions are deterministic and the second-stage decisions are
dependent on random scenario s. X denotes the feasible set of first stage decisions
and Y s is the feasible set of second stage decisions for random scenario s ∈ S. The
second-stage problem (1.8) may be infeasible for some first stage decisions x ∈ X .
To deal with the uncertainty in the second stage, a scenario-based modeling
approach is proposed that has been widely used in stochastic programming. In the
second stage, let us consider random scenario s ∈ S to have a discrete distribution,
where Ps is the probability of occurrence for scenario s ∈ S, and S is a finite set of sce-
narios. Given a finite set of scenarios, S, with associated probabilities, Ps , s ∈ S, the
1.5 Risk-Neutral, Risk-Averse and Mean-Risk Decision-Making 11


expected value E[Q(x, ξ s )] can be evaluated as E[Q(x, ξ s )] = s∈S Ps Q(x, ξ s ).
Hence, we can present the deterministic equivalent of the stochastic formulation
(1.7).


min cT x + Ps (qs )T ys (1.9)
s∈S
s.t. x ∈ X, ys ∈ Y s ; s ∈ S.

Model (1.9) is also known as the wait-and-see model (e.g., Birge and Louveaux
2011; Kall and Mayer 2011). In contrast to the two-stage approach with the recourse
model (1.7), in the wait-and-see approach both, the decisions on the first stage vari-
ables x and the second stage variables ys , are taken simultaneously only when the
outcome of ξ s = (qs , Y s ) is known.
Risk-Averse Decision-Making
In the model proposed below, CVaR is represented by an auxiliary function (1.10)
introduced by Rockafellar and Uryasev (2000), for a set of pre-defined scenarios
s ∈ S with corresponding probabilities, Ps . Using the wait-and-see approach, the
stochastic formulation of the risk-averse decision-making problem aimed at CVaR
of loss (cost) minimization, given confidence level α, can be written as

Minimize

C V a R = V a R + (1 − α)−1 Ps τs (1.10)
s∈S

subject to

τs ≥ c T x + (qs )T ys − V a R; s ∈ S (1.11)
x∈X (1.12)
y ∈Y ; s∈S
s s
(1.13)
τs ≥ 0; s ∈ S. (1.14)

In the above formulation, constraints (1.11) compute the tail cost, τs , for scenario
s and condition (1.14) indicates that the scenarios in which the loss exceeds VaR are
considered only.
Mean-Risk Decision-Making
In the mean-risk formulation for the wait-and-see approach, λ is a non-negative
trade-off coefficient representing the decision-maker risk preference. For a given
12 1 Introduction

confidence level α one can construct the mean-risk efficient frontier by the parame-
terization on λ the weighted-sum program presented below.

Minimize
 
λ(c T x + Ps (qs )T ys ) + (1 − λ)(V a R + (1 − α)−1 Ps τs ), (1.15)
s∈S s∈S

where 0 ≤ λ ≤ 1
subject to

τs ≥ c T x + (qs )T ys − V a R; s ∈ S (1.16)
x∈X (1.17)
ys ∈ Y s ; s ∈ S (1.18)
τs ≥ 0; s ∈ S. (1.19)

The resulting decision vector x is efficient in the mean-risk sense, i.e., it has the
lowest possible CVaR for a given expected cost, and for a given CVaR it has the
lowest possible expected cost.
 Part I
Selection of Supply Portfolio
Chapter 2
Selection of Static Supply Portfolio

2.1 Introduction

In a customer-driven supply chain manufacturers should be prepared to produce dif-

ferent products to meet different customer needs. Each product is typically composed
of many common and non-common (custom) parts that can be sourced from different
suppliers with different supply capacity. An important issue is how to best allocate
the orders for parts among various part suppliers to fulfill all customer orders for
products and to achieve a high customer service level at a low cost and, in addition,
to mitigate the impact of supply chain disruption risks. Supply chain management,
in particular, deals with selection of supply portfolio, i.e., selection of suppliers and
order quantity allocation under uncertain quality of supplied materials and reliabil-
ity of on-time delivery. The decision maker needs to decide from which supplier to
purchase parts required to meet customer demand. The decision is based on price,
quality and reliability criteria that may conflict each other. For example, the sup-
plier offering the lowest price may not have the best quality or the supplier with the
best quality may not deliver on time. In stochastic supply settings, supplier selec-
tion allows the producer do decide whether it should cooperate with a low cost, yet
risky suppliers over more expensive but possibly more reliable suppliers. A common
risk-neutral objective of minimizing expected cost or maximizing expected service
level is therefore influenced by uncertainty and risk. As a result, new non-risk-neutral
objectives of minimizing and maximizing the number of outcomes that could occur
above an acceptable cost level and below an acceptable service level, respectively,
are observed in practice. Furthermore, to reduce the fixed ordering costs of creating
contracts and maintaining relationships with suppliers, the number of suppliers and
the total number of orders should be minimized. On the other hand, however, the
selection of more suppliers may divert the risk of unreliable supplies. In global supply
chains a multi-regional suppliers base is a frequent solution, where suppliers from
different geographic regions are selected. Then, in addition to independent local dis-
ruptions (i.e., equipment breakdown, fires, etc.) that are uniquely associated with a

© Springer International Publishing AG 2018 15

T. Sawik, Supply Chain Disruption Management Using Stochastic Mixed
Integer Programming, International Series in Operations Research
& Management Science 256, DOI 10.1007/978-3-319-58823-0_2
16 2 Selection of Static Supply Portfolio

particular supplier, the supplies of parts are also subject to regional disruptions (e.g.,
floods, hurricanes, earthquakes, economic crisis, etc.) simultaneously of all suppliers
in the same region.
This chapter deals with selection of a static supply portfolio under disruption
risks, i.e., for determining a single-period supply portfolio. The proposed portfolio
approach allows the two popular in financial engineering percentile measures of
risk, Value-at-Risk (VaR) and Conditional Value-at-Risk (CVaR) to be applied for
managing the risk of supply disruptions. For a finite number of scenarios, CVaR
allows the evaluation of worst-case costs (or worst-case service level) and shaping
of the resulting cost (service level) distribution through optimal supplier selection
and order allocation decisions, i.e., the selection of optimal supply portfolio. Since
common parts can be efficiently managed by material requirement planning methods,
the focus in this chapter is on supplies of custom parts that can be critical in a make-
to-order manufacturing.
The following SMIP models are presented in this chapter:

SP_E(c) for risk-neutral selection of supply portfolio to minimize expected

cost;
SP_E(sl) for risk-neutral selection of supply portfolio to maximize expected
service level;
SP_CV(c) for risk-averse selection of supply portfolio to minimize CVaR
of cost;
SP_CV(sl) for risk-averse selection of supply portfolio to maximize CVaR
of service level;
SP_ECV(c) for mean-risk selection of supply portfolio to optimize trade-off
between expected cost and CVaR of cost;
SP_ECV(sl) for mean-risk selection of supply portfolio to optimize trade-
off between expected service level and CVaR of service level.

In the computational experiments described in Sect. 2.6, single-region and multi-

region sourcing subject to local and regional disruption risks are illustrated with
numerical examples.
In the next chapter the portfolio approach will be to enhanced for a multi-period
supplier selection and order allocation in the presence of supply chain disruption
and delay risks, where in the scenario analysis the low probability and high impact
supply disruptions are combined with the high probability and low impact supply
delays. Unlike for a single-period problem considered in this chapter, in a multi-
period setting the decision maker needs to decide from which supplier and when to
purchase custom parts required for each customer order to meet customer requested
due dates at a low cost and a high customer service level and to mitigate the impact
of disruption and delay risks.
2.2 Problem Description 17

2.2 Problem Description

In the supply chain under consideration various types of products are assembled
by a single producer to satisfy customer orders, using custom parts purchased from
multiple suppliers (for notation used, see Table 2.1). Each supplier can provide the
producer with custom parts for all customer orders. However, the suppliers have
different limited capacity and, in addition, differ in price and quality of offered parts
and in reliability of delivery of parts. Let I = {1, . . . , I} be the set of I suppliers and
J = {1, . . . , J} the set of J customer orders for the products, known ahead of time.

Table 2.1 Notation: static supply portfolio

Indices
i = supplier, i ∈ I
j = customer order, j ∈ J
r = geographic region, r ∈ R
s = disruption scenario, s ∈ S
Input Parameters
ci = capacity of supplier i
dj = demand for parts required for customer order j

D = j∈J dj - total demand for parts
ei = cost of ordering parts from supplier i
hj = per unit penalty cost of unfulfilled customer order j
oij = unit price of parts for customer order j purchased from supplier i
pi = local disruption probability for supplier i
pr = regional disruption probability for all suppliers in region r
α = confidence level
ρi = expected defect rate of supplier i

Each order j ∈ J is described by the quantity dj of required custom parts. Denote

by ci the capacity of supplier i ∈ I, by ei the cost of ordering parts from supplier i ∈ I,
and by oij the unit purchasing price of parts for customer order j ∈ J from supplier
i ∈ I.
The ordered parts are dispatched to the producer after the completion time of their
manufacturing. For each supplier, however, the quality of delivered part may vary
randomly. When the suppliers are selected the risk of defective parts can be considered
using past observations. Since quality of parts vary among different suppliers, a
different average defect rate can be associated with each supply portfolio. Let ρi be
the expected defect rate of supplier i.
The suppliers are assumed to be located in R disjoint geographic regions.  Denote
by I r ⊆ I the subset of suppliers in region r ∈ R = {1, . . . , R}, where r∈R I r = I.
The supplies of parts are subject to random local disruptions that are uniquely asso-
ciated with a particular supplier, which may arise from equipment breakdowns, local
labor strike, fires, etc. Denote by pi the local disruption probability for supplier i, i.e.,
18 2 Selection of Static Supply Portfolio

the parts ordered from supplier i are delivered without disruptions with probability
(1 − pi ), or not at all with probability pi . In addition to independent local disruptions
of each supplier individually, the supplies of parts are also subject to regional corre-
lated disruption of all suppliers in the same region simultaneously, with probability
pr for region r ∈ R.
Denote by πi the disruption probability of every supplier i ∈ I r , r ∈ R

πi = pr + (1 − pr )pi ; i ∈ I r , r ∈ R. (2.1)

Let S = {1, . . . , S} be the index set of S = 2I disruption scenarios, where each

scenario s ∈ S defines a subset Is ⊂ I of non-disrupted suppliers. The supplies from
every supplier, i ∈ I \ Is , can be independently disrupted either by a local or by a
regional disaster event. The probability Ps of each disruption scenario s ∈ S is a
product over all regions r ∈ R of probabilities Psr of realizing disruption scenario s
for suppliers in I r , 
Ps = Psr , (2.2)
r∈R

where Psr is (cf. Sect. 1.3)

   
(1 − pr ) i∈I r  Is (1 − pi ) i∈I r \Is pi if I r Is = ∅
Psr =   (2.3)
pr + (1 − pr ) i∈I r pi if I r Is = ∅.

The producer does not need to pay for ordered and defective or undelivered parts.
However, the producer may be charged with a much higher cost of unfulfilled cus-
tomer orders for products, caused by the shortage of parts, undelivered due to supply
disruptions. Let hj be the per unit penalty cost of unfulfilled customer order j.
The decision maker needs to decide from which suppliers to purchase custom
parts required for each customer order to achieve a minimum cost of ordering, pur-
chasing and shortages and to mitigate the impact of disruption risks by minimizing
the potential worst-case cost or maximizing the potential worst-case service level.

2.3 Models for Risk-Neutral Decision-Making

In this section two SMIP models SP_E(c) and SP_E(sl) are proposed for risk-neutral
selection of a static supply portfolio, i.e., for determining a single-period supply port-
folio to minimize expected cost and maximize expected service level, respectively.
The static supply portfolio is defined below, (for definition of problem variables, see
Table 2.2).
(V1 , . . . , VI ),

where 
Vi = 1
i∈I
2.3 Models for Risk-Neutral Decision-Making 19

and 0 ≤ Vi ≤ 1 is the fraction of the total demand for parts ordered from supplier i,
and Vi is determined by the custom parts allocation variables vij

Table 2.2 Variables: static supply portfolio

First stage variables
ui = 1, if an order for parts is placed on supplier i; otherwise ui = 0 (supplier selection)
vij = the fraction of demand for parts required for customer order j ordered from supplier i
(allocation of demand for custom parts)
Auxiliary variables
Vi = the fraction of total demand for parts allocated to supplier i (supply portfolio: allocation
of total demand for parts)
VaRc Cost-at-Risk, the targeted cost such that for a given confidence level α, for 100α% of
the scenarios, the outcome is below VaRc
VaR sl Service-at-Risk, the targeted service level such that for a given confidence level α, for
100α% of the scenarios, the outcome is above VaRsl
Cs ≥ 0, the tail cost for scenario s, i.e., the amount by which costs in scenario s exceed VaRc
Ss ≥ 0, the tail service level for scenario s, i.e., the amount by which VaRsl exceeds service
level in scenario s


Vi = dj vij /D; i ∈ I. (2.4)
j∈J

When deciding on a static supply portfolio it is assumed that the orders for all
parts are simultaneously placed on selected suppliers (e.g., at time 0), and each
supplier delivers all the ordered parts at the earliest possible delivery date. Therefore,
the allocation of orders for parts among the suppliers is not combined with the
allocation of orders among the planning periods. Nevertheless, the static portfolio
should be checked against the risk of supply disruptions across all potential disruption
scenarios.
Notice that Table 2.2 does not explicitly define the second stage variables for the
SMIP problem considered. The seconds stage variables are simply demand alloca-
tion variables for realized disruption scenarios s, ṽijs ; i ∈ I, j ∈ J, s ∈ S, defined as
follows

v if i ∈ Is , j ∈ J, s ∈ S
ṽij = ij
s
0 if i ∈ / Is , j ∈ J, s ∈ S.

In view of the above definition, an explicit introduction of the second stage variables
ṽijs into the SMIP model formulations is not required.
In a risk-neutral operating conditions the overall quality of the supply portfolio
can be measured by the expected cost per part, E c , (2.5), or expected service level
E sl , (2.6).
20 2 Selection of Static Supply Portfolio
 
Ec = ei ui /D + oij dj vij /D
i∈I i∈I j∈J

+ Ps (hj − oij )dj vij /D (2.5)
s∈S i∈I
/ s j∈J

E sl = Ps dj vij /D (2.6)
s∈S i∈Is j∈J

The expected cost E c includes, cost of ordering,


i∈I ei ui /D,
cost
 ofpurchasing non defective parts,
i∈I j∈J oij dj vij /D,
and cost of shortage of parts due to supply disruptions (cost of unfulfilled customer
orders
  less cost
 of non delivered parts),
s∈S i∈I
/s j∈J Ps (hj − oij )dj vij /D.
The purchase orders for parts are assumed to be inflated by the reject rates ρi of
defective parts, i.e., are equal to (1 + ρi )dj vij for all i ∈ I, j ∈ J. However, since
the producer does not need to pay for ordered and defective parts in the amount of
ρi dj vij , the
corresponding purchasing
  cost  per part for delivered parts is simply given
by i∈I j∈J oij dj vij /D − s∈S i∈I /s j∈J Ps oij dj vij /D.
The expected cost per part, E c , (2.5), can also be written as follows
 
Ec = ei ui /D + Ps oij dj vij /D
i∈I s∈S i∈Is j∈J

+ Ps hj dj vij /D, (2.7)
s∈S i∈I
/ s j∈J

  
where s∈S i∈Is j∈J Ps oij dj vij /D is the expected purchasing cost per part for
delivered parts.
The expected service level, E sl , (2.6), is a surrogate measure of expected customer
demand fulfillment rate and represents the expected fraction of fulfilled demand for
required parts.
The SMIP models SP_E(c) and SP_E(sl) are formulated below. The supply port-
folio will be optimized by minimizing expected cost per part, E c , (2.5) or by maxi-
mizing expected service level, E sl , (2.6).

SP_E(c): Selection of risk-neutral Supply Portfolio to minimize expected cost

Minimize (2.5)
subject to
1. Supply portfolio selection constraints:
- the total demand for parts required for each customer order must be fully
allocated among suppliers,
2.3 Models for Risk-Neutral Decision-Making 21

- for each selected supplier the total quantity of ordered parts cannot exceed
the supplier capacity,
- parts cannot be ordered from non-selected suppliers,
- at least one customer order should be assigned to each selected supplier,

vij = 1; j ∈ J (2.8)
i∈I

(1 + ρi )dj vij ≤ ci ui ; i ∈ I (2.9)
j∈J
vij ≤ ui ; i ∈ I, j ∈ J (2.10)

vij ≥ ui ; i ∈ I (2.11)
j∈J

2. Non-negativity and integrality conditions

ui ∈ {0, 1}; i ∈ I (2.12)

vij ∈ [0, 1]; i ∈ I, j ∈ J. (2.13)

Notice that if hj = h∀j ∈ J, i.e., per unit penalty cost of unfulfilled customer order
is identical for all orders j, then E c , (2.7), can be expressed by the following simplified
formula
 
Ec = ei ui /D + Ps oij dj vij /D + h(1 − E sl ), (2.14)
i∈I s∈S i∈Is j∈J

where E sl is the expected service level (2.6).

SP_E(sl): Selection of risk-neutral supply portfolio to maximize expected

service level
Maximize (2.6)
subject to (2.8)–(2.13).

If total available
 capacity of
all suppliers is less than total demand for required
parts, i.e., i∈I ci /(1 + ρi ) ≤ j∈J dj , then the demand allocation equality con-
straints (2.8) should be replaced by inequalities

vij ≤ 1; j ∈ J; (2.15)
i∈I

otherwise no feasible solution exists.

A simple upper bound on the expected service level, E sl , (2.6), is derived below.
22 2 Selection of Static Supply Portfolio

Proposition 2.1

E sl ≤ min{1, (1 − pr )(1 − pi )ci /(1 + ρi )D}. (2.16)
r∈R i∈I r

Proof The supply portfolio selection constraints (2.9) imply that


Ps dj vij /D ≤
s∈S i∈Is j∈J

(1 − pr )(1 − pi )dj vij /D ≤
r∈R i∈I r j∈J

(1 − pr )(1 − pi )ci ui /(1 + ρi )D ≤
r∈R i∈I r

(1 − pr )(1 − pi )ci /(1 + ρi )D,
r∈R i∈I r

where (1 − pr )(1 − pi ) = 1 − πi , (2.1), is non-disruption probability of supplier

i ∈ Ir.
 
Since E sl cannot be greater than 1, its upper bound is 1, if r∈R i∈I r (1 − pr )(1 −
pi )ci /(1 + ρi )D > 1.

In the proposed models the parts required for each customer order are assumed
to be partially provided by one or more suppliers and the customer order allocation
variable vij represents the fraction of all parts required for order j provided by supplier
i. In some practical cases, all custom parts of the same type that are required for
a customer order are purchased from a single supplier. Then, the corresponding
continuous allocation variable vij should be redefined as a binary assignment variable
denoting whether or not all parts required for order j are provided by supplier i. If
all vij are defined to be binary variables, then SP_E(c) and SP_E(sl) become pure
stochastic binary programs.

2.4 Models for Risk-Averse Decision-Making

In the risk-averse selection of supply portfolio under disruption risks, the confidence
level α is fixed by the decision maker to control the risk of losses due to supply
disruptions. We assume that the decision maker is willing to accept only portfolios
for which the total probability of scenarios with costs greater than VaRc or with
service level lower than VaRsl is not greater than 1 − α. Furthermore, a risk aversive
decision maker wants to minimize the expected worst-case costs exceeding VaRc or
to maximize the expected worst-case service level below VaRsl .
2.4 Models for Risk-Averse Decision-Making 23

Define by Cs the tail cost for scenario s, where tail cost is defined as the amount by
which costs in scenario s exceed VaRc . In a similar way, define by Ss the tail service
level for scenario s, where tail service level is defined as the nonnegative amount by
which VaRsl exceeds service level in scenario s.
The portfolio will be optimized by calculating VaRc and minimizing CVaRc simul-
taneously or by calculating VaRsl and maximizing CVaRsl , respectively. By measur-
ing CVaRc or CVaRsl , the magnitude of the tail costs or the tail service level is
considered to achieve a more accurate estimate of the risks of minimizing cost or
maximizing service level, respectively. When using CVaRc to minimize worst-case
costs and CVaRsl to maximize worst-case service level, CVaRc is always not less
than VaRc and CVaRsl is always not greater than VaRsl , respectively.
In the proposed model CVaR is represented by an auxiliary function (2.17) and
(2.20) introduced by Rockafellar and Uryasev (2000). The SMIP models SP_CV(c)
and SP_CV(sl) for selection of risk-averse supply portfolio to reduce the risk of high
costs and the risk of low service level, respectively, is formulated below.

SP_CV(c): Selection of risk-averse supply portfolio to minimize CVaR of cost

Minimize

CV aRc = V aRc + (1 − α)−1 Ps Cs (2.17)
s∈S

subject to
1. Supply portfolio selection constraints: (2.8)–(2.11)
2. Risk constraints:
- the tail cost for scenario s is defined as the nonnegative amount by which
cost per part in scenario s exceeds VaRc ,
 
Cs ≥ ei ui /D + oij dj vij /D
i∈I i∈I j∈J

+ (hj − oij )dj vij /D − V aRc ; s ∈ S (2.18)
i∈I
/ s j∈J

3. Non-negativity and integrality conditions: (2.12), (2.13) and

Cs ≥ 0; s ∈ S. (2.19)
24 2 Selection of Static Supply Portfolio

SP_CV(sl): Selection of risk-averse supply portfolio to maximize CVaR of

service level
Maximize 
CV aRsl = V aRsl − (1 − α)−1 Ps Ss (2.20)
s∈S

subject to
1. Supply portfolio selection constraints: (2.8)–(2.11)
2. Risk constraints:
- the tail service level for scenario s is defined as the nonnegative amount
by which VaRsl exceeds service level in scenario s,

Ss ≥ V aRsl − dj vij /D; s ∈ S (2.21)
i∈Is j∈J

3. Non-negativity and integrality conditions: (2.12), (2.13) and

Ss ≥ 0; s ∈ S. (2.22)

Note that as Cs and Ss are constrained of being positive, the model SP_CV(c)
tries to decrease VaRc and the model SP_CV(sl) tries to increase VaRsl , respectively.
Hence they positively impact the objective functions. However, large reduction in
VaRc and large increase in VaRsl may result in more scenarios with positive tail costs
and with positive tail service levels, respectively.
If for some customer order j all required parts must be supplied by a single supplier,
then the corresponding nonnegative allocation variable vij should be redefined as a
binary assignment variable denoting whether or not all parts required for order j are
provided by supplier i, similarly as for the risk-neutral models SP_E(c) and SP_E(sl).

2.5 Models for Mean-Risk Decision-Making

In the single objective approach the supply portfolio is selected by minimizing either
the expected cost per part, E c , (2.5), the expected service level, E sl , (2.6), the expected
worst-case cost per part, CV aRc , (2.17) or the expected worst-case service level,
CV aRsl , (2.20). In this section the two cost functions and the two service level func-
tions are considered simultaneously, and a bi-objective selection of supply portfolio
is presented aimed at minimizing both objective functions to balance expected costs
or expected service level with the risk tolerance. This trade-off model is known as
the mean-risk model (e.g., Ogryczak and Ruszczynski 2002), formulated as the opti-
mization of a composite objective consisting of the expected cost (service level) and
the CVaR as a risk measure.
2.5 Models for Mean-Risk Decision-Making 25

The nondominated solution set of the bi-objective supply portfolio can be found by
the parameterization on λ the weighted-sum programs SP_ECV(c) and SP_ECV(sl)
presented below. The mean-risk program SP_ECV(c) is based on model SP_CV(c)
with the addition of objective (2.5) of model SP_E(c). Similarly, the mean-risk
program SP_ECV(sl) is based on model SP_CV(sl) with the addition of objective
(2.6) of model SP_E(sl).

SP_ECV(c): Selection of mean-risk supply portfolio to minimize weighted

sum of expected cost and CVaR of cost
Minimize
λE c + (1 − λ)CV aRc (2.23)

where 0 ≤ λ ≤ 1,
subject to (2.5), (2.8)–(2.13), (2.17)–(2.19).

SP_ECV(sl): Selection of mean-risk supply portfolio to maximize weighted

sum of expected service level and CVaR of service level
Maximize
λE sl + (1 − λ)CV aRsl (2.24)

where 0 ≤ λ ≤ 1,
subject to (2.6), (2.8)–(2.13), (2.20)–(2.22).

Steuer (1996) proved that for mixed integer programs, there may be portions of the
nondominated set (nearby weakly nondominated solution) that the above approach
is unable to compute, even if the complete parameterization on λ is attempted.

2.6 Computational Examples

In this section some computational examples are presented to illustrate possible

applications of the proposed SMIP approach for selection of static supply portfolio
under disruption risks. First, a single-region sourcing case will be illustrated, where
all suppliers are located in a single geographic region, and then examples of multi-
region sourcing with subsets of suppliers in different geographic regions, each subject
to different regional disruption risks. For the single-region sourcing, minimization
of cost is considered only, whereas for the multi-region sourcing, both minimization
of cost and maximization of service level are considered.
26 2 Selection of Static Supply Portfolio

2.6.1 Single-Region Sourcing

In this subsection all suppliers are assumed to be located in the same geographic
region, and hence the regional disruption can be called a global disruption. Denote
by, p∗ , the global disruption probability for the entire region. Now, the probability, Ps
(2.3), of each disruption scenario s, can be calculated using the following formula

(1 − p∗ )P̂s  if Is = ∅
Ps = (2.25)
p∗ + (1 − p∗ ) i∈I pi if Is = ∅,

where P̂s is the probability of disruption scenario s in the presence of independent

local disruptive events only
 
P̂s = (1 − pi ) · pi . (2.26)
i∈Is i∈I
/s

If the probability of regional disruption p∗ = 0, then the probability Ps reduces to P̂s

for independent local disruptive events.
The following parameters have been used for the example problems:
• I, the number of suppliers, was equal to 7, 10 or 14 and the corresponding number
S = 2I of disruption scenarios, was equal to 128, 1024 or 16384, respectively;
• J, the number of customer orders, was equal to 50;
• dj , the numbers of required parts for each customer order, were integers uniformly
distributed over [100, 500], i.e., generated from a U[100;500]
 distribution;
• ci , the capacity of each supplier i, was equal to 2 j∈J dj /I (· denotes the
smallest integer not less than ·), i.e., the total capacity of all suppliers was equal
to the double total demand for parts;
• ei , the cost of ordering parts from supplier i, was equal to 500 for each supplier i;
• hj , the per unit shortage cost for customer order j, was equal to 100 for all customer
orders j;
• ρi , the expected defect rate of each supplier i, was exponentially distributed, rang-
ing from 0.0003 to 0.03;
• oij , the unit price of parts required for each customer order j purchased from each
supplier i, was uniformly distributed over [10,15] (i.e., drawn from U[10;15]) and
reduced by the factor (1 − ρi ) to get a lower price for parts from the suppliers with
a higher defect rate;
• pi , the local disruption probability was uniformly distributed over [0,0.06], i.e.,
the disruption probabilities were drawn independently from U[0;0.06];
• p∗ , the global disruption probability was equal to 0.01;
• α, the confidence level, was equal to 0.50, 0.75, 0.90, 0.95 or 0.99.

For the example problems, the total demand for parts is D = j∈J dj = 14750
parts. Solution results for the risk-neutral model SP_E(c) are shown in Table 2.3, and
for the risk-averse model SP_CV(c) with different confidence levels, in Table 2.4. The
2.6 Computational Examples 27

size of the mixed integer programs for different number I of suppliers is represented
by the total number of variables, Var., number of binary variables, Bin., number
of constraints, Cons, and number of nonzero coefficients in the constraint matrix,
Nonz. Table 2.4 also presents the probability 1 − F(V aRc ) of outcomes with worst-
case cost above VaRc . Note that the number of variables and constraints in the mixed
integer program SP_CV grows exponentially in the number I of suppliers. The table
demonstrate that the number of selected suppliers increases with the confidence level
α, which indicates that the impact of disruption risks is mitigated by diversification
of the supply portfolio. Note that VaRc becomes smaller than expected cost when
α = 0.50 and α = 0.75.

Table 2.3 Risk-neutral solutions for model SP_E(c): single-region sourcing

No. of Suppliers Expected Cost No. of Selected Suppliers
7 12.38 4
Var. = 364, Bin. = 7, Cons. = 72, Nonz. = 1428 (a)
10 13.18 5
Var. = 520, Bin. = 10, Cons. = 81, Nonz. = 2040 (a)
14 12.03 7
Var. = 728, Bin. = 14, Cons. = 93, Nonz. = 2856 (a)
(a)Var. = no. of variables, Bin. = no. of binary variables, Cons. = no. of constraints, Nonz. = no.
of nonzero coefficients

The optimal risk-neutral supply portfolio for model SP_E(c) and 10 suppliers is
shown in Fig. 2.1. In addition, the  figure presents for each supplier i the expected
defect rate ρi , the average unit price j∈J oij /J, and disruption probabilities, πi , (2.1).
For the optimal supply portfolio the total demand was equally allocated among five
suppliers with the lowest disruption probabilities.
In the computational experiments the confidence level α is set at five levels of
0.5, 0.75, 0.90, 0.95, and 0.99, which means that focus is on minimizing the highest
50%, 25%, 10%, 5%, and 1% of all scenario outcomes, i.e., costs per part.
Figure 2.2 shows the probability mass functions and the cumulative distribution
functions for the optimal risk-averse portfolios with different confidence levels for
10 suppliers. Figure 2.2 indicates that the mass function of cost per part is concen-
trated in a few points and the resulting cumulative distribution is a discontinues step
function with jumps at those points. Such results are typical for the scenario-based
optimization under uncertainty, where the probability measure is concentrated in
finitely many points. The resulting discontinuity (vertical jumps) of the distribution
function leads to probability intervals of confidence level α with the same VaR. The
discrete distributions of cost per part for the optimal supply portfolios with four
different confidence levels and the corresponding probabilities concentrated at each
level of cost are presented also in Table 2.5. The table shows that the probabilities
are concentrated at 6, 10, 10, 10 points, respectively for the confidence level α =
0.5, 0.9, 0.989, 0.99. In the examples, a large probability atom is concentrated at
the highest cost. As a consequence, a slight increase of the confidence level from
28 2 Selection of Static Supply Portfolio

Table 2.4 Risk-averse solutions: single-region sourcing

Confidence level α 0.50 0.75 0.90 0.95 0.99
Model SP_CV(c): 7 suppliers
Var. = 493, Bin. = 7, Cons. = 200, Nonz. = 47380 (a)
CVaRc 14.15 17.69 28.21 39.01 100.14
VaRc 10.61 10.61 13.90 21.19 100.14
Ec 12.38 12.38 12.59 12.94 14.24
1 − F(V aRc ) 0.053 0.053 0.039 0.018 0
No. of suppliers selected 4 4 7 7 4
Model SP_CV(c): 10 suppliers
Var. = 1545, Bin. = 10, Cons. = 1115, Nonz. = 526338 (a)
CVaR c 16.02 21.69 31.67 42.05 100.17
VaRc 10.35 10.35 20.29 23.19 100.17
Ec 13.18 13.18 14.06 13.96 15.89
1 − F(V aRc ) 0.113 0.113 0.041 0.032 0
No. of suppliers selected 5 5 9 10 5
Model SP_CV(c): 14 suppliers
Var. = 17113, Bin. = 14, Cons. = 16477, Nonz. = 11733800 (a)
CVaRc 13.60 16.70 25.59 35.94 100.24
VaRc 10.51 10.51 13.75 16.89 100.24
E c 12.05 12.05 12.50 12.43 15.50
1 − F(V aRc ) 0.115 0.097 0.058 0.029 0
No. of suppliers selected 7 7 13 11 7
(a)Var. = no. of variables, Bin. = no. of binary variables, Cons. = no. of constraints, Nonz. = no.
of nonzero coefficients

α = 0.989 to α = 0.99 results in a significant change in VaR from 37.33 to 100.17,

while only a slight increase of CVaRc from 94.81 to 100.17 is observed. Moreover,
the optimal portfolio has been changed significantly; for α = 0.989 the total demand
has been equally allocated among all ten suppliers (Vi = 0.1 forall i), whereas for
α = 0.99 among five suppliers only (Vi = 0.2 for i = 1, 2, 6, 7, 8). This degree of
instability of the optimal supply portfolio due to the discontinuity in the distribution
function may be distressing in practice, when a slightly higher confidence level is
required. Despite the limited change in CVaRc , the above results demonstrate that
the well known misbehaviour in the dependence of VaRc and optimal supply port-
folio on the confidence level can as well be encountered when CVaRc is applied as
a risk measure.
The computational results indicate that the smaller is the number of concentration
points and the greater are probability atoms concentrated at those points, the greater
can be the positive difference F(V aRc ) − α, i.e., the smaller than 1 − α can be the
probability of outcomes with cost higher than VaRc . For example (see, Table 2.4),
for I = 10 and α = 0.5, V aRc = 10.35 and F(V aRc ) = 0.88666 > 0.5, which indi-
cates a high concentration of probability measure at point 10.35 for the optimal
2.6 Computational Examples 29

supply portfolio. Actually, the probability that cost per part is 10.35 is 0.88666 (see,
Table 2.5), which indicates that V aRc = 10.35 is the lowest cost that may occur and

Fig. 2.1 Risk-neutral supply portfolio for model SP_E(c): 10 suppliers

30 2 Selection of Static Supply Portfolio

Fig. 2.2 Risk-averse supply portfolios and cost distributions for model SP_CV(c): 10 suppliers
2.6 Computational Examples 31

Table 2.5 Probability of cost per part for optimal risk-averse supply portfolios: 10 suppliers
Cost interval α = 0.5 α = 0.9 α = 0.989 α = 0.99
[10, 11) 0.886661618 0.736768317 0.693881575 0
[11, 12) 0 0 0 0.821176973
[19, 20) 0 0 0.251829646 0
[20, 21) 0 0.221857039 0 0
[28, 29) 0.098887038 0 0.040311352 0
[29, 30) 0 0 0 0.156904596
[30, 31) 0 0.029090556 0 0
[37, 38) 0 0 0.003744839 0
[40, 41) 0 0.002178297 0 0
[46, 47) 0.004355651 0 0.000223405 0
[47, 48) 0 0 0 0.011507829
[50, 51) 0 0.000102579 0 0
[55, 56) 0 0 8.94e-06 0
[60, 61) 0 3.15e-06 0 0
[64, 65) 9.47e-05 0 2.43e-07 0.0004038
[70, 71) 0 6.30e-08 0 0
[73, 74) 0 0 4.41e-09 0
[80, 81) 0 7.91e-10 0 0
[82, 83) 1.01e-06 0 5.14e-11 6.76e-06
[90, 91) 0 5.66e-12 0 0
[91, 92) 0 0 3.46e-13 0
[100, 101) 0.010000004 0.01 0.01 0.010000043

that for the confidence level α = 0.5, less than 11.33% of the cost outcomes are
above VaRc .
Moreover, if the highest cost probability is greater than 1 − α, then CVaRc and
VaRc are identical and both equal to the highest cost. In the example for ten suppliers
and α = 0.99, the highest cost per part is 100.17 and the probability concentrated
at 100.17 is 0.01000004 > 1 − α, then V aRc = 100.17 is the highest cost per part
that may occur and hence CV aRc = V aRc = 100.17 (see, Table 2.5 and Fig. 2.2).
Similar results CV aRc = V aRc = 100.14 and CV aRc = V aRc = 100.24 have been
obtained for α = 0.99, respectively for seven and 14 suppliers (see, Table 2.4), which
indicates that the corresponding probabilities of the highest cost per part are greater
than 1 − α = 0.01.
If the probability measure is concentrated at the highest cost and is greater than
1 − α, so that CVaRc and VaRc are identical with the highest cost, then for a higher
confidence level α, a smaller number of suppliers are selected, which indicates that
diversification of the supply portfolio is not necessary any more. For instance, the
optimal risk-averse supply portfolio selected for α = 0.99 consists of five suppliers
only, the same number as that for a much lower α (cf. Table 2.4, Fig. 2.2).
32 2 Selection of Static Supply Portfolio

In the computational experiments the local disruption probabilities pi were

assumed to be very low and were drawn from U[0;0.06]. To study the effect of
the increasing range of disruption probabilities, the probabilities have been drawn
also from U[0;0.25], U[0;0.5] or U[0;1]. The effect of varying distribution of local
disruption probabilities is illustrated in Fig. 2.3, where the optimal risk-averse supply
portfolios are presented. Figure 2.3 indicates that for a greater range of disruption
probabilities, the suppliers with the highest disruption probabilities are not selected.
For example, for U[0;1] and α = 0.75, suppliers i = 4, 6, 7, 10 with the four highest
disruption rates were not selected, whereas for α = 0.985, suppliers i = 7, 10 with
the two highest disruption rates were not selected. Similar results are observed for
the other distributions of disruption probability.

Table 2.6 Solutions results for model SP_CV(c) with binary assignment variables vij ∈ {0, 1}: 10
suppliers
Confidence level α 0.50 0.75 0.90 0.95 0.99
Var. = 1545, Bin. = 510, Cons. = 1105, Nonz. = 526328 (a)
CVaRc 16.08 21.80 31.75 42.08 100.20
VaRc 10.36 10.40 20.16 23.21 100.20
Ec 13.22 13.25 14.07 13.94 15.74
1 − F(V aRc ) 0.155 0.155 0.094 0.053 0
No. of suppliers selected 6 6 9 10 6
(a)Var. = no. of variables, Bin. = no. of binary variables, Cons. = no. of constraints, Nonz. = no.
of nonzero coefficients

Finally, Table 2.6 presents solution results for model SP_CV(c) applied to opti-
mization of a single sourcing, with binary assignment variables vij , i.e., when for each
customer order, the required parts must be provided by a single supplier only. Com-
parison of the results shown in Table 2.6 with the corresponding results presented in
Table 2.4 for continuous allocation variables vij , indicates that in the former case both
the expected cost per part and CVaRc were slightly higher and, in addition, for a low
α the number of selected suppliers was greater. Such results were expected, since
model SP_CV(c) with the continuous allocation variables is a partial LP relaxation
of that model with the binary assignment variables.
For the mean-risk approach, the subsets of nondominated solutions were computed
by parameterization on λ ∈ {0.01, 0.10, 0.25, 0.50, 0.75, 0.90, 0.99} the weighted-
sum program SP_ECV(c). The subset of nondominated solutions found for the
selected seven levels of trade-off parameter λ is: (E c , CV aRc ) = (13.18, 36.32),
(13.33, 33.97), (13.49, 32.65), (13.84, 31.84), (13.86, 31.82), (14.06, 31.67). The
trade-off between the expected cost and the expected worst-case cost is clearly shown
in Fig. 2.4, where the convex efficient front for the mean-risk model SP_ECV(c) with
α = 0.9 is presented. The results emphasize the effect of varying cost/risk preference
of the decision maker.
2.6 Computational Examples 33

Fig. 2.3 Risk-averse supply portfolios for different local disruption probabilities for model
SP_CV(c): 10 suppliers
34 2 Selection of Static Supply Portfolio

Fig. 2.4 Pareto front for mean-risk model SP_ECV(c): 10 suppliers, α = 0.9

Note that solutions to single objective models SP_E(c) and SP_CV(c) are equiv-
alent to the nondominated solutions of the weighted-sum program SP_ECV(c) for
λ = 1 and λ = 0, respectively.
The computational experiments were performed using the AMPL programming
language and the CPLEX solver. The solver was capable of finding proven optimal
solutions within CPU seconds for all examples.

2.6.2 Multi-region Sourcing

In this subsection the suppliers are assumed to be located in multiple geographic

regions subject to different regional disruption risks. The following parameters used
for the example problems are different from those in Sect. 2.6.1:
• I, the number of suppliers, was equal to 10 and the corresponding number S = 2I
of disruption scenarios, was equal to 1024;
• R, the number of geographic regions, was equal to 3, and the subsets of suppliers
were I 1 = {1, 2, 3}, I 2 = {4, 5, 6} and I 3 = {7, 8, 9, 10}, respectively;
• J, the number of customer orders, was equal to 25;
• dj , the numbers of required parts for each customer order, were integers uniformly
distributed over [1000, 15000] for all customer orders j. and the resulting total
demand for parts was D = 132500;
• ei , the cost of ordering parts, were integers in {5000, 6000, . . . , 10000}, {10000,
11000, . . . , 15000} and {15000, 16000, . . . , 30000}, respectively for suppliers i ∈
I 1 , i ∈ I 2 and i ∈ I 3 ;
• hj , the per unit shortage cost for each customer order j, was integer uniformly
distributed over [maxi∈I (oij ), 4 maxi∈I (oij )], i.e., ranging from one to four times of
maximum purchasing cost of required parts;
2.6 Computational Examples 35

• oij , the unit price of parts for customer order j purchased from supplier i, was
uniformly distributed over [13,15], [11,13] and [9,11], respectively for suppliers
i ∈ I 1 , i ∈ I 2 and i ∈ I 3 ;
• pi , the local disruption probability was uniformly distributed over [0.005,0.01],
[0.01,0.05] and [0.05;0.10], respectively for suppliers i ∈ I 1 , i ∈ I 2 and i ∈ I 3 ,
i.e., the disruption probabilities were drawn independently from U[0.005;0.01],
U[0.01,0.05] and U[0.05;0.10], respectively;
• pr , the regional disruption probability was 0.001, 0.005 and 0.01, respectively for
region r = 1, r = 2 and r = 3;

Fig. 2.5 Suppliers


Basic characteristics of suppliers: average unit price j∈J oij /J, and disruption
probability, πi , (2.1), are presented in Fig. 2.5. Figure 2.5 indicates that the most
reliable and most expensive are suppliers i = 1, 2, 3 in region r = 1, while suppliers
i = 7, 8, 9, 10 in region r = 3 are most competitive and most unreliable. In particular,
supplier i = 8 is the cheapest and most unreliable among all suppliers.
The solution results for the risk-averse models SP_CV(c) and SP_CV(sl) with
different confidence levels are shown in Table 2.7. For both models the number of
selected suppliers increases with the confidence level. While for model SP_CV(c) the
cheapest, yet most unreliable supplier i = 8 was never selected, for model SP_CV(sl)
all 10 suppliers are selected for α = 0.9, 0.95 and α = 0.99.
Figure 2.6 shows the optimal risk-averse supply portfolios for models SP_CV(c)
and SP_CV(sl) and the three confidence levels, α = 0.75, 0.9, 0.99. For both mod-
els and α = 0.75 the most unreliable suppliers i = 7, 8, 9, 10 in region r = 3 are
not selected and for all confidence levels most demand for parts is allocated among
the three most reliable, yet most expensive suppliers, i = 1, 2, 3, in region r = 1, in
particular for α = 0.75, 0.9. Similar properties of the risk-averse supply portfolios
were observed in case of single-region sourcing (see, Fig. 2.3).
36 2 Selection of Static Supply Portfolio

Table 2.7 Risk-averse solutions: multi-region sourcing

Confidence level α 0.50 0.75 0.90 0.95 0.99
Model SP_CV(c)
Var. = 1285, Bin. = 10, Cons. = 1319, Nonz. = 269558 (a)
CVaRc 15.35 17.15 20.49 23.90 30.43
VaRc 13.29 13.68 16.44 17.19 26.21
1 − F(V aRc ) 0.149 0.125 0.055 0.036 0.005
Ec 14.32 14.55 14.59 14.51 14.61
E (b)
sl 97.51 97.98 97.30 97.56 96.15
Suppliers Selected(% of total demand) 1(20) 1(20) 1(20) 1(20) 1(19)
2(19) 2(19) 2(20) 2(19) 2(13)
3(20) 3(20) 3(20) 3(20) 3(12)
4(14) 4(8) 4(8) 4(6)
5(8) 5(7) 5(6) 5(9) 5(6)
6(13) 6(20) 6(6) 6(8) 6(8)
7(20) 7(12) 7(16) 7(9)
9(8) 9(12)
10(15)
Model SP_CV(sl)
Var. = 1285, Bin. = 10, Cons. = 1319, Nonz. = 131318 (a)
CVaRsl % 96.01 92.02 85.29 80.12 69.82
VaRsl % 100 100 91.87 87.50 77.77
1 − F(V aRsl ) 0.125 0.125 0.064 0.033 0.007
E sl (b) 98.00 98.00 97.29 97.43 97.04
E c 15.56 15.48 15.86 15.85 15.61
Suppliers Selected(% of total demand) 1(20) 1(20) 1(20) 1(20) 1(20)
2(19) 2(19) 2(20) 2(20) 2(18)
3(20) 3(20) 3(20) 3(20) 3(18)
4(20) 4(20) 4(8) 4(8) 4(10)
5(1) 5(1) 5(8) 5(9) 5(4)
6(20) 6(20) 6(8) 6(9) 6(9)
7(4) 7(3) 7(9)
8(4) 8(3) 8(4)
9(4) 9(3) 9(4)
10(4) 10(3) 10(4)
(a) Var. = no. of variables, Bin. = no. of binary variables, Cons. = no. of constraints, Nonz. = no.
of nonzero
  coefficients

(b) (
s∈S i∈Is j∈J Ps dj vij /D)100%

The computational experiments indicate that

• probability of disruption a supply is a key determinant in the decision of allocation

of demand among the suppliers. In a risk averse model, an order for delivery of
2.6 Computational Examples 37

parts from a particular supplier is selected more on the supply non-disruption

likelihood than on its purchasing cost or defect rate.
• The suppliers associated with the highest disruption rates are rarely selected.

Fig. 2.6 Risk-averse supply portfolios: a model SP_CV(c), b model SP_CV(sl)

• In most cases the number of selected suppliers increases with the confidence level
α, which indicates that the impact of disruption risks is mitigated by diversification
of the supply portfolio.
• The greater is the range of disruption probabilities, the higher are both expected
and worst-case costs.
• The closer are disruption rates for different suppliers, the closer are the corre-
sponding quantities of ordered parts in the optimal portfolio.

The probability measure for the scenario-based optimization under uncertainty is

concentrated in finitely many points and the resulting discrete distribution of cost may
38 2 Selection of Static Supply Portfolio

have different effect on the optimal portfolio. In particular, the well known misbe-
haviour in the dependence of VaR on the confidence level can as well be encountered
when CVaR is applied as a risk measure. For instance, if a large probability atom
is concentrated at some cost, a slight increase of the confidence level may results in
a significant change in VaR as well as in the optimal portfolio, while only a slight
change of CVaR may occur. Such an instability of the optimal portfolio due to the dis-
continuity in the distribution function may be distressing in practice, when a slightly
higher confidence level is required.
On the other hand the computational results indicate that the smaller is the number
of concentration points and the greater are probability atoms concentrated at those
points, the greater can be the positive difference F(V aR) − α, i.e., the smaller than
1 − α can be the probability of outcomes with cost higher (service level lower) than
VaR.
The computational experiments prove that the proposed exact solution approach
based on MIP approach provides the decision maker with a simple tool for evaluating
the relationship between expected and worst-case costs. For a finite number of sce-
narios, the proposed models allow the evaluation of worst-case costs and shaping of
the resulting cost distribution through the selection of optimal supply portfolio. The
optimal risk averse supply portfolio can be found within CPU seconds for a limited
number of scenarios considered, using commercially available solvers for MIP.

2.7 Notes

The supply chain risk management has been extensively studied over the past decade.
Research addresses the two risk levels (e.g., Tang 2006): operational risks or disrup-
tion risks. Operational risks are referred to the inherent uncertainties arising from the
problems of coordinating supply and demand such as uncertain customer demand,
uncertain supply, and uncertain cost. Disruption risks are referred to the major dis-
ruptions to normal activities caused by natural and man-made disasters such as earth-
quakes, floods, hurricanes, etc., or equipment breakdowns, economic crises such as
currency evaluation, labor strikes, terrorist attacks. In most cases, the business impact
associated with disruption risks is much greater than that of the operational risks.
In practice four basic approaches can be applied to mitigate the impact of supply
chain risks (Tang 2006): supply management, demand management, product man-
agement, and information management. In particular, to ensure efficient supply of
materials along a supply chain, supply chain management deals with selection of a
supply portfolio, i.e., supplier selection and order quantity allocation under uncertain
quality of supplied materials and reliability of on-time delivery. The supplier selec-
tion and order quantity allocation problem is a complex stochastic combinatorial
optimization problem, however the research on supplier selection under disruption
risks is limited. For example, chance-constrained programming models were devel-
oped by Kasilingam and Lee (1996) to account for stochastic demand and by Wu
and Olson (2008) to consider expected losses from quality acceptance inspection or
2.7 Notes 39

late delivery. Parlar and Perry (1996) present a continuous time model in which the
availability of each of the m suppliers is uncertain because of disruptions such as
equipment breakdown. By considering the case that each supplier is either “on” or
“off”, there are 2m possible number of states for the whole system. For each of these
2m states, they analyze a state-specific (q, Q) ordering policy so that the buyer would
order Q units when the on-hand inventory reaches q. The risks associated with a
supplier network was studied by Berger et al. (2004), who considered catastrophic
super events that affect all suppliers, as well as unique events that impact only one
single supplier, and then a decision-tree based model was presented to help determine
the optimal number of suppliers needed for the buying firm. Ruiz-Torres and Mah-
moodi (2007) considered unequal failure probabilities for all the suppliers. Berger
and Zeng (2006) studied the optimal supply size in a single or multiple sourcing strat-
egy context, under a number of scenarios that are determined by various financial
loss functions, the operating cost functions and the probabilities of all the suppliers
being down. Yu et al. (2009) considered the impacts of supply disruption risks on
the choice between the single and dual sourcing methods in a two-echelon supply
chain with a non-stationary and price-sensitive demand. Yue et al. (2010) introduced
frontier sourcing portfolios to support manufacturers sourcing decisions, which con-
sider the cost and probability of finishing the order on time. Ravindran et al. (2012)
developed multi-criteria supplier selection models incorporating supplier risks. In
the multi-objective formulation, price, lead-time, disruption risk due to natural event
and quality risk are explicitly considered as four conflicting objectives that have to be
minimized simultaneously. Four different variants of goal programming were used
to solve the multi-objective optimization problem. Xanthopoulos et al. (2012) devel-
oped newsvendor-type inventory models for capturing the trade-off between inven-
tory policies and disruption risks in a dual-sourcing supply chain network, where
both supply channels are subject to disruption risks. The models were developed for
both risk-neutral and risk-averse decision-making. Li and Zabinsky (2011) developed
a two-stage stochastic programming model and a chance-constrained programming
model to determine a minimal set of suppliers and optimal order quantities. Both
models include several objectives and strive to balance a small number of suppliers
with the risk of not being able to meet demand. The stochastic programming model
is scenario-based and uses penalty coefficients whereas the chance-constrained pro-
gramming model assumes a probability distribution and constrains the probability of
not meeting demand. Hammami et al. (2014) proposed a scenario-based stochastic
model for supplier selection in the presence of uncertain fluctuations of currency
exchange rates and price discounts.
The vast majority of the decision models are mathematical programming models
either single objective, e.g., Kasilingam and Lee (1996), Basnet and Leung (2005),
Sawik (2005) or multiple objectives, e.g., Weber and Current (1993), Xia and Wu
(2007), Demirtas and Ustun (2008), Ustun and Demirtas (2008). The models devel-
oped for supplier selection and order allocation can be either single-period models
(e.g., Weber and Current 1993, Demirtas and Ustun 2008) that do not consider inven-
tory management or multi-period models (e.g., Ghodsypour 2001, Basnet and Leung
40 2 Selection of Static Supply Portfolio

2005, Ustun and Demirtas 2008, Che and Wang 2008) which consider the inventory
management by lot-sizing and scheduling of orders.
The material presented in this chapter is based on research reported by Sawik
(2011b, c), who proposed a portfolio approach for the supplier selection and order
quantity allocation under disruption risks and under operational risks, respectively.
The author applied the two popular in financial engineering percentile measures
of risk, value-at-risk (VaR) and conditional value-at-risk (CVaR) (e.g., Sarykalin
et al. 2008) for managing the risk of supply disruptions or supply delays. The pro-
posed models were further enhanced in this chapter for maximization of expected or
expected worst-case service level and for a multi-region sourcing subject to regional
disruption risks.
Various simplifying assumptions that have been used in the models presented in
this chapter can be relaxed. For example, it has been assumed that each supplier
is capable of manufacturing all required part types. In a more general setting, each
supplier may only be prepared to manufacture a subset of part types and provide
with the parts the corresponding subset of customer orders. The proposed models
can be enhanced also for a discount environment, where the suppliers offer discounts
based on quantity or business volume of ordered parts, e.g., Sawik (2010). A critical
issue that need to be considered before any practical application of the proposed
models is attempted, however, is the estimation of probabilities and the resulting
costs associated with each type of disaster event, for which different approaches are
suggested in the literature, such as expert systems, game theory, utilization of large
simulation models, etc. (e.g., Knemeyer et al. 2009).

Problems

2.1 Modify the probability for disruption scenarios (2.2) to account for correlated
regional disruptions that may affect simultaneously suppliers in different regions.

2.2 Modify the SMIP models presented in this chapter for multiple part types with
subsets of part types required for each customer order and subsets of suppliers avail-
able for each part type.

2.3 Enhance the SMIP models presented in this chapter for selection of static supply
portfolio
(a) with total quantity discount for all ordered parts.
(b) with total business volume discount.
(see, Sawik 2010).

2.4 Mixed mean-risk static supply portfolio

(a) Modify model SP_ECV(c) to optimize expected cost and CVaR of service level
and model SP_ECV(sl) to optimize expected service level and CVaR of cost.
(b) How should the values of the optimized objective functions be scaled into the
interval [0,1] to avoid dimensional inconsistency among the two objectives and how
Problems 41

should the trade-off parameter be selected?

(c) How would you interpret the mixed mean-risk supply portfolio?

2.5 Explain why for a greater range of disruption probabilities, both expected and
expected worst-case costs are higher.
Chapter 3
Selection of Dynamic Supply Portfolio

3.1 Introduction

In this chapter the static portfolio approach and the SMIP formulations presented
in Chap. 2 are enhanced for a multi-period supplier selection and order quantity
allocation in the presence of both the low probability and high impact supply chain
disruption risks and the high probability and low impact supply chain delay risks. The
suppliers are subject to local delivery delay risks, and both local and regional delivery
disruption risks. In the delivery scenario analysis, both types of the supply chain
risks are simultaneously considered. The high probability supply delays may have a
significant impact in a make-to-order manufacturing and just-in-time environment.
In particular, when all suppliers are similarly exposed to high impact disruption risks,
then the low impact delay risk may predominate the supplier selection decision.
The following SMIP models are presented in this chapter:

DSP_E(c) for risk-neutral selection of dynamic supply portfolio to minimize

expected cost;
DSP_E(sl) for risk-neutral selection of dynamic supply portfolio to maximize
expected service level;
DSP_CV(c) for risk-averse selection of dynamic supply portfolio to minimize
CVaR of cost;
DSP_CV(sl) for risk-averse selection of dynamic supply portfolio to maximize
CVaR of service level;
DSP_ECV(c) for mean-risk selection of dynamic supply portfolio to optimize
trade-off between expected cost and CVaR of cost;
DSP_ECV(sl) for mean-risk selection of dynamic supply portfolio to optimize
trade-off between expected service level and CVaR of service level.

© Springer International Publishing AG 2018 43

T. Sawik, Supply Chain Disruption Management Using Stochastic Mixed
Integer Programming, International Series in Operations Research
& Management Science 256, DOI 10.1007/978-3-319-58823-0_3
44 3 Selection of Dynamic Supply Portfolio

In the computational experiments described in Sect. 3.6, the general ‘on-time,

delay or never’ delivery scenarios are illustrated with numerical examples and com-
pared with ‘longest delay or never’ worst-case scenarios.

3.2 Problem Description

Table 3.1 Notation: dynamic supply portfolio

Indices
i = supplier, i ∈ I
j = customer order, j ∈ J
r = geographic region, r ∈ R
s = delivery scenario, s ∈ S
t = planning period, t ∈ T
Input Parameters
cit = capacity of supplier i in period t
dj = demand for parts required for customer order j

D = j∈J d j - total demand for parts
δj = the earliest delivery date of parts for customer order j
δj = the latest delivery date of parts for customer order j
ei = cost of ordering parts from supplier i
gj = per unit and per period penalty cost of delayed customer order j caused by delayed delivery
of required parts
hj = per unit penalty cost of unfulfilled customer order j caused by shortage of required parts
oi j = unit price of parts for customer order j purchased from supplier i
εi j = per unit price reduction for each delayed part for customer order j from supplier i
pi,τ = probability of τ periods delay for delivery of parts from supplier i, τ ∈ {0, . . . , τ }
pi,τ +1 = local disruption probability of supplier i
pr = regional disruption probability of all suppliers i ∈ I r in region r
α = confidence level
τ = 0, . . . , τ , delivery delay
ρi = expected defect rate of supplier i

The approach proposed in this chapter can be applied for a very common type of sup-
ply chain, with a producer of single product, who obtains raw materials from several
different suppliers with limited supply capacity to meet customer orders by customer
requested due dates. However, the approach is also applicable to the case of a single
producer, who assembles different types of products using product-specific parts pur-
chased from multiple suppliers (for notation used, see Table 3.1). In order to simplify
further considerations it is assumed that for each product type, one product-specific
part type (e.g. a critical custom part type) needs to be supplied in required amount
by custom parts manufacturers. For example, in the electronics industry producer of
different electronic devices needs to be supplied by electronics manufacturers with
3.2 Problem Description 45

printed wiring boards of different device-specific design. However, the last assump-
tion can be easily relaxed to consider supplies of different product-specific part types
required for each product type.
Let I = {1, . . . , I } be the set of I suppliers and J = {1, . . . , J } the set of J
customer orders for the finished products, known ahead of time. Each customer
order j ∈ J is described by the quantity d j of required custom parts and their latest
delivery date, δ j , to ensure meeting the customer requested due date for products.
The planning horizon consists of T periods and denote by T = {1, . . . T } the set of
planning periods.
Each supplier can provide the producer with custom parts for all customer orders.
However, the suppliers have different capacity and, in addition, differ in price and
quality of purchased parts and in reliability of delivery. Let cit be the capacity of
supplier i in period t, ei , cost of ordering parts from supplier i, oi j , per unit price of
custom parts for customer order j purchased from supplier i and ρi , the expected
defect rate for supplier i.
The suppliers are assumed to be located in R disjoint geographic regions.  Denote
by I r ⊆ I the subset of suppliers in region r ∈ R = {1, . . . , R}, where r ∈R I r = I .
The supplies of parts are subject to random local delays or disruptions that are
uniquely associated with a particular supplier, which may arise from equipment
breakdowns, local labor strike, fires, etc. In addition to independent local delays and
independent local disruptions of each supplier individually there are also potential
regional disasters that may result in correlated regional disruption of all suppliers
in the same region simultaneously. For example, such regional disaster events may
include floods, hurricanes, earthquakes, widespread labor strikes in a transportation
sector, etc.
The delivery of parts from each supplier i ∈ I is subject to random delay of differ-
ent length, τ ∈ T = {0, . . . , τ , τ + 1}, i.e., parts ordered for period t are delivered
in period t + τ , where τ = 0, represents on time delivery and τ = τ , represents max-
imum delay that may occur. The delivery of parts is also subject to random disrup-
tions, i.e., parts are not delivered at all. By convention, the dummy delay τ = τ + 1,
represents disruption of supplies, i.e., no delivery of parts. If the delivery of parts
for customer order j occurs in period t + τ > δ j , then the delivery is delayed and
the producer is charged by the customer with contractual penalty, g j d j (t + τ − δ j ),
where g j is the per unit and per period penalty cost of delayed customer order j
caused by the delayed delivery of required parts. If supplies of parts for customer
order j are disrupted (i.e., dummy delay of τ = τ + 1 periods), then the producer
is charged with the contractual penalty, h j d j , for unfulfilled order j, where h j is
the per unit penalty cost of unfulfilled customer order j caused by the shortage of
required parts.
Denote by S = {1, . . . , S} the index set of all delivery scenarios, and by Ps the
probability of delivery scenario s ∈ S. While supply disruptions are typically mod-
elled by binomial random variables, the combined disruptions and different length
delays of supply are modelled by multinomial random variables. Each scenario s ∈ S
can be represented by an integer-valued vector τs = {τ1s , . . . , τ I s }, where τis ∈ T
is the length of delay from supplier i ∈ I under scenario s ∈ S. When all potential
46 3 Selection of Dynamic Supply Portfolio

delivery scenarios are considered, then S = (τ + 2) I . For each scenario s ∈ S, the

supplies from every supplier can be delayed or disrupted either by a local or a regional
event. Denote by Is ⊂ I the subset of non-shutdown (non-disrupted) suppliers, who
can deliver orders on time or delayed under scenario s. The probability Ps for delivery
scenario s ∈ S with the subset Is of non-shutdown suppliers is

Ps = Psr . (3.1)
r ∈R

Psr is the probability of realizing of delivery scenario s for suppliers in I r

  
(1 − pr ) i∈I r 
( pi,τis ) if I r  Is = ∅
Psr = (3.2)
pr + (1 − pr ) i∈I r pi,τ +1 if I r Is = ∅,

where pi,τis is the probability of occurrence the delay of length τis periods from sup-
plier i under scenario s. By convention, pi,τ +1 , is the probability of local disruption
of supplier i. The parts ordered fromsupplier i can be delivered on-time with proba-
bility, pi0 , delayed with probability ττ =1 piτ , or not delivered at all with probability

pi,τ +1 = 1 − ττ =0 piτ .
Denote by πi the total disruption (shutdown) probability of every supplier i ∈
Ir,r ∈ R
πi = pr + (1 − pr ) pi,τ +1 ; i ∈ I r , r ∈ R. (3.3)

The decision maker needs to decide on the selection of part suppliers, on quantities
of various custom parts to be ordered from each selected supplier and on delivery
dates to minimize expected cost or maximize expected service level or to mitigate the
impact of disruption risks by minimizing the potential worst-case cost or maximizing
the potential worst-case service level. Hence, the decision maker needs to select a
risk-neutral or risk-averse dynamic supply portfolio, i.e. the allocation of demand
parts among the suppliers and among the planning periods.

3.3 Models for Risk-Neutral Decision-Making

In this section two SMIP models DSP_E(c) and DSP_E(sl) are presented for selection
of a risk-neutral multi-period supply portfolio in the presence of supply delay and
disruption risks to minimize expected cost or to maximize expected service level,
respectively.
The decision maker needs to select a dynamic supply portfolio, i.e. the allocation
of orders for parts among the suppliers and among the planning periods. The dynamic
supply portfolio is defined below, (for definition of problem variables, see Table 3.2).

{Vit : i ∈ I, t ∈ T },
3.3 Models for Risk-Neutral Decision-Making 47

Table 3.2 Variables: dynamic supply portfolio

First stage variables
ui = 1, if an order for parts is placed on supplier i; otherwise u i = 0 (supplier selection)
vi jt = the fraction of total demand for parts required for customer order j to be delivered by
supplier i in period t (allocation of demand for custom parts)
Auxiliary variables
Vit = the fraction of total demand for parts ordered from supplier i to be delivered in period t
(dynamic supply portfolio: allocation of total demand for parts among suppliers and over
time)
VaRc Cost-at-Risk, the targeted cost such that for a given confidence level α, for 100α% of the
scenarios, the outcome is below VaRc
VaR sl Service-at-Risk, the targeted service level such that for a given confidence level α, for
100α% of the scenarios, the outcome is above VaRsl
Cs ≥ 0, the tail cost for scenario s, i.e., the amount by which costs in scenario s exceed VaRc
Ss ≥ 0, the tail service level for scenario s, i.e., the amount by which VaRsl exceeds service
level in scenario s

where 
Vit = 1
i∈I t∈T

and 0 ≤ Vit ≤ 1 is the fraction of the total demand for parts ordered from supplier i
for period t and Vit is determined by the custom parts allocation variables vi jt

Vit = d j vi jt /D; i ∈ I. (3.4)
j∈J

The supply delays and disruptions may result in the shortage of required parts and
the corresponding delay and shortage penalty costs of delayed or unfulfilled customer
orders should be incorporated into the model. Clearly, the producer does not need to
pay for ordered and defective or undelivered parts, whereas parts delivered late may
be paid for at a reduced price. However, the producer can be charged with a much
higher penalty cost for delayed or unfulfilled customer orders for products, caused
by the shortage of required parts due to defective, delayed or undelivered parts. In
a make-to-order manufacturing no inventory of custom parts can be kept on hand
and the parts are requisitioned with each customer order. In addition, custom parts
required for each customer order j are to be delivered within a time window [δ j , δ j ]
derived from the customer requested due date to further reduce any inventory holding
cost. The delivery cannot be earlier than the earliest date δ j and the required parts
are assumed to be processed as soon as they are delivered. As a result no overage
costs can be considered. On the other hand, the required parts can be delivered late,
beyond the latest date δ j , or not delivered at all. Then, delay or disruption penalty
costs represent the underage costs.
Notice that Table 3.2 does not explicitly define the second stage variables for the
SMIP problem considered. The second stage variables are simply demand allocation
48 3 Selection of Dynamic Supply Portfolio

variables for realized disruption scenarios s, ṽisjt ; i ∈ I, j ∈ J, t ∈ T, s ∈ S, defined

as follows

v if i ∈ Is , j ∈ J, t ∈ T, s ∈ S
ṽisjt = i jt−τis
0 / Is , j ∈ J, t ∈ T, s ∈ S.
if i ∈

In view of the above definition, an explicit introduction of the second stage variables
ṽisjt into the SMIP model formulations is not required.
For each delivery scenario s ∈ S and each customer order j ∈ J to be supplied
with parts by non-disrupted supplier i ∈ Is in period t (i.e., with vi jt > 0), the penalty
cost of delivery delayed by 1 ≤ τis ≤ τ periods is given by

g j max{0, t + τis − δ j }d j ,

i.e., deliveries later than δ j are penalized only.

The resulting total expected delay penalty cost over all delivery scenarios can be
expressed by
 
Ps (g j (t + τis − δ j ) − εi j )d j vi jt ,
s∈S i∈Is t∈T j∈J :δ j <t+τis

where the producer does not need to pay full price for late delivery of parts and hence
the purchasing costs of those parts are reduced by εi j per unit for each delayed part
for customer order j from supplier i.
The total expected penalty cost of unfulfilled customer orders due to shortage of
required parts caused by supply disruptions is

Ps (h j − oi j )d j vi jt ,
s∈S i ∈I
/ s j∈J t∈T

where the producer does not need to pay for parts ordered and non delivered and
hence the purchasing costs of those parts are deducted from the cost of unfulfilled
customer orders.
In a risk-neutral operating conditions the overall quality of the supply portfolio
can be measured by the expected cost per part, E c , (3.5), or expected service level
E sl , (3.6).
 
Ec = ei u i /D + oi j d j vi jt /D
i∈I i∈I j∈J t∈T
 
+ Ps (g j (t + τis − δ j ) − εi j )d j vi jt /D
s∈S i∈Is t∈T j∈J :δ j <t+τis

+ Ps (h j − oi j )d j vi jt /D (3.5)
s∈S i ∈I
/ s j∈J t∈T

E sl = Ps d j vi jt /D (3.6)
s∈S i∈Is j∈J t∈T
3.3 Models for Risk-Neutral Decision-Making 49

The expected cost E c includes, cost of ordering,


i∈I ei u i /D,
cost
 ofpurchasing  non defective parts,
i∈I j∈J t∈T j oi j vi jt /D,
d
cost
   deliveries
of delayed  of parts,
s∈S i∈Is t∈T j∈J :δ j <t+τis Ps (g j (t + τis − δ j ) − εi j )d j vi jt /D,
and cost of shortage of parts due to supply disruptions (cost of unfulfilled customer
orders
  less cost
 of non
 delivered parts),
s∈S i ∈I
/ s j∈J t∈T Ps (h j − oi j )d j vi jt /D.

The purchase orders for parts are assumed inflated by the reject rates ρi , i ∈ I
 to be
of defective parts, i.e., are equal to i∈I j∈J t∈T (1 + ρi )d j vi jt . However, since
the
 producer
  does not need to pay for ordered and defective parts in the amount of
 j∈J
i∈I  t∈T  ρi d j vi jt , and pays a reduced price for delayed parts in the amount
of i∈I t∈T j∈J :δ j <t+τ d j vi jt , the corresponding purchasing cost per part for
delivered parts is simply given by

oi j d j vi jt /D
i∈I j∈J t∈T
  
− Ps εi j d j vi jt /D − Ps oi j d j vi jt /D.
s∈S i∈Is t∈T j∈J :δ j <t+τis s∈S i ∈I
/ s j∈J t∈T

The expected service level, E sl , (3.6), is a surrogate measure of expected customer

demand fulfillment rate and represents the expected fraction of fulfilled demand for
required partss.
The SMIP models DSP_E(c) and DSP_E(sl) for selection of risk-neutral dynamic
supply portfolio to minimize expected cost or maximize expected service level,
respectively, are presented below.

DSP_E(c): Selection of risk-neutral Dynamic Supply Portfolio to minimize

expected cost
Minimize (3.5)
subject to
1. Dynamic supply portfolio selection constraints:
- for each customer order the required parts must be delivered not earlier
than the earliest and not later than the latest delivery date,
- parts cannot be ordered from non-selected suppliers,
- for each selected supplier the total quantity of ordered parts cannot exceed
the supplier capacity,
- at least a minimum order size should be assigned to each selected supplier,
 
vi jt = 1; j ∈ J (3.7)
i∈I t∈T :δ j ≤t≤δ j

vi jt ≤ u i ; i ∈ I, j ∈ J, t ∈ T (3.8)
50 3 Selection of Dynamic Supply Portfolio


(1 + ρi )d j vi jt ≤ cit u i ; i ∈ I, t ∈ T (3.9)
j∈J

vi jt ≥ u i ; i ∈ I (3.10)
j∈J t∈T

4. Non-negativity and integrality conditions:

u i ∈ {0, 1}; i ∈ I, t ∈ T (3.11)
vi jt ∈ [0, 1]; i ∈ I, j ∈ J, t ∈ T. (3.12)

DSP_E(sl): Selection of risk-neutral dynamic supply portfolio to maximize

expected service level
Maximize (3.6)
subject to (3.7)–(3.12).

If total 
available
 capacity of all suppliers
 is less than total demand for required
parts, i.e., i∈I t∈T cit /(1 + ρi ) ≤ j∈J d j , then the demand allocation equality
constraints (3.7) should be replaced by inequalities
 
vi jt ≤ 1; j ∈ J ; (3.13)
i∈I t∈T :δ j ≤t≤δ j

otherwise no feasible solution exists.

A simple upper bound on the expected service level, E sl , (3.6), is derived below.

Proposition 3.1

E sl ≤ min{1, (1 − pr )(1 − pi,τ +1 )cit /(1 + ρi )D}. (3.14)
r ∈R i∈I r t∈T

Proof The dynamic supply portfolio selection constraints (3.9) imply that

Ps d j vi jt /D ≤
s∈S i∈Is j∈J t∈T

(1 − pr )(1 − pi,τ +1 )d j vi jt /D ≤
r ∈R i∈I r j∈J t∈T

(1 − pr )(1 − pi )cit u i /(1 + ρi )D ≤
r ∈R i∈I r t∈T

(1 − pr )(1 − pi,τ +1 )cit /(1 + ρi )D,
r ∈R i∈I r t∈T
3.3 Models for Risk-Neutral Decision-Making 51

where (1 − pr )(1 − pi,τ +1 ) = 1 − πi , (3.3), is a non-disruption probability of sup-

plier i ∈ I r .   
Since E sl cannot be greater than 1, its upper bound is 1, if r ∈R i∈I r t∈T (1 −
pr )(1 − pi,τ +1 )cit /(1 + ρi )D > 1.

If for each customer order, all custom parts should be provided by a single delivery
from one supplier only, then the continuous demand allocation variables vi jt should
be redefined as binary assignment variables denoting whether or not all parts required
for customer order j are provided by supplier i in period t, (e.g., Chap. 2).

3.3.1 Dynamic Supply Portfolio for All-or-Nothing Delivery

Scenarios

The dynamic supply portfolio approach is particularly useful when supply delay risks
need to be considered. Otherwise, the static supply portfolio presented in Chap. 2 may
be a more appropriate approach. For the pure all-or-nothing delivery scenarios (i.e.,
for disruption scenarios), with supply disruptions only and no delays (see Sects. 2.2,
2.3 in Chap. 2), the proposed multi-period models for I suppliers and T planning
periods can be transformed into equivalent single-period models for m = (I )(T )
suppliers. To this end, each pair of indices (i, t), i ∈ I, t ∈ T should be replaced
with a single index k ∈ K , where K = {1, . . . , m} represents the set of equivalent
m single-period suppliers. In the proposed SMIP models, variables u i , vi jt should
be replaced by u k , vk j , respectively. Then, the dynamic supply portfolio selection
constraints (3.7) are replaced by

vk j = 1; j ∈ J,
k∈K j

where K j is the subset of equivalent single-period suppliers that are capable of

delivering parts required for customer order j and it represents the subset {(i, t) :
i ∈ I, δ j ≤ t ≤ δ j } of pairs (i, t) feasible for customer order j.

3.4 Models for Risk-Averse Decision-Making

In the risk-averse selection of supply portfolio under disruption risks, the confidence
level α is fixed by the decision maker to control the risk of losses due to supply
disruptions. We assume that the decision maker is willing to accept only portfolios
for which the the total probability of scenarios with costs greater than VaRc or with
service level lower than VaRsl is not greater than 1 − α. Furthermore, a risk aversive
decision maker wants to minimize the expected worst-case costs exceeding VaRc or
to maximize the expected worst-case service level below VaRsl .
52 3 Selection of Dynamic Supply Portfolio

Define by Cs the tail cost for scenario s, where tail cost is defined as the amount
by which costs in scenario s exceed VaRc . In a similar way, define by Ss the tail
service level for scenario s, where tail service level is defined as the nonnegative
amount by which VaRsl exceeds service level in scenario s.
The portfolio will be optimized by calculating VaRc and minimizing CVaRc simul-
taneously or by calculating VaRsl and maximizing CVaRsl , respectively.
In the proposed models CVaR is represented by an auxiliary function (3.15) and
(3.18) introduced by Rockafellar and Uryasev (2000). The SMIP models DSP_CV(c)
and DSP_CV(sl) for the risk-averse selection of supply portfolio to reduce the risk
of high costs and the risk of low service level, respectively, is formulated below.

DSP_CV(c): Selection of risk-averse dynamic supply portfolio to minimize

CVaR of cost
Minimize 
C V a R c = V a R c + (1 − α)−1 Ps Cs (3.15)
s∈S

subject to
1. Dynamic supply portfolio selection constraints: (3.7)–(3.10)
2. Risk constraints:
- the tail cost for scenario s is defined as the nonnegative amount by which
cost per part in scenario s exceeds VaRc ,
 
Cs ≥ ei u i /D + oi j d j vi jt /D
i∈I i∈I j∈J t∈T
 
+ (g j (t + τis − δ j ) − εi j )d j vi jt /D
i∈Is t∈T j∈J :δ j <t+τis

+ (h j − oi j )d j vi jt /D − V a R c (3.16)
i ∈I
/ s j∈J t∈T

3. Non-negativity and integrality conditions: (3.11), (3.12) and

Cs ≥ 0; s ∈ S. (3.17)

DSP_CV(sl): Selection of risk-averse dynamic supply portfolio to maximize

CVaR of service level
Maximize 
C V a R sl = V a R sl − (1 − α)−1 Ps Ss (3.18)
s∈S

subject to
3.4 Models for Risk-Averse Decision-Making 53

1. Dynamic supply portfolio selection constraints: (3.7)–(3.10)

2. Risk constraints:
- the tail service level for scenario s is defined as the nonnegative amount
by which VaRsl exceeds service level in scenario s,

Ss ≥ V a R sl − d j vi jt /D; s ∈ S (3.19)
i∈Is j∈J t∈T

3. Non-negativity and integrality conditions: (3.11), (3.12) and

Ss ≥ 0; s ∈ S. (3.20)

Note that variables V a R c and V a R sl do not need to be constrained of being

nonnegative. As Cs and Ss are constrained of being positive, the model tries to
decrease V a R c and increase V a R sl , respectively, and hence positively impact the
objective function. However, large reduction in V a R c and increase in V a R sl may
result in more scenarios with positive tail costs and service levels, respectively.

3.5 Models for Mean-Risk Decision-Making

In the single objective approach the supply portfolio is selected by minimizing either
the expected cost per part, E c , (3.5), or the expected worst-case cost per part, C V a R c ,
(3.15) or by maximizing either the expected service level, E sl , (3.6), or the expected
worst-case service level, C V a R sl , (3.18). In this section the two cost functions and
the two service level functions are considered simultaneously, and a bi-objective
selection of supply portfolio is presented aimed at minimizing both objective func-
tions to balance expected costs or expected service level with the risk tolerance. This
trade-off model is known as the mean-risk model, formulated as the optimization of
a composite objective consisting of the expected cost (service level) and the CVaR
as a risk measure.
A subset of nondominated solutions for the bi-objective dynamic supply portfolio
can be found by the parameterization on λ the weighted-sum programs DSP_ECV(c)
and DSP_ECV(sl) presented below. The mean-risk program DSP_ECV(c) is based
on model DSP_CV(c) with the addition of objective (3.5) of model SP_E(c). Sim-
ilarly, the mean-risk program DSP_ECV(sl) is based on model DSP_CV(sl) with
the addition of objective (3.6) of model DSP_E(sl).

DSP_ECV(c): Selection of mean-risk dynamic supply portfolio to minimize

weighted sum of expected cost and CVaR of cost
54 3 Selection of Dynamic Supply Portfolio

Minimize
λE c + (1 − λ)C V a R c , (3.21)

where 0 ≤ λ ≤ 1
subject to (3.5), (3.7)–(3.12), (3.15)–(3.17).

DSP_ECV(sl): Selection of mean-risk dynamic supply portfolio to maximize

weighted sum of expected service level and CVaR of service level
Maximize
λE sl + (1 − λ)C V a R sl (3.22)

where 0 ≤ λ ≤ 1,
subject to (3.6)–(3.12), (3.18)–(3.20).

3.6 Computational Examples

In this section computational examples are presented to illustrate possible applica-

tions of the proposed SMIP models for selection of a dynamic supply portfolio in the
presence of delay and disruption risks. Although the input data for the examples are

Fig. 3.1 Suppliers

3.6 Computational Examples 55

hypothetical, their relations to each other are real. In the cost-oriented selection of
supply portfolio a combination of disruption and delay risks simultaneously impacts
both shortage and delay penalties. In the service-oriented decisions, demand fulfill-
ment rate (the fraction of demand for parts fulfilled during the planning horizon) is
impacted.
The following parameters have been used for the example problems:
• I , the number of suppliers, was equal to 5.
• R, the number of geographic regions, was equal to 2, and the subsets of suppliers
were I 1 = {1, 2} and I 2 = {3, 4, 5}, respectively.
• J , the number of customer orders, was equal to 50.
• T , the planning horizon, was equal to 14 periods.
• τ , maximum delivery delay was equal to 4 periods for each supplier, so that the set
of all possible delays is T = {0, 1, 2, 3, 4, 5}, where a dummy delay, τ + 1 = 5,
represents local disruption of a supplier. The corresponding number of delivery
scenarios, was equal to S = 6 I = 7776.
• d j , the numbers of required parts for each customer order, were integers uni-
formly distributed over [500, 5000] for all customer orders j and the resulting
total demand for parts was D = 138000.
• ei , the cost of ordering parts, were integers in {8000, 9000} and {12000, 13000,
14000}, respectively for suppliers i ∈ I 1 and i ∈ I 2 .
• oi j , the unit price of parts for customer order j purchased from supplier i, was
uniformly distributed over [13,15] and [9,11], respectively for suppliers i ∈ I 1 and
i ∈ I 2.
• εi j =0.05oi j , per unit price reduction for each delayed part of customer order j
from supplier i was 5% of the unit price.
• g j = 1, per unit and per period delay penalty cost was equal to 1 for all customer
orders j.
• h j = 4 maxi∈I (oi j ), per unit shortage cost for each customer order j, was equal
to four times of the maximum purchasing cost of required parts.
• cit = 2D/((T )(I ))
, ( ·
denotes the smallest integer not less than ·), i.e., the
total capacity of all suppliers was equal to the double total demand for parts;
• pi5 , the local disruption probability (dummy delay, τ + 1 = 5, represents local
disruption) was uniformly distributed over [0.01,0.05] and [0.05;0.10], respec-
tively for suppliers i ∈ I 1 and i ∈ I 2 , i.e., the disruption probabilities were drawn
independently from U[0.01,0.05] and U[0.05;0.10], respectively.
Given local disruption probabilities, pi5 , i ∈ I , the probabilities for different delay
length τ = 0, 1, 2, 3, 4 were calculated as follows:
probability of on time delivery (τ = 0), pi0 = 0.3(1 − pi5 );
probability of 1-period delay (τ = 1), pi1 = 0.25(1 − pi5 );
probability of 2-period delay (τ = 2), pi2 = 0.2(1 − pi5 );
probability of 3-period delay (τ = 3), pi3 = 0.15(1 − pi5 );
probability of maximum, 4-period delay (τ = τ = 4), pi4 = 0.1(1 − pi5 );
for all suppliers i ∈ I , i.e., the longer delay the lower probability of its occurrence.
56 3 Selection of Dynamic Supply Portfolio

Notice that for each supplier i, the expected delay is equal to 2 periods ( 0.25(1 −
pi5 ) + 2(0.2(1 − pi5 )) + 3(0.15(1 − pi5 )) + 4(0.1(1 − pi5 ))/(1 − pi5 )
=2)
• pr , the regional disruption probability was 0.005 and 0.01, respectively for region
r = 1 and r = 2.
• δ j , the latest delivery date of parts required for customer order j, was integer
uniformly distributed over [3, T ], i.e., generated from a U[3;14] distribution.
• δ j = δ j − 2, the earliest delivery date of parts required for each customer order j,
was two periods earlier (i.e., the expected delay) than the latest delivery date.
• ρi , the expected defect rate of each supplier i was exponentially distributed.
• α, the confidence level, was equal to 0.50, 0.75, 0.90, 0.95 or 0.99.
Note that cost of lost customer orders, h j , is set to be much higher than the corre-
sponding cost, g j , of delayed orders, which is typical for industrial practice.
 Figure 3.1 shows basic characteristics of each supplier i, average price per part,
j∈J oi j /J , and disruption probability, πi , (3.3). Suppliers i = 1, 2 in region r = 1
are most reliable and most expensive, supplier i = 4 is most unreliable, while supplier
i = 5 is the cheapest one.
The risk-neutral solutions for models DSP_E(c) and DSP_E(sl) are shown in
Table 3.3, and the risk-averse solutions for models DSP_CV(c) and DSP_CV(sl)
with different confidence levels, in Table 3.4. Table 3.3 indicates that for both objec-
tive functions, the risk-neutral supply portfolios do not include the most unreliable
supplier i = 4. The corresponding risk-neutral dynamic supply portfolios for each
supplier i, ( j∈J d j vi jt /D, t ∈ T ), are shown in Fig. 3.2. The cost-based dynamic
supply portfolio is concentrated in a shorter time interval than the supply portfolio
for the service level objective, which may reduce delay penalty costs. Both dynamic
portfolios are leveled with respect to the three most reliable suppliers i = 1, 2, 3,
where, supplier i = 3, is, in addition, one of the cheapest suppliers. The full capacity
of these suppliers is fully utilized most of the time.
Figure 3.3 shows distribution of the expected risk-neutral dynamic supply portfo-
lio, Ṽit , i ∈ I, t ∈ {1, . . . , T + τ }, defined below.
 
Ṽit = Ps d j ṽisjt 
s∈S j∈J t  ∈T :t=t  +τis

where

vi jt  if τis ≤ τ
ṽisjt  =
0 if τis = τ + 1.

The expected fraction of total demand for parts is unevenly distributed over the
planning horizon. For both objective functions, the highest delivery levels appear in
mid-horizon and the lowest levels at the beginning and at the end of the horizon.
In the computational experiments for the risk-averse portfolios, the confidence
level α is set at five levels of 0.5, 0.75, 0.90, 0.95, and 0.99, which means that focus
is on minimizing the highest 50%, 25%, 10%, 5%, and 1% of all scenario outcomes,
3.6 Computational Examples 57

i.e., costs per part or service level. When α increases, a more risk-aversive decision-
making focuses on a smaller set of outcomes and the number of selected suppliers
also is increasing to mitigate the impact of disruptions risks by diversification of the
supply portfolio. Table 3.4 demonstrates that supplier i = 4 is not selected for the
lowest confidence level α = 0.5, for which the risk-averse supply portfolio is similar
to the risk-neutral portfolio. For larger α, all suppliers are selected, except for the
largest α = 0.99 and model DSP_CV(sl), for which supplier i = 4 is not selected
again. This indicates that diversification of the supply portfolio is no longer required
to mitigate the impact of disruption risks. Note that for α = 0.50 and α = 0.75, VaRc
is smaller than expected cost, E c , whereas VaRsl is greater than expected service level,
E sl .

Fig. 3.2 Risk-neutral dynamic supply portfolio: a model DSP_E(c), b model DSP_E(sl)
58 3 Selection of Dynamic Supply Portfolio

The risk-averse dynamic supply portfolios for α = 0.9 and α = 0.99 and the two
models are shown in Figs. 3.4 and 3.5.
Unlike the risk-neutral dynamic supply portfolios that are leveled over the plan-
ning horizon, in particular with respect to the three most reliable suppliers i = 1, 2, 3,
the corresponding risk-averse portfolios are more unevenly distributed over the hori-
zon. The most unleveled are the risk-averse supply portfolios for service level objec-
tive. The only exception is the risk-averse dynamic supply portfolios for the cost-
based objective and the highest confidence level α = 0.99, which is very similar to
the risk-neutral portfolio (cf. Figs. 3.2a and 3.4b). The main difference is the selec-
tion of supplier i = 4 as a supportive supplier for the risk-averse portfolio, with a
small fraction of total demand allotted.

Fig. 3.3 Expected risk-neutral dynamic supply portfolio: a model DSP_E(c), b model DSP_E(sl)
3.6 Computational Examples 59

Table 3.3 Risk-neutral solutions

Model DSP_E(c) Model DSP_E(sl)
Var. = 3505, Bin. = 5, Cons. = 3675, Nonz. = 18325 (a)
E c = 15.76 E sl = 95.54
E sl = 94.76 E c = 16.51
Suppliers Selected(% of total demand)(b)
1(26) 1(30)
2(28) 2(34)
3(28) 3(28)
5(18) 5(8)
(a) Var. = number of variables, Bin. = number of binary variables,

Cons. = number of constraints, Nonz. = number of nonzero coefficients.

(b) 1(
 
d j v1 jt /D × 100),
 j∈J  t∈T
2( j∈J t∈T d j v2 jt /D × 100),
 
3( j∈J t∈T d j v3 jt /D × 100),
 
4( j∈J t∈T d j v4 jt /D × 100),
 
5( j∈J t∈T d j v5 jt /D × 100).

The optimal risk-averse cost distributions for confidence level α = 0.9 and α =
0.99 are shown in Fig. 3.7. For α = 0.9, larger probability atoms are concentrated
at higher costs than for α = 0.99, i.e., more risk-aversive supply portfolios better
mitigate the risk of high costs.
The computational results (e.g., Sawik 2011d) indicate that

• both cost- and service-oriented dynamic supply portfolios that simultaneously

mitigate the impact of low-probability disruption risk and high-probability delay
risk lead to a high expected demand fulfillment rate;
• as shortage-to-delay unit penalty ratio increases, more demand is moved from
unreliable, low-cost suppliers to reliable, high-cost suppliers to minimize shortage
cost and better mitigate the impact of disruption risk;
• a cost-oriented dynamic supply portfolio is better leveled over the planning horizon
and is concentrated in a fewer number of periods;
• neglecting potential delay risks in supplier selection may lead to greater supply
fluctuations and manufacturing delays,which is particularly harmful in a make-
to-order and just-in-time environment.

The proposed dynamic portfolio approach leads to time-indexed stochastic mixed

integer programs with a strong LP relaxation, which has proven to be computationally
very efficient.
60 3 Selection of Dynamic Supply Portfolio

3.6.1 Worst-case Scenarios with Longest Delays

and Disruptions

In this subsection a special subset of delivery scenarios is considered such that deliv-
ery of parts from each supplier is either delayed by maximum delay length, τ , or
parts are not delivered at all, because of supply disruptions. Let us call such delivery
scenarios, LDN (Longest Delay-or-Never) scenarios. The total number of LDN sce-
narios is 2 I = 32. Given local disruption probabilities, pi5 , i ∈ I , the probability for
the longest delay of delivery is simply pi4 = (1 − pi5 ), while the remaining proba-

Table 3.4 Risk-averse solutions

Confidence level α 0.50 0.75 0.90 0.95 0.99
Model DSP_CV(c)
Var. = 11282, Bin. = 5, Cons. = 11451, Nonz. = 27288757 (a)
CVaRc 18.92 24.22 30.29 35.93 45.66
VaRc 13.10 15.27 21.05 30.14 43.17
Ec 15.88 16.06 16.05 15.95 15.86
E sl (b) 95.49 95.52 95.20 94.88 94.53
Suppliers Selected(% of total demand)(c) 1(31) 1(31) 1(31) 1(30) 1(26)
2(33) 2(33) 2(33) 2(29) 2(26)
3(26) 3(28) 3(12) 3(14) 3(23)
4(4) 4(12) 4(14) 4(7)
5(10) 5(4) 5(12) 5(13) 5(18)
Model DSP_CV(sl)
Var. = 11282, Bin. = 5, Cons. = 11451, Nonz. = 22713877 (a)
CVaRsl % 96.01 92.02 85.29 80.12 69.82
VaRsl % 100 100 91.87 87.50 77.77
E sl (b) 98.00 98.00 97.29 97.43 97.04
Ec 15.56 15.48 15.86 15.85 15.61
Suppliers Selected(% of total demand)(c) 1(30) 1(30) 1(30) 1(29) 1(25)
2(34) 2(34) 2(34) 2(29) 2(25)
3(28) 3(28) 3(12) 3(14) 3(25)
4(4) 4(12) 4(14)
5(8) 5(4) 5(12) 5(14) 5(25)
(a) Var. = no. of variables, Bin. = no. of binary variables,

Cons. = no. of constraints, Nonz. = no. of nonzero coefficients.

(b)
   
t∈T Ps d j vi jt /D × 100%.
(c) 1(
s∈S

i∈Is j∈J
d v
j 1 jt /D × 100),
 j∈J  t∈T
2( j∈J t∈T d j v2 jt /D × 100),
 
3( j∈J t∈T d j v3 jt /D × 100),
 
4( j∈J t∈T d j v4 jt /D × 100),
 
5( j∈J t∈T d j v5 jt /D × 100).
3.6 Computational Examples 61

bilities, pi0 , pi1 , pi2 , pi3 are zero for all i ∈ I . Notice that LDN delivery scenarios
can be considered as pure all-or-nothing disruption scenarios, where ordered parts
are either delivered with a fixed delay, τ , or not delivered at all.
The risk-neutral solutions for LDN scenarios and models DSP_E(c) and
DSP_E(sl) are summarized in Table 3.5. The solution results for model DSP_E(c)
are similar and for model DSP_E(sl) are identical with the corresponding results for
the general delivery scenarios. The probabilities Ps for scenarios s ∈ S with non-
disrupted subsets of suppliers, Is , are identical for both general and LDN scenarios
and the service level objective is independent of delivery delays. Table 3.5 shows that
the expected costs are greater for LDN scenarios (cf. Table 3.3).

Fig. 3.4 Risk-averse dynamic supply portfolio for model DSP_CV(c): α = 0.9 and α = 0.99
62 3 Selection of Dynamic Supply Portfolio

The risk-averse solutions for LDN scenarios and model DSP_CV(c) are summa-
rized in Table 3.6, while Fig. 3.6 presents examples of optimal risk-averse dynamic
supply portfolios for confidence levels α = 0.9 and α = 0.99. Table 3.6 shows that
solution values for the LDN delivery scenarios are greater than the corresponding
values for the general delivery scenario, while the corresponding expected service
level and supply portfolios are similar. The main difference is no selection of most
unreliable supplier i = 4 for the highest confidence level α = 0.99 (cf. Table 3.4).
The dynamic risk-averse supply portfolio for the highest confidence level, α = 0.99,
is again better leveled, however, not as well as for the general delivery scenario.

Fig. 3.5 Risk-averse dynamic supply portfolio for model DSP_CV(sl): α = 0.9 and α = 0.99
3.6 Computational Examples 63

Table 3.5 Risk-neutral solutions for LDN scenarios

Model DSP_E(c) Model DSP_E(sl)
Var. = 3505, Bin. = 5, Cons. = 3675, Nonz. = 18325 (a)
Ec = 17.02 E sl = 95.54
E = 94.71
sl E c = 18.22
Suppliers Selected(% of total demand) (b)

1(26) 1(30)
2(27) 2(34)
3(27) 3(28)
5(20) 5(8)
(a) Var. = number of variables, Bin. = number of binary variables,

Cons. = number of constraints, Nonz. = number of nonzero coefficients.

(b) 1(
 
d j v1 jt /D × 100),
  t∈T
j∈J
2( j∈J t∈T d j v2 jt /D × 100),
 
3( j∈J t∈T d j v3 jt /D × 100),
 
4( j∈J t∈T d j v4 jt /D × 100),
 
5( j∈J t∈T d j v5 jt /D × 100).

Table 3.6 Risk-averse solutions for LDN scenarios: model DSP_CV(c)

Confidence level α 0.50 0.75 0.90 0.95 0.99
Var. = 3538, Bin. = 5, Cons. = 3707, Nonz. = 130549 (a)
CVaRc 20.08 25.45 31.18 36.72 46.12
VaRc 14.49 18.72 22.00 32.22 43.85
Ec 17.28 17.39 17.43 17.25 17.06
E sl (b) 95.44 95.41 95.20 94.87 94.56
Suppliers Selected(% of total demand)(c) 1(31) 1(31) 1(31) 1(30) 1(26)
2(33) 2(33) 2(33) 2(29) 2(26)
3(23) 3(23) 3(12) 3(14) 3(24)
4(6) 4(12) 4(14)
5(13) 5(7) 5(12) 5(13) 5(24)
(a) Var. = no. of variables, Bin. = no. of binary variables,

Cons. = no. of constraints, Nonz. = no. of nonzero coefficients.

(b)
   
t∈T Ps d j vi jt /D × 100%.
(c) 1(
s∈S

i∈Is j∈J
d v
j 1 jt /D × 100),
 j∈J  t∈T
2( j∈J t∈T d j v2 jt /D × 100),
 
3( j∈J t∈T d j v3 jt /D × 100),
 
4( j∈J t∈T d j v4 jt /D × 100),
 
5( j∈J t∈T d j v5 jt /D × 100).

The solution results for LDN scenarios and model DSP_CV(sl) are not shown,
since they are identical with those for the general delivery scenarios presented in
Table 3.4.
64 3 Selection of Dynamic Supply Portfolio

Fig. 3.6 Risk-averse dynamic supply portfolios for model DSP_CV(c) and LDN scenarios:
a α = 0.9, b α = 0.99

The optimal risk-averse cost distributions for confidence level α = 0.9 and
α = 0.99 are shown in Figs. 3.7 and 3.8, respectively for general delivery scenar-
ios and LDN disruption scenario. Figures 3.7 and 3.8 indicate that the probability
mass function of cost per part is concentrated in a few points, which is typical for
the scenario-based optimization under uncertainty, where the probability measure is
concentrated in finitely many points. Comparison of probability mass functions in
Figs. 3.7 and 3.8 shows that for the binomial LDN disruption scenarios the prob-
ability measure is more concentrated in finitely many points than for the general
multinomial delivery scenarios.
3.6 Computational Examples 65

Fig. 3.7 Probability mass function for model DSP_CV(c): α = 0.9 and α = 0.99

Fig. 3.8 Probability mass function for model DSP_CV(c) and LDN scenarios: α = 0.9 and α =
0.99

The computational experiments were performed using the AMPL programming

language and the Gurobi 7.0.0 solver on a MacBookPro laptop with Intel Core i7
processor running at 2.8 GHz and with 16GB RAM. The solver was capable of finding
66 3 Selection of Dynamic Supply Portfolio

proven optimal solution for all examples with CPU time ranging from fraction of a
second for risk-neutral solutions to several seconds for risk-averse solutions.

3.7 Notes

The problem of a multi-period supplier selection and order quantity allocation in the
presence of supply chain delay and disruption risks is very rarely reported in the liter-
ature. The decision-making requires to simultaneously consider the high probability
and low impact supply delays and the low probability and high impact supply dis-
ruptions. While the delay risks may frequently arise from problems in coordinating
supply and demand, (e.g., Oke and Gopalakrishnan 2009; Sawik 1977, 2011c), the
disruption risks are referred to the major disruptions to normal activities caused by
natural and man-made disasters. In most cases, the business impact associated with
disruption risks is much greater than that of the delay risks. The high probability
supply delays may have a significant impact in a make-to-order manufacturing and
just-in-time environment. In particular, when all suppliers are similarly exposed to
high impact disruption risks, then the low impact delay risk may predominate the
supplier selection decision. For example, in Toyota supply chain, many suppliers
are similarly exposed to low probability regional disruption risks due to seismic
hazard in many Japanese prefectures (Marszewska 2016). Then, the supplier relia-
bility of on-time delivery and the associated delay risk become important criteria for
supplier selection in just-in-time environment. For example, to reduce dependency
on external suppliers and attain the capacity to absorb supply fluctuations Toyota
invested in the technology necessary to produce higher-end electronic components
in-house (Ahmadjian and Lincoln 2001). The majority of models developed for sup-
plier selection and order quantity allocation are static (single-period) models (e.g.,
Weber and Current 1993; Demirtas and Ustun 2008) that do not consider inven-
tory management. The dynamic (multi-period) models are capable of considering
the inventory management by lot-sizing and scheduling of orders (e.g., Ghodsypour
2001; Basnet and Leung 2005; Ustun and Demirtas 2008; Che and Wang 2008). For
custom-engineered products, however, no inventory of custom parts can be kept on
hand. Instead, the custom parts often need to be requisitioned with each customer
order and hence the custom parts inventory need not to be considered.
The idea of a dynamic supply portfolio approach in the presence of both the low
probability and high impact supply disruptions and the high probability and low
impact supply delays was presented by Sawik (2011d) for a multi-period supplier
selection and order quantity allocation in a make-to-order environment. For the selec-
tion of a dynamic supply portfolio, a SMIP was proposed to incorporate risks via
scenario analysis. In the scenario analysis, both types of the supply chain risks are
simultaneously considered. The disruption and delay risks were incorporated utiliz-
ing the concepts of percentile measures of risk, VaR and CVaR. In Sawik (2011d),
the cost-based objective function was considered only and general delivery scenarios
with on time, delayed or no supplies. The delays were modeled as statistically inde-
3.7 Notes 67

pendent, discrete multivariate random variables. In addition to the general scenarios,

the two special subsets of the scenarios were considered, with at most one disrup-
tion of each supplier over the horizon and with multiple consecutive disruptions of
each supplier over the horizon since its first disruption. The best-case and worst-case
scenarios, respectively with no delays or with the longest delays were analyzed.
In the future research, service level, (3.6), represented by the expected demand
for parts fulfilled during the entire planning horizon can be replaced by the expected
demand fulfilled within delivery due windows.

Problems

3.1 Modify the probability for delivery scenarios (3.2) to account for regional dis-
ruptions that may affect suppliers during a fixed time interval.

3.2 Modify the SMIP models presented in this chapter for multiple part types with
subsets of part types required for each customer order and subsets of suppliers avail-
able for each part type.

3.3 Modify the SMIP models presented in this chapter to replace multiple deliveries
from each supplier by a single delivery at a fixed delivery date or delivery time
window.

3.4 Mixed mean-risk dynamic supply portfolio

(a) Modify model DSP_ECV(c) to optimize expected cost and CVaR of service level
and model DSP_ECV(sl) to optimize expected service level and CVaR of cost.
(b) How should the values of the optimized objective functions be scaled into the
interval [0,1] to avoid dimensional inconsistency among the two objectives and how
should the trade-off parameter be selected?
(c) How would you interpret the mixed mean-risk dynamic supply portfolio?

3.5 In the computational examples in Sect. 3.6, Figs. 3.4 and 3.5 indicate that
the cost-optimal risk-averse dynamic supply portfolio for the highest confidence
level α = 0.99 is better balanced than the corresponding service-optimal portfolio,
while for a lower α, both portfolios are more similar. Why do the optimal risk-
averse service-based dynamic supply portfolios have such properties for the example
problem?
Chapter 4
Selection of Resilient Supply Portfolio

4.1 Introduction

In this chapter, the portfolio approach and SMIP models presented in Chap. 2 are
enhanced for the combined selection and protection of part suppliers and order quan-
tity allocation in a supply chain with disruption risks. The protection decisions include
the selection of suppliers to be protected against disruptions and the allocation of
emergency inventory of parts to be pre-positioned at the protected suppliers so as to
maintain uninterrupted supplies in case of natural or man-made disruptive events. The
decision maker needs to decide which supplier to select for parts delivery and how
to allocate orders quantity among the selected suppliers, and which of the selected
suppliers to protect against disruptions and how to allocate emergency inventory
among the protected suppliers. The problem objective is to achieve a minimum cost
of suppliers protection, emergency inventory pre-positioning, parts ordering, pur-
chasing, transportation and shortage and to mitigate the impact of disruption risks
by minimizing the potential worst-case cost. The resulting supply portfolio can be
called resilient, with the supplier’s resiliency defined as its capability of supplying
parts in the face of disruptive events. The portfolio includes protected suppliers that
are capable of fully or partially supplying parts in the face of disruptive events as well
as the emergency inventory pre-positioned at the protected suppliers. Depending on
the level of supplier protection, the capacity of each protected supplier remains fully
or partially available under a disruptive event. The emergency inventory is used to
compensate for the loss of capacity of suppliers hit by disruptions, unprotected or
insufficiently protected, and to partially or fully replace non-delivered parts ordered
from the disrupted suppliers. For the selection of risk-neutral, risk-averse or mean-
risk supply portfolio, SMIP formulations are developed with single- or multi-level
protection of fortified suppliers. In the former case, full remaining capacity of a
disrupted supplier is maintained by its fortification (protection against disruptions),
whereas in the latter case the remaining fraction of full capacity depends on the pro-
tection level applied. A simple protection index is introduced to evaluate the trade-off

© Springer International Publishing AG 2018 69

T. Sawik, Supply Chain Disruption Management Using Stochastic Mixed
Integer Programming, International Series in Operations Research
& Management Science 256, DOI 10.1007/978-3-319-58823-0_4
70 4 Selection of Resilient Supply Portfolio

between the cost of suppliers protection and the estimated losses caused by supply
disruptions, if no protective countermeasure is applied.
The following SMIP models are presented in this chapter:

RSP_E for risk-neutral selection of resilient supply portfolio to minimize

expected cost;
RSP_CV for risk-averse selection of resilient supply portfolio to minimize
CVaR of cost;
RSP_ECV for mean-risk selection of resilient supply portfolio to optimize
trade-off between expected cost and CVaR of cost.
RSP(mlp)_E model RSP_E for multi-level protection;
RSP(mlp)_CV model RSP_CV for multi-level protection;
RSP(mlp)_ECV model RSP_ECV for multi-level protection.

The resilient supply portfolio is selected ahead of time to optimize average

(risk-neutral models, RSP_E and RSP(mlp)_E), worst-case (risk-averse models,
RSP_CV and RSP(mlp)_CV) or combined, average - worst-case (mean-risk mod-
els RSP_ECV and RSP(mlp)_ECV) performance of a supply chain under disruption
risks.
Computational examples are provided in Sect. 4.6 to illustrate the proposed SMIP
approach.

4.2 Problem Description: Single-Level Protection

In the supply chain under consideration various types of products are assembled by a
single producer to satisfy customer orders, using different part types purchased from
multiple suppliers (for notation used, see Table 4.1). The suppliers have different
limited capacity and, in addition, differ in price and quality of offered parts. Let
I = {1, . . . I} be the set of I suppliers and J = {1, . . . J} the set of J part types
required for the products. Denote by dj the demand for each part type j ∈ J and
assume that dj is known ahead of time.
The usage per part of supplier’s capacity is assumed to be different for different
part types and suppliers. Let ci be the total capacity of supplier i ∈ I (e.g., the total
number of available machine-hours) and aij the unit capacity consumption of supplier
i for part type j, i.e., the amount of capacity of supplier i used to manufacture one
part type j (e.g., the unit processing time).
Denote by oij the unit purchasing price, including shipping cost of part type j ∈ J
from supplier i ∈ I (assume that for suppliers incapable of providing some part types,
the corresponding unit price and unit capacity consumption are very large numbers).
Let ρij be the expected defect rate (reject rate) of supplier i for part type j. The rate
4.2 Problem Description: Single-Level Protection 71

Table 4.1 Notation: single-level protection

Indices
i = supplier, i ∈ I
j = part type, j ∈ J
s = disruption scenario, s ∈ S
Input Parameters
aij = per unit capacity consumption of supplier i for part type j
ci = capacity of supplier i
dj = demand for part type j

D = j∈J dj - total demand for parts
ei = cost of ordering parts from supplier i
fi = protection cost for supplier i
gj = per unit shortage cost of part type j
hij = per unit cost of pre-positioning emergency inventory of parts type j at supplier i
oij = unit price of part type j purchased and shipped from supplier i
pi = local disruption probability for supplier i
p∗ = global disruption probability for all suppliers
α = confidence level
ν = minimum order size
ρij = expected defect rate of supplier i for part type j

is based on past observations. The fixed cost of ordering parts from supplier i ∈ I is
denoted by ei .
The supplies of parts are subject to independent random local disruptions that
are uniquely associated with a particular supplier, which may arise from equip-
ment breakdowns, local labor strike, bankruptcy, terrorist attack, from local natural
disasters such as earthquakes, fires, floods, hurricanes, etc. Denote by pi the local dis-
ruption probability for supplier i, i.e., the parts ordered from supplier i are delivered
without disruptions with probability (1 − pi ), or not at all with probability pi .
In addition to independent local disruptions of each supplier, there are potential
global disasters that may result in all suppliers disruption simultaneously. For exam-
ple, such global super events may include economic crisis, widespread labor strike in
a transportation sector, etc. Although the probability of such disaster events usually
is very low, its consequences may be very high. Denote by p∗ the probability of
simultaneous correlated global disruption of all suppliers due to some disaster super
event.
Let Ps be the probability that disruption scenario s is realized, where each scenario
s ∈ S is comprised of a unique subset Is ⊂ I of suppliers who deliver parts without
disruptions, and S = {1, . . . , S} is the index set of all scenarios. There are a total of
S = 2I potential disruption scenarios.
72 4 Selection of Resilient Supply Portfolio

The global disaster and the local disruptive events at each supplier are assumed to
be independent events, therefore the probability Ps of each disruption scenario s ∈ S
under the risks of both type of events is

(1 − p∗ )P̂s  if Is = ∅
Ps = p∗ + (1 − p∗ ) pi if Is = ∅,
i∈I

where P̂s is the probability of disruption scenario s in the presence of independent

local disruptive events only
 
P̂s = (1 − pi ) · pi .
i∈Is i∈I
/s

If the probability of global disruption p∗ = 0, then the probability Ps reduces to P̂s

for independent local disruptive events.
The producer does not need to pay for ordered and undelivered parts. However,
the producer is charged with a much higher cost of unfulfilled customer orders for
products, caused by the shortage of parts, undelivered due to supply disruptions. Let
gj be the per unit cost of shortage of part type j.
In order to mitigate the impact of disruption risks, managers may consider pro-
tective countermeasures to fully fortify suppliers against disruptions and to maintain
the suppliers normal capability. The capacity of a protected supplier is assumed to
remain unchanged under a disruptive event. Therefore, for a protected supplier i its
disruption probability pi is in practice changed to zero.
The protective countermeasure of each supplier may be combined with pre-
positioning of emergency inventory of parts manufactured by the protected sup-
plier. The emergency inventory is used to compensate for the loss of capacity of
the other suppliers, unprotected and hit by disruptions, and to fulfill non-delivered
orders placed on the disrupted suppliers. The emergency inventory pre-positioned at
a protected supplier is limited, e.g., by available budget, by available storage space
or by its base capacity, if the inventory is an overtime production. The inventory is
replenished once it is used. Assume that the inventory quantity is linked to capacity
and cannot be greater than the protected supplier normal capacity. As a result each
protected supplier is capable of supplying twice as many parts as its base capacity.
Let fi be the cost required for full protection of supplier i against disruptions, and
denote by hij the per unit cost of pre-positioning the emergency inventory of part
type j at the protected supplier i.
The decision maker needs to decide which supplier to select for purchasing the
required parts, which of the selected suppliers to protect against disruptive events
and how to allocate order quantity among the selected suppliers and emergency
inventory among the protected suppliers to achieve a minimum total cost of sup-
pliers protection, emergency inventory pre-positioning, parts ordering, purchasing,
transportation, defects and shortage and to mitigate the impact of disruption risks by
minimizing the potential worst-case cost.
4.3 Resilient Supply Portfolio with Single-Level Protection 73

4.3 Resilient Supply Portfolio with Single-Level Protection

In this section three SMIP models are formulated for supplier selection and protec-
tion, and order quantity allocation problem, i.e., for determining a resilient supply
portfolio. For definition of problem variables, see Table 4.2.

Table 4.2 Variables: resilient supply portfolio with single-level protection

First stage variables
qi = 1, if selected supplier i is protected against disruptions; otherwise qi = 0 (supplier
protection variable)
ui = 1, if an order for parts is placed on supplier i; otherwise ui = 0 (supplier selection
variable)
vij = the fraction of demand for parts type j ordered from unprotected supplier i (part
type demand unprotected allocation variable)
wij = the fraction of demand for parts type j ordered from protected supplier i (part type
demand protected allocation variable)
xij = emergency inventory of parts type j pre-positioned at protected supplier i, in frac-
tion of supplier’s capacity ci (inventory allocation variable)
Second stage variables
yijs = emergency inventory of parts type j pre-positioned at protected supplier i, used
under disruption scenario s (inventory usage variable)
zjs = inventory shortage of parts type j under disruption scenario s
Auxiliary variables
VaRc Cost-at-Risk, the targeted cost such that for a given confidence level α, for 100α%
of the scenarios, the outcome is below VaRc
Cs ≥ 0, the tail cost for scenario s, i.e., the amount by which costs in scenario s exceed
VaRc

The resilient supply portfolio is selected ahead of time in such a way as to minimize
the potential average or worst-case cost (or a combination of both) under different
disruption scenarios. The capacity of protected suppliers and the emergency inventory
pre-positioned at the protected suppliers are used to reduce the potential highest costs.
When deciding on a supply portfolio it is assumed that the orders for all parts
are simultaneously placed on selected suppliers, and each protected and each non-
disrupted unprotected supplier delivers all the ordered parts. However, unprotected
suppliers hit by disruptions fail to deliver the ordered parts, and then the non deliv-
ered orders are partially or fully replaced by the emergency inventory of parts pre-
positioned at the protected suppliers. The decision on the protection of selected
suppliers and the pro-positioning of emergency inventory at the protected suppliers
is a part of the resilient supply portfolio decision-making.
74 4 Selection of Resilient Supply Portfolio

4.3.1 Model for Risk-Neutral Decision-Making

In this subsection a SMIP model RSP_E is presented for selection of risk-neutral

resilient supply portfolio. In the risk-neutral decision-making the overall qual-
c
ity of the supply portfolio can  be measured by the expected cost per part, E
(4.1),
  of suppliers protection, i∈I fi qi /D, and emergency inventory pre-positioning,
i∈I j∈J hij ci x ij /a ij D, of parts ordering, i∈I ei ui /D, of parts purchasing and
shipping, including defects, plus replacement delivery of parts pre-positioned, in
which
 the cost of a defective part is assumedto be identical with its price,
i∈I j∈J oij (dj vij /(1 − ρij ) + dj wij /(1 − ρij ) + s∈S Ps ci yijs /aij )/D, and finally
cost
 of shortage
 of parts
 due to supply disruptions less cost of non delivered parts,
s∈S P s ( j∈J g j z js − i∈I
/s j∈J oij dj vij /(1 − ρ ij ))/D.
The purchase orders for parts are assumed to be inflated by the reject rates ρij of
defective parts, i.e., are equal to dj vij /(1 − ρij ) and dj wij /(1 − ρij ), for all i ∈ I, j ∈ J,
respectively for unprotected and protected suppliers.
Note, that if an unprotected supplier is subject to a disruptive event and fails
to deliver the ordered parts, the non delivered parts can be replaced by emergency
inventory pre-positioned at protected suppliers. Then, the amount dj vij /(1 − ρij ) of
part type j ordered from unprotectedand failed supplier i ∈ / Is can be fully or partially
met with the emergency inventory i∈I ci xij /aij . The resilient supply portfolio aims
at reducing potential losses from non delivered parts through the optimal selection of
suppliers for protection, the allocation among them the emergency inventory and the
allocation of orders for parts among both the unprotected and protected suppliers.
The SMIP model RSP_E for selection of risk-neutral resilient supply portfo-
lio is formulated below. In the proposed model, the portfolio will be optimized by
minimizing expected cost per part E c , (4.1).

RSP_E: Selection of risk-neutral Resilient Supply Portfolio to minimize

expected cost
Minimize Expected Cost per Part
  
Ec = ei ui /D + fi qi /D + hij ci xij /aij D
i∈I i∈I i∈I j∈J

+ oij dj (vij + wij )/(1 − ρij )D
i∈I j∈J
   
+ Ps ( gj zjs + oij ci yijs /aij − oij dj vij /(1 − ρij ))/D
s∈S j∈J i∈I j∈J i∈I
/ s j∈J
(4.1)

subject to
4.3 Resilient Supply Portfolio with Single-Level Protection 75

1. Supplier selection and protection constraints:

– only selected suppliers can be protected, and emergency inventory can be
pre-positioned only at protected suppliers,
– the orders for parts can be allocated among selected suppliers only,
– each selected supplier is either protected or unprotected and so are the
corresponding types of orders for parts allocated among the suppliers,

xij ≤ qi ≤ ui ; i ∈ I (4.2)
j∈J
vij ≤ ui ; i ∈ I, j ∈ J (4.3)
vij ≤ 1 − qi ; i ∈ I, j ∈ J (4.4)
wij ≤ qi ; i ∈ I, j ∈ J (4.5)

2. Order quantity and emergency inventory allocation constraints:

– orders for all required parts of each type must be allocated among selected
suppliers,
– for each selected supplier (unprotected and protected), the total capacity
required to manufacture the ordered quantities of parts cannot exceed available
capacity,
– the ordered quantity assigned to each selected supplier (unprotected or
protected) cannot be less than the minimum order size,
– the emergency inventory used to replace non-delivered units cannot
exceed the pre-positioned emergency inventory,
– for each part type and disruption scenario, non delivered orders can (no
inventory shortage, zjs = 0) or cannot (inventory shortage, zjs > 0) be fully
replaced by using the emergency inventory,

(vij + wij ) = 1; j ∈ J (4.6)
i∈I

aij dj vij /(1 − ρij ) ≤ ci (ui − qi ); i ∈ I (4.7)
j∈J

aij dj wij /(1 − ρij ) ≤ ci qi ; i ∈ I (4.8)
j∈J

dj vij /(1 − ρij ) ≥ ν(ui − qi ); i ∈ I (4.9)
j∈J

(dj wij /(1 − ρij ) + ci yijs /aij ) ≥ νqi ; i ∈ I, s ∈ S (4.10)
j∈J
yijs ≤ xij ; i ∈ I, j ∈ J, s ∈ S (4.11)
 
zjs = dj vij /(1 − ρij ) − ci yijs /aij ; j ∈ J, s ∈ S (4.12)
i∈I
/s i∈I
76 4 Selection of Resilient Supply Portfolio

3. Non-negativity and integrality conditions

qi ∈ {0, 1}; i ∈ I (4.13)

ui ∈ {0, 1}; i ∈ I (4.14)
vij ∈ [0, 1]; i ∈ I, j ∈ J (4.15)
wij ∈ [0, 1]; i ∈ I, j ∈ J (4.16)
xij ∈ [0, 1]; i ∈ I, j ∈ J (4.17)
yijs ∈ [0, 1]; i ∈ I, j ∈ J, s ∈ S (4.18)
zjs ≥ 0; j ∈ J, s ∈ S. (4.19)

Denote by
(V1 , . . . , VI ),

the supply portfolio, where i∈I Vi = 1 and 0 ≤ Vi ≤ 1 is the fraction of the total
demand for parts ordered from supplier i. The resilient supply portfolio is based on
demand allocation, does not account for reject rates and is determined by the order
quantity allocation variables vij , wij

Vi = dj (vij + wij )/D; i ∈ I. (4.20)
j∈J

Alternatively, to account for the actual supplies of parts and the reject rates, the
supply portfolio (V1 , . . . , VI ) can be calculated as below.

dj (vij + wij )/(1 − ρij )
Vi
j∈J
=  ; i ∈ I. (4.21)
i∈I j∈J dj (vij + wij )/(1 − ρij )

Note that the nonnegative inventory shortage variables zjs can be eliminated from
the model using Eq. (4.12). Then, variables zjs in the objective function (4.1) should
be replaced by  
dj vij /(1 − ρij ) − ci yijs /aij ,
i∈I
/s i∈I

and equality constraints (4.12), replaced with inequality constraints

 
ci yijs /aij ≤ dj vij /(1 − ρij ); j ∈ J, s ∈ S.
i∈I i∈I
/s

The inventory shortage variables zjs are slack variables for the last constraints.
Similarly, slack variables for inequality constraints (4.11) represent surplus of the
pre-positioned inventory (xij − yijs ) of each part type j at each supplier i, under each
4.3 Resilient Supply Portfolio with Single-Level Protection 77

disruption scenario s ∈ S. Given optimal supply portfolio, the actual usage of emer-
gency inventory depends on the realized disruption scenario and the corresponding
needs for a replacement of non delivered orders from disrupted suppliers. Depending
on the realized disruption scenario, the pre-positioned inventory can be used fully,
partially or not at all.
In the proposed model the required parts of each type are assumed to be partially
provided by one or more suppliers and the order allocation variables vij or wij represent
the fraction of all required parts of type j provided, respectively by unprotected or
protected supplier i. In some practical cases all parts of the same type are purchased
from a single supplier. Then, the corresponding continuous allocation variables vij
and wij should be redefined as binary assignment variables denoting whether or not
all parts of type j are provided, respectively by unprotected or protected supplier i.

4.3.2 Model for Risk-Averse Decision-Making

In the selection of a resilient supply portfolio under disruption risks, the decision
maker controls the risk of high losses due to supply disruptions by choosing the
confidence level α. We assume that the decision maker is willing to accept only
portfolios for which the total probability of scenarios with costs greater than VaRc is
not greater than 1 − α. The greater the confidence level α, the more risk aversive is
the decision maker and the smaller percent of the highest cost outcomes is focused
on. Moreover, a risk aversive decision maker wants to minimize the expected worst-
case costs exceeding VaRc , by minimizing CVaRc . When using CVaRc to minimize
worst-case costs, CVaRc is always not less than VaRc .
Define Cs as the tail cost for scenario s, where tail cost is defined as the amount by
which costs in scenario s exceed VaRc . The portfolio will be optimized by calculating
VaRc and minimizing CVaRc simultaneously. By measuring CVaRc , the magnitude
of the tail costs is considered to achieve a more accurate estimate of the risks of min-
imizing cost. In the proposed model, CVaRc is represented by an auxiliary function
(4.22) introduced by Rockafellar and Uryasev (2000). The SMIP model RSP_CV
for selection of risk-averse resilient supply portfolio to reduce the risk of high costs
is formulated below.

RSP_CV: Selection of risk-averse resilient supply portfolio to minimize

CVaR of cost
Minimize 
CVaRc = VaRc + (1 − α)−1 Ps Cs (4.22)
s∈S

subject to
78 4 Selection of Resilient Supply Portfolio

1. Supplier selection and protection constraints: (4.2)–(4.5)

2. Order quantity and emergency inventory allocation constraints:
(4.6)–(4.12)
3. Risk constraints:
– the tail cost for scenario s is defined as the nonnegative amount by which
cost per part in scenario s exceeds VaRc ,
  
Cs ≥ ei ui /D + fi qi /D + hij ci xij /aij D
i∈I i∈I i∈I j∈J

+ oij dj (vij + wij )/(1 − ρij )D
i∈I j∈J
  
+( gj zjs + oij ci yijs /aij − oij dj vij /(1 − ρij ))/D
j∈J i∈I j∈J i∈I
/ s j∈J
−VaRc ; s ∈ S
(4.23)

4. Non-negativity and integrality conditions: (4.13)–(4.19) and

Cs ≥ 0; s ∈ S. (4.24)

Note that, if for some part type j all required parts must be supplied by a single sup-
plier, then the corresponding nonnegative allocation variables vij , wij , i ∈ I should
be redefined to be binary assignment variables, similarly as for RSP_E model.

4.3.3 Model for Mean-Risk Decision-Making

In the single objective approach the resilient supply portfolio is selected by mini-
mizing either the expected cost per part, E c , (4.1) or the expected worst-case cost
per part, CVaRc , (4.22). In this subsection the two cost functions are considered
simultaneously, and a bi-objective selection of resilient supply portfolio is presented
aimed at minimizing both objective functions to balance expected costs with the
risk tolerance. This type of trade-off model is known as the mean-risk model (e.g.
Ogryczak and Ruszczynski 2002), formulated as the optimization of a composite
objective consisting of the expected cost and the conditional-cost-at-risk as a risk
measure.
The nondominated solution set of the bi-objective resilient supply portfolio can be
found by the parameterization on λ the weighted-sum program RSP_ECV presented
below. The scalarizing SMIP model is based on RSP_CV model with the addition
of objective (4.1) of model RSP_E.
4.3 Resilient Supply Portfolio with Single-Level Protection 79

RSP_ECV: Selection of mean-risk resilient supply portfolio to minimize

weighted sum of expected cost and CVaR of cost
Minimize
λE c + (1 − λ)CVaRc (4.25)

where 0 ≤ λ ≤ 1,
subject to (4.1)–(4.19), (4.22)–(4.24).

4.4 Protection Index

There are many alternative countermeasures that decision maker must consider to
manage the risk of supply disruptions. For each protective countermeasure being
considered, it is important to determine the cost of the countermeasure and the esti-
mated loss to be incurred if a supply disruption occurs. For example, if a protective
wall is built to protect a supplier against flooding, the decision maker must consider
the construction cost against the estimated loss caused by potential flooding that may
disrupt supplies of materials, if no protective wall is build. The trade-off between
the cost of suppliers protection against potential disruptions and the losses caused by
supply disruptions can be evaluated, for example, by the following unit protection
cost

fi
ϕ =  i∈I . (4.26)
j∈J gj dj

The unit protection cost is the ratio of total protection cost to total loss resulted from
the shortage of parts caused by the potential supply disruptions. Given demand for
parts, the unit protection cost, ϕ, increases with suppliers protection costs, fi , i ∈ I,
and decreases with parts shortage costs, gj , j ∈ J.
To ensure delivery of all required parts under disruption risks it is not necessary
to protect all suppliers against disruptions and only a subset of suppliers needs to be
protected. As a result the actual value of unit protection cost can be less than (4.26),
and its optimal value ϕ o can be found as a solution to the following mixed integer
program.

fi qi 
ϕ = min{  i∈I
o
: aij dj wij /(1 − ρij ) ≤ 2ci qi ; i ∈ I,
j∈J gj dj j∈J

wij = 1; j ∈ J, wij ≤ qi ; i ∈ I, j ∈ J,
i∈I
qi ∈ {0, 1}; i ∈ I, wij ∈ [0, 1]; i ∈ I, j ∈ J}. (4.27)
80 4 Selection of Resilient Supply Portfolio

The optimal value ϕ o is calculated under global disruption scenario, assuming no

supplies from unprotected suppliers and a full level of emergency inventory pre-
positioned at each protected supplier. As a result a double capacity is available at
each protected supplier and no capacity at each unprotected supplier.
The unit protection cost, however, does not involve any information on the esti-
mated probability of potential disruptive events. It is intuitively believed, that the
higher the probability of a disruptive event and the higher the potential losses, the
higher the protection cost, the decision maker is willing to cover. In addition to
the above unit protection cost, ϕ, the following protection index Φ is introduced in
which the protection cost of each supplier i, fi , is additionally divided by its disruption
probability, (p∗ + (1 − p∗ )pi )
 ∗
i∈I fi /(p + (1 − p∗ )pi )
Φ=  . (4.28)
j∈J gj dj

The protection index is the ratio of total weighted (multiplied by 1/(p∗ + (1 − p∗ )pi ))
protection cost to total loss resulted from the shortage of parts caused by supply
disruptions. Given demand for parts, the protection index, Φ, increases with supplier
protection cost per disruption probability, fi /(p∗ + (1 − p∗ )pi ), i ∈ I, and decreases
with parts shortage costs, gj , j ∈ J.
The optimal value Φ o of protection index can be found as a solution to the fol-
lowing mixed integer program, in which no supplies from unprotected suppliers and
a full level of emergency inventory pre-positioned at each protected supplier are
assumed.

(fi /(p∗ + (1 − p∗ )pi ))qi
Φ = min{ i∈I
o
 :
j∈J gj dj

aij dj wij /(1 − ρij ) ≤ 2ci qi ; i ∈ I,
j∈J

wij = 1; j ∈ J, wij ≤ qi ; i ∈ I, j ∈ J,
i∈I
qi ∈ {0, 1}; i ∈ I, wij ∈ [0, 1]; i ∈ I, j ∈ J}. (4.29)

Both the unit protection cost and the protection index can be used for a rough
evaluation of different protection strategies.

4.5 Resilient Supply Portfolio with Multi-Level Protection

One of the main assumptions in the models presented in Sect. 4.3 is that the capac-
ity of a fortified supplier remains unchanged after any disruptive event and hence
the impacted supplier is capable of full delivering all ordered parts. However, this
assumption may not be always realistic in practice. For example, fortifying a supplier
4.5 Resilient Supply Portfolio with Multi-Level Protection 81

against an earthquake may not protect it sufficiently well against a much stronger
earthquake or a combined earthquake and flooding, etc., and hence the capacity of
an impacted supplier can be actually reduced.
In this section, an enhancement of proposed models is considered assuming that
the impact of disruptive events on suppliers capacity may not be fully mitigated by
the fortification of a supplier, based on the level of protection investments. In the
models presented below, the suppliers can be fortified at different discrete levels of
protection, each of which has its own fortification cost and the associated capacity
available after a disruptive event. The remaining capacity available depends on the
protection level applied. The higher the protection level the higher the remaining
capacity in the aftermath of a disruptive event. However, the maximum amount of
emergency inventory pre-positioned at a fortified supplier is bounded by its full
capacity, independent of the protection level.
The new parameters and variables are introduced in Table 4.3. Denote by l ∈ Li =
{1, . . . , Li }, the protection level of supplier i, which refers to the fortification cost
and the fraction of supplier’s full capacity available after occurrence of a disruptive
event. Level l = Li represents the highest available protection for supplier i against
disruptions and refers to the highest fortification cost and the highest fraction of
remaining capacity. Let fˆil and γil be the fortification cost and the remaining fraction
of full capacity for supplier i protected at level l, where

0 < fˆi1 < fˆi2 < · · · < fˆiLi

and
0 < γi1 < γi2 < · · · < γiLi ≤ 1.

Table 4.3 Notation: multi-level protection

Indices
l = protection level, l ∈ Li , i ∈ I
Parameters
fˆil = fortification cost for supplier i protected at level l
γil = the fraction of full capacity of supplier i protected at level l, available under disruptive event
First stage variables
Qil = 1, if selected supplier i is protected at level l; otherwise (supplier protection level variable)
Wijl = the fraction of demand for parts type j ordered from supplier i protected at level l (part type
demand allocation variable)
82 4 Selection of Resilient Supply Portfolio

A supplier i protected at level l delivers fraction γil of all ordered parts, under a
disruptive event. The non delivered parts can be replaced by the emergency  inven-
tory pre-positioned at protected suppliers. The non delivered amount, i∈I dj (1 −
γil )Wijl /(1
− ρij ), of each part type j can be fully or partially met with the emergency
inventory i∈I ci xij /aij .
The enhancement RSP(mlp)_E of model RSP_E is shown below.

RSP(mlp)_E: Selection of risk-neutral resilient supply portfolio with

multi-level protection
Minimize
  
Ê c = ei ui /D + fˆil Qil /D + hij ci xij /aij D
i∈I i∈I l∈Li i∈I j∈J
 
+ oij dj (vij + Wijl )/(1 − ρij )D
i∈I j∈J l∈Li
   
+ Ps ( gj zjs + oij ci yijs /aij − oij dj vij /(1 − ρij )
s∈S j∈J i∈I j∈J i∈I
/ s j∈J

− oij dj (1 − γil )Wijl /(1 − ρij ))/D
i∈I
/ s j∈J l∈Li
(4.30)

subject to (4.3), (4.11), (4.14), (4.15), (4.17)–(4.19) and

 
xij ≤ Qil ≤ ui ; i ∈ I (4.31)
j∈J l∈Li

vij ≤ 1 − Qil ; i ∈ I, j ∈ J (4.32)
l∈Li
Wijl ≤ Qil ; i ∈ I, j ∈ J, l ∈ Li (4.33)
 
(vij + Wijl ) = 1; j ∈ J (4.34)
i∈I l∈Li
 
aij dj vij /(1 − ρij ) ≤ ci (ui − Qil ); i ∈ I (4.35)
j∈J l∈Li

aij dj Wijl /(1 − ρij ) ≤ ci Qil ; i ∈ I, l ∈ Li (4.36)
j∈J
 
dj vij /(1 − ρij ) ≥ ν(ui − Qil ); i ∈ I (4.37)
j∈J l∈Li
 
(dj Wijl /(1 − ρij ) + ci yijs /aij ) ≥ ν Qil ; i ∈ I, s ∈ S (4.38)
j∈J l∈Li l∈Li
4.5 Resilient Supply Portfolio with Multi-Level Protection 83

 
zjs = dj vij /(1 − ρij ) + dj (1 − γil )Wijl /(1 − ρij )
i∈I
/s i∈I
/ s l∈Li

− ci yijs /aij ; j ∈ J, s ∈ S (4.39)
i∈I
Qil ∈ {0, 1}; i ∈ I, l ∈ Li (4.40)
Wijl ∈ [0, 1]; i ∈ I, j ∈ J, l ∈ Li . (4.41)

 
In the objective function (4.30), i∈I l∈Li fˆil Qil , is the 
total fortification
  cost
of protected suppliers, and the last subtracted term, s∈S Ps i∈I /s j∈J l∈Li ij dj
o
(1 − γil )Wijl /(1 − ρij ), is the expected cost of non-delivered parts by protected sup-
pliers.
In model RSP(mlp)_E, constraints (4.31) and (4.33) ensure that each supplier
can be protected at most at one level, and the allocation of demand for parts among
the fortified suppliers accounts for their protection levels. Since each supplier i can
be protected at most at one level l, (4.31), at most one variable Wijl , (4.33), may
take on a positive value for each part type j. Equation (4.39) defines shortage of each
part type inventory under each disruptive event, due to undelivered parts by both
unprotected and protected suppliers. Equation (4.39) is equivalent to (4.12) in model
RSP_E, where a fortified supplier was capable of full delivering of ordered parts
under disruptive event.
Notice that model RSP(mlp)_E can be derived from RSP_E by replacing vari-
ables qi , i ∈ I and wij , i ∈ I, j ∈   levels l ∈ Li of the
J by sums over all protection
new variables Qil and Wijl : qi = l∈Li Qil ; i ∈ I and wij = l∈Li Wijl ; i ∈ I, j ∈ J.
In addition, the objective function (4.1) and constraint (4.12) have been modified to
account for undelivered parts by protected suppliers, and Eqs. (4.5) and (4.8) rede-
fined for each protection level.
The enhancements, RSP(mlp)_CV and RSP(mlp)_ECV, respectively of risk-
averse model RSP_CV and mean-risk model RSP_ECV are presented below.

RSP(mlp)_CV: Selection of risk-averse resilient supply portfolio with

multi-level protection
Minimize (4.22)
subject to (4.3), (4.11), (4.14), (4.15), (4.17)–(4.19), (4.22), (4.31)–(4.41)
and

  
Cs ≥ ei ui /D + fˆil Qil /D + hij ci xij /aij D
i∈I i∈I l∈Li i∈I j∈J
 
+ oij dj (vij + Wijl )/(1 − ρij )D
i∈I j∈J l∈Li
84 4 Selection of Resilient Supply Portfolio

  
+( gj zjs + oij ci yijs /aij − oij dj vij /(1 − ρij )
j∈J i∈I j∈J i∈I
/ s j∈J

− oij dj (1 − γil )Wijl /(1 − ρij ))/D − VaRc ; s ∈ S. (4.42)
i∈I
/ s j∈J l∈Li

RSP(mlp)_ECV: Selection of mean-risk resilient supply portfolio with

multi-level protection
Minimize
λÊ c + (1 − λ)CVaRc (4.43)

where 0 ≤ λ ≤ 1,
subject to (4.3), (4.11), (4.14), (4.15), (4.17)–(4.19), (4.22), (4.24),
(4.30)–(4.41).

4.6 Computational Examples

In this section some computational examples are presented to illustrate possible

applications of the proposed SMIP approach for the selection and protection of sup-
pliers and order quantity allocation in a supply chain under disruption risks. First, the
examples with single protection levels are presented and then, for comparison, appli-
cations of models for multi-level protection are illustrated. The following parameters
have been used for the example problems:
• I, the number of suppliers, was equal to 10 and the corresponding number S = 2I
of disruption scenarios, was equal to 1024;
• J, the number of part types, was equal to 25;
• aij , the unit capacity consumptions were integers in {1, 2, 3} drawn from int(U[1;3])
distribution, for all suppliers i and part types j;
• ci , the capacity
 of each supplier i, was integer drawn from
1000
( j∈J aij dj /I)U[0.75; 1.25]/1000 distribution (
· denotes the smallest
integer not less than ·);
• dj , the required numbers of parts, were integers in {5000, 10000, 15000,
20000, 25000, 30000, 35000, 40000, 45000, 50000} drawn from 5000int(U[1;
10]) distribution, for all part types j;
• ei , the cost of ordering parts, were integers in {500, 600, 700, 800, 900, 1000}
drawn from 100int(U[5;10]) distribution, for all suppliers i;
4.6 Computational Examples 85

• fi , the protection costs, were integers in {500000, 550000, 600000, 650000,

700000, 750000, 800000, 850000, 900000} drawn from 50000int(U[10;18]) dis-
tribution, for all suppliers i;
• gj , the unit shortage costs were integers in {70, 80, 90, 100} drawn from
10int(U[7;10]) distribution, for all part types j;
• hij , the unit cost of emergency inventory of part type j pre-positioned at supplier
i, was equal to 0.4oij for all suppliers i and part types j;
• oij , the unit price (including shipping cost) of part type j purchased and transported
from each supplier i, was uniformly distributed over [10,15] (i.e. drawn from
U[10;15]) and reduced by the factor (1 − ρij ) to get a lower price for parts from
the suppliers with a higher defect rate;
• ρij , the expected defect rate of each supplier i for each part type j, was exponentially
distributed, ranging from 0.0003 to 0.03;
• ν, the minimum order size, was equal to 500;
• pi , the local disruption probability was uniformly distributed over [0,0.06] or
over [0.06,0.15], i.e., the disruption probabilities were drawn independently from
U[0;0.06] or from U[0.06;0.15]. The global disruption probability p∗ , was either
0.001 or 0.01, respectively;
• α, the confidence level, was equal to 0.50, 0.75, 0.90, 0.95 or 0.99.
The computational experiments were performed for the same replication  of the
above input data set. The resulting total demand for parts was D = j∈J dj =
740, 000, for all test examples. The two different combinations of global and local dis-
ruption probabilities were considered: p∗ = 0.001 and pi ∈ U[0; 0.06], i ∈ I for reli-
able suppliers and p∗ = 0.01 and pi ∈ U[0.06; 0.15], i ∈ I for unreliable suppliers.

4.6.1 Single-Level Protection

The solution results are presented in Tables 4.4 and 4.5, respectively for the risk-
neutral model RSP_E and the risk-averse model RSP_CV with different confidence
levels. The confidence level α is set at five levels of 0.5, 0.75, 0.90, 0.95, and 0.99,
which means that focus is on minimizing the highest 50%, 25%, 10%, 5%, and 1% of
all scenario outcomes, i.e., costs per part. The size of the corresponding mixed integer
programs is represented by the total number of variables, Var., number of binary
variables, Bin., number of constraints, Cons, and number of nonzero coefficients
in the constraint matrix, Nonz. Table 4.5 also presents the probability 1 − F(VaRc )
of outcomes with worst-case cost above VaRc as well as the expected cost for the
optimal risk-averse supply portfolio.
86 4 Selection of Resilient Supply Portfolio

Table 4.4 Solution results for model RSP_E

Disruption probability p∗ = 0.001 pi ∈ [0, 0.06], i ∈ I p∗ = 0.01, pi ∈ [0.06, 0.15], i ∈ I

Var. = 282105, Bin. = 20, Cons. = 282151, Nonz. = 924155 †
Expected Cost 12.63 14.97
No.of Suppliers Selected 9 10
(incl. Suppliers Protected) (0) (2)
Pre-Positioned Emergency Inventory 0 141 888
CPU‡ 5 3243
† Var. = number of variables, Bin. = number of binary variables,
Cons. = number of constraints, Nonz. = number of nonzero coefficients.
‡ CPU seconds for proving optimality on a MacBookPro6.2, Intel Core i7, 2.66 GHz,
RAM 8 GB/CPLEX 12.4.

The optimal risk-neutral supply portfolio, (V1 , . . . , VI ), (4.20), for model RSP_E
that aims at minimization of total expected cost per part is shown in Fig. 4.1 for two
scenarios:
(a) with global disruption probability p∗ = 0.001 and reliable suppliers with local
disruption probabilities pi ∈ [0, 0.06], i ∈ I, and
(b) with global disruption probability p∗ = 0.01 and unreliable suppliers with local
disruption probabilities pi ∈ [0.06, 0.15], i ∈ I.
In addition, total disruption probability p∗ + (1 − p∗ )pi for each supplier i is also
presented. The optimal supply portfolio for scenario with reliable suppliers, allocates
the total demand for parts among nine unprotected suppliers with the lowest disrup-
tion probabilities, except for supplier 4 with the highest disruption probability. The
lower disruption probability of a supplier, the higher percentage of the total demand
allocated. The above observations indicate that when the shortage cost of parts dom-
inates the purchasing cost, disruption probability becomes a key determinant in the
decision of demand allocation among the suppliers to minimize total expected cost.
Similar results are observed for scenario with unreliable suppliers. However, the
total demand is now allocated among all ten suppliers, with the largest orders placed
on protected suppliers 1 and 10, where the emergency inventory was pre-positioned
(see, Table 4.4).
The optimal risk-averse supply portfolios for RSP_CV model and the three confi-
dence levels: 0.90, 0.95, and 0.99 are shown in Fig. 4.2. Table 4.5 and Fig. 4.2 indicate
that when α increases and a more risk-averse decision-making focuses on a smaller
set of outcomes, the number of protected suppliers also increases to mitigate the
impact of disruption risks. Simultaneously, the pre-positioned emergency inventory
at the protected suppliers increases, while the total number of all selected suppli-
ers decreases and more orders are placed on the protected suppliers. This indicates
that instead of further diversification of supplies by selecting of more suppliers, the
impact of disruption risks is rather mitigated by selecting of less suppliers and by
protecting most of the selected suppliers and, in addition, by the pre-positioning of
emergency inventory at the protected suppliers.
4.6 Computational Examples 87

(a)

(b)

Fig. 4.1 Optimal supply portfolio for model RSP_E: a p∗ = 0.001, pi ∈ [0, 0.06], i ∈ I, b p∗ =
0.01, pi ∈ [0.06, 0.15], i ∈ I (P - protected supplier)

Unless protected, the suppliers with the highest disruption probability are rarely
selected. For example (see, Fig. 4.2), in the optimal supply portfolio for scenario
with reliable suppliers and α = 0.99 or for scenario with unreliable suppliers with
α = 0.9, 0.95, 0.99, supplier 4 (cf. Fig. 4.1) is not selected, while supplier 10 is
selected and protected.
The discrete distributions of cost per part for the optimal risk-averse supply
portfolios (Fig. 4.2a) with three different confidence levels, for global disruption
probability p∗ = 0.001 and reliable suppliers with local disruption probabilities
pi ∈ [0, 0.06], i ∈ I are presented in Table 4.6. The table demonstrates that the proba-
bility mass function is concentrated in a few points, which is typical for the scenario-
based optimization under uncertainty, where the probability measure is concentrated
in finitely many points. Large probability atoms are concentrated at eight points, six
points, one point, respectively for the confidence level α = 0.9, 0.95, 0.99. In particu-
lar, for α = 0.99 the whole probability is concentrated at one point only: at cost 16.00.
88 4 Selection of Resilient Supply Portfolio

Table 4.5 Solutions results for model RSP_CV

Confidence level α 0.50 0.75 0.90 0.95 0.99
Var. = 283130, Bin. = 20, Cons. = 283175, Nonz. = 1 868 008 †
p∗ = 0.001, pi ∈ U[0; 0.06]
CVaR c 13.43 14.02 14.68 15.21 16.00
VaRc 12.78 12.93 14.01 14.36 16.00
Ec 13.11 13.20 14.08 14.35 16.00
1 − F(VaRc ) 0.151 0.151 0.033 0.025 0.000
No. of Suppliers Selected (incl. Protected) 10(1) 10(1) 10(2) 10(2) 8(4)
Pre-Positioned Emergency Inventory 70 139 65 652 128 192 133 349 193 950
CPU‡ 1037 3079 4770 16752 1852
p∗ = 0.01, pi ∈ U[0.06; 0.15]
CVaRc 15.55 15.89 16.00 16.00 16.00
VaRc 15.02 15.44 16.00 16.00 16.00
Ec 15.28 15.54 16.00 16.00 16.00
1 − F(VaRc ) 0.302 0.067 0.000 0.000 0.000
No. of Suppliers Selected (incl. Protected) 10(3) 10(3) 8(4) 8(4) 8(4)
Pre-Positioned Emergency Inventory 167 357 213 402 193 950 193 950 193 950
CPU‡ 23931 28856 2354 4030 4357
† Var. = number of variables, Bin. = number of binary variables,

Cons. = number of constraints, Nonz. = number of nonzero coefficients.

VaRc = CVaRc = E c = 16.00 is the highest cost that may occur

for the optimal supply portfolio, 1-F(16) = 0.

‡ CPU seconds for proving optimality on a MacBookPro 6.2, Intel Core i7, 2.66 GHz,

RAM 8 GB/CPLEX 12.4.

Similar results were obtained for the global disruption probability p∗ = 0.01 and the
unreliable suppliers with local disruption probabilities pi ∈ [0.06, 0.15], i ∈ I. For
the optimal resilient supply portfolio (Fig. 4.2b), the whole probability is concen-
trated at a single cost level 16.00 for each confidence level α = 0.9, 0.95, 0.99, i.e.,
the cost distributions were identically shaped by the optimal resilient supply portfolio
for the three different confidence levels. The optimal resilient supply portfolios for
scenario with reliable suppliers with α = 0.99 as well as for scenario with unreliable
suppliers with α = 0.9, 0.95, 0.99, are identical (see, Table 4.5 and Fig. 4.2).
The computational results indicate that the smaller is the number of concentration
points and the greater are probability atoms concentrated at those points, the greater
can be the positive difference F(VaRc ) − α, i.e., the smaller than 1 − α can be the
probability 1 − F(VaRc ) of outcomes with worst-case cost higher than VaRc . For
example (see, Table 4.5), for α = 0.9, VaRc =14.01 and 1 − F(VaRc ) = 0.033 < 1 −
α = 0.1, which indicates a high concentration of probability measure at point 14.01
for the optimal supply portfolio. Actually, the probability that cost per part is 14.01
is 0.967, which indicates that VaRc = 14.01 is the lowest cost that may occur and
4.6 Computational Examples 89

(a)

(b)

Fig. 4.2 Optimal supply portfolio for model RSP_CV: a p∗ = 0.001, pi ∈ [0, 0.06], i ∈ I,
b p∗ = 0.01, pi ∈ [0.06, 0.15], i ∈ I

that for the confidence level α = 0.9, less than 3.3% of the cost outcomes are above
VaRc .
Moreover, if the highest cost probability is greater than 1 − α, then CVaRc and
VaRc are identical and both equal to the highest cost. In the example for scenario with
reliable suppliers and α = 0.99 (see, Table 4.6), the highest cost per part is 16.00 and
the probability concentrated at 16.00 is 0.999 > 1 − α = 0.01, then VaRc = 16.00
is the highest cost per part that may occur (1 − F(VaRc ) = 0.000, in Table 4.5) and
hence CVaRc = VaRc = 16.00. Similar results with CVaRc = VaRc = 16.00 were
obtained for the optimal supply portfolios with α = 0.9, 0.95, 0.99, for scenarios
with unreliable suppliers with global and local disruption probabilities, respectively
p∗ = 0.01, pi ∈ [0.06, 0.15], i ∈ I, (see, Table 4.5). The results presented in Table 4.5
indicate that when for the optimal supply portfolio with a confidence level α, the
90 4 Selection of Resilient Supply Portfolio

Table 4.6 Probability of cost Cost interval α = 0.90 α = 0.95 α = 0.99

per part for optimal supply
portfolios: p∗ = 0.001, [14, 15) 0.9899 0.9950 0
pi ∈ [0, 0.06], i ∈ I [15, 16) 0.0032 0.0014 1
[16, 17) 0.0040 0.0020 0
[17, 18) 0.0005 0.0003 0
[18, 19) 0.0011 0.0003 0
[19, 20) 0.0002 0 0
[20, 21) 0.0001 0 0
[47, 48) 0 0.0010 0
[53, 54) 0.0010 0 0

probability measure is concentrated at the highest cost and hence is greater than
1 − α, so that CVaRc and VaRc are identical with the highest cost, then the number
of selected suppliers is smaller than for a lower confidence level, however more
suppliers are protected.
Table 4.5, also demonstrates that the higher the protection index Φ (4.28) (or its
optimal value Φ o (4.29)), the less the number of protected suppliers in the optimal
supply portfolio. For example, for scenario with reliable suppliers (Φ = 8.353 and
Φ o = 1.033) only a few of the selected suppliers are protected and only for higher
confidence levels α, whereas for scenario with unreliable suppliers (Φ = 1.141 and
Φ o = 0.295) more of the selected suppliers are protected, even for lower confidence
levels.

Table 4.7 Nondominated solutions for mean-risk model RSP_ECV:α = 0.99, p∗ = 0.001, pi ∈
[0, 0.06], i ∈ I
λ 0.01 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 0.99
p∗ = 0.001, pi ∈ U[0; 0.06], i ∈ I (ϕ o = 0.032, Φ o = 1.033)
Ec 16.00 15.98 15.43 15.03 14.39 14.36 13.15 12.63
CVaRc 16.00 16.00 16.44 16.89 18.29 18.40 25.71 33.70
VaRc 16.00 15.99 15.42 15.07 14.55 14.52 19.08 23.77
No. of Suppliers Selected 8 8 10 10 10 10 10 9
(incl. Protected) (4) (4) (3) (3) (2) (2) (1) (0)
Emergency Inventory 193950 193997 218988 177867 144381 141565 70594 0
CPU‡ 5186 4200 1531 1659 4720 6902 7961 10345 8250 2997 217
12l‡ CPU seconds for proving optimality on a MacBookPro 6.2, Intel Core i7, 2.66 GHz,
12lRAM 8 GB/CPLEX 12.4.
4.6 Computational Examples 91

Fig. 4.3 Pareto front for model RSP_ECV: α = 0.99, p∗ = 0.001, pi ∈ [0, 0.06], i ∈ I

For the bi-objective mean-risk approach, the subsets of nondominated solutions

were computed by parameterization on λ ∈ {0.01, 0.10, 0.20, 0.30, 0.40, 0.50, 0.60,
0.70, 0.80, 0.90, 0.99} the weighted-sum program RSP_ECV. The results obtained
for the confidence level α = 0.99 and reliable suppliers with global and local dis-
ruption probability, respectively p∗ = 0.001 and pi ∈ [0, 0.06], i ∈ I, are presented
in Table 4.7. The subsets of nondominated solutions that were found for the selected
11 levels of weight λ consists of eight different nondominated solutions. The trade-
off between the expected cost and the expected worst-case cost is clearly shown
in Fig. 4.3, where the convex efficient front of Mean Cost - CVaR, for the con-
fidence level α = 0.99 is presented. The results emphasize the effect of varying
average/worst-case cost preference of the decision maker; the lower the trade-off
parameter λ, the more risk-oriented the decision making. Note that solutions to single
objective models RSP_E and RSP_CV are equivalent to the nondominated solutions
of the weighted-sum program RSP_ECV with λ = 1 and λ = 0, respectively.
Similar computations for the confidence level α = 0.99 and unreliable sup-
pliers with global and local disruption probability, respectively p∗ = 0.01 and
pi ∈ [0.06, 0.15], i ∈ I, produce the three nondominated solutions only, (E c , CVaRc ):
(16.00,16,00) for λ ∈ {0.01, 0.10, 0.20, 0.30, 0.40}, (15.98,16.01) for λ ∈ {0.50,
0.60, 0.70, 0.80, 0.90} and (14.97,47.25) for λ = 0.99. For the first two solutions
eight suppliers were selected, including two protected, and for the last solution ten
suppliers were selected of them two were protected. Figure 4.4 presents the supply
portfolio for the last nondominated solution as well as the corresponding optimal
cost distribution, which indicates a large probability atom of 0.01 concentrated at the
highest cost of 47.25. Note that the risk-averse supply portfolio for λ = 0.99 is an
average performance-oriented portfolio (cf., objective function (4.25)) and virtually
neglects the risk of high costs (0.01 weight of CVaRc in (4.25)). Therefore, it is very
close to the risk-neutral portfolio presented in Fig. 4.1b. The cost distribution shaped
with a risky supply portfolio that focuses on an average performance of supply chain
more often includes probability atoms at high costs.
92 4 Selection of Resilient Supply Portfolio

Fig. 4.4 Optimal supply portfolio and probability mass function for model RSP_ECV: p∗ = 0.01,
pi ∈ [0.06, 0.15], i ∈ I, α = 0.99, λ = 0.99

Comparison of risk-neutral and risk-averse supply portfolios indicates that most

risk-averse portfolios include protected suppliers and hence, can be resilient. In
contrast, the risk-neutral model that focuses on the expected cost only, rarely selects
a resilient portfolio with protected suppliers. Moreover, the mean-risk model that
minimizes the weighted sum of the expected cost and the expected worst-case cost
produces a subset of nondominated supply portfolios that contains both protected
and unprotected suppliers.
The computational experiments with the single-level protection models indicate
that:

• when the shortage cost of parts dominates the purchasing cost, probability of
disrupting a supply is a key determinant in the selection of risk-neutral supply
portfolio.
• a particular supplier is selected more on the supplier non-disruption likelihood
than on the other factors such as purchasing cost or defect rate,
4.6 Computational Examples 93

• the suppliers associated with the highest disruption rates, unless protected, are
rarely selected.
• in the risk-averse decision-making, the number of protected suppliers increases
with the confidence level α, and simultaneously the number of all selected suppliers
decreases,
• in the risk-averse decision-making instead of diversification of supplies by select-
ing of more suppliers, the impact of disruption risks is rather mitigated by selecting
of less suppliers and by protecting most of the selected suppliers and, in addition,
by the pre-positioning of emergency inventory at the protected suppliers.

The computational experiments were performed using the AMPL programming

language and the CPLEX 12.4 solver (with the default traditional branch-and-cut set-
tings) on a laptop MacBookPro 6.2 with Intel Core i7 processor running at 2.66 GHz
and with 8GB RAM. The CPLEX solver was capable of finding proven optimal
solutions for all examples with CPU time ranging from several to several thousands
seconds.
In most cases the CPU time required to find proven optimal solutions for scenario
with unreliable suppliers (p∗ = 0.01 and pi ∈ [0.06, 0.15], i ∈ I) was greater than
that required for scenario with reliable suppliers (p∗ = 0.001 and pi ∈ [0, 0.06], i ∈
I). For a higher disruption probability, the protection of suppliers and the pre-
positioning and usage of emergency inventory occur more frequently. For example,
the optimal risk-neutral solution with the protected suppliers and the pre-positioned
emergency inventory, required much longer CPU time than the low probability case
with no protected suppliers and no emergency inventory (see, Table 4.4).

4.6.2 Multi-level Protection

This subsection presents computational examples to illustrate the SMIP approach to

selection of resilient supply portfolio with multi-level protection of fortified suppliers.
For comparison, the basic input data used are the same as for the examples with
single protection levels. The local disruption probabilities of suppliers, pi , were
distributed exactly the same as for the examples with single protection levels, however
slightly different values were generated for the examples in this subsection.
 Figure 4.5
shows basic characteristics of each supplier i: average price per part, j∈J oij /J, and
disruption probability, p∗ + (1 − p∗ )pi , for two scenarios:
(a) with global disruption probability p∗ = 0.001 and reliable suppliers with local
disruption probabilities pi ∈ [0, 0.06], i ∈ I, and
(b) with global disruption probability p∗ = 0.01 and unreliable suppliers with local
disruption probabilities pi ∈ [0.06, 0.15], i ∈ I.
94 4 Selection of Resilient Supply Portfolio

Fig. 4.5 Suppliers

Table 4.8 Solution results for model RSP(mlp)_E

Disruption probability p∗ = 0.001, pi ∈ [0, 0.06], i ∈ I p∗ = 0.01, pi ∈ [0.06, 0.15], i ∈ I
Var. = 282875, Bin. = 50, Cons. = 282920, Nonz. = 1439505 †
Expected Cost 12.79 14.53
No.of Suppliers Selected 10 10
Protected Suppliers(protection level) 9(1) 1(1), 6(1), 9(1)
Pre-Positioned Emergency Inventory 85030 261 909
CPU‡ 17 1044
† Var. = number of variables, Bin. = number of binary variables,
Cons. = number of constraints, Nonz. = number of nonzero coefficients.
‡ CPU seconds for proving optimality on a MacBookPro, Intel Core i7, 2.8 GHz,

RAM 16 GB/Gurobi 7

The only new parameters, the number of protection levels, L i , the fortification
cost, fˆil , and the remaining fraction of capacity, γil , for supplier i protected at level l,
are shown below.
L i = 4 for all suppliers i ∈ I.
fˆi1 = fi /4, fˆi2 = fi /2, fˆi3 = 3fi /4, fˆi4 = fi for all suppliers i ∈ I, where fi is the
fortification cost for the case with single protection levels.
γi1 = 0.2, γi1 = 0.5, γi3 = 0.7, γi4 = 0.9 for all suppliers i ∈ I.
Thus, the fortification cost for the highest protection level, l = 4, is identical with
the fortification cost, fi , for the examples with single protection levels. However,
unlike for single protection levels, where 100% of protected supplier’s capacity is
available after disruption, now only 90% is the highest fraction of remaining capacity.
4.6 Computational Examples 95

Table 4.9 Solutions results for model RSP(mlp)_CV

Confidence level α 0.50 0.75 0.90 0.95 0.99
Var. = 283900, Bin. = 50, Cons. = 283944, Nonz. = 3 182 078 †
p∗ = 0.001, pi ∈ U[0; 0.06]
CVaR c 13.32 13.77 14.45 14.93 16.17
VaRc 12.34 13.00 13.80 13.90 16.15
Ec 12.83 13.19 13.87 13.94 16.12
1 − F(VaRc ) 0.281 0.162 0.029 0.019 0.006
No. of Suppliers Selected 10 10 9 10 10
Protected Suppliers(protection level) 1(1) 1(1) 1(1) 1(2)
2(1)
4(1)
6(1) 6(1)
7(1)
9(1) 9(1) 9(1) 9(1) 9(2)
10(1)
Pre-Positioned Emergency Inventory 91574 173 542 259 787 264 526 551 749
CPU‡ 565 1271 6304 4230 2190
p∗ = 0.01, pi ∈ U[0.06; 0.15]
CVaRc 15.21 15.99 16.17 16.17 16.18
VaRc 14.34 14.89 16.16 16.18 16.18
Ec 14.77 15.16 16.14 16.16 16.18
1 − F(VaRc ) 0.222 0.134 0.047 0.005 0.000
No. of Suppliers Selected 10 9 9 10 8
Protected Suppliers(protection level) 1(1) 1(1) 1(2) 1(2) 1(2)
2(1) 2(1) 2(1)
4(1) 4(1) 4(1)
6(1) 6(1)
7(1) 7(1) 7(1)
9(1) 9(2) 9(2) 9(2) 9(2)
10(1) 10(1)
Pre-Positioned Emergency Inventory 336 809 377 519 543 054 541 012 538 451
CPU‡ 6865 11133 1598 4084 6035
† Var. = number of variables, Bin. = number of binary variables,
Cons. = number of constraints, Nonz. = number of nonzero coefficients.
VaRc = CVaRc = E c = 16.18 is the highest cost that may occur

for the optimal supply portfolio, 1-F(16.18) = 0.

‡ CPU seconds for proving optimality on a MacBookPro, Intel Core i7, 2.8 GHz,

RAM 16 GB/Gurobi 7

The solution results are presented in Tables 4.8 and 4.9, respectively for the risk-
neutral model RSP(mlp)_E and the risk-averse model RSP(mlp)_CV with different
confidence levels. The tables show results for two scenarios: (a) with reliable sup-
pliers and (b) with unreliable suppliers.
96 4 Selection of Resilient Supply Portfolio

Fig. 4.6 Optimal supply portfolio for model RSP(mlp)_CV: a p∗ = 0.001, pi ∈ [0, 0.06], i ∈ I,
b p∗ = 0.01, pi ∈ [0.06, 0.15], i ∈ I

In general, risk-neutral supply portfolios for single- and multi-level protection are
similar, (cf. Tables 4.4 and 4.8), as well as the corresponding risk-averse portfolios,
(cf. Tables 4.5 and 4.9). For both the risk-neutral and the risk averse portfolio, the
number of protected suppliers is greater for scenario (b) with unreliable suppliers,
while it additionally increases with the confidence level for the risk-averse portfolio.
For example, for scenario (b) with unreliable suppliers, the risk-averse subset of
protected suppliers for confidence levels, α = 0.9, 0.95, 0.99, is identical with the
subset of protected suppliers for scenario (a) with reliable suppliers and the highest
confidence level, α = 0.99. The corresponding solution values are very close each
other (see, Table 4.9). The optimal risk-averse supply portfolios, (V1 , . . . , VI ), (4.20),
for scenario (a) with reliable suppliers and scenario (b) with unreliable suppliers and
confidence levels, α = 0.9, 0.95, 0.99, as well as the disruption probability, p∗ +
(1 − p∗ )pi , for each supplier i, are shown in Fig. 4.6. The resilient portfolios for
4.6 Computational Examples 97

scenario (b) and different confidence levels are very close each other. The largest
portion of demand for parts has been allocated among the five protected suppliers
i = 1, 2, 4, 7, 9, while remaining small orders for parts are placed among unprotected
suppliers.
Comparison of the corresponding portfolios for single- and multi-level protection
demonstrate that only lower protection levels are selected to fortify suppliers. Instead
of selection more costly, higher protection levels, the resilient solution rather pre-
positions more emergency inventory at fortified suppliers to minimize expected and
expected worst-case cost, respectively for model RSP(mlp)_E and RSP(mlp)_CV.

4.7 Notes

Supply chain resilience is a relatively new concept that can be defined as the adap-
tive capability of the supply chain to prepare for unexpected events, respond to
disruptions, and recover from them by maintaining continuity of operations at the
desired level of connectedness and control over structure and function (Ponomarov
and Holcomb 2009). Falasca et al. (2008) defined resilience as the ability of a supply
chain system to reduce the probabilities of disruptions, to reduce the consequences
of those disruptions and the time to recover disrupted operations to their normal per-
formance. Also Fiksel (2006) refers resiliency to a firm’s capacity to survive, adapt,
and grow in the face of change and uncertainty. According to Sheffi (2005), the com-
panies can develop the resilience in three general ways: (1) creating redundancies
throughout the supply chain; for example with holding extra inventory, maintain low
capacity utilization, and contracting with multiple suppliers, (2) increasing supply
chain flexibility; for example with adoption of standardized processes, using con-
current instead of sequential processes, plan to postpone, align procurement strategy
with supplier relationships, and (3) changing the corporate culture.
Despite the abundant literature on supplier selection and order quantity alloca-
tion problems (e.g., Aissaoui et al. 2007, Ho et al. 2010), the research on quanti-
tative approaches for building a resilient portfolio of suppliers has not often been
reported in the literature. Torabi et al. (2015) proposed a bi-objective mixed possi-
bilistic, two-stage stochastic programming model for supplier selection and order
allocation problem to build the resilient supply base under operational and disrup-
tion risks. The model accounts for uncertainty of critical data and applies several
proactive strategies such as suppliers business continuity plans, fortification of sup-
pliers and contracting with backup suppliers to enhance the resilience level of the
selected supply base. The issue of linking risk assessment with risk mitigation for
low-probability high-consequence events such as disruptions of supplies is discussed
by Kleindorfer and Saad (2005) and Cohen and Kunreuther (2007) and the need to
build resiliency to disruptive events in supply chains is discussed by Knemeyer et al.
(2009), who considered a proactive planning, based on methodology used by the
insurance industry. They quantify the risk of multiple types of catastrophic events on
key supply chain locations. The proposed proactive planning process involves four
98 4 Selection of Resilient Supply Portfolio

critical steps: identification of key supply chain locations and threats, estimation of
probabilities and loss for each location, evaluation of alternative countermeasures
for each location, and selection of countermeasures that may prevent or mitigate dis-
ruption risks. Examples of such countermeasures include relocation of facility away
from high risk location, e.g., moving a warehouse to a hurricane-free area, redesign
of facility to increase storm preparedness, building storm walls to help protect against
flooding, maintaining excess inventory, etc. In practice, the fortification of suppliers
to protect them against disruptions is recently observed. For example, to prevent
flooding during monsoons, a giant flood wall with sealable aluminum flood barriers
across entrance points has been constructed in Thailand, around the perimeter of
Nava Nakorn industrial zone, where over 220 factories of electronics and computer
components suppliers are located (Fuller 2012).
In a related stream of research, the design of supply chain networks that are
resilient to disruptions is considered and the fortification models are developed to
improve the reliability of the existing infrastructure systems for which a complete
reconfiguration would be cost prohibitive, e.g., Snyder et al. (2005). The objective
of such fortification models is to identify optimal strategies for allocating limited
resources among possible mitigation investments. For example, Church and Scaparra
(2006) introduced the r-interdiction median problem with fortification in a distribu-
tion network with p operating facilities and a set of system users who receive service
from their nearest facility. The problem objective is to determine the optimal alloca-
tion of a limited amount of protective resources in such a way that the accessibility
reduction due to a worst-case loss of r unprotected facilities is minimized. They
formulate the problem as a mixed integer program. The proposed model presents the
limitation of requiring a complete enumeration of all possible ways of interdicting
r of the p facilities. In order to alleviate the size restrictions, in a subsequent work
Scaparra and Church (2008) presented a bi-level formulation of the r-interdiction
median problem with fortification. The top level problem involves the decisions about
which facilities to fortify in order to minimize the worst-case efficiency reduction
due to the loss of unprotected facilities and worst-case scenario losses are modeled
in the lower-level interdiction problem.
In the disaster management literature a limited number of studies focus on the pre-
positioning of the different types of emergency supplies in the presence of uncertainty.
For example, Rawls and Turnquist (2010) and Noyan (2012) consider the problem of
determining the locations of the response facilities and the inventory levels of disaster
relief supplies at each facility to effectively manage the response operations when a
disaster occurs. There is also a growing body of literature addressing the quantities of
pre-positioned emergency inventory, also called, strategic inventory reserves, Schmitt
(2011) or just-in-case inventory, Sheffi (2005), Sheffi and Rice (2005), which should
be held throughout the supply chain to protect against disruption risks.
The material presented in this chapter is based in part on research reported in
Sawik (2013a), where SMIP models were proposed for risk-neutral, risk averse and
a bi-objective mean-risk selection of resilient supply portfolio under disruption risks.
In the proposed models a single protection level was considered only with a full
remaining capacity of a fortified supplier available after disruption. In this chapter,
4.7 Notes 99

however, the SMIP approach was enhanced for multi-level protection with the amount
of capacity remaining after disruption, dependent on the fortification cost and the
protection level applied. In Sawik (2013b) a resilient supply portfolio was considered
with fortified suppliers and regular inventory pre-positioned at the fortified suppliers.
The regular inventory can be fully used under each disruption scenario to fulfill
regular orders placed on the protected suppliers.
It is worth to note that the number of variables and constraints in the proposed
models grow exponentially in the number I of suppliers, if all S = 2I potential scenar-
ios are considered. The 10-supplier examples for 25-part types have approximately
280,000 variables and 280,000 constraints, while for a 20-supplier problem with all
potential disruption scenarios considered and a single part type only, this increases
to over 20,000,000 variables and 20,000,000 constraints. Solving the problem with
such huge number of scenarios is cumbersome. Even the construction of problems of
this size may become intractable, e.g., Chahara and Taaffe (2009). However, the total
number of potential disruption scenarios that need to be considered can be reduced
by eliminating scenarios that are unlikely to realize. Moreover, using of scenario
management approaches (e.g., Jenkins 2000 can also be considered to identify a
subset of potential disruption scenarios for which a detailed analysis of their impact
on supply chain performance may provide information on the impact of all potential
disruption scenarios. Another approach used to reduce the number of random disrup-
tion scenarios is the Fuzzy C-Mean clustering technique (e.g., Izakian and Abraham
2011) and the possibilistic scenario-based model (e.g., Torabi et al. 2015). In order
to reduce the number of scenarios, the FCM technique is used to cluster possible
disruptive events at suppliers to different clusters by which the centers of clusters are
used as representatives of disruptive events.
In this chapter, the local and regional supply disruptions are assumed to occur inde-
pendently. The future research should consider new resilience strategies for depen-
dent disruptive events (e.g., Li et al. 2013) or multiple concurrent disruptions (Zobel
and Khansa 2014).

Problems

4.1 Modify the probability, Ps , for disruption scenarios to account for suppliers
located in different regions, subject to regional disruptions that may affect all sup-
pliers in the same region simultaneously. Modify the SMIP models presented in this
chapter for a joint fortification of all suppliers in one region.

4.2 Define service level for the SMIP models presented in this chapter and modify
the models for the service level objective function.
100 4 Selection of Resilient Supply Portfolio

4.3 Limited storage space for emergency inventory

(a) Modify the SMIP models presented in Sect. 4.3 for the resilient supply portfolio
with single protection levels to account for limited storage space for emergency
inventory at protected suppliers.
(b) Modify the SMIP models presented in Sect. 4.5 for the resilient supply portfolio
with multiple protection levels to account for limited storage space for emergency
inventory at protected suppliers, dependent on protection level.

4.4 Modify the protection index introduced in Sect. 4.4 for the resilient supply port-
folios with multiple protection levels.

4.5 In the computational examples for unreliable suppliers, Table 4.9 and Fig. 4.6b
indicate that the risk-averse resilient solutions for confidence levels α = 0.9, 0.95,
0.99 are very close each other. How, would you explain the reason for that, and what
would be the explanation for the case with identical solutions?
Hint: compare with solution results presented in Table 4.5 for a single protection
level.
 Part II
Integrated Selection of Supply Portfolio
and Scheduling
Chapter 5
Integrated Selection of Supply Portfolio
and Scheduling of Production

5.1 Introduction

The supplier selection and order quantity allocation are a medium- to short-term
decision, driven by the time-varying customer demand. The scheduling horizon for
supplies of parts coincides with the scheduling horizon for customer orders and to
achieve the best results the supplier selection and order quantity allocation decisions
should also be made for the same time horizon. The advantage of a joint decision
making can be shown especially in the presence of supply chain disruption risks.
This chapter proposes a SMIP approach to integrated supplier selection and customer
order scheduling in the presence of supply chain disruption risks, for a single, dual
or multiple sourcing strategy. The suppliers are assumed to be located in two or more
disjoint geographic regions: in the producer’s region (domestic suppliers) and outside
the producer’s region (foreign suppliers). The supplies are subject to independent
random local disruptions that are uniquely associated with a particular supplier and
to random regional disruptions that may result in disruption of all suppliers in the
same geographic region simultaneously. The domestic suppliers are relatively reliable
but more expensive, while the foreign suppliers offer competitive prices. However
the foreign suppliers are more prone to breakdowns and material flows from these
suppliers are more exposed to unexpected disruptions due to natural or man made
disasters and longer shipping time and distance. Given a set of customer orders
for products, the decision maker needs to decide which single supplier, which two
suppliers (one from each region) or which multiple suppliers to select for purchasing
parts required to complete the customer orders and how to schedule the orders over the
planning horizon, to mitigate the impact of disruption risks. The problem objective
is either to minimize total cost of ordering and purchasing of parts plus penalty cost
of delayed and unfulfilled customer orders due to the parts shortages or to maximize
customer service level. The resulting allocation of total demand for parts among the
selected suppliers and the schedule of customer orders for every potential disruption

© Springer International Publishing AG 2018 103

T. Sawik, Supply Chain Disruption Management Using Stochastic Mixed
Integer Programming, International Series in Operations Research
& Management Science 256, DOI 10.1007/978-3-319-58823-0_5
104 5 Integrated Selection of Supply Portfolio and Scheduling of Production

scenario should be determined ahead of time, either to minimize the average or

worst-case cost or to maximize the average or worst-case customer service level.
The following time-indexed SMIP models are presented in this chapter:

SPSm_E(c), SPS2_E(c) and SPS1_E(c) for risk-neutral selection of supply

portfolio and scheduling of customer orders to minimize expected cost:
multiple sourcing, dual sourcing and single sourcing, respectively;
SPSm_E(sl), SPS2_E(sl) and SPS1_E(sl) for risk-neutral selection of supply
portfolio and scheduling of customer orders to maximize expected service
level: multiple sourcing, dual sourcing and single sourcing, respectively;
SPSm_CV(c), SPS2_CV(c) and SPS1_CV(c) for risk-averse selection of
supply portfolio and scheduling of customer orders to minimize CVaR of
cost: multiple sourcing, dual sourcing and single sourcing, respectively;
SPSm_CV(sl), SPS2_CV(sl) and SPS1_CV(sl) for risk-averse selection of
supply portfolio and scheduling of customer orders to maximize CVaR of
service level: multiple sourcing, dual sourcing and single sourcing, respec-
tively;
SPSm_E(c)CV(sl) for mixed mean-risk selection of supply portfolio and
scheduling of customer orders to trade-off expected cost and CVaR of ser-
vice level: multiple sourcing.

For a single, dual or multiple sourcing strategy and for the two different objec-
tive functions, the risk-neutral, risk-averse and mean-risk solutions that optimize,
respectively average, worst-case and trade-off between average and worst-case per-
formance of a supply chain are illustrated with computational examples and com-
pared in Sect. 5.6. In addition, the two risk-averse service level measures: expected
worst-case order fulfillment rate and expected worst-case demand fulfillment rate
were computationally compared in Sect. 5.6.3.

5.2 Problem Description

In this section the problem of integrated supplier selection, order quantity allocation
and customer orders scheduling in the presence of supply chain disruption risks
is described. While the supplier selection is considered to be a long-term strategic
decision, the order quantity allocation and customer order scheduling are short- to
medium-term tactical decisions. In particular, in a make-to-order manufacturing, all
the above decisions can be made for a short- to medium-term planning horizon. Given
a set of part suppliers, the supply portfolio determines an allocation of demand for
parts among a subset of selected suppliers and simultaneously, for each disruption
scenario an assignment of customer orders to time periods over the planning horizon
is found. To justify the integration of supplier selection, order quantity allocation
and customer order aggregate scheduling, assume that all these decisions are made
5.2 Problem Description 105

Fig. 5.1 A three-echelon supply chain

for a medium-term planning horizon. However, the order quantity allocation among
the suppliers selected for a longer time horizon, in a make-to-order environment can
also be made for a short-term horizon with no orders placed on some suppliers, if a
shorter planning horizon needs to be considered.
Consider a three-echelon customer driven supply chain (Fig. 5.1) in which various
types of products are assembled by a single producer to meet customer demand, using
the same critical part type that can be manufactured and provided by many suppliers
(for notation used, see Table 5.1).
Let I = {1, . . . , I } be the set of I suppliers, J = {1, . . . , J } the set of J customer
orders for products, and T = {1, . . . , T } the set of T planning periods.
Denote by b j and d j , respectively the size and the due date of customer order
j ∈ J , i.e., the number units of ordered product type and the latest period of their
completion required to deliver the products to the customer by requested date. The
customer orders are single-period orders such that each order can be completed in
one planning period, e.g., Sawik (2007).
Let a j be the unit requirement for the critical part  of each product in customer
order j ∈ J . The total demand
 for all parts is A = j∈J a j b j and the total demand
for all products is B = j∈J b j .
The orders for parts are assumed to be placed at the start of the planning horizon,
when all customer orders for products are known. Let oi be the unit purchasing price
of parts from supplier i ∈ I and denote by ei the fixed ordering cost of creating
contracts and maintaining relationships with supplier i ∈ I . Each supplier has suf-
ficient capacity to meet total demand for parts and to complete and prepare orders
for shipping. Then, all parts ordered from a supplier are shipped together in a single
delivery. The order preparation and transportation time of a shipment from supplier
i ∈ I to the producer is constant and equals to τi periods so that the parts ordered
from supplier i ∈ I are delivered in period τi and then can be used for the assembly
of products in period τi + 1, at the earliest.
The suppliers are assumed to be located in R disjoint geographic regions.  Denote
by I r ⊆ I the subset of suppliers in region r ∈ R = {1, . . . , R}, where r ∈R I r = I .
The supplies of parts are subject to random independent local disruptions that are
uniquely associated with a particular supplier, which may arise from equipment
breakdowns, local labor strike, fires, etc. Denote by pi the local disruption probability
for supplier i, i.e., the parts ordered from supplier i are delivered without disruptions
with probability (1 − pi ), or not at all with probability pi .
106 5 Integrated Selection of Supply Portfolio and Scheduling of Production

Table 5.1 Notation: selection of supply portfolio and scheduling

Indices
i = supplier, i ∈ I
j = customer order, j ∈ J
r = geographic region, r ∈ R
s = disruption scenario, s ∈ S
t = planning period, t ∈ T
Input Parameters
a j = per unit requirement for parts of each product in customer order j
b j = size (number of products) of customer order j
A = total demand for parts
B = total demand for products
c j = per unit capacity consumption of producer for customer order j
Ct = capacity of producer in period t
d j = due date for customer order j
ei = fixed cost of ordering parts from supplier i
g j = per unit and per period penalty cost of delayed customer order j
h j = per unit penalty cost of unfulfilled customer order j
I r = subset of suppliers in geographic region r
oi = per unit price of parts purchased from supplier i
pi = local disruption probability for supplier i
pr = regional disruption probability for all suppliers in region r
p ∗ = global disruption probability for all suppliers
α = confidence level
τi = delivery lead time from supplier i

In addition to independent local disruptions of each supplier individually, there

are also potential regional disasters that may result in correlated regional disruption
of all suppliers in the same region and global disruptive super events that may simul-
taneously impact all suppliers. For example, such global disaster super events may
include an economic crisis, widespread labor strike in a transportation sector.
Denote by pr and p ∗ the probability of regional disruption, simultaneously of
all suppliers i ∈ I r in region r ∈ R, and global disruption, simultaneously of all
suppliers i ∈ I , respectively. The global disaster, the regional disasters in each region
and the local disruptive events at each supplier are assumed to be independent events.
Let πi be the disruption probability of every supplier i ∈ I r , r ∈ R

πi = p ∗ + (1 − p ∗ ) pr + (1 − p ∗ )(1 − pr ) pi ; i ∈ I r , r ∈ R. (5.1)

Denote by Ps the probability that disruption scenario s is realized, where each

scenario s ∈ S is comprised of a unique subset Is ⊂ I of suppliers who deliver parts
without disruptions, and S = {1, . . . , S} is the index set of all disruption scenarios
(there are a total of S = 2 I potential disruption scenarios). For each scenario s ∈ S,
5.2 Problem Description 107

the supplies from every supplier, i ∈ I \ Is , can be disrupted either by a local, regional
or global disruptive event.
The probability Ps for each disruption scenario s ∈ S with the subset Is of non-
disrupted suppliers, and with all possible combinations of different disaster events
considered, is (cf. Sect. 1.3)
 
(1 − p ∗ ) r ∈R 
Psr if Is = ∅
Ps = ∗ ∗ (5.2)
p + (1 − p ) r ∈R Psr if Is = ∅,

where Psr is the probability of realizing of disruption scenario s for suppliers in I r

   
(1 − pr ) i∈I r  Is (1 − pi ) i∈I r \Is pi if I r Is = ∅
Psr =   (5.3)
pr + (1 − pr ) i∈I r pi if I r Is = ∅.

Assume that the producer has limited time-varying capacity, and denote by Ct
the producer capacity available in planning period t ∈ T , and by c j the unit capacity
consumption for each product in customer order j ∈ J .
The producer can be charged with a contractual, order specific penalty cost for
delayed or unfulfilled customer orders, caused by the shortage of parts, that are
delivered late or not at all due to supply disruptions. Let g j and h j be, respectively the
per unit and per period penalty cost of delayed customer order j ∈ J and the per unit
total penalty cost of unfulfilled customer order j ∈ J . The penalty cost h j is assumed
to also include the producer lost profit. Therefore, the unit penalty costs g j and h j
are selected in such a way that each product in an unfulfilled order is penalized much
higher than the corresponding product in a delayed order, i.e., h j  g j , j ∈ J . In
some cases the contractual penalty cost for unfulfilled customer orders can be shared
between the supplier and the producer to compensate the latter for the lost profit due
to the undelivered parts. This can be modeled by choosing appropriate lower values
of the corresponding unit penalty costs, h j , to reduce the producer direct losses.
Then, the remaining part of the producer penalty cost is assumed to be covered by
the contractual compensation from the disrupted supplier.
The producer has three sourcing alternatives to select: single, dual or multiple
sourcing. The objective of the integrated supplier selection and customer order
scheduling is to select a single supplier (single sourcing), two suppliers from two
different regions (dual sourcing), or multiple suppliers (multiple sourcing), allocate
the total demand for parts among the selected suppliers and schedule the customer
orders over the planning horizon to complete the orders and mitigate the impact of
disruption risks. The three sourcing strategies will be compared with respect to the
two alternative optimality criteria: cost and service level. The objective function is
either to minimize expected cost, expected worst-case cost, or maximize expected
service level or expected worst-case service level. The supply portfolio (the allocation
of total demand for parts among the selected suppliers) and the schedule of customer
orders for every potential disruption scenario are determined at the beginning of the
planning horizon.
108 5 Integrated Selection of Supply Portfolio and Scheduling of Production

5.3 Models for Risk-Neutral Decision-Making

In this section six time-indexed SMIP models, SPS1_E(c), SPS2_E(c), SPSm_E(c),

SPS1_E(sl), SPS2_E(sl), SPSm_E(sl), are proposed for the integrated supplier
selection and customer order scheduling to optimize average performance of a sup-
ply chain in the presence of disruption risks. The objective of models SPS1_E(c),
SPS2_E(c), SPSm_E(c) and SPS1_E(sl), SPS2_E(sl), SPSm_E(sl) is to minimize
the expected cost per product (“E(c)” in the model name) and to maximize the
expected service level (“E(sl)” in the model name), respectively.
While the objective of supplier selection is to determine a supply portfolio, i.e., an
allocation of demand for parts among the suppliers, the objective of customer orders
scheduling is to determine an aggregate production schedule, i.e., an assignment
of orders to planning periods over a planning horizon, subject to capacity and parts
availability constraints. For the selected supply portfolio and each disruption scenario
the optimal schedule of customer orders is determined in such a way as to minimize
the penalty cost for delayed and unscheduled (rejected) orders, and by this minimize
total cost or maximize service level, respectively.
The following three different sourcing strategies will be considered:
• Single sourcing (“1” in the model name), where the total demand for parts is
assigned to exactly one supplier.
• Dual sourcing (“2” in the model name), where the total demand for parts is either
assigned to one supplier or allocated between two suppliers, from two different
geographic regions. For example, the total demand for parts is split between the
two suppliers, one more expensive and more reliable domestic supplier, and one
less expensive and less reliable foreign supplier.
• Multiple sourcing (“m” in the model name), where the total demand for parts is
allocated among many suppliers.
The problem variables are defined Table 5.2.

5.3.1 Minimization of Cost

In this subsection the three time-indexed SMIP models are presented for the risk-
neutral selection of supply portfolio and scheduling of customer orders to minimize
expected cost: model SPSm_E(c) for multiple sourcing, model SPS2_E(c) for dual
sourcing and model SPS1_E(c) for single sourcing. The single sourcing strategy can
be considered to be a special case of a dual sourcing, while the dual sourcing can be
considered a special case of a multiple sourcing. Therefore, first model SPSm_E(c)
is presented, then model SPS2_E(c) and finally model SPS1_E(c).
5.3 Models for Risk-Neutral Decision-Making 109

Table 5.2 Variables: selection of supply portfolio and scheduling

First stage variables
ui = 1, if supplier i is selected; otherwise u i = 0 (supplier selection)
vi ∈ [0, 1], the fraction of total demand for parts ordered from supplier i (supply portfolio for
dual or multiple sourcing)
Second stage variables
wsjt = 1, if under disruption scenario s customer order j is scheduled for period t; otherwise
wsjt = 0 (production scheduling)
Auxiliary variables
VaRc Cost-at-Risk, the targeted cost such that for a given confidence level α, for 100α% of the
scenarios, the outcome is below VaRc
VaR sl Service-at-Risk, the targeted service level such that for a given confidence level α, for
100α% of the scenarios, the outcome is above VaRsl
Cs ≥ 0, the tail cost for scenario s, i.e., the amount by which costs in scenario s exceed VaRc
Ss ≥ 0, the tail service level for scenario s, i.e., the amount by which VaRsl exceeds service
level in scenario s

In the risk-neutral decision-making, the overall quality of the supply portfolio and
the schedule of customer orders can be measured by the expected cost per product,
(5.4), of parts ordering, i∈I ei u i /B, and purchasing ( s∈S Ps ( i∈Is Aoi vi )/B),
where the producer is not charged with ordered and undelivered parts, plus penalty
cost of delayed and unfulfilled  (rejected)
 customer
 orders due to delays anddis-
ruptions of part supplies, P ( t∈T :t>d j g j b j (t − d j )w jt )/B +
s
  s∈S s j∈J s∈S
Ps ( j∈J h j b j (1 − t∈T w jt ))/B.
s

The SMIP model SPSm_E(c) for a multiple sourcing selection of supply portfolio
and scheduling of customer orders to minimize expected cost is formulated below.

SPSm_E(c): Risk-neutral selection of Supply Portfolio and Scheduling of

customer orders to minimize expected cost: multiple sourcing
Minimize
  
Ec = ei u i /B + Ps ( Aoi vi )/B
i∈I s∈S i∈Is
  
+ Ps ( g j b j (t − d j )wsjt )/B
s∈S j∈J t∈T :t>d j
  
+ Ps ( h j b j (1 − wsjt ))/B (5.4)
s∈S j∈J t∈T

subject to
Demand allocation constraints:
- the total demand for parts must be fully allocated among the suppliers,
110 5 Integrated Selection of Supply Portfolio and Scheduling of Production

vi = 1 (5.5)
i∈I

Multiple sourcing strategy constraints:

- demand for parts cannot be assigned to non-selected suppliers,

vi ≤ u i ; i ∈ I (5.6)

Order-to-period assignment constraints:

- for each disruption scenario s, each customer order j is either sched-
t∈T w jt = 1), or unscheduled and rejected
s
uled
 during the planning horizon (
( t∈T wsjt = 0),
- for each disruption scenario s and each planning period t, the cumulative
demand for parts of all customer orders scheduled in periods 1 through t cannot
exceed the cumulative deliveries of parts in periods 1 through t − 1, from the
non-disrupted suppliers i ∈ Is ,
- for each disruption scenario s, the total requirement for parts of scheduled
customer orders is not greater than the total supplies from the non-disrupted
suppliers i ∈ Is ,

wsjt ≤ 1; j ∈ J, s ∈ S (5.7)
t∈T
  
a j b j wsjt  ≤ A vi ; t ∈ T, s ∈ S (5.8)
j∈J t  ∈T :t  ≤t i∈Is :τi ≤t−1
 
a j b j wsjt ≤ A vi ; s ∈ S (5.9)
j∈J t∈T i∈Is

Producer capacity constraints:

- for any period t and each disruption scenario s, the total demand on
capacity of all customer orders scheduled in period t must not exceed the
producer capacity available in this period,

b j c j wsjt ≤ Ct ; t ∈ T, s ∈ S (5.10)
j∈J

Non-negativity and integrality conditions:

u i ∈ {0, 1}; i ∈ I (5.11)

vi ∈ [0, 1]; i ∈ I (5.12)
wsjt ∈ {0, 1}; j ∈ J, t ∈ T, s ∈ S. (5.13)
5.3 Models for Risk-Neutral Decision-Making 111

In the SMIP model SPS2_E(c) formulated below for a dual sourcing, all suppliers
are assumed to be located in two different geographic regions and at most one supplier
can be selected from each region.

SPS2_E(c): Risk-neutral selection of supply portfolio and scheduling of

customer orders to minimize expected cost: dual sourcing
Minimize (5.4)
subject to (5.5)–(5.13) and
Dual sourcing strategy constraints:
- at most one supplier can be selected from each region,

u i ≤ 1; r = 1, 2. (5.14)
i∈I r

Note that dual and multiple sourcing allows, respectively two and more suppliers
to be selected, however an optimal solution may assign total demand for parts to a
single supplier only, if such a solution optimizes the objective function.
The single-sourcing model SPS1_E(c) is a special case of the dual-sourcing model
SPS2_E(c). In order to derive SPS1_E(c) from SPS2_E(c), it is sufficient to replace
inequality vi ≤ u i ; i ∈ I , (5.6), with an equality constraint vi = u i ; i ∈ I and by
this eliminate variables vi , i ∈ I . Simultaneously, constraint (5.14) becomes weaker
than (5.5) and can be removed, while constraints (5.8) and (5.9) should be replaced
by   
wsjt  ≤ J u i ; t ∈ T, s ∈ S
j∈J t  ∈T :t  ≤t i∈Is :τi ≤t−1

or  
wsjt  ≤ u i ; j ∈ J, t ∈ T, s ∈ S,
t  ∈T :t  ≤t i∈Is :τi ≤t−1

and  
wsjt ≤ J ui ; s ∈ S
j∈J t∈T i∈Is

or  
wsjt ≤ u i ; j ∈ J, s ∈ S,
t∈T i∈Is

respectively.
Constraints (5.8) can also be replaced by the following alternative constraints
 
twsjt ≥ (τi + 1)u i ; j ∈ J, s ∈ S,
t∈T i∈Is
112 5 Integrated Selection of Supply Portfolio and Scheduling of Production

which sometimes may appear to be more efficient computationally.

The stochastic binary program SPS1_E(c) for a single sourcing selection of supply
portfolio and scheduling of customer orders is formulated below.

SPS1_E(c): Risk-neutral selection of supply portfolio and scheduling of

customer orders to minimize expected cost: single sourcing
Minimize
  
ei u i /B + Ps ( Aoi u i )/B
i∈I s∈S i∈Is
  
+ Ps ( g j b j (t − d j )wsjt )/B
s∈S j∈J t∈T :t>d j
  
+ Ps ( h j b j (1 − wsjt ))/B (5.15)
s∈S j∈J t∈T

subject to (5.10), (5.11), (5.13) and

Supplier selection constraints:
- exactly one supplier of parts is selected,

ui = 1 (5.16)
i∈I

Order-to-period assignment constraints: (5.7) and

- for each disruption scenario s, all customer orders can be scheduled only
after the delivery of required parts, purchased from a non disrupted supplier
i ∈ Is ,
- each customer order j can be scheduled during the planning horizon under
disruption scenario s, only if required parts are ordered from a non disrupted
supplier i ∈ Is ,
 
wsjt  ≤ u i ; j ∈ J, t ∈ T, s ∈ S (5.17)
t  ∈T :t  ≤t i∈Is :τi ≤t−1
 
wsjt ≤ u i ; j ∈ J, s ∈ S. (5.18)
t∈T i∈Is

The solution to supplier selection and customer order scheduling problem deter-
mines
 for every disruption scenario s ∈ S, the aggregate production
  schedule
{ j∈J b j wsjt ; t ∈ T } as well as the corresponding service level, j∈J t∈T :t≤d j
wsjt /J , (the fraction of customer orders fulfilled by their due dates). In addition,
for every disruption scenario s ∈ S, the subset of customer
 orders scheduled by
the requested due dates d j , { j ∈ J : 1 + mini∈I τi ≤ t∈T twsjt ≤ d j }, the subset
5.3 Models for Risk-Neutral Decision-Making 113


of delayed customer orders, { j ∈ J : t∈T
 tw jt s> d j }, and the subset of rejected
s

(unscheduled) customer orders, { j ∈ J : t∈T w jt = 0}, are determined.

5.3.2 Maximization of Customer Service Level

The customer service level can be measured either by order fulfillment rate or by
demand fulfillment rate, where the order fulfillment rate and the demand fulfillment
rate is the fraction of customer orders (irrespective of their size) and the fraction of
customer demand, respectively, that is fulfilled by the customer requested due dates.
When the focus is on fulfilling customer orders rather than total demanded quantity,
the order fulfillment rate would be the preferred service level measure. Otherwise,
the demand fulfillment rate would be selected. Since the order fulfillment rate does
not account for the size of customer orders, a high service level can be achieved by
fulfilling a large number of small size orders, while leaving the unfulfilled demand
relatively high. On the other hand, a high demand fulfillment rate can be achieved
by fulfilling a few large size orders, while leaving relatively high, the number of
unfulfilled small size orders.
The aim of the next three SMIP models is to achieve the best average customer
service level by maximizing the expected fraction of customer orders fulfilled by
their due dates, i.e., by maximizing the expected order fulfillment rate.

SPSm_E(sl): Risk-neutral selection of supply portfolio and scheduling of

customer orders to maximize expected service level: multiple sourcing
Maximize
  
Ps wsjt /J (5.19)
s∈S j∈J t∈T :t≤d j

subject to (5.5)–(5.13).

SPS2_E(sl): Risk-neutral selection of supply portfolio and scheduling of

customer orders to maximize expected service level: dual sourcing
Maximize (5.19)
subject to (5.5)–(5.14).
114 5 Integrated Selection of Supply Portfolio and Scheduling of Production

SPS1_E(sl): Risk-neutral selection of supply portfolio and scheduling of

customer orders to maximize expected service level: single sourcing
Maximize (5.19)
subject to (5.7), (5.10), (5.11), (5.13), (5.16)–(5.18).

Note that maximization of the customer service level simultaneously leads to

reduction of the penalty costs for delayed and unfulfilled customer orders, that rep-
resent an important part of the total cost structure. However, as the objective func-
tion (5.19) does not directly include any cost components, the optimal solution to
SPSm_E(sl), SPS2_E(sl) and SPS1_E(sl) that maximizes the customer service level
depends only on the delivery lead time from suppliers and distribution among the
suppliers of local and regional disruption probabilities. Therefore, the solution can
be considered to be an ideal solution with respect to the customer service level in
a risk-neutral decision-making, and may be used for comparisons with the optimal
cost-based solutions.
If instead of the number of customer orders fulfilled on time (i.e., the order,
fulfillment rate) customer service level is measured by the fraction of customer
demand fulfilled on time (i.e., by the demand fulfillment rate), then (5.19) in mod-
els SPSm_E(sl), SPS2_E(sl) and SPS1_E(sl) should be replaced by the following
objective function to be maximized
  
Ps b j wsjt /B. (5.20)
s∈S j∈J t∈T :t≤d j

5.4 Models for Risk-Averse Decision-Making

In the selection of supply portfolio and scheduling of customer orders under disrup-
tion risks, the decision maker controls the risk of high losses due to supply disruptions
by choosing the confidence level α. For a given confidence level, let VaRc be the
acceptable cost level above which we want to minimize the number of outcomes and
CVaRc considers those portfolio outcomes, where costs exceed VaRc . In a similar
way, denote by VaRsl the acceptable service level below which we want to maximize
the number of outcomes and CVaRsl considers those portfolio outcomes, where ser-
vice levels are below VaRsl . We assume that the decision maker is willing to accept
only portfolios for which the total probability of scenarios with costs greater than
VaRc (or with service levels lower than VaRsl ) is not greater than 1 − α. The greater
the confidence level α, the more risk aversive is the decision maker and the smaller
percent of the highest cost (of the lowest service level, respectively) outcomes is
focused on.
5.4 Models for Risk-Averse Decision-Making 115

In this section the six time-indexed SMIP models: SPSm_CV(c), SPS2_CV(c),

SPS1_CV(c), SPSm_CV(sl), SPS2_CV(sl), SPS1_CV(sl) are proposed for the
integrated supplier selection and customer order scheduling to optimize worst-case
performance of a supply chain in the presence of disruption risks. The objective
of models SPSm_CV(c), SPS2_CV(c), SPS1_CV(c) and SPSm_CV(sl), SPS2_
CV(sl), SPS1_CV(sl) is to minimize CVaRc of cost per product (“CV(c)” in the
model name) and to maximize CVaRsl of service level (“CV(sl)” in the model name),
respectively. The models are risk-averse equivalents, respectively, of SPSm_E(c),
SPS2_E(c), SPS1_E(c) and SPSm_E(sl), SPS2_E(sl), SPS1_E(sl) formulations
presented in Sect. 5.3 for the risk-neutral decision-making.

5.4.1 Minimization of Cost

Define Cs as the tail cost for scenario s, where tail cost is defined as the amount
by which costs in scenario s exceed VaRc . The risk-averse supply portfolio and
production schedule will be optimized by calculating VaRc and minimizing CVaRc
simultaneously. In the proposed model, CVaRc is represented by an auxiliary function
(5.21) introduced by Rockafellar and Uryasev (2000).
The SMIP models SPSm_CV(c), SPS2_CV(c) and SPS1_CV(c) for supplier
selection and customer order scheduling to optimize worst-case performance of a
supply chain and reduce the risk of high costs, respectively for multiple, dual and
single sourcing, are formulated below.

SPSm_CV(c): Risk-averse selection of supply portfolio and scheduling of

customer orders to minimize CVaR of cost: multiple sourcing
Minimize

C V a R c = V a R c + (1 − α)−1 Ps Cs (5.21)
s∈S

subject to (5.5)–(5.13) and

Risk constraints:
- the tail cost for scenario s is defined as the nonnegative amount by which
cost in scenario s exceeds VaRc ,
 
Cs ≥ ei u i /B + Aoi vi /B
i∈I i∈Is
   
+ g j b j (t − d j )wsjt /B + h j b j (1 − wsjt )/B
j∈J t∈T :t>d j j∈J t∈T

−V a R ; s ∈ S
c
(5.22)
Cs ≥ 0. (5.23)
116 5 Integrated Selection of Supply Portfolio and Scheduling of Production

SPS2_CV(c): Risk-averse selection of supply portfolio and scheduling of

customer orders to minimize CVaR of cost: dual sourcing
Minimize (5.21)
subject to (5.5)–(5.14), (5.22), (5.23).

SPS1_CV(c): Risk-averse selection of supply portfolio and scheduling of

customer orders to minimize CVaR of cost: single sourcing
Minimize (5.21)
subject to (5.7), (5.10), (5.11), (5.13), (5.16)–(5.18), (5.23) and
Risk constraints:
- the tail cost for scenario s is defined as the nonnegative amount by which
cost in scenario s exceeds VaRc ,
 
Cs ≥ ei u i /B + Aoi u i /B
i∈I i∈Is
   
+ g j b j (t − d j )wsjt /B + h j b j (1 − wsjt )/B
j∈J t∈T :t>d j j∈J t∈T

−V a R ; s ∈ S.
c
(5.24)

5.4.2 Maximization of Customer Service Level

Define Ss as the tail service level for scenario s, where tail service level is defined
as the amount by which VaRsl exceeds service level in scenario s. The risk-averse
supply portfolio and production schedule will be optimized by calculating VaRsl and
maximizing CVaRc simultaneously.
The SMIP models SPSm_CV(sl), SPS2_CV(sl) and SPS1_CV(sl) for supplier
selection and customer order scheduling to optimize worst-case performance of a
supply chain and reduce the risk of low service levels, respectively for multiple, dual
and single sourcing, are formulated below.

SPSm_CV(sl): Risk-averse selection of supply portfolio and scheduling of

customer orders to maximize CVaR of service level: multiple sourcing
Maximize

C V a R sl = V a R sl − (1 − α)−1 Ps Ss (5.25)
s∈S
5.4 Models for Risk-Averse Decision-Making 117

subject to (5.5)–(5.13) and

Risk constraints:
- the tail service level for scenario s is defined as the nonnegative amount
by which VaRsl exceeds service level in scenario s,
 
Ss ≥ V a R sl − wsjt /J ; s ∈ S (5.26)
j∈J t∈T :t≤d j

Ss ≥ 0. (5.27)

SPS2_CV(sl): Risk-averse selection of supply portfolio and scheduling of

customer orders to maximize CVaR of service level: dual sourcing
Maximize (5.25)
subject to (5.5)–(5.14), (5.26), (5.27).

SPS1_CV(sl): Risk-averse selection of supply portfolio and scheduling of

customer orders to maximize CVaR of service level: single sourcing
Maximize (5.25)
subject to (5.7), (5.10), (5.11), (5.13), (5.16)–(5.18), (5.26), (5.27).

If instead of the number of customer orders fulfilled on time (i.e., the order,
fulfillment rate) customer service level is measured by the fraction of customer
demand fulfilled on time (i.e., by the demand fulfillment rate), then (5.26) in models
SPSm_CV(sl), SPS2_CV(sl) and SPS1_CV(sl) should be replaced by the following
constraints  
Ss ≥ V a R sl − b j wsjt /B; s ∈ S. (5.28)
j∈J t∈T :t≤d j

Model Enhancements

The SMIP models SPSm_CV(sl), SPS2_CV(sl) and SPS1_CV(sl) can be strength-

ened by the addition of valid inequalities to precisely determine VaRsl of service level
for a given confidence level α. As a result, tighter LP relaxations of the corresponding
mixed integer programs are achieved.
First, introduce the additional binary variable:
•   selections variable: z s = 1, if for slscenario s, customer service level,
scenario
j∈J t∈T :t≤d j w jt /J , is not less than VaR ; otherwise z s = 0.

The valid inequalities that define VaRsl of service level are shown below
118 5 Integrated Selection of Supply Portfolio and Scheduling of Production
 
zs ≥ wsjt /J − V a R sl ; s ∈ S (5.29)
j∈J t∈T :t≤d j
 
zs ≤ 1 + wsjt /J − V a R sl ; s ∈ S (5.30)
j∈J t∈T :t≤d j

Ps z s ≥ α, (5.31)
s∈S

where (5.29) and (5.30) determine scenarios s with the customer service level not
less than VaRsl , and (5.31) ensures that the total probability of all such scenarios is
not less than the confidence level α.
The introduction of valid inequalities will increase the size of the mixed inte-
ger programs SPSm_CV(sl), SPS2_CV(sl) and SPS1_CV(sl). The total number of
additional binary variables z s and constraints (5.29)–(5.31) is respectively, 2 I and
2 I +1 + 1, and hence they grow exponentially with the number I of suppliers.
Note that both VaRsl and Ss can be restricted to being not greater that one and,
in addition, Ss ≤ 1 − z s , ∀s ∈ S, which is equivalent to (5.29).
Denote by SPSm_CV(sl)+, SPS2_CV(sl)+ and SPS1_CV(sl)+, the stochastic
mixed integer programs SPSm_CV(sl), SPS2_CV(sl) and SPS1_CV(sl) strength-
ened with added valid inequalities (5.29)–(5.31), respectively.
If in models SPSm_CV(sl), SPS2_CV(sl) and SPS1_CV(sl), customer service
level is measured by the fraction of customer demand fulfilled on time (i.e., by
  of scenario selection variable z s and
demand fulfillment rate), then in the definition
in constraints (5.29), (5.30), the term j∈J t∈T :t≤d j wsjt /J , should be replaced by
 
t∈T :t≤d j b j w jt /B.
s
j∈J

5.5 Model for Mixed Mean-Risk Decision-Making

In this section a trade-off model SPSm_E(c)CV(sl) for a multiple sourcing strategy

is developed to evaluate the impact on average cost of optimization the expected
worst-case service level. The expected cost per product, E c (5.4), and the expected
worst-case service level, C V a R sl (5.25), will be simultaneously optimized using the
weighted-sum approach.
In order to avoid dimensional inconsistency among the two conflicting objective
functions, their valuesc arec scaled into the interval [0,1].
Denote by f c = EE c −E
−E c
, the normalized expected cost per product (E c , E c are the
C V a R sl −C V a R sl
minimum, the maximum values of E c , respectively), and by f sl = , the
C V a R sl −C V a R sl
normalized expected worst-case service level (C V a R sl , C V a R sl are the minimum,
the maximum values of C V a R sl for a given confidence level α, respectively).
The normalized objective functions f c and f sl are defined below
5.5 Model for Mixed Mean-Risk Decision-Making 119
    
fc = ( ei u i + Ps ( Aoi vi + g j b j (t − d j )wsjt
i∈I s∈S i∈Is j∈J t∈T :t>d j
 
+ h j b j (1 − wsjt )))/B − E c )/(E c − E c ) (5.32)
j∈J t∈T

C V a R sl − (V a R − (1 − α)−1 Ps Ss )
f sl
= s∈S
. (5.33)
C V a R sl − C V a R sl

The maximum value of expected worst-case service level (5.25), C V a R sl , for

a given confidence level α, and the associated maximum value of expected cost
(5.4), E c , are obtained as the optimal solution of SPSm_CV(sl) (with constraint
(5.26)/(5.28), respectively for order/demand fulfillment rate as a service level mea-
sure). The minimum value of expected cost (5.4), E c , and the associated minimum
value of expected worst-case service level, C V a R sl , for a given confidence level
α are obtained as the optimal solution to the following SMIP problem (with con-
straint (5.26)/(5.28), respectively for order/demand fulfillment rate as a service level
measure).

Minimize 0.99999E c -0.00001C V a R sl

subject to (5.5)–(5.13), (5.27), (5.26)/(5.28).

The much greater weight assigned to the objective function, E c , reflects the pri-
ority given to minimization of expected cost.
Finally, the trade-off model SPSm_E(c)CV(sl) for a mixed mean-risk optimiza-
tion of expected cost per product and expected worst-case service level is shown
below (with constraint (5.26)/(5.28), respectively for order/demand fulfillment rate
as a service level measure).

SPSm_E(c)CV(sl): Mixed mean-risk selection of supply portfolio and

scheduling of customer orders to trade-off expected cost and CVaR of service
level: multiple sourcing
Minimize

λ f c + (1 − λ) f sl , (5.34)

where 0 ≤ λ ≤ 1,
subject to (5.5)–(5.13), (5.27), (5.26)/(5.28), (5.32), (5.33).
120 5 Integrated Selection of Supply Portfolio and Scheduling of Production

5.6 Computational Examples

In this section some computational examples are presented to illustrate possible

applications of the proposed SMIP approach for the selection of suppliers, order
quantity allocation and customer orders scheduling for a single, dual and multiple
sourcing strategy and the two different service level measures: order fulfillment rate
and demand fulfillment rate. First, single versus dual sourcing and single versus
multiple sourcing strategy are compared, then order fulfillment rate versus demand
fulfillment rate are compared as worst-case service level performance measures and
finally the impact of optimization the worst-case service on average cost is evaluated.

5.6.1 Single Versus Dual Sourcing

We assume that suppliers are located in two different geographic regions r = 1, 2.

Suppliers i ∈ I 1 are domestic suppliers, relatively reliable but more expensive. Sup-
pliers i ∈ I 2 are located outside the producer’s geographic region and offer compet-
itive prices. For the domestic suppliers i ∈ I 1 , the unit prices, oi , are higher than for
the foreign suppliers i ∈ I 2 , while the fixed ordering costs ei are lower. However
the foreign suppliers are more prone to breakdowns and material flows from these
suppliers are more exposed to unexpected disruptions due to natural or man made
disasters and longer shipping times and distance to the producer. The regional dis-
ruption probability, p 1 , for the producer domestic region r = 1, is much lower than
the probability, p 2 , for the foreign region r = 2.
The probability Ps of each disruption scenario s ∈ S under the risks of local and
regional disaster events is
⎧ 1 2 
⎪
⎪ p p + (1 − p 1 )(1 − p 2 ) i∈I pi if Is = ∅
⎪
⎨ (1 − p 1 ) p 2   1 2
i∈Is (1 − pi ) i∈I 1 \Is pi + (1 − p )(1 − p ) P̂s if Is ⊆ I 1
Ps =   (5.35)
⎪ p (1 − p ) i∈Is (1 − pi ) i∈I 2 \I pi + (1 − p )(1 − p ) P̂s
⎪
1 2 1 2 if Is ⊆ I 2
⎪
⎩ s  1  2
1 2
(1 − p )(1 − p ) P̂s if Is I  = ∅, Is I  = ∅,

where P̂s is the probability of disruption scenario s in the presence of independent

local disruptive events only
 
P̂s = (1 − pi ) · pi . (5.36)
i∈Is i ∈I
/ s

The following parameters have been used for the example problems:
• I , the number of suppliers, was equal to 10 and the number of disruption scenarios,
was equal to the total number of all potential scenarios S = 2 I = 1024;
• J , the number of customer orders, was equal to 25;
5.6 Computational Examples 121

• R, the number of geographic regions, was equal to 2, and the subsets of domestic
and foreign suppliers were I 1 = {1, 2, 3, 4, 5} and I 2 = {6, 7, 8, 9, 10}, respec-
tively;
• T , the number of planning periods, was equal to 10;
• a j , the unit requirements for parts of products in customer orders were integers in
{1, 2, 3} drawn from int(U[1;3]) distribution, for all orders j;
• b j , the size of customer orders (required numbers of products), were integers in
{500, 1000, . . . , 5000} drawn from 500int(U[1;10]) distribution, for all customer
orders j;
• c j , the unit capacity consumptions of producer, were integers in {1, 2, 3} drawn
from int(U[1;3]) distribution, for all customer orders j;
• Ct , the capacity
 of producer in each period t, was integer drawn from
1000 (2 j∈J b j c j /(T − maxi∈I τi ))U [0.75; 1.25]/1000 distribution, i.e., in
each period the producer capacity was from 75% to 125% of the double capacity
required to complete all customer orders during the planning horizon, after the
latest delivery of parts;
• d j , the due dates for customer orders, were integers in {1 + mini∈I (τi ), . . . , T }
drawn from int(U[2;10]) distribution, for all customer orders j;
• ei , the cost of ordering parts, were integers in {5000, 6000, . . . , 10000} and integers
in {15000, 16000, . . . , 30000}, respectively for domestic suppliers i ∈ I 1 and for
foreign suppliers i ∈ I 2 ;
• g j , the unit daily penalty cost of delayed customer orders, was equal to a j
maxi∈I (oi )/350 for all orders j, i.e., was approximately 0.28% of the maximum
unit price of required parts;
• h j , the unit penalty cost of unfulfilled customer orders, was to 2 a j maxi∈I (oi )
for all orders j, i.e., was approximately twice as large as the maximum unit price
of required parts;
• oi , the unit price of parts purchased from supplier i, was uniformly distributed over
[11,16] and over [1,6], respectively for domestic suppliers i ∈ I 1 and for foreign
suppliers i ∈ I 2 .
• α, the confidence level, was equal to 0.50, 0.75, 0.90, 0.95 or 0.99.
• pi , the local disruption probability was uniformly distributed over [0.005,0.01] for
domestic suppliers i ∈ I 1 and over [0.05,0.10] for foreign suppliers i ∈ I 2 , i.e., the
disruption probabilities were drawn independently from U[0.005;0.01] and from
U[0.05;0.10], respectively for domestic and foreign suppliers.
The regional disruption probability was p 1 = 0.001 for domestic suppliers i ∈ I 1
and p 2 = 0.01 for foreign suppliers i ∈ I 2 .
p ∗ , the global disruption probability was 0, i.e., no global disaster super event is
considered.
• τi , delivery lead time from domestic suppliers i ∈ I 1 , were integers in {1, 2} drawn
from int(U[1;2]) distribution, and from foreign suppliers i ∈ I 2 , were integers in
{2, 3, 4} drawn from int(U[2;4]) distribution.
The computational experiments were performed for the same replication of the
above input data set. For all test examples, the resulting total demand for parts and
122 5 Integrated Selection of Supply Portfolio and Scheduling of Production

products is A = 132500 and B = 66000, respectively. The unit price per part oi and
disruption probability pr + (1 − pr ) pi of each supplier i ∈ I r , r = 1, 2 are shown
in Fig. 5.2.
The solution results for the two sourcing strategies and the two objective func-
tions are presented in Tables 5.3, 5.4 and 5.5, respectively for the risk-neutral
models and the risk-averse models with different confidence levels, where for the
risk-averse maximization of service level the enhanced models SPS1_CV(sl)+ and
SPS2_CV(sl)+ were applied. The confidence level α is set at five levels of 0.5, 0.75,
0.90, 0.95, and 0.99, which means that focus is on minimizing the highest (maximiz-
ing the lowest) 50%, 25%, 10%, 5%, and 1% of all scenario outcomes, i.e., costs per
product (service levels, respectively). The size of the corresponding mixed integer
programs is represented by the total number of variables, Var., number of binary
variables, Bin., number of constraints, Cons, and number of nonzero coefficients in

Suppliers Characteristics
14 0.1
0.09
12
0.08

Probability
10 0.07
8 0.06
Price

0.05
6 0.04
4 0.03
0.02
2
0.01
0 0
1 2 3 4 5 6 7 8 9 10
Domestic Suppliers: 1-5 Foreign Suppliers: 6-10

-Price per Part Disruption Probability

Fig. 5.2 Suppliers

Table 5.3 Risk-neutral solutions: single versus dual sourcing

Sourcing strategy Single sourcing Dual sourcing
Minimization of expected cost (5.4) and (5.15)
Expected Cost 11.92 11.92
Expected service level (a) 61.80 61.80
Selected Supplier 6 6
CPU(b) 144 45
Maximization of expected service level (5.19)
Expected service level 99.35 99.35
Expected Cost 23.19 23.19
Selected Supplier 4 4
CPU(b) 61 89
(a) (
  
t∈T :t≤d j w jt /J )100%
s
s∈S Ps j∈J
(b)
CPU seconds for proving optimality on a MacBookPro 6.2, Intel Core i7, 2.66GHz, RAM
8GB/CPLEX 12.5
5.6 Computational Examples 123

the constraint matrix, Nonz. In addition to the optimal solution values for the pri-
mary objective functions and the allocation of demand among the selected suppliers,
Tables 5.3, 5.4 and 5.5 present the associated values of the other objective function.
Table 5.3 indicates that for the risk-neutral models the same optimal solution was
found for each sourcing strategy, and the obtained risk-neutral solutions are identical
with the corresponding risk-averse solutions for α = 0.5 (see, Tables 5.4 and 5.5).
Note that a low price, unreliable foreign supplier i = 6 was selected to minimize the
expected cost per product, while a high price, reliable domestic supplier i = 4 was
selected to maximize the expected service level.
For the risk-averse cost minimization, Table 5.4 indicates that the solution results
for single and dual sourcing are identical for all confidence levels α except for
the highest level α = 0.99, where for a dual sourcing the total demand is allocated
between two suppliers to reduce the worst-case cost outcomes. For the low confidence
level α = 0.5, a less reliable, low price foreign supplier i = 6 is selected, while for a
higher α more reliable and more expensive domestic supplier i = 4 is chosen. When
α = 0.99, the expected worst-case cost is lower for a dual sourcing, for which the
total demand is allocated between two suppliers.

Table 5.4 Risk-averse minimization of cost: single versus dual sourcing

Confidence level α 0.50 0.75 0.90 0.95 0.99
Model SPS1_CV(c): Var. = 230 060, Bin. = 229 035, Cons. = 45 945, Nonz. = 2 115 890 (a)
CVaRc 16.05 23.78 24.96 26.92 42.64
VaRc 7.79 22.99 22.99 22.99 22.99
Supplier Selected 6 4 4 4 4
Expected Cost 11.92 23.19 23.19 23.19 23.19
Expected Service Level (b) 61.80 99.35 99.35 99.35 99.35
CPU(c) 88 103 204 60 880
Model SPS2_CV(c): Var. = 230 070, Bin. = 229 035, Cons. = 45 957, Nonz. = 2 121 040 (a)
CVaRc 16.05 23.78 24.96 26.92 38.29
VaRc 7.79 22.99 22.99 22.99 37.20
Suppliers Selected(% of total demand) 6(100) 4(100) 4(100) 4(100) 4(60)
10(40)
Expected Cost 11.92 23.19 23.19 23.19 23.19
Expected Service Level (b) 61.80 99.35 99.35 99.35 55.32
CPU(c) 34 363 55 44 3473
(a) Var. = number of variables, Bin. = number of binary variables, Cons. = number of constraints,
Nonz.
(b) (
= number
 of nonzero coefficients
t∈T :t≤d j w jt /J )100%
s
s∈S Ps j∈J
(c)
CPU seconds for proving optimality on a MacBookPro 6.2, Intel Core i7, 2.66GHz, RAM
8GB/CPLEX 12.5

For the risk-averse service level maximization, for which the supplier selection
is independent on any cost parameters and the solution depends only on disruption
probability, Table 5.5 demonstrates that for a single sourcing, the same expensive
124 5 Integrated Selection of Supply Portfolio and Scheduling of Production

but reliable domestic supplier i = 4 is selected for all confidence levels. The same
results are obtained for a dual sourcing and α = 0.5, 0.75, 0.9, while for a higher
confidence level, the total demand is allocated between two suppliers to reduce
the worst-case service level outcomes. Then, the expected worst-case service level
is higher than the corresponding level for a single sourcing. In particular, for the
highest confidence level α = 0.99, the expected worst-case service level for a dual
sourcing is twice as large as that for a single sourcing. Table 5.5 also shows the results
for the original models SPS1_CV(sl) and SPS2_CV(sl). The comparison of CPU
times clearly indicates the advantage of the enhanced models, in particular, model
SPS2_CV(sl)+.
To identify the possible impact of the basic cost parameter setting on the achieved
solution results, the additional computational experiments were performed with the
much smaller differences of the fixed ordering cost, ei , and the unit price, oi , for the
two groups of suppliers. The following values of parameters ei and oi , i = 1, . . . , 10,
were selected: e = (6000, 8000, 7000, 8000, 8500, 8500, 9500, 9000, 10000, 9500)
and o = (6, 6, 5, 7, 6, 4, 5, 4, 4, 3). Since the service level-based objective function
is independent on any cost parameters, the optimal solutions have not changed, while

Table 5.5 Risk-averse maximization of service level: single versus dual sourcing
Confidence level α 0.50 0.75 0.90 0.95 0.99
Model SPS1_CV(sl)+: Var. = 231 082, Bin. = 230 053, Cons. = 49 035, Nonz. = 2 242 779 (a)
CVaRsl % 98.70 97.39 93.48 86.95 34.77
VaRsl % 100 100 100 100 100
Expected Service Level (b) 99.35 99.35 99.35 99.35 99.35
Expected Cost 23.19 23.19 23.19 23.19 23.19
Supplier Selected 4 4 4 4 4
CPU(c) 10 10 12 14 258
Model SPS1_CV(sl): Var. = 231 060, Bin. = 229 035, Cons. = 46 028, Nonz. = 1 934 798 (a)
CPU(c) 19 20 15 18 293
Model SPS2_CV(sl)+: Var. = 231 088, Bin. = 230 053, Cons. = 48 017, Nonz. = 2 240 749 (a)
CVaRsl % 98.70 97.39 93.48 87.35 66.17
VaRsl % 100 100 100 96 72
Expected Service Level(b) 99.35 99.35 99.35 99.16 95.01
Expected Cost 23.19 23.19 23.19 26.73 25.04
Suppliers Selected(% of total demand) 4(100) 4(100) 4(100) 4(89) 4(45)
7(11) 7(55)
CPU(c) 354 153 192 5900 349
Model SPS2_CV(sl): Var. = 230 070, Bin. = 229 035, Cons. = 45 968, Nonz. = 1 996 339 (a)
CPU(c) 465 4380 1409 11920 3725
(a) Var. = number of variables, Bin. = number of binary variables, Cons. = number of constraints,

Nonz.
(b) (
= number
 of nonzero coefficients
t∈T :t≤d j w jt /J )100%
s
s∈S Ps j∈J
(c)
CPU seconds for proving optimality on a MacBookPro 6.2, Intel Core i7, 2.66GHz, RAM
8GB/CPLEX 12.5
5.6 Computational Examples 125

Table 5.6 Risk-averse supply portfolios for suppliers offering similar prices
Confidence level α 0.50 0.75 0.90 0.95 0.99
Model SPS1_CV(c)
Supplier Selected 3 3 3 3 3
Model SPS2_CV(c)
Suppliers Selected(% of total demand) 3(100) 3(100) 3(100) 3(100) 3(55%)
10(45)

for the cost-based objective function only slight changes were observed (see, optimal
supply portfolios in Table 5.6). The cheapest supplier, i = 3, from among the most
reliable domestic suppliers is selected as the main one (for a dual sourcing) or the
only one (for a single sourcing). The results have indicated that the supplier reliability
is a key selection parameter, even for a cost-based objective function.
For the optimal risk-averse supply portfolio with α = 0.99 and the two objec-
tive functions, Figs. 5.3 and 5.4 show the distribution of cost per product and the

Distribution of Cost
for Minimum CVaR of Cost (alpha=0.99)
1

0.8
Probability

0.6

0.4

0.2

0
10 20 30 40 50 60 70 80 90 100
Cost per Product
Single-sourcing Dual-sourcing

Distribution of Cost
for Maximum CVaR of Service Level (alpha=0.99)
1

0.8
Probability

0.6

0.4

0.2

0
10 20 30 40 50 60 70 80 90 100
Cost per Product
Single-sourcing Dual-sourcing

Fig. 5.3 Distribution of cost per product for risk-averse supply portfolio with α = 0.99
126 5 Integrated Selection of Supply Portfolio and Scheduling of Production

distribution of customer service level, respectively. Figure 5.3 demonstrates that for
both objective functions, the distribution of cost is concentrated at the lowest cost
per product for both single and dual sourcing, which indicates that a risk-averse
maximization of service level that aims at reducing the expected worst-case fraction
of delayed and unscheduled customer orders, implicitly reduces the correspond-
ing expected worst-case penalty costs. Figure 5.4 shows that for a single sourcing,
the distribution of service level is concentrated at the highest percent of customer
orders fulfilled by their due dates, because for both objective functions the same
single, reliable supplier (supplier 4) was selected to minimize the worst-case cost
of unfulfilled customer orders or maximize the worst-case customer service level,
respectively For a dual sourcing, however, the probability measure is concentrated at
the highest service level, only if the CVaR of service level is maximized. When the
CVaR of cost is minimized however, the highest probability measure is concentrated

Distribution of Service Level

for Maximum CVaR of Service Level (alpha=0.99)
1

0.8
Probability

0.6

0.4

0.2

0
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Percent of Customer Orders Fulfilled by Due Dates
Single-sourcing Dual-sourcing

Distribution of Service Level

for Minimum CVaR of Cost (alpha=0.99)
1

0.8
Probability

0.6

0.4

0.2

0
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Percent of Customer Orders Fulfilled by Due Dates
Single-sourcing Dual-sourcing

Fig. 5.4 Distribution of customer service level for risk-averse supply portfolio with α = 0.99
5.6 Computational Examples 127

at a lower service level, which indicates that a risk-averse dual sourcing solution
that selects a low price, less reliable supporting supplier (supplier 10 vs. supplier 7,
see Fig. 5.2 and Tables 5.3, 5.4) to reduce the worst-case costs, may simultaneously
decrease the worst-case service levels. For a single sourcing, however, small prob-
ability atoms are concentrated at the highest cost and at the lowest service level
(0% - all customer orders delayed or unscheduled), respectively for the risk-averse
minimization of cost and maximization of service level. This additionally indicates
that for a single sourcing, all customer orders may be either fulfilled by their due
dates or delayed/unscheduled, with a small probability for the latter event, whereas
for a dual sourcing, the probability measure is distributed over a range of cost or
service level outcomes. 
Figure 5.5 presents aggregated demand pattern for products,
  j∈J :d j =t b j ,
t ∈ T , and expected aggregated production schedule, s∈S Ps j∈J b j wsjt , t ∈ T
for the optimal risk-averse supply portfolios, the two objective functions and the two
sourcing strategies. For a dual sourcing and maximum CVaR of service level when
no cost components are included in the objective function, the expected production

Expected Production Schedule

for Minimum CVaR of Cost (alpha=0.99)
25000 25000

20000 20000
Production

Demand
15000 15000

10000 10000

5000 5000

0 0
1 2 3 4 5 6 7 8 9 10
Day

Expected Production (single sourcing) Expected Production (dual sourcing) Demand

Expected Production Schedule

for Maximum CVaR of Service Level (alpha=0.99)
25000 25000

20000 20000
Demand
Production

15000 15000

10000 10000

5000 5000

0 0
1 2 3 4 5 6 7 8 9 10
Day

Expected Production (single sourcing) Expected Production (dual sourcing) Demand

Fig. 5.5 Expected production schedule for risk-averse supply portfolio with α = 0.99
128 5 Integrated Selection of Supply Portfolio and Scheduling of Production

schedule approximately follows the aggregated demand and for a single sourcing and
both objective functions, the production schedules begin with the highest production
level, that may help to reduce the penalty cost or the fraction of delayed customer
orders.
Finally, the impact of penalty costs imposed on the producer by the customers
is evaluated, assuming that a disrupted supplier is contractually obliged to partially
cover the producer charges for unfulfilled customer orders due to undelivered parts.
Table 5.7 shows the solution results for the unit penalty cost of unfulfilled customer
orders, h j , reduced by half with respect to their original values, i.e., for h j approx-
imately as large as the maximum unit price of required parts. Thus, the disrupted
suppliers are assumed to cover the remaining 50% of the producer penalty charges.
Table 5.7 demonstrates that for a lower producer penalty for unfulfilled customer
orders, in a dual sourcing no supporting supplier is needed for the highest confidence
level. For both single and dual sourcing, the same single supplier was selected. This
indicates that the lower the penalty for unfulfilled customer orders, the less reliable
and cheaper supplier can be selected by the producer.
Comparison of single and dual sourcing strategies indicates that for both the risk-
neutral and the risk-averse solutions with a low confidence level, the same single
supplier is selected only; a low price, risky supplier to minimize cost or an expen-
sive, reliable supplier to maximize customer service level. In order to minimize
the expected cost per product or CVaR of cost per product for low confidence lev-
els, the cheapest supplier is usually selected. In contrast, to maximize the expected
service level or CVaR of service level for low confidence levels, the most reliable
supplier (with the lowest disruption probability) is mostly selected. For a higher con-
fidence level, both single and dual sourcing model selects a single, reliable supplier to
minimize worst-case cost of unfulfilled customer orders or maximize worst-case cus-
tomer service level. A difference between single and dual sourcing solutions arises

Table 5.7 Risk-averse solutions for reduced penalty costs: single versus dual sourcing
Confidence level α 0.50 0.75 0.90 0.95 0.99
Model SPS1_CV(c)
CVaRc 11.52 15.25 23.25 23.52 25.61
VaRc 7.79 7.79 22.99 22.99 22.99
Suppliers Selected 6 6 4 4 4
Expected Cost 9.65 9.65 23.02 23.02 23.02
Expected Service Level (a) 61.62 59.46 99.35 99.35 99.35
Model SPS2_CV(c)
CVaRc 11.52 15.25 23.25 23.52 25.61
VaRc 7.79 7.79 22.99 22.99 22.99
Suppliers Selected 6 6 4 4 4
Expected Cost 9.65 9.65 23.02 23.02 23.02
Expected Service Level (a) 61.62 59.46 99.35 99.35 99.35
(a) (
  
t∈T :t≤d j w jt /J )100%
P s
s∈S s j∈J
5.6 Computational Examples 129

only for the highest confidence levels. Then, for a dual sourcing and the risk-averse
solutions with the highest confidence levels, an expensive, reliable supplier selected
for lower confidence levels is additionally supported with a low price, risky supplier
to allocate the total demand for parts between the two suppliers. However, the selec-
tion of the supporting risky supplier depends on the objective function: a cheaper and
less reliable supplier is selected to reduce the risk of high costs and a more expensive
and more reliable supplier is selected to reduce the risk of low service level.
In general, the computational results indicate that the supplier reliability is a key
selection parameter. In order to maximize service level the most reliable supplier is
selected as the main one (for a dual sourcing) or the only one (for a single sourcing),
whereas to minimize cost, the cheapest one is selected from among most reliable
suppliers, respectively.
For the limited number of scenarios considered, the proven optimal solution can
be found, using the CPLEX solver for mixed integer programming. However, in the
proposed models the number of scheduling variables wsjt is O((J )(S)(T )) and the
number of constraints is O((J + T )S), i.e., they grow linearly in the number S of
disruption scenarios and hence exponentially in the number I of suppliers, if all,
S = 2 I , potential scenarios are considered.
The computational experiments were performed using the AMPL programming
language and the CPLEX 12.5 solver on a laptop MacBookPro 6.2 with Intel Core i7
processor running at 2.66 GHz and with 8GB RAM. The CPLEX solver was capable
of finding proven optimal solutions for all examples with CPU time ranging from
several seconds for the cost-based objectives to around 6000 s for the service level
objectives. However, the enhanced models SPS1_CV(sl)+ and SPS2_CV(sl)+ had
to be applied for the risk-averse maximization of service level, otherwise the much
longer CPU time (up to 12000 s) was required to prove optimality.

5.6.2 Single Versus Multiple Sourcing

In this subsection the risk-averse strategies are compared for single and multiple
sourcing. The following parameters used for the example problems are different
from those in Sect. 5.6.1:
• R, the number of geographic regions, was equal to 3, and the subsets of suppliers
were I 1 = {1, 2, 3}, I 2 = {4, 5, 6} and I 3 = {7, 8, 9, 10}, respectively;
• τi , the order preparation and shipping times from suppliers were 2, 3 and 4 time
periods, respectively for suppliers i ∈ I 1 , i ∈ I 2 and i ∈ I 3 ;
• d j , the due dates for customer orders, were integers in {3, . . . , T } drawn from
int(U[3;10]) distribution, for all customer orders j;
• ei , the cost of ordering parts, were integers in {5000, 6000, . . . , 10000}, {10000,
11000, . . . , 15000} and {15000, 16000, . . . , 30000}, respectively for suppliers i ∈
I 1 , i ∈ I 2 and i ∈ I 3 ;
130 5 Integrated Selection of Supply Portfolio and Scheduling of Production

• oi , the unit price of parts purchased from supplier i, was uniformly distributed
over [11,16], [6,11] and [1,6], respectively for suppliers i ∈ I 1 , i ∈ I 2 and i ∈ I 3 ;
• pi , the local disruption probability was uniformly distributed over [0.005,0.01],
[0.01,0.05] and [0.05;0.10], respectively for suppliers i ∈ I 1 , i ∈ I 2 and i ∈ I 3 ,
i.e., the disruption probabilities were drawn independently from U[0.005;0.01],
U[0.01,0.05] and U[0.05;0.10], respectively;
• pr , the regional disruption probability was 0.001, 0.005 and 0.01, respectively for
region r = 1, r = 2 and r = 3;
The computational experiments were performed for the same replication of the
above input data set. The following data set was generated for all test examples:
a = (2, 1, 3, 3, 1, 3, 2, 1, 2, 2, 2, 2, 3, 2, 1, 3, 2, 1, 3, 3, 2, 1, 1, 2, 1);
b = (1, 2, 9, 7, 8, 5, 1, 7, 5, 4, 7, 4, 10, 6, 8, 1, 4, 2, 4, 8, 6, 3, 8, 7, 3) × 500;
The resulting total demand for parts and products is A = 132500 and B = 66000,
respectively.
c = (2, 1, 1, 2, 3, 3, 1, 3, 2, 1, 2, 1, 3, 1, 1, 3, 2, 3, 1, 1, 3, 2, 2, 1, 2);
Ct = 38000, ∀t = 1, . . . , 10;
d = (7, 4, 10, 10, 3, 8, 7, 5, 5, 6, 8, 7, 10, 5, 3, 9, 6, 4, 9, 9, 7, 4, 5, 7, 4);
e = (8, 7, 10, 13, 14, 11, 19, 17, 19, 27) × 1000;
g j = 1∀ j = 1, . . . , 25;
h = (52, 26, 78, 78, 26, 78, 52, 26, 52, 52, 52, 52, 78, 52, 26, 78, 52, 26, 78, 78,
52, 26, 26, 52, 26);
o = (13, 12, 12, 8, 6, 6, 2, 5, 4, 4);

Suppliers Characteristics
14 0.1
0.09
12
0.08
10 0.07
Probability

0.06
Price

8
0.05
6 0.04
4 0.03
0.02
2
0.01
0 0
1 2 3 4 5 6 7 8 9 10
Region 1: suppliers 1 - 3 Region 2: suppliers 4 - 6 Region 3: suppliers 7 - 10

-Price per Part Disruption Probability

Fig. 5.6 Suppliers

p = (0.00513571, 0.00666354, 0.00902974, 0.0356206, 0.040175, 0.0294692,

0.0519967, 0.0827215, 0.0739062, 0.0656449).
5.6 Computational Examples 131

The corresponding disruption probabilities (5.1) of all suppliers are:

π = (0.00613057, 0.00765688, 0.0100207, 0.0404425, 0.0449741, 0.0343219,
0.0614767, 0.0918943, 0.0831672, 0.0749885).
All potential disruption scenarios were considered and to calculate the disruption
probability for each scenario, the formulae (5.2) and (5.3) were applied. The unit price
per part oi and the disruption probability πi , (5.1), of each supplier i ∈ I r , r = 1, 2, 3
are shown in Fig. 5.6. The figure indicates that the most reliable (with the lowest dis-
ruption probability) is supplier 1, the least reliable (with the highest disruption prob-
ability) is supplier 8, the most expensive (with the highest price per part) is supplier 1,
and the cheapest (with the lowest price per part) is supplier 7. Note that geographic
regions are numbered in such a way that the unit prices are nonincreasing with r ,
while the fixed ordering costs and the disruption probabilities are nondecreasing
with r , i.e.,

oi 1 ≥ oi 2 ≥ oi 3 , ei 1 ≤ ei 2 ≤ ei 3 and πi 1 ≤ πi 2 ≤ πi 3 ; ∀i 1 ∈ I 1 , i 2 ∈ I 2 , i 3 ∈ I 3 .

The solution results for the two sourcing strategies and the two objective functions
are presented in Tables 5.8 and 5.9, respectively. For the risk-averse maximization of
service level, models SPS1_CV(sl) and SPSm_CV(sl) were enhanced with the valid
inequalities (5.29)–(5.31). The confidence level α is set at five levels of 0.5, 0.75,
0.90, 0.95, and 0.99, which means that focus is on minimizing the highest (maxi-
mizing the lowest) 50%, 25%, 10%, 5%, and 1% of all scenario outcomes, i.e., costs
per product (service levels, respectively). In addition to the optimal solution values
for the primary objective functions and the allocation of demand among the selected
suppliers, Tables 5.8 and 5.9 present the associated values of the other objective func-
tion. In particular, the expected cost and the expected service level associated with
the optimal risk-averse solutions are presented to evaluate an average performance
of the supply chain.
For the minimization of worst-case cost, Table 5.8 indicates that for low confidence
levels α, the solution results for single and multiple sourcing are identical, while for
the higher levels a subset of suppliers is selected for the multiple sourcing. For a
single sourcing and low confidence level α = 0.5, 0.75, the cheapest supplier i = 7
is selected, then for a higher α, more reliable and expensive suppliers are chosen
and finally, the most reliable and expensive supplier i = 1 is selected for α = 0.99.
For a multiple sourcing and the high confidence levels α = 0.9, 0.95, 0.99, the total
demand is allocated among five or six suppliers, including the cheapest one.
For the maximization of worst-case service level, where the supplier selection is
independent on any cost parameters and the solution depends only on the distribu-
tion of disruption probabilities, Table 5.9 demonstrates that for a single sourcing, the
same most reliable supplier i = 1 is selected for all confidence levels. For a multi-
ple sourcing, however, and the confidence level α = 0.5, 0.75, 0.9, similar optimal
solutions were found with the total demand for parts allocated among the three most
reliable and most expensive suppliers i = 1, 2, 3 in region r = 1, while to reduce
132 5 Integrated Selection of Supply Portfolio and Scheduling of Production

Table 5.8 Risk-averse minimization of cost: single versus multiple sourcing

Confidence level α 0.50 0.75 0.90 0.95 0.99
Model SPS1_CV(c): Var. = 202095, Bin. = 201060, Cons. = 43717, Nonz. = 1754761 (a)
CVaRc 10.60 16.47 26.07 28.50 42.22
VaRc 4.73 4.73 12.33 24.20 26.22
Supplier Selected 7 7 6 2 1
Expected Cost 7.66 7.66 13.70 24.42 26.38
Expected Service Level (b) 67.55 67.55 88.84 99.23 99.39
CPU(c) 969 416 1143 613 2456
Model SPSm_CV(c): Var. = 202095, Bin. = 201060, Cons. = 43717, Nonz. = 1754761 (a)
CVaRc 10.60 16.47 23.53 26.51 30.74
VaRc 4.73 4.73 19.00 23.33 29.27
Suppliers Selected(% of total demand) 7 7 2(26.04) 2(21.89)
3(20.38)
4(17.35) 4(20.38)
5(17.73) 5(18.87) 5(15.47)
6(18.12) 6(18.87) 6(15.47)
7(14.72) 7(15.85) 7(12.83)
9(16.23)
10(15.85) 10(13.96)
Expected Cost 7.66 7.66 16.60 22.21 26.65
Expected Service Level (b) 67.55 67.55 76.03 38.22 31.43
CPU(c) 183 1035 2600 1965 13640
(a) Var. = number of variables, Bin. = number of binary variables,

Cons.
(b) (
= number
 of constraints,
 Nonz. = number of nonzero coefficients
t∈T :t≤d j w jt /J )100%
s
s∈S Ps j∈J
(c)
CPU seconds for proving optimality on a MacBookPro 6.2, Intel Core i7, 2.66GHz, RAM
8GB/CPLEX 12.5

the worst-case service level outcomes for the high confidence level α = 0.95 and
α = 0.99, total demand is allocated among seven and ten suppliers, respectively,
including the most reliable i = 1, 2, 3.
Figures 5.7 and 5.8 present the distribution of cost per product and the distribu-
tion of customer service level for the two sourcing strategies and the confidence level
α = 0.99. For a single sourcing and the highest confidence level α = 0.99, the high-
est probability measure of 0.9977 is concentrated at 100% of customer orders fulfilled
by their due dates (see, Fig. 5.8), because for both objective functions, the same, most
reliable supplier i = 1 was selected to minimize the worst-case cost of unfulfilled cus-
tomer orders or maximize the worst-case customer service level, respectively. Both
the distribution of cost and the distribution of service level contain also large proba-
bility atoms of 0.0061 at the highest cost, 52.32 (Fig. 5.7) and at the lowest service
level, 0% (Fig. 5.8). Similar results are not observed for a multiple sourcing, which
eliminates the probability atoms at the highest cost and at the lowest service level. As
5.6 Computational Examples 133

Table 5.9 Risk-averse maximization of service level: single versus multiple sourcing
Confidence level α 0.50 0.75 0.90 0.95 0.99
Model SPS1_CV(sl): Var. = 202095, Bin. = 201060, Cons. = 43727, Nonz. = 1649427 (a)
CVaRsl % 98.76 97.54 93.86 87.73 38.68
VaRsl % 100 100 100 100 100
Supplier Selected 1 1 1 1 1
Expected Service Level (b) 99.39 99.39 99.39 99.39 99.39
Expected Cost 26.38 26.38 26.38 26.38 26.38
Model SPSm_CV(sl): Var. = 202095, Bin. = 201060, Cons. = 43727, Nonz. = 1649427 (a)
CVaRsl % 99.21 98.47 96.22 92.45 86.19
VaRsl % 100 100 100 100 92
Suppliers Selected(% of total demand) 1(44.15) 1(47.92) 1(47.92) 1(47.92) 1(21.51)
2(37.74) 2(30.57) 2(30.57) 2(30.57) 2(21.51)
3(18.11) 3(21.51) 3(21.51) 3(21.51) 3(14.34)
4(6.79)
5(7.17)
6(7.17)
7(7.17)
9(7.17)
10(7.17)
Expected Service Level (b) 99.61 99.62 99.62 99.62 95.64
Expected Cost 25.68 25.64 25.64 25.66 22.76
(a) Var. = number of variables, Bin. = number of binary variables,

Cons.
(b) (
= number
 of constraints,
 Nonz. = number of nonzero coefficients
t∈T :t≤d j w jt /J )100%
s
s∈S Ps j∈J

a consequence, the corresponding optimal solution values are much better than for a
single sourcing (worst-case cost 30.74 vs. 42.22 and worst-case service level 86.19%
vs. 38.68%). On the other hand, for a multiple sourcing and the highest confidence
level α = 0.99, the optimal solution with the minimum, C V a R c = 30.74, simulta-
neously produces the lowest expected service level, E sl = 31.43%, (see, Table 5.8),
while the optimal solution with the maximum, C V a R sl = 86.19%, generates the
average expected cost E c = 22.76 (see, Table 5.9). The above conflicting results
for a multiple sourcing are not observed for lower confidence levels α, for which
additional supporting suppliers are not selected.
It should be pointed out that the cost and the service level objectives are in con-
flict, which is clearly indicated by the results in Tables 5.8 and 5.9. The expected
service level is much lower for the cost-based objective function (Table 5.8), and
vice versa the expected cost is much higher for the service level-based objective
function (Table 5.9). The low-cost and low-reliability suppliers dominate the cost-
based supply portfolios, while the high-reliability and high-cost suppliers dominate
the service level-based portfolios. Furthermore, the average and the worst-case per-
formance measures and the corresponding optimal solutions are also in conflict. The
134 5 Integrated Selection of Supply Portfolio and Scheduling of Production

Distribution of Cost
for Minimum CVaR of Cost (alpha=0.99)
1

0.8
Probability
0.6

0.4

0.2

0
10 20 30 40 50 60 70 80 90 100
Cost per Product
Single-sourcing Multiple-sourcing

Distribution of Cost
for Maximum CVaR of Service Level (alpha=0.99)
1

0.8
Probability

0.6

0.4

0.2

0
10 20 30 40 50 60 70 80 90 100
Cost per Product

Single-sourcing Multiple-sourcing

Fig. 5.7 Distribution of cost per product for risk-averse solution with α = 0.99

higher the confidence level α, the more risk aversive the decision-making and the
smaller percent of the highest cost (or the lowest service level) outcomes is focused
on. As a result, the average outcomes are getting neglectful and the average per-
formance associated with the optimal risk-averse solution may become worse. The
optimal risk-averse solution with a low worst-case cost may produce a high average
cost. Similarly, the optimal risk-averse solution with a high worst-case service level,
may yield a low average service level. For a high confidence level α, the impact of
disruption risks is usually mitigated by diversification of the supply portfolio, i.e.,
by selecting of more suppliers. In particular, more expensive or for maximization
of demand fulfillment rate less reliable supporting suppliers may be added to the
risk-averse supply portfolio, which deteriorates the associated average performance
measures, i.e., the expected values of cost or service level.
As an illustrative
 example, Fig. 5.9 presents the cumulative demand pattern for

products, t  ∈T :t  ≤t j∈J :d j =t  b j , t ∈ T , and the expected cumulative production
  
schedules, s∈S Ps t  ∈T :t  ≤t j∈J b j wsjt  , t ∈ T for the optimal risk-averse solu-
tions with the confidence level α = 0.9. Figure 5.9 indicates that for a minimum
worst-case cost objective, the demand of customers is met with a small expected
fraction of the rejected demand (0.034) for the single sourcing, while for the mul-
5.6 Computational Examples 135

Distribution of Service Level

for Maximum CVaR of Service Level (alpha=0.99)
1

Probability 0.8

0.6

0.4

0.2

0
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Percent of Customer Orders Fulfilled by Due Dates
Single-sourcing Multiple-sourcing

Distribution of Service Level

for Minimum CVaR of Cost (alpha=0.99)
1

0.8
Probability

0.6

0.4

0.2

0
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Percent of Customer Orders Fulfilled by Due Dates
Single-sourcing Multiple-sourcing

Fig. 5.8 Distribution of customer service level for risk-averse solution with α = 0.99

tiple sourcing, the expected fraction is much greater (0.122). The expected rejected
demand, however, is much smaller for the worst-case service level objective and both
sourcing strategies, when no cost components are included in the objective function
and more reliable suppliers are selected. Then, the expected production follows the
demand pattern with a very small expected fraction of the rejected demand, 0.0061
and 0.0038, respectively for single and multiple sourcing.
The risk-averse solutions have been compared for the two sourcing strategies and
the two objective functions. Some of the basic features of the obtained solutions are
discussed below (see, Table 5.10).
For the worst-case cost objective function and a single sourcing, the higher the
confidence level, the more reliable supplier is selected to reduce a higher risk of
penalty cost for unfulfilled customer orders. In particular, to minimize the expected
worst-case cost per product for low confidence levels, the cheapest suppliers are
usually selected. For the worst-case cost objective function and low confidence levels,
the optimal solutions for a single and a multiple sourcing are identical. A difference
136 5 Integrated Selection of Supply Portfolio and Scheduling of Production

Expected Production Schedule

for Minimum CVaR of Cost (alpha=0.9)
70000 70000

Cumulative Production
60000 60000

Cumulative Demand
50000 50000

40000 40000

30000 30000

20000 20000

10000 10000

0 0
1 2 3 4 5 6 7 8 9 10
Period

Cumulative Production (single sourcing) Cumulative Production (multiple sourcing)

Cumulative Demand

Expected Production Schedule

for Maximum CVaR of Service Level (alpha=0.9)
70000 70000
Cumulative Production

60000 60000

Cumulative Demand
50000 50000

40000 40000

30000 30000

20000 20000

10000 10000

0 0
1 2 3 4 5 6 7 8 9 10
Period

Cumulative Production (single sourcing) Cumulative Production (multiple sourcing)

Cumulative Demand

Fig. 5.9 Expected cumulative production for risk-averse solution with α = 0.9

Table 5.10 Single versus multiple sourcing solution

Objective function Single sourcing Multiple sourcing
Worst-case cost The higher is the confidence For low confidence levels, the
level, the more reliable supplier solution for a single sourcing
is selected
Worst-case service level The most reliable supplier is The higher is the confidence
selected for all confidence levels level the more suppliers are
selected

between single and multiple sourcing solutions arises only for high confidence levels,
whereas for the worst-case service level objective, the corresponding single and
multiple sourcing solutions are different for all confidence levels.
5.6 Computational Examples 137

In order to maximize the expected worst-case service level when no cost compo-
nents are considered in the objective function, the most reliable supplier (with the
lowest disruption probability) is selected for a single sourcing, even for low con-
fidence levels. For a multiple sourcing and the worst-case service level, the total
demand for parts is allocated among more suppliers for each confidence level to
reduce the risk of unfulfilled customer orders. For high confidence levels and both
cost and service level objectives, more suppliers are selected to mitigate the impact of
disruption risks by diversification of the supply portfolio. The most reliable suppliers
are selected to minimize the expected worst-case cost of unfulfilled customer orders
or to maximize the expected worst-case customer service level. Furthermore, for the
expected worst-case service level objective and both sourcing strategies, the expected
production schedule approximately follows the product demand pattern with a very
small expected fraction of rejected demand.
Comparison of the probability mass functions for the two sourcing strategies
and the two objective functions has indicated that a multiple sourcing strategy bet-
ter shapes the distribution of cost or customer service level. The multiple sourcing
eliminates large probability atoms at the highest cost or at the lowest service level
that otherwise may occur, if a single sourcing strategy is applied. Thus, a multiple
sourcing strategy better mitigates the risk of high costs or low service levels.
The computational experiments were performed using the AMPL programming
language and the CPLEX 12.5 and Gurobi 5.1 solvers on a laptop MacBookPro
6.2 with Intel Core i7 processor running at 2.66 GHz and with 8GB RAM. The
CPLEX solver outperformed the Gurobi solver when the worst-case cost was mini-
mized, while the Gurobi solver outperformed the CPLEX solver when the worst-case
service level was maximized. The solvers were capable of finding proven optimal
solutions for all examples with CPU time ranging from several seconds to over four
hours. However, if valid inequalities (5.29)–(5.31) were not added to tighten models
SPS1_CV(sl) and SPSm_CV(sl) for optimization of the worst-case service level,
then CPU time required to find proven optimal solutions for the example problems
was up to 15% and up to 6% longer, respectively.

5.6.3 Order Versus Demand Fulfillment Rate

This subsection focuses on comparison of the two alternative risk-averse service level
measures: the expected worst-case order fulfillment rate and the expected worst-case
demand fulfillment rate. In the risk-averse models for maximization of service level,
CVaRsl , for a given confidence level, α, is represented by an auxiliary function
(5.25) introduced by Rockafellar and Uryasev (2000), where the probability of tail
distribution (i.e., the probability of outcomes with worst-case service level below
VaRsl ) is fixed to(1 − α). However, the actual probability of the tail distribution
of service level, s∈S:Ss >0 Ps , can be less than (1 − α). As a result, the optimized
value of CVaRsl (5.25) can be greater than:
138 5 Integrated Selection of Supply Portfolio and Scheduling of Production

- the actual expected worst-case fraction of customer orders fulfilled on time (e.g.,
model SPSm_CV(sl))
  
s∈S:Ss >0 Ps j∈J t∈T :t≤d j wsjt /J
 , (5.37)
s∈S:Ss >0 Ps

or
- the actual expected worst-case fraction of demand fulfilled on time (e.g., model
SPSm_CV(sl) with (5.26) replaced by (5.28))
  
s∈S:Ss >0 Ps j∈J t∈T :t≤d j b j wsjt /B
 . (5.38)
s∈S:Ss >0 Ps

The above results are typical for the scenario-based optimization under uncer-
tainty, where the probability measure is concentrated in finitely many points, called
“probability atoms”, i.e., singletons which have positive probability measure. The
smaller is the number of concentration pointsand the greater are probability atoms,
the smaller than 1 − α can be the probability, s∈S:Ss >0 Ps , of outcomes with service
level below VaR, e.g., Sawik (2011b), see also Chap. 2.
The solution results for the two models SPSm_CV(sl) and SPSm_CV(sl) with
(5.26) replaced by (5.28) are compared in Table 5.11. For the service level defined
by order fulfillment rate (model SPSm_CV(sl)), Table 5.11 demonstrates that for
the confidence level α = 0.5, 0.75, 0.9, 0.95, similar optimal solutions were found
with the total demand for parts allocated among the three most reliable suppliers
i = 1, 2, 3 in region r = 1, and for the highest confidence level α = 0.99, total
demand is allocated among nine suppliers, except for the least reliable supplier i = 8.
For the service level defined by demand fulfillment rate (model SPSm_CV(sl) with
(5.26) replaced by (5.28)), Table 5.11 demonstrates that for the confidence level
α = 0.5, 0.75, 0.9, 0.95, similar optimal solutions were found with the total demand
for parts allocated among the two most reliable suppliers i = 1, 2 in region r = 1,
while to reduce the worst-case service level outcomes for the highest confidence
level α = 0.99, total demand is allocated among all ten suppliers.
Figure 5.10 presents the distribution of demand fulfillment rate, for the two objec-
tive functions of model SPSm_CV(sl) and the two confidence levels α = 0.9 and
α = 0.99. The probability mass functions are concentrated in a few points, which is
typical for the scenario-based optimization under uncertainty.
For α = 0.9, the distribution of service level is similar for both objective func-
tions. The highest probability measure, 0.987, is concentrated at 100% of customer
demand fulfilled by due dates. The most reliable suppliers i = 1, 2 and i = 1, 2, 3
were selected to maximize the expected worst-case service level, for maximization
of demand fulfillment rate and order fulfillment rate, respectively. However, the dis-
tribution of service level contains also large probability atoms at lower service levels:
0.012 at 70% and 0.001 at the lowest service level, 0%, for maximization of demand
fulfillment rate, and 0.009 at 90% 0.006 at 80%, 0.005 at 70% and 0.001 at 0%, for
5.6 Computational Examples 139

Table 5.11 Order versus demand fulfillment rate: risk-averse multiple sourcing
Confidence level α 0.50 0.75 0.90 0.95 0.99
Model SPSm_CV(sl) for order fulfillment rate
CVaRsl % 99.21 98.47 96.22 92.45 86.19
VaRsl % 100 100 100 100 92
Suppliers Selected(% of total demand) 1(44.15) 1(47.92) 1(47.92) 1(47.92) 1(21.51)
2(37.74) 2(30.57) 2(30.57) 2(30.57) 2(21.51)
3(18.11) 3(21.51) 3(21.51) 3(21.51) 3(14.34)
4(6.79)
5(7.17)
6(7.17)
7(7.17)
9(7.17)
10(7.17)
Expected Service Level (a) 99.61 99.62 99.62 99.62 95.64
Expected Cost 25.68 25.64 25.64 25.66 22.76
Expected Fulfilled Demand (c) 99.44 99.45 99.45 99.45 95.86
Expected Worst-Case Fulfilled Orders (5.37) 81.93 82.45 82.49 82.58 83.35
Model SPSm_CV(sl) for demand fulfillment rate
CVaRsl % 98.97 97.98 95.00 90.05 78.82
VaRsl % 100 100 100 100 84.85
Suppliers Selected(% of total demand) 1(50.19) 1(50.19) 1(51.32) 1(51.32) 1(21.13)
2(49.81) 2(49.81) 2(48.68) 2(48.68) 2(16.98)
3(16.60)
4(5.66)
5(5.66)
6(11.32)
7(5.66)
8(5.66)
9(5.66)
10(5.66)
Expected Service Level (b) 99.49 99.49 99.50 99.50 87.90
Expected Cost 25.52 25.52 25.52 25.54 24.90
Expected Fulfilled Demand (c) 99.49 99.49 99.50 99.50 87.90
Expected Worst-Case Fulfilled Demand (5.38) 60.70 60.84 60.94 61.03 75.56
(a) (
  
t∈T :t≤d j w jt /J )100%
s
s∈S Ps j∈J
(b) (
  
Ps b j wsjt /B)100%
(c) (
 s∈S  j∈J  t∈T :t≤d j s
s∈S Ps j∈J t∈T b j w jt /B)100%
140 5 Integrated Selection of Supply Portfolio and Scheduling of Production

(a) Distribution of Customer Service Level

1

Probability 0.8

0.6

0.4

0.2

0
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

Demand Fulfillment Rate

alpha=0.9 alpha=0.99

(b) Distribution of Customer Service Level

1

0.8
Probability

0.6

0.4

0.2

0
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

Demand Fulfillment Rate

alpha=0.9 alpha=0.99

Fig. 5.10 Distribution of demand fulfillment rate: a model SPSm_CV(sl) for demand fulfillment
rate, b model SPSm_CV(sl) for order fulfillment rate

maximization of order fulfillment rate. Thus, for α = 0.9, the solution maximizing
demand fulfillment rate with a greater probability atom at 70% may be outperformed
by the solution maximizing order fulfillment rate.
For α = 0.99, the distribution of service level depends on the objective function
selected. For maximization of demand fulfillment rate, the highest probability mea-
sure of 0.71 is concentrated at 100% and 0.29 at 90% of customer demand fulfilled
by their due dates, while for maximization of order fulfillment rate, 0.938 is concen-
trated at 90%, 0.0562 at 100% and 0.005 at 80% of customer demand fulfilled by due
dates. However, for α = 0.99, no probability atoms occur at lower service levels.
5.6 Computational Examples 141

For a high confidence level α, the impact of disruption risks is usually mitigated by
diversification of the supply portfolio, i.e., by selecting more suppliers. In particular,
for maximization of demand fulfillment rate less reliable supporting suppliers may be
added to the risk-averse supply portfolio, which deteriorates the associated average
performance measure, i.e., the expected value of service level.

5.6.4 Conditional Service-At Risk Versus Expected Cost

In this subsection the impact on the cost in the process of optimization of worst-
case service level is illustrated with additional computational examples. The cost
includes the cost of ordering and purchasing of parts plus penalty cost of delayed and
unfulfilled customer orders due to parts shortages. For example, the impact on the cost
in the process of optimization of worst-case service level is illustrated in Table 5.11
with the expected cost per product, E c , associated with the risk-averse solution for
model SPSm_CV(sl). The expected costs associated with the risk-averse solutions
for α = 0.5, 0.75, 0.9, 0.95 are similar. To maximize the worst-case service level,
the parts are purchased from high-reliability and high-cost suppliers only. As a result
the purchasing cost dominates, while the cost of parts shortage is not significant.
However, for the highest confidence level α = 0.99, when the supply portfolio is
most diversified to mitigate the impact of worst 1% of all scenarios outcomes, the
orders for parts are also placed on low-cost and low-reliability suppliers. The resulting
total purchasing cost decreases as well as the resulting total expected cost per product,
E c = 22.76 for model SPSm_CV(sl) and E c = 24.90 for model SPSm_CV(sl) with
(5.26) replaced by (5.28).
Table 5.12 presents a subset of nondominated solutions for the mixed mean-risk
model SPSm_E(c)CV(sl) and confidence level α = 0.9, obtained for a subset of
trade-off parameter, λ = 0, 0.01, 0.25, 0.5, 0.75, 0.99, 1. Note that a single-objective
solution with maximum conditional service-at-risk (cf. Table 5.11) and with mini-
mum expected cost is obtained for λ = 0 and λ = 1, respectively.
Table 5.12 clearly shows that the two objective functions: expected worst-case
service level and expected cost are in conflict. The higher the trade-off parameter λ,
the more cost-oriented the decision-making and the lower the expected cost, E c , the
lower the conditional service-at-risk, C V a R sl , as well the associated expected ser-
vice level for both measures: order fulfillment rate and demand fulfillment rate. As λ
increases from 0 to 1, the supply portfolio is changing from the one based on a subset
of most reliable and expensive suppliers, i = 1, 2, 3 (model SPSm_E(c)CV(sl)) and
i = 1, 2 (model SPSm_E(c)CV(sl) with (5.26) replaced by (5.28)) for λ = 0, 0.01,
followed with more diversified subsets of less reliable and cheaper suppliers selected
for λ = 0.25, 0.5, 0.75, through the single, cheapest supplier, i = 7, for λ = 0.99, 1.
While CVaRsl and expected service level, E sl , are significantly decreasing with
the trade-off parameter λ, the associated expected fraction of fulfilled (on time
or delayed) demand decreases at the much lower rate. This indicates that as λ
142 5 Integrated Selection of Supply Portfolio and Scheduling of Production

Table 5.12 Nondominated solutions for mean-risk model SPSm_E(c)CV(sl): α = 0.9

Trade-off parameter λ 0 0.01 0.25 0.50 0.75 0.99 1
Model SPSm_E(c)CV(sl) for order fulfillment rate
Ec 25.64 25.50 14.55 12.94 8.49 7.77 7.66
CVaRsl % 96.22 96.16 88.19 84.88 53.63 28.94 27.66
VaRsl % 100 100 92 92 92 76 72
Suppliers Selected(% of total demand) 1(47.92) 1(40.75)
2(30.57) 2(37.74) 2(7.55) 2(6.04)
3(21.51) 3(21.51) 3(6.41)
5(21.51) 5(20.38) 5(4.90)
6(21.51) 6(21.51) 6(5.67)
7(21.51) 7(30.56) 7(89.43) 7(100) 7(100)
10(21.51) 10(21.51)
Expected Service Level (a) 99.62 99.62 98.02 97.74 88.16 71.29 67.56
Expected Fulfilled Demand (c) 99.45 99.44 96.70 96.31 93.94 93.82 93.85
Model SPSm_E(c)CV(sl) for demand fulfillment rate
Ec 25.52 25.52 18.52 13.11 8.23 7.67 7.66
CVaRsl % 95.00 95.00 85.79 77.44 43.02 29.91 28.08
VaRsl % 100 100 90.91 84.85 87.88 78.79 73.48
Suppliers Selected(% of total demand) 1(51.32) 1(50.19) 1(4.91)
2(48.68) 2(49.81) 2(16.98) 2(6.04)
3(13.58)
4(13.21)
5(13.58) 5(22.64)
6(13.58) 6(21.51) 6(10.57)
7(13.21) 7(27.17) 7(89.43) 7(100) 7(100)
9(10.95) 9(22.64)
Expected Service Level (b) 99.50 99.50 97.27 96.26 83.39 73.90 68.96
Expected Fulfilled Demand (c) 99.50 99.50 97.27 96.26 94.38 93.84 93.85
(a)
  
( s∈S Ps j∈J t∈T :t≤d j wsjt /J )100%
(b) (
  
s∈S Ps b j wsjt /B)100%
(c) (
  j∈J
 t∈T :t≤d j s
s∈S Ps j∈J t∈T b j w jt /B)100%

approaches 1, the fulfilled demand becomes more delayed and the resulting service
level decreases.
Examples of nondominated schedules for the confidence level α = 0.9 and
the trade-off parameter λ = 0.5 are shown in Fig. 5.11. Figure 5.11 compares the
expected cumulative production with the expected worst-case cumulative production
for model SPSm_E(c)CV(sl) maximizing either demand fulfillment rate or order ful-
fillment rate. The corresponding production schedules are similar for both objectives
with a slightly greater expected and expected worst-case cumulative production for
maximization of demand fulfillment rate.
Overall, the obtained solution results for the service level objectives are in line
with the other approaches used in the area of supply chain risk management. For
5.6 Computational Examples 143

(a) Expected and Expected Worst-Case Production Schedule

(alpha=0.9, lambda=0.5)
70000 70000

Cumulative Production 60000 60000

Cumulative Demand
50000 50000

40000 40000

30000 30000

20000 20000

10000 10000

0 0
1 2 3 4 5 6 7 8 9 10
Period
Expected Cumulative Production- Expected Worst-Case Cumulative Production

Cumulative Demand-

(b) Expectedand Expected Worst-Case Production Schedule

(alpha=0.9, lambda=0.5)
70000 70000

60000 60000
Cumulative Production

Cumulative Demand
50000 50000

40000 40000

30000 30000

20000 20000

10000 10000

0 0
1 2 3 4 5 6 7 8 9 10
Period
Expected Cumulative Production Expected Worst-Case Cumulative Production
Cumulative Demand

Fig. 5.11 Expected and expected worst-case cumulative production for α = 0.9 and λ = 0.5:
a model SPSm_E(c)CV(sl) for demand fulfillment rate, b model SPSm_E(c)CV(sl) for order
fulfillment rate

example, the higher the confidence level, the more diversified the supply portfolio
and the smaller the probability atoms at the lowest service level, i.e., the risk of
low service level is more efficiently mitigated. Furthermore, the two service level
measures are related in magnitude, while the worst-case service level and the expected
cost are in opposition. Some additional properties of the obtained optimal risk-averse
solutions are listed and discussed below.
144 5 Integrated Selection of Supply Portfolio and Scheduling of Production

• Given confidence level α, the worst-case order fulfillment rate shows a higher ser-
vice performance than the worst-case demand fulfillment rate. The supply portfolio
is more diversified and the expected worst-case fraction of fulfilled orders is greater
for most confidence levels.
• The expected service level exceeds the corresponding expected worst-case service
level (conditional service-at-risk), i.e., the average performance is better than the
worst-case performance of a supply chain.
• For most confidence levels α, the total demand for parts is allocated among the
most reliable suppliers only. For the highest level α = 0.99, however, the impact of
disruption risks is usually mitigated by diversification of the supply portfolio, i.e.,
by selecting more suppliers. In particular, less reliable supporting suppliers may
be added to the risk-averse supply portfolio, which deteriorates the associated
average performance measure, i.e., the expected service level.
• For both service level measures, the expected production schedule approximately
follows the product demand pattern, with a relatively small fraction of unfulfilled
customer demand for low confidence levels and greater for the highest confidence
level. However, the expected worst-case production depends on the optimized ser-
vice level measure and the confidence level. The largest unfulfilled demand has
been observed for low confidence levels and model SPSm_CV(sl) for maximiza-
tion of demand fulfillment rate, for which the supply portfolio is less diversified
than the corresponding portfolio for maximization of order fulfillment rate.
• Comparison of solution results for the two service level measures indicates that
maximization of the expected worst-case order fulfillment rate may better mitigate
the impact of disruption risks. For most confidence levels, the expected worst-case
order fulfillment rate is greater than the expected worst-case demand fulfillment
rate.
Comparison of optimal CVaRsl (5.25), with the actual expected worst-case service
level, (5.37) and (5.38), clearly demonstrates that the optimal value of the auxiliary
objective function (5.25) used to maximize conditional service-at-risk, overestimates
the actual value of CVaRsl calculated using (5.37) and (5.38), which is typical for
the scenario-based optimization under uncertainty, when the probability measure is
concentrated in finitely many points.
The computational experiments were performed using the AMPL programming
language and Gurobi 6.0 solver on a laptop MacBookPro with Intel Core i7 processor
running at 2.8GHz and with 16GB RAM. The solver was capable of finding proven
optimal solutions for all examples with CPU time ranging from a few seconds to
around one hour. CPU time increases with the confidence level α and approaches
one hour for α = 0.99, when the supply portfolio is most diversified to mitigate the
impact of worst 1% of all scenarios outcomes. Given confidence level, the compu-
tational effort required to find proven optimal solution is smaller for maximization
of demand fulfillment rate than the order fulfillment rate, for which the size of cus-
tomer order is not considered and hence all orders are equally important. For model
5.6 Computational Examples 145

SPSm_E(c)CV(sl), given confidence level, the greatest CPU time is required for
medium values of the trade-off parameter λ = 0.25, 0.5, 0.75, when both criteria are
significant.

5.7 Notes

As firms expand their business globally, their supply chains involve more global
partners. According to an empirical study conducted by Shin et al. (2000) and
Sounderpandian et al. (2008) dual or multiple sourcing is a common business prac-
tice to mitigate the impact of various operational and disruption risks. In view of the
recent trend of outsourcing and globalization, integrated selection of part suppliers
and allocation of order quantities and scheduling of customer orders may signifi-
cantly improve performance of a multi-echelon supply chain under disruption risks.
However, the research on quantitative approaches to the integrated supplier selec-
tion and customer order scheduling in the presence of supply chain disruption risks
has not been often reported in the literature. Most work on integrated supply chain
scheduling focuses on coordinating the flows of supply and demand over a sup-
ply chain network to minimize the inventory, transportation and shortage costs. For
example, Chen and Vairaktarakis (2005), Chen and Pundoor (2006) and Pundoor and
Chen (2009) studied simplified models for integrated scheduling of production and
distribution operations. The authors have analyzed computational complexity of var-
ious cases of the problem and have developed heuristics for NP-hard cases. Lei et al.
(2006) considered an integrated production, inventory and distribution routing prob-
lem involving heterogeneous transporters with non-instantaneous traveling times
and many capacitated customer demand centers. A MIP approach combined with a
heuristic routing algorithm was proposed to coordinate the production, inventory and
transportation operations. Bard and Nananukul (2009) developed a MIP model and
a reactive tabu search-based algorithm for a transportation scheduling problem that
included a single production facility, a set of customers with time-varying demand
and a fleet of vehicles. Wang and Lei (2012) considered the problem of operations
scheduling for a capacitated multi-echelon shipping network with delivery deadlines,
where semi-finished goods are shipped from suppliers to customers through process-
ing centers, with the objective of minimizing the shipping and penalty cost. The three
polynomial-time solvable cases of this problem were reported: with identical order
quantities; with designated suppliers; and with divisible customer order sizes. Liu
and Papageorgiou (2013) developed a multi objective MIP approach to address pro-
duction, distribution and capacity planning of global supply chains considering cost,
responsiveness and customer service level simultaneously. An integrated approach to
deterministic coordinated supply chain scheduling was proposed by Sawik (2007) to
simultaneously schedule manufacturing and supply of parts and assembly of finished
products. Given a set of part suppliers and a set of customer orders for finished prod-
ucts, the problem objective was to determine which orders were provided with parts
by each supplier, to schedule manufacturing of parts at each supplier and delivery
146 5 Integrated Selection of Supply Portfolio and Scheduling of Production

of parts from each supplier to the producer, and to schedule customer orders at the
producer, such that a high customer service level was achieved and the total cost was
minimized. The selection of part supplier for each customer order was combined
with a due date setting for some orders to maximize the number of orders that can be
completed by customer requested due dates. A monolithic MIP model was presented
and compared with a hierarchy of mixed integer programs for a sequential selection
of suppliers and scheduling of manufacturing and delivery of parts and assembly of
products. Different enhancements of the above MIP approach for the coordinated
scheduling in multi-echelon supply chains were presented in Sawik (2009a). Var-
ious perspectives on supply chain coordination issues were reported and reviewed
by Arshinder (2008) and the gaps existing in the literature were identified. Li and
Wang (2007) reviewed coordination mechanisms of supply chain systems in a frame-
work that was based on supply chain decision structure and nature of demand and
a review of methods and literature on supply chain coordination through contracts
was provided by Hezarkhani and Kubiak (2010).
The customer service level in a make-to-order environment was studied by
Altendorfer and Jodlbauer (2011). Shao and Dong (2012) compared and analyzed
order fulfillment performance measures for two different production control systems:
make-to-order versus make-to-stock. They formulated service-maximization models
with inventory cost budget constraints. For the make-to-stock production, Larsen and
Thorstenson (2008, 2014) differentiated between an order fill rate and a volume fill
rate and specified their performance for different inventory control systems. They
showed how the order and volume fill rates are related in magnitude. Bijulal et al.
(2011) analyzed the production-inventory system performance in terms of order fill
rate and average system costs, affected by the control parameters.
The material presented in this chapter is based on the research results reported
in Sawik (2013c, 2014a, b) where single, dual and multiple sourcing strategies
were compared for risk-neutral and risk-averse decision-making. In addition, Sawik
(2016b) compared the two risk-averse service level measures: expected worst-case
order fulfillment rate and expected worst-case demand fulfillment rate. The former
corresponds to order fill rate and the latter to volume fill rate in the make-to-stock
production (Larsen and Thorstenson 2008, 2014). The future research should focus
on comparison of the proposed integrated supplier selection and customer order
scheduling with a common hierarchical approach, where first the supplier selection
and order quantity allocation subject to disruption risks is accomplished and then,
given a schedule of part supplies, the optimal schedule of customer orders is deter-
mined for each disruption scenario, subject to parts availability constraints.

Problems

5.1 Modify the SMIP models presented in this chapter for multiple part types with
subsets of part types required for each customer order and subsets of suppliers avail-
able for each part type.
Problems 147

5.2 Modify the SMIP models presented in this chapter for finite capacity suppliers.

5.3 Mixed mean-risk supply portfolio and scheduling

(a) Formulate model SPSm_E(sl)CV(c) for a mixed mean-risk selection of supply
portfolio and scheduling of customer orders to trade-off expected service level and
CVaR of cost.
(b) How should the values of the optimized objective functions be scaled into the
interval [0,1] to avoid dimensional inconsistency among the two objectives and how
should the trade-off parameter be selected?
(c) How would you interpret the mixed mean-risk solution?

5.4 Explain why order fulfillment rate shows a higher service performance than
demand fulfillment rate.

5.5 Looking into the computational examples try to compare dual versus multiple
solution results.
Chapter 6
Integrated Selection of Supply Portfolio
and Scheduling of Production and
Distribution

6.1 Introduction

The key operational functions in a supply chain are supply, production and distri-
bution operations. To achieve a high performance of supply chain, it is crucial to
integrate these three functions and jointly schedule supply, production and distribu-
tion in a coordinated manner. For example, in customer-driven supply chains, where
customer orders are executed immediately or shortly after arrival of material supplies
and the ordered products are delivered to customers immediately or shortly after their
completion, the impact of disruption risks can be best mitigated when an integrated
decision-making is applied. At the same time, the integrated decision-making allows
reaching various conflicting objectives, such as reduction in total cost and increase in
service level. The purpose of this chapter is to study the integrated decision-making to
simultaneously select suppliers of parts, allocate order quantity and schedule produc-
tion and delivery of finished products to customers in a supply chain under disruption
risks. In addition to supplier selection, order quantity allocation and scheduling of
customer orders, distribution of finished products to customers is simultaneously con-
sidered with different shipping methods to optimize the trade-off between cost and
service level. The service level in this chapter denotes demand fulfillment rate (see,
Sect. 5.3.2). The three different shipping methods will be modelled and compared
for the distribution of products: batch shipping with a single shipment of different
customer orders, batch shipping with multiple shipments of different customer orders
and individual shipping of each customer order immediately after its completion. In
addition, the SMIP formulation based on the wait-and-see approach to optimize the
trade-off between expected cost and expected service level will be compared with a
deterministic MIP model based on the expected value approach, in which random
parameters are replaced by their expected values.

© Springer International Publishing AG 2018 149

T. Sawik, Supply Chain Disruption Management Using Stochastic Mixed
Integer Programming, International Series in Operations Research
& Management Science 256, DOI 10.1007/978-3-319-58823-0_6
150 6 Integrated Selection of Supply Portfolio …

The following time-indexed SMIP and MIP models are presented in this chapter:

SCS1_E for risk-neutral scheduling in a supply chain with single batch ship-
ping of products to distribution centers;
SCS2_E for risk-neutral scheduling in a supply chain with individual shipping
of products for each customer order to distribution centers;
SCS3_E for risk-neutral scheduling in a supply chain with multiple batch
shipping of products to distribution centers;
SCS1_CV for risk-averse scheduling in a supply chain with single batch ship-
ping of products to distribution centers;
SCS2_CV for risk-averse scheduling in a supply chain with individual ship-
ping of products for each customer order to distribution centers;
SCS3_CV for risk-averse scheduling in a supply chain with multiple batch
shipping of products to distribution centers;
ESCS1 for the expected value-based supply chain scheduling, corresponding
to model SCS1_E.

The models incorporate supply-production, production-distribution and supply-

distribution coordinating constraints to efficiently coordinate supply, production and
distribution schedules.
Numerical examples and computational results are reported in Sect. 6.6 for risk-
neutral, risk-averse and expected value - based decision-making.

6.2 Problem Description

Fig. 6.1 A multi-echelon supply chain

6.2 Problem Description 151

A multi-echelon supply chain under disruption risk is studied with multiple suppliers
of a critical part type, producer of one product type, and multiple distribution centers
that provide with products a set of customers (see, Fig. 6.1).
Let I = {1, . . . , I } be the set of I suppliers, J = {1, . . . , J } the set of J customers,
K = {1, . . . , K } the set of K distribution centers, and T = {1, . . . , T } the set of T
planning periods (for notation used, see Table 6.1).

Table 6.1 Notation: supply chain scheduling

Indices
i = supplier, i ∈ I
j = customer order, j ∈ J
k = distribution center, k ∈ K
r = region, r ∈ R
s = disruption scenario, s ∈ S
t = planning period, t ∈ T
Input Parameters
ak = fixed transportation cost of each shipment to distribution center k ∈ K
b j = size (number of products) of customer order j

B = j∈J b j , total demand for products
C P = per period capacity of producer
C V = vehicle capacity
d j = due date for customer order j
ei = fixed cost of ordering parts from supplier i
g j = per unit and per period penalty cost of delayed customer order j
h j = per unit penalty cost of unfulfilled customer order j
I r = subset of suppliers in region r
Jk = subset of customers (customer orders) served by distribution center k
oi = per unit price of parts purchased from supplier i
pi = local disruption probability for supplier i
pr = regional disruption probability for all suppliers in region r
τi = delivery lead time from supplier i
σk = transportation time to distribution center k

Denote by Jk ⊂ J the subset of customers (or equivalently customer orders)

served by the distribution center k ∈ K . The demand of each customer j ∈ Jk is
defined by the number of products, b j and the latest period of their delivery to the
distribution center 
k, d j . The total demand for parts is identical with the total demand
for products, B = j∈J b j . For each customer order j ∈ J , denote by g j and h j , the
unitary delay penalty cost and unfulfillment penalty cost, respectively. The producer
aggregate capacity in each planning period is denoted by C P . The capacity depends
on machine configuration of production facility and the processing time available
on each machine in a single planning period, e.g., Sawik (2009c, 2011a). (In the
152 6 Integrated Selection of Supply Portfolio …

computational examples provided in Sect. 6.6, the production facility is a flexible

flow shop with parallel batch machines.)
Each supplier i ∈ I is characterized by the four parameters: the unit purchasing
price, oi ; the fixed ordering cost, ei ; the constant lead time (manufacturing and
transportation time to producer), τi ; and the local disruption probability, pi , i.e., the
probability that parts ordered from supplier i are not delivered. The parts ordered
from supplier i can be used by the producer not earlier than in period τi + 1.
Three distribution strategies of customer orders will be considered:
• The finished products for all customers j ∈ Jk are delivered to distribution center
k in a single batch shipment, after completion of all scheduled customer orders
j ∈ Jk .
• The finished products for each customer order j ∈ Jk are individually shipped to
distribution center k, immediately after its completion.
• The finished products for customers j ∈ Jk are delivered to distribution center k
in multiple batch shipments, subject to limited transportation capacity.
The transportation time to distribution center k is constant and equals to σk periods
so that the products shipped in period t are delivered in period t + σk − 1.
For the multiple batch shipping strategy, additionally denote by C V , vehicle lim-
ited capacity defined as the maximum number of finished products that can be shipped
together, and by ak , a fixed transportation cost of each shipment to distribution center
k ∈ K.
The above shipping methods can be easily justified from a practical point of view.
The individual and immediate shipping are commonly used for distribution of time-
sensitive products to achieve a fast delivery. The single and multiple batch shipping
methods are used for orders going to the same distribution center, The shipping
methods with unlimited transportation capacity, e.g., infinitely many vehicles, are
applicable when the delivery is handled by a third-party logistics provider that typi-
cally owns a large number of vehicles. On the other hand, the multiple batch shipping
with limited transportation capacity accounts for limited capacity of a single vehicle.
The shipping methods considered do not account for transportation time and costs to
customers since routing (e.g., vehicle routing) is not a part of the decision-making,
and the customers are assumed to be individually served by the corresponding distri-
bution centers. Alternatively, for each distribution center k ∈ K , all customers j ∈ Jk
can be assumed to be co-located.
The suppliers are assumed to be located in R disjoint geographic regions. Denote
by I r ⊆ I the subset of suppliers in region r ∈ R = {1, . . . , R}, where r ∈R I r = I .
The supplies of parts are subject to random local disruptions that are uniquely asso-
ciated with a particular supplier, which may arise from equipment breakdowns, local
labor strike, fires, etc. Denote by pi the local disruption probability for supplier i, i.e.,
the parts ordered from supplier i are delivered without disruptions with probability
(1 − pi ), or not at all with probability pi . In addition to independent local disrup-
tions of each supplier, the supplies of parts are also subject to correlated regional
disruption of all suppliers in the same region simultaneously, with probability pr for
region r ∈ R.
6.2 Problem Description 153

Let S = {1, . . . , S} be the index set of S = 2 I disruption scenarios, where each

scenario s ∈ S defines a subset Is ⊂ I of non-disrupted suppliers. The supplies from
every supplier, i ∈ I \ Is , can be independently disrupted either by a local or by a
regional disaster event. The probability Ps of each disruption scenario s ∈ S is a
product over all regions r ∈ R of probabilities Psr of realizing disruption scenario s
for suppliers in I r ,

Ps = Psr ,
r ∈R

where Psr is (cf. Sect. 1.3)

   
(1 − pr ) i∈I r  Is (1 − pi ) i∈I r \Is pi if I r Is = ∅
Psr =  
pr + (1 − pr ) i∈I r pi if I r Is = ∅.

The problem of the integrated supply, production and distribution scheduling will
be formulated under the following simplified assumptions:
• orders for parts are placed at the start of the planning horizon, when all customer
orders for products are known;
• each supplier capacity is sufficient to meet total demand for parts and to complete
and prepare orders for shipping in a single planning period;
• all parts ordered from a supplier are shipped together in a single delivery;
• customers are assigned to distribution centers ahead of time and each customer is
served by exactly one distribution center;
• each customer order can be completed by the producer in a single planning period;
• customer requested due dates are replaced with delivery due dates for distribution
centers;
• transportation times to distribution centers are constant, and the delivery dates are
determined by the shipping dates to distribution centers;
• transportation times and costs to individual customers are not considered and
vehicle routing is not a part of the decision-making;
• inventory of parts and products are not considered;
• three shipping methods: single batch shipping, multiple batch shipping and indi-
vidual and immediate shipping are considered separately.
Some of the above assumptions can be easily relaxed, while others need more
advanced models to be developed.

6.3 Models for Risk-Neutral Decision-Making

In this section three SMIP models SCS1_E, SCS2_E and SCS3_E are presented for
the integrated supplier selection, order quantity allocation and scheduling produc-
tion of finished products and distribution to customers to minimize expected cost
154 6 Integrated Selection of Supply Portfolio …

per product and maximize expected service level (expected demand fulfillment rate).
Model SCS1_E is formulated for supply chain scheduling with a single batch ship-
ping of finished products to each distribution center for all customers served by that
center. Model SCS2_E accounts for individual and immediate shipment of products
to distribution centers for each customer. In models SCS1_E and SCS2_E trans-
portation cost and capacity are not considered. Model SCS3_E is formulated for
multiple batch shipping of finished products to each distribution center and accounts
for transportation cost and capacity.
The problem variables are defined in Table 6.2.

Table 6.2 Variables: supply chain scheduling

First stage variables
ui = 1, if supplier i is selected; otherwise u i = 0 (supplier selection)
vi ∈ [0, 1], the fraction of total demand for parts ordered from supplier i (supply portfolio)
Second stage variables
wsjt = 1, if under disruption scenario s customer order j is scheduled for period t; otherwise
wsjt = 0 (production scheduling, single and multiple batch shipping)
= 1, if under disruption scenario s customer order j is scheduled for period t and shipped
to distribution center in period t + 1; otherwise wsjt = 0 (production and distribution
scheduling, individual shipping)
s
xkt = 1, if under disruption scenario s, batch shipment of products to distribution center k is
scheduled for period t; otherwise xkt s = 0 (distribution scheduling, single and multiple

batch shipping)
y sj = 1, if under disruption scenario s customer order j is delivered by its due date; otherwise
y sj = 0 (customer order non-delay delivery, single batch shipping)
z jt = 1, if under disruption scenario s customer order j is shipped to distribution center in
s

period t; otherwise z sjt = 0 (distribution scheduling, multiple batch shipping)

Auxiliary variables
VaRc Cost-at-Risk, the targeted cost such that for a given confidence level α, for 100α% of
the scenarios, the outcome is below VaRc
VaR sl Service-at-Risk, the targeted service level such that for a given confidence level α, for
100α% of the scenarios, the outcome is above VaRsl
Cs ≥ 0, the tail cost for scenario s, i.e., the amount by which costs in scenario s exceed VaRc
Ss ≥ 0, the tail service level for scenario s, i.e., the amount by which VaRsl exceeds service
level in scenario s

6.3.1 Scheduling with Single Batch Shipping


The total customer demand, j∈J b j , under each disruption scenario s ∈ S, can be
split into the three satisfaction levels: non-delayed, delayed or unscheduled (rejected).
For a single batch shipping, these portions of the customer demand are defined below.
6.3 Models for Risk-Neutral Decision-Making 155


• Non-delayed,
 j∈Jb j y sj ,
• Delayed, j∈J b j ( t∈T wsjt − y sj ), 
• Unscheduled (rejected), j∈J b j (1 − t∈T wsjt ).

The expected cost per product, E c , and service level, E sl , as well their normalized
values, respectively f c and f sl , are defined below.
  
Ec = ( ei u i + Ps ( Boi vi
i∈I s∈S i∈Is
   
+ gjbj( wsjt − y sj ) + h j b j (1 − wsjt )))/B. (6.1)
j∈J t∈T j∈J t∈T

E sl = Ps b j y sj /B. (6.2)
s∈S j∈J

Ec − Ec
fc = c , (6.3)
E − Ec
c
where E c and E are the minimum and the maximum values of E c , respectively.
sl
E − E sl
f sl
= sl
, (6.4)
E − E sl
sl
where E sl and E are the minimum and the maximum values of E sl , respectively.
A subset of nondominated solutions can be obtained by the parameterization on
λ, the model presented below.

SCS1_E: Risk-neutral Supply Chain Scheduling: single batch shipping

Minimize

λ f c + (1 − λ) f sl , (6.5)

where 0 ≤ λ ≤ 1,
subject to (6.1)–(6.4) and
Supply portfolio selection constraints:

vi = 1 (6.6)
i∈I
vi ≤ u i ; i ∈ I (6.7)
156 6 Integrated Selection of Supply Portfolio …

Customer order scheduling constraints:


wsjt ≤ 1; j ∈ J, s ∈ S (6.8)
t∈T

b j wsjt ≤ C P ; t ∈ T, s ∈ S (6.9)
j∈J

Supply-production coordinating constraints:

  
b j wsjt ≤ B vi ; t ∈ T, s ∈ S (6.10)
j∈J t ∈T :t ≤t i∈Is :τi ≤t−1

Single batch shipping constraints:


xkts ≤ 1; k ∈ K , s ∈ S (6.11)
t∈T K
 
xkts ≥ wsjt ; k ∈ K , j ∈ Jk , s ∈ S (6.12)
t∈T K t∈T
 
xkts ≤ wsjt ; k ∈ K , s ∈ S, (6.13)
t∈T K j∈Jk t∈T

where T K = {mini∈I τi + 2, . . . , T + 1} is the set of shipping periods.

Production-distribution coordinating constraints:
 
t xkts ≥ (t + 1)wsjt ; k ∈ K , j ∈ Jk , s ∈ S (6.14)
t∈T K t∈T

Non-delayed delivery constraints:


y sj ≤ wsjt ; k ∈ K , j ∈ Jk , s ∈ S (6.15)
t∈T :t≤d j −σk

y sj ≤ xkts ; k ∈ K , j ∈ Jk , s ∈ S (6.16)
t∈T K :t≤d j −σk +1
 
wsjt + xkts − 1 ≤ y sj ; k ∈ K , j ∈ Jk , s ∈ S (6.17)
t∈T :t≤d j −σk t∈T K :t≤d j −σk +1

Non-negativity and integrality conditions:

u i ∈ {0, 1}; i ∈ I (6.18)

vi ∈ [0, 1]; i ∈ I (6.19)
6.3 Models for Risk-Neutral Decision-Making 157

wsjt ∈ {0, 1}; j ∈ J, t ∈ T, s ∈ S (6.20)

xkts ∈ {0, 1}; k ∈ K , t ∈ T K , s ∈ S (6.21)
y sj ≥ 0; j ∈ J, s ∈ S. (6.22)

Constraints (6.6)–(6.7) define a feasible supply portfolio, and (6.8)–(6.9), a feasi-

ble assignment of customer orders to planning periods for each disruption scenario.
The supply-production coordinating constraints (6.10) ensure that for each disruption
scenario s, the cumulative demand for required parts of all customer orders scheduled
by period t is not greater than the cumulative supplies by period t − 1 from the non-
disrupted suppliers i ∈ Is . Equation (6.11) ensures that for each disruption scenario,
at most one batch shipment can be scheduled to each distribution center. Constraint
(6.12) denotes that for each disruption scenario, shipment to distribution center k is
scheduled, only if at least one customer order j ∈ Jk is completed, and Eq. (6.13)
that no shipment to distribution center k is scheduled, if no customer order j ∈ Jk is
completed. The production-distribution coordinating constraints (6.14) ensure that
for each disruption scenario, a shipment to distribution center k can be scheduled
only after the latest completion period of scheduled customer orders j ∈ Jk . Finally,
constraints (6.15)–(6.17) denote that for each disruption scenario s ∈ S, customer
order j ∈ Jk can be delivered without delay (i.e., y sj = 1), unless it is scheduled not
later than d j − σk and shipped to distribution center k not later than d j − σk + 1;
otherwise the customer order is delayed or unscheduled (i.e., y sj = 0),
In the above formulation y sj is not restricted to being binary. However, for any
feasible solution satisfying customer due-date meeting constraints (6.15)–(6.17),
y sj ∈ {0, 1}; j ∈ J, s ∈ S. Note that y sj = 1 only if both t∈T :t≤d j −σk wsjt = 1; j ∈

Jk and t∈T K :t≤d j −σk +1 xkts = 1; j ∈ Jk . Then both constraints (6.15) and (6.16)
become y sj ≤ 1, while (6.17) becomes y sj ≥ 1, hence y sj = 1. Otherwise; i.e., if right-
hand side of either (6.15) or (6.16), or of both (6.15) and (6.16) is 0, then y sj = 0.
Since the cost and the service level objectives are in conflict, the minimum and
c
maximum values of expected cost E c , E , and expected customer service level,
sl
E sl , E , are obtained by solving the two mixed integer programs presented below.

Minimize E c , (6.1), subject to (6.6)–(6.22).

Maximize E sl , (6.2), subject to (6.6)–(6.22).

sl
The values E c and E sl are determined by solving the first problem, and E and
c
E , by solving the second problem.
158 6 Integrated Selection of Supply Portfolio …

6.3.2 Scheduling with Individual and Immediate Shipping

In model SCS2_E presented below the variables, xkts ∈ {0, 1}; k ∈ K , t ∈ T K , s ∈

S and y sj ∈ {0, 1}; j ∈ J, s ∈ S are not required any more, however the production
scheduling variable, wsjt ∈ {0, 1}; j ∈ J, t ∈ T, s ∈ S, now becomes a joint produc-
tion and distribution scheduling variable, redefined as follows:
wsjt = 1, if under disruption scenario s customer order j is processed in period t
and shipped to the distribution center in period t + 1; otherwise wsjt = 0.
Accordingly, the objective functions E c , (6.1) and E sl , (6.2) are rewritten as
below.
  
Ec = ( ei u i + Ps ( Boi vi
i∈I s∈S i∈Is
    
+ gjbj( wsjt − wsjt ) + h j b j (1 − wsjt )))/B (6.23)
j∈J t∈T t∈T :t≤d j −σk j j∈J t∈T
 
E =
sl
Ps b j wsjt /B, (6.24)
s∈S j∈J t∈T :t≤d j −σk j

where σk j is the transportation time to the distribution center k that serves customer
j ∈ Jk
In model SCS2_E presented below the objective functions E c , (6.1) and E sl , (6.2)
have been replaced with (6.23) and (6.24), respectively.

SCS2_E: Risk-neutral supply chain scheduling: individual shipping

Minimize (6.5)
subject to (6.3), (6.4), (6.6)–(6.10), (6.18)–(6.20), (6.23)–(6.24).

c sl
For model SCS2_E, the values of E c , E , and E sl , E , are obtained by solving
the two mixed integer programs:

Minimize E c , (6.23), subject to (6.6)–(6.10), (6.18)–(6.20),

Maximize E sl , (6.24), subject to (6.6)–(6.10), (6.18)–(6.20).

6.3 Models for Risk-Neutral Decision-Making 159

6.3.3 Scheduling with Multiple Batch Shipping and Limited

Transportation Capacity

In this subsection a multiple batch shipping method is modelled, where each batch
size is limited by vehicle capacity C V and, in addition, a fixed transportation cost,
ak , is incurred for each shipment to distribution center k ∈ K .
In model SCS3_E presented below the variables, y sj ∈ {0, 1}; j ∈ J, s ∈ S, are
not required any more. Instead a new customer order shipping variable is introduced,
z sjt ∈ {0, 1}; j ∈ J, t ∈ T K , s ∈ S, defined as follows:
z sjt = 1, if under disruption scenario s customer order j is shipped to the distrib-
ution center in period t; otherwise z sjt = 0.
The objective functions E c and E sl are defined below.
   
Ec = ( ei u i + Ps ( Boi vi + ak xkts
i∈I s∈S i∈Is k∈K t∈T K
    
+ gjbj( wsjt − z sjt ) + h j b j (1 − wsjt )))/B (6.25)
j∈J t∈T t∈T K :t≤d j −σk j +1 j∈J t∈T
 
E sl = Ps b j z sjt /B, (6.26)
s∈S j∈J t∈T K :t≤d j −σk j +1

   s
where s∈S k∈K t∈T KPs ak x kt is the expected
 transportation
 cost of shipping
to distribution centers and s∈S j∈J Ps g j b j ( t∈T wsjt − t∈T K :t≤d j −σk +1 z sjt ) is
j
the expected penalty for delayed orders.

SCS3_E: Risk-neutral supply chain scheduling: multiple batch shipping

Minimize (6.5)
subject to (6.3), (6.4), (6.6)–(6.10), (6.18)–(6.21), (6.25)–(6.26) and
Multiple batch shipping constraints:
 
z sjt = wsjt ; j ∈ J, s ∈ S (6.27)
t∈T K t∈T

b j z sjt ≤ C V xkts ; k ∈ K , t ∈ T K , s ∈ S (6.28)
j∈Jk
 
xkts ≥ b j wsjt /C V ; k ∈ K , s ∈ S (6.29)
t∈T K j∈Jk t∈T

z sjt ≥ xkts ; k ∈ K , t ∈ T K , s ∈ S (6.30)
j∈Jk
160 6 Integrated Selection of Supply Portfolio …

Production-distribution coordinating constraints:

 
t z sjt ≥ (t + 1)wsjt ; j ∈ J, s ∈ S (6.31)
t∈T K t∈T

Supply-distribution coordinating constraints:

  
b j z sjt ≤ B vi ; t ∈ T K , s ∈ S (6.32)
j∈J t ∈T K :t ≤t i∈Is :τi ≤t−2
z sjt ∈ {0, 1}; j ∈ J, t ∈ T K , s ∈ S. (6.33)

For each disruption scenario, Eq. (6.27) ensures that each completed order is
shipped to distribution center. Constraint (6.28) ensures that each batch shipment
size cannot exceed vehicle limited capacity, and Eq. (6.29) determines minimum
number of batch shipments to each distribution center. Eq. (6.30) guarantees that at
least one order is shipped in every batch. The production-distribution coordinating
constraints (6.31) ensure that each order can be shipped only after its completion.
Finally, the supply-distribution coordinating constraints (6.32) ensure that for each
disruption scenario s, the cumulative distribution of products shipped to customers
by period t is not greater than the cumulative supplies of required parts delivered
by the non-disrupted suppliers i ∈ Is by period t − 2. (Note that parts delivered by
supplier i ∈ Is in period τi can be used for production not earlier than in period
τi + 1, and then the completed products can be shipped to distribution center not
earlier than in period τi + 2.)
c sl
For model SCS3_E, the values of E c , E , and E sl , E , are obtained by solving
the two mixed integer programs:

Minimize E c , (6.25), subject to (6.6)–(6.10), (6.18)–(6.21), (6.27)–(6.33).

Maximize E sl , (6.26), subject to (6.6)–(6.10), (6.18)–(6.21), (6.27)–(6.33).

6.4 Models for Risk-Averse Decision-Making

In this section, three time-indexed SMIP models SCS1_CV, SCS2_CV and

SCS3_CV are proposed for risk-averse supply chain scheduling to optimize the
weighted-sum of worst-case cost and worst-case service level under disruption risks.
The models are based on the risk-neutral models SCS1_E, SCS2_E and SCS3_E,
respectively.
6.4 Models for Risk-Averse Decision-Making 161

Let VaRc be the Value-at-Risk of cost per product, i.e., the targeted cost such that
for a given confidence level, α, for 100α% of disruption scenarios, the outcome is
below VaRc and let CVaRc be Conditional Value-at-Risk of cost per product, i.e., the
expected cost in the worst 100(1 − α)% of the scenarios with the cost above VaRc

CVaRc = VaRc + (1 − α)−1 Ps Cs . (6.34)
s∈S

In a similar way, define by VaRsl the Value-at-Risk of service level, i.e., the targeted
service level such that for a given confidence level α, for 100α% of disruption
scenarios, the outcome is above VaRsl , and by CVaRsl the Conditional Value-at-Risk
of service level, i.e., the expected service level in the worst 100(1 − α)% of scenarios
with the service level below VaRsl

CVaRsl = VaRsl − (1 − α)−1 Ps Ss . (6.35)
s∈S

The risk-averse integrated supply, production and distribution schedule will be

optimized by simultaneously calculating VaRc and VaRsl and minimizing weighted
difference of CVaRc and CVaRsl . Model SCS1_CV is presented below.

SCS1_CV: Risk-averse supply chain scheduling to minimize trade-off

between CVaR of cost and CVaR of service level: single batch shipping
Minimize

λC V a R c − (1 − λ)C V a R sl , (6.36)

where 0 ≤ λ ≤ 1,
subject to (6.6)–(6.22), (6.34), (6.35) and
Risk constraints:
– the tail cost for scenario s is defined as the nonnegative amount by which
cost in scenario s exceeds VaRc ,
– the tail service level for scenario s is defined as the nonnegative amount
by which VaRsl exceeds service level in scenario s,
 
Cs ≥ ( ei u i + Boi vi
i∈I i∈Is
 
+ gjbj( wsjt − y sj )
j∈J t∈T
 
+ h j b j (1 − wsjt ))/B − V a Rc ; s ∈ S (6.37)
j∈J t∈T
162 6 Integrated Selection of Supply Portfolio …


Ss ≥ V a R sl − b j y sj /B; s ∈ S (6.38)
j∈J
Cs ≥ 0; s ∈ S (6.39)
Ss ≥ 0; s ∈ S, (6.40)

where Cs is the tail cost and Ss is the tail service level, for scenario s.

In a similar way we can formulate the risk-averse versions SCS2_CV and

SCS3_CV of the risk-neutral models SCS2_E and SCS3_E. The models are pre-
sented below.

SCS2_CV: Risk-averse supply chain scheduling to minimize trade-off

between CVaR of cost and CVaR of service level: individual shipping
Minimize (6.36)
subject to (6.6)–(6.10), (6.18)–(6.20), (6.23), (6.24), (6.34), (6.35), (6.39),
(6.40) and
Risk constraints:
 
Cs ≥ ( ei u i + Boi vi
i∈I i∈Is
   
+ ak xkts + gjbj( wsjt − z sjt )
k∈K t∈T K j∈J t∈T t∈T K :t≤d j −σk j +1
 
+ h j b j (1 − wsjt ))/B − V a R c ; s ∈ S (6.41)
j∈J t∈T
 
Ss ≥ V a R sl − b j wsjt /B; s ∈ S. (6.42)
j∈J t∈T :t≤d j −σk j

SCS3_CV: Risk-averse supply chain scheduling to minimize trade-off

between CVaR of cost and CVaR of service level: multiple batch shipping
Minimize (6.36)
subject to (6.6)–(6.10), (6.18)–(6.21), (6.27)–(6.35), (6.39), (6.40) and
Risk constraints:
 
Cs ≥ ( ei u i + Boi vi
i∈I i∈Is
   
+ ak xkts + gjbj( wsjt − z sjt )
k∈K t∈T K j∈J t∈T t∈T K :t≤d j −σk j +1
6.4 Models for Risk-Averse Decision-Making 163

 
+ h j b j (1 − wsjt ))/B − V a R c ; s ∈ S (6.43)
j∈J t∈T
 
Ss ≥ V a R sl − b j z sjt /B; s ∈ S. (6.44)
j∈J t∈T K :t≤d j −σk j +1

6.5 Expected Value Problem

Stochastic mixed integer programs are usually hard to solve because they are large-
scale optimization problems when applied to real-world problems. A simplified
approach, which may sometimes be useful in practice, is to consider a simpler deter-
ministic program, known as expected value problem, in which the random parame-
ters are replaced by their expected values (e.g., Birge and Louveaux 2011, Kall and
Mayer 2011, Sawik 2016d). As an example of such approach this section presents
the expected value problem corresponding to the risk-neutral scheduling in supply
chain with single batch shipping of products to distribution centers, described by
SMIP model SCS1_E. In model SCS1_E, where the randomness is characterized
by a set of disruption scenarios, the only random parameters are suppliers all-or-
nothing fulfillment rates, which appear both in the objective function (6.1) and in
constraints (6.10). Denote by ESCS1, a deterministic MIP model of the expected
value problem corresponding to model SCS1_E. In model ESCS1, suppliers prob-
abilistic all-or-nothing fulfillment rates (1, for a non disrupted supplier and 0, for a
disrupted supplier) defined for each disruption scenario are replaced by the expected
fulfillment rates of each supplier

1 − πi = (1 − pr )(1 − pi ); i ∈ I r , r ∈ R,

where

πi = pr + (1 − pr ) pi , i ∈ I r , r ∈ R (6.45)

is total disruption probability of supplier i.

Notice that the expected fulfillment rate of a supplier is identical with his non-
disruption probability.
Accordingly, stochastic binary decision variables, wsjt , xkts , y sj , (6.20)–(6.22),
defined for each disruption scenario s ∈ S are replaced by their deterministic equiv-
alents W jt , X kt , Y j (see, Table 6.3)
164 6 Integrated Selection of Supply Portfolio …

Table 6.3 Notation: expected value problem

Parameters
πi = pr + (1 − pr ) pi , total disruption probability of supplier i ∈ I r , r ∈ R
1 − πi = (1 − pr )(1 − pi ), the expected fraction of an order delivered by supplier i ∈
I r , r ∈ R (supplier expected fulfillment rate)
Variables
W jt = 1, if customer order j is scheduled for period t; otherwise W jt = 0 (production
scheduling)
X kt = 1, if batch shipment of products to distribution center k is scheduled for period t;
otherwise X kt = 0 (distribution scheduling)
Yj = 1, if customer order j is delivered by its due date; otherwise Y j = 0 (customer
order non-delay delivery)

Now, the expected cost per product, E c , and the expected service level, E sl , are
defined as follows
 
Ec = ( ei u i + Boi (1 − πi )vi
i∈I i∈I
   
+ gjbj( W jt − Y j ) + h j b j (1 − W jt ))/B (6.46)
j∈J t∈T j∈J t∈T

E sl = b j Y j /B (6.47)
j∈J

Model ESCS1 is presented below.

Model ESCS1
Minimize (6.5)
subject to (6.3), (6.4), (6.6), (6.7), (6.18), (6.19), (6.46), (6.47) and

W jt ≤ 1; j ∈ J (6.48)
t∈T

b j W jt ≤ C; t ∈ T (6.49)
j∈J
  
b j W jt ≤ B (1 − πi )vi ; t ∈ T (6.50)
j∈J t ∈T :t ≤t i∈I :τi ≤t−1
 
t X kt ≥ (t + 1)W jt ; k ∈ K , j ∈ Jk (6.51)
t∈T K t∈T

X kt ≤ 1; k ∈ K (6.52)
t∈T K
6.5 Expected Value Problem 165


Yj ≤ W jt ; k ∈ K , j ∈ Jk (6.53)
t∈T :t≤d j −σk

Yj ≤ X kt ; k ∈ K , j ∈ Jk (6.54)
t∈T K :t≤d j −σk +1
 
W jt + X kt − 1 ≤ Y j ; k ∈ K , j ∈ Jk (6.55)
t∈T :t≤d j −σk t∈T K :t≤d j −σk +1

W jt ∈ {0, 1}; j ∈ J, t ∈ T (6.56)

X kt ∈ {0, 1}; k ∈ K , t ∈ T K (6.57)
Y j ≥ 0; j ∈ J. (6.58)

Notice that unlike SMIP model SCS1_E which is formulated to determine optimal
schedules for all potential disruption scenarios, model ESCS1 accounts for a single
scenario only, representing the expected supplies. Except for the expected values
of the random parameters, this model does not take into account any distribution
information and the solution remains the same as long as the expectations do not
change. In contrast to model SCS1_E, where the selection of supply portfolio is
combined with supply chain scheduling for all disruption scenarios considered, now
the portfolio is determined along with a single schedule.

6.6 Computational Examples

The examples presented in this section are modeled after a real world electronics
supply chain (e.g., Sawik 2011a). The supply chain consists of multiple manufac-
turers/suppliers of electronic components, a single producer where finished products
(e.g., cellular phones) are assembled to meet customer orders and a set of distribu-
tion centers that deliver the products to customers (e.g., carriers) who generate final
demand for products. The completed customer orders are shipped to the distribu-
tion centers either in batches of different customer orders or each customer order is
shipped individually, immediately after its completion using a direct shipping line.
The line consists of a surface mount technology line, where printed wiring boards are
assembled, material preparation stage, where all materials required for each product
are prepared, postponement stage, where products for some orders are customized,
flashing/flexing stations, where required software is downloaded and packing sta-
tions, where products and required accessories are packed for shipping. The input
data for the example problems were prepared considering monthly production.
The following basic parameters have been used for the example problems:
• I = 9, J = 25, K = 3, R = 3, T = 10, and S = 2 I = 512;
• K = {1, 2, 3}, and J1 = {1, . . . , 10}, J2 = {11, . . . , 20}, J3 = {21, . . . , 25};
166 6 Integrated Selection of Supply Portfolio …

• R = {1, 2, 3}, and I 1 = {1, 2, 3}, I 2 = {4, 5, 6}, I 3 = {7, 8, 9};

• b j ∈ {500, 1000, . . . , 5000} ∀ j = 1, . . . , 25, B = 66000;
• C P = 20000;
• C V = max j∈J b j = 5000;
• e = (8, 7, 10, 13, 14, 11, 19, 17, 19) × 1000;
• g j = 1, h j = 26 ∀ j = 1, . . . , 25;
• o = (13, 12, 12, 8, 6, 6, 2, 5, 4);
• τi , was 2, 3 and 4 time periods, respectively for suppliers i ∈ I 1 , i ∈ I 2 and i ∈ I 3 ;
• σk , was 1, 2 and 3 time periods, respectively for transportation to distribution center
k = 1, 2 and 3;
• ak = C V σk = 5000, 10000 and 15000, respectively for transportation to distribution
center k = 1, 2 and 3;
• d j ∈ {2 + mini∈I (τi ) + mink∈K (σk ), . . . , T + maxk∈K (σk )} ∀ j = 1, . . . , 25;

Suppliers characteristics
0.1 14
Disruption probability

12
0.08

Price per part

10
0.06 8

0.04 6
4
0 .02
2
0 0
1 2 3 4 5 6 7 8 9
Supplier
Disruption Probability Purchasing Price

Fig. 6.2 Suppliers

The local and regional disruption probabilities, pi , i ∈ I and pr , r ∈ R, were:

p = (0.00513571, 0.00666354, 0.00902974, 0.0356206, 0.040175, 0.0294692,
0.0519967, 0.0827215, 0.0739062) and p 1 = 0.001, p 2 = 0.005, p 3 = 0.01.
The corresponding total disruption probabilities, πi = pr + (1 − pr ) pi , i ∈ I r ,
r ∈ R, (6.45), of all suppliers were: π =(0.00613057, 0.00765688, 0.0100207, 0.040
4425, 0.0449741, 0.0343219, 0.0614767, 0.0918943, 0.0831672).
The unit price per part, oi , and the disruption probability, πi , of each supplier
i ∈ I are shown in Fig. 6.2. The figure indicates that the most reliable (with the
lowest disruption probability) is supplier 1, the least reliable (with the highest dis-
ruption probability) is supplier 8, the most expensive (with the highest price per part)
is supplier 1, and the cheapest (with the lowest price per part) is supplier 7. Note
that geographic regions are numbered in such a way that the unit prices are nonin-
creasing with r , while the fixed ordering costs and the disruption probabilities are
6.6 Computational Examples 167

nondecreasing with r , i.e.,

oi1 ≥ oi2 ≥ oi3 , ei1 ≤ ei2 ≤ ei3 and πi1 ≤ πi2 ≤ πi3 ; ∀i 1 ∈ I 1 , i 2 ∈ I 2 , i 3 ∈ I 3 .

6.6.1 Risk-Neutral Decision-Making

Table 6.4 Risk-neutral solutions: single batch shipping

λ 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Var. = 124207, Bin. = 112723, Cons. = 84406, Nonz. = 1628119 (a)
c
Exp.Cost 16.42( E ) 10.96 10.67 10.40 7.73 7.72 7.71 5.25 4.70 4.30 4.28( E c )
Exp.Service Level (b)
sl
78.65 ( E ) 78.36 78.30 78.04 74.30 74.29 74.17 56.67 51.71 41.96 41.94( E sl )
Exp.Fulfilled Demand (c) 78.65 88.82 88.85 88.63 96.17 96.19 95.83 94.56 94.10 93.78 93.85
Suppliers Selected 1(21)
(% of total demand) 2(21) 2(37) 2(30) 2(30) 2(29) 2(29) 2(29)
3(18)
4(9)
5(5) 5(21) 5(20) 5(21)
6(8) 6(21) 6(20) 6(21) 6(30) 6(30) 6(30) 6(29) 6(11)
7(12) 7(21) 7(21) 7(28) 7(41) 7(41) 7(41) 7(71) 7(89) 7(100) 7(100)
8(2)
9(4) 9(9)
CPU [sec] (d) 15285 74074 21819 60464 31004 42468 14903 37542 26009 69686 113
(a) Var. = number of variables, Bin. = number of binary variables,

 = number
Cons.
(b)
 of constraints, Nonz. = number of nonzero coefficients.
j∈J Ps b j y j /B × 100%.
s
s∈S
(c)
  
t∈T Ps b j w jt /B × 100%.
s
s∈S j∈J
(d) CPU seconds for proven optimal solutions, except for λ = 0.1, 0.9

with GAP<1%

The solution results for the risk-neutral decision-making are presented in Tables 6.4,
6.5 and 6.6, respectively for single batch shipping, individual shipping and multiple
batch shipping with limited transportation capacity. For comparison, the expected
cost per product in Table 6.6 does not include the expected transportation cost per
product, which is shown separately. The results indicate that for λ = 1 (minimization
of cost) the cheapest supplier i = 7 is selected only, while for λ = 0 (maximization
of customer service level) the total demand for parts is allocated among all nine
suppliers for batch shipping and eight suppliers, except of the least reliable supplier
8, for individual shipping. The highest proportion of demand is allotted among the
three most reliable and most expensive suppliers i = 1, 2, 3. As λ increases from
0 to 1, i.e., the decision maker preference shifts from customer service level to
cost, more demand is moved from expensive and reliable suppliers to low-cost,
168 6 Integrated Selection of Supply Portfolio …

Table 6.5 Risk-neutral solutions: individual shipping

λ 0 0.1 0.2 0.3 0.4, 0.5, 0.6, 0.7 0.8, 0.9 1
Var. = 100470, Bin. = 100459, Cons. = 21343, Nonz. = 909247 (a)
c
Exp.Cost 15.38(E ) 10.59 7.87 6.19 5.57 4.77 4.15(E c )
(b) sl
Exp.Service Level 87.56(E ) 87.36 85.98 84.52 83.48 74.45 55.45(E sl )
Exp.Fulfilled Demand (c) 87.56 97.36 95.74 95.06 94.86 94.37 93.85
Suppliers Selected 1(52)
(% of total demand) 2(12) 2(40) 2(9) 2(9) 2(9)
3(12) 3(12)
4(4) 4(12)
5(4) 5(12) 5(12) 5(12)
6(4) 6(12) 6(55) 6(20) 6(20) 6(20)
7(8) 7(12) 7(12) 7(59) 7(71) 7(80) 7(100)
9(4) 9(12)
CPU [sec] 1 79 61 409 112, 8, 10, 12 11, 26 3
(a) Var. = number of variables, Bin. = number of binary variables,
Cons.
 = number
 of constraints, Nonz. = number of nonzero coefficients.

(b)
t∈T :t≤d j −σk j Ps b j w jt /B × 100%.
s
s∈S j∈J
(c)
  
t∈T Ps b j w jt /B × 100%
s
s∈S j∈J

unreliable suppliers. The results for individual and multiple batch shipping, where
more shipments are scheduled, are similar.
For all the shipping methods and most λ, the total demand for parts is allocated
among four suppliers i = 2, 5, 6, 7, where supplier i = 2 is the second most reliable
and i = 6 is the most reliable in regions r = 2, 3, while i = 7 is the cheapest among
all suppliers. For small λ, however, more suppliers are selected to improve individual
service level.
In general, for all shipping methods the service-oriented supply portfolio (λ ≤ 0.3)
is more diversified than the cost-oriented portfolio (λ ≥ 0.7).
In addition to the expected customer service level, i.e., the expected fraction of
customer demand fulfilled ontime,Tables 6.4, 6.5 and 6.6 also show the expected
fraction of fulfilled demand, s∈S j∈J t∈T Ps b j wsjt /B, i.e., demand fulfilled on
time or delayed. The solution results demonstrate that for the maximum service
level objective, which is independent of any cost parameters, the largest expected
fraction of non-delayed customer demand is associated with the smallest expected
fraction of fulfilled demand. This indicates that to maximize expected service level,
customer orders that cannot be fulfilled by requested due dates are rejected. Note that
the difference between the expected fraction of fulfilled demand and the expected
customer service level is the expected fraction of delayed customer demand. The
latter measure is equal to zero for the maximum service level objective, increases
6.6 Computational Examples 169

Table 6.6 Risk-neutral solutions: multiple batch shipping

λ 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7, 0.8 0.9 1
Var. = 124207, Bin. = 112723, Cons. = 84406, Nonz. = 1628119 (a)
Exp.Cost (b)
c
17.85 E ) 10.35 8.95 7.10 6.58 6.55 5.97 5.41 4.78 6.07( E c )
Exp.Transportation Cost (c) 2.04 2.00 1.97 1.96 1.95 1.95 1.93 1.93 1.92 1.92
Exp.Service Level (d)
sl
85.31( E ) 85.23 84.66 83.11 82.46 82.40 80.42 77.73 70.96 55.44( E sl )
Exp.Fulfilled Demand (e) 85.31 97.04 95.78 95.56 95.44 95.43 95.07 94.69 94.36 93.85
Suppliers Selected 1(42) 1(14)
(% of total demand) 2(14) 2(14) 2(21) 2(21) 2(20) 2(20) 2(14) 2(7)
3(14) 3(14)
4(4) 4(15) 4(15)
5(5) 5(14) 5(14) 5(14)
6(6) 6(14) 6(21) 6(14) 6(20) 6(19) 6(19) 6(19) 6(19)
7(8) 7(15) 7(15) 7(51) 7(60) 7(61) 7(67) 7(74) 7(81) 7(100)
8(2)
9(5) 9(14)
CPU [sec] 36 24354 22107 32958 15961 14689 24447 4560, 14392 25220 315
(a) Var. = number of variables, Bin. = number of binary variables,

Cons. = number of constraints, Nonz. = number of nonzero coefficients.

(b) without expected transportation cost, except for λ = 0 and λ = 1.
(c)
  
s∈S k∈K  t∈T K Ps ak x kt /B.
s
(d)

t∈T K :t≤d j −σk j +1 Ps b j z jt /B × 100%.
s
s∈S j∈J
(e)
  
t∈T Ps b j w jt /B × 100%
s
s∈S j∈J

with the trade-off parameter λ, and reaches its maximum for the minimum cost
objective.
The results presented for the three shipping methods in Tables 6.4, 6.5 and 6.6
are illustrated and compared in Figs. 6.3 and 6.4. Figure 6.3 compares the expected
values of cost per product, service level and fulfilled demand for 11 levels of trade-off
parameter λ, while Fig. 6.4 compares the obtained subsets of nondominated supply
portfolios (the allocation of total demand for parts among selected suppliers), with
the four major suppliers i = 2, 5, 6, 7 indicated. The corresponding portfolios for
different shipping methods are similar, with the cheapest suppliers dominating in the
individual shipping. The results point out that individual and multiple batch shipping
lead to better solution values, i.e., a higher expected service level and a lower expected
ordering, purchasing and shortage cost.
For single batch shipping, Figs. 6.5, 6.6 and 6.7 show the expected supply, produc-
tion and delivery schedules, respectively for λ = 0 (i.e., for the maximum expected
service), λ = 0.5 and λ = 1 (i.e., for the minimum expected cost). The expected
schedules were computed using the formulae presented below:
170 6 Integrated Selection of Supply Portfolio …

Expected Solution Values

(a) 100 100

Customer Demand %
80 80
Cost per Product

60 60

40 40

20 20

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
lambda
Expected Cost Expected Service Level % Expected Fulfilled Demand %

Expected Solution Values

(b) 100 100

Customer Demand %
80 80
Cost per Product

60 60

40 40

20 20

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
lambda
Expected Cost Expected Service Level % Expected Fulfilled Demand %

Expected Solution Values

(c) 100 100
Customer Demand %
Cost per Product

80 80

60 60

40 40

20 20

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
lambda
Expected Cost Expected Service Level % Expected Fulfilled Demand %

Fig. 6.3 Expected solution values: a single batch shipping, b individual shipping, c multiple batch
shipping
6.6 Computational Examples 171

Fig. 6.4 Nondominated supply portfolios: a single batch shipping, b individual shipping, c multiple
batch shipping
172 6 Integrated Selection of Supply Portfolio …

Expected Supplies
40000
from
Region 1
30000
Parts

20000 from
Region 2
from
10000
Region 3

0
1 2 3 4 5 6 7 8 9 10
Period

Expected Production
20000

16000
Products

12000

8000

4000

0
1 2 3 4 5 6 7 8 9 10
Period

Expected Deliveries
20000
to to DC2
DC1
16000
to DC3
Products

12000

8000

4000

0
1 2 3 4 5 6 7 8 9 10
Period

Fig. 6.5 Expected supplies, production and deliveries for maximum service level (λ = 0): single
batch shipping

• Expected schedule of supplies of parts to the producer

 
Ps Bvi ; t ∈ T (6.59)
s∈S i∈Is :τi =t
6.6 Computational Examples 173

Expected Supplies
30000 from
Region 3
25000 from
from
Region 1 Region 2
20000
Parts

15000

10000

5000

0
1 2 3 4 5 6 7 8 9 10
Period

Expected Production
20000

15000
Products

10000

5000

0
1 2 3 4 5 6 7 8 9 10
Period

Expected Deliveries
25000 to DC1
to DC2
20000
Products

15000 to DC3

10000

5000
to DC2
to DC3
0
1 2 3 4 5 6 7 8 9 10
Period

Fig. 6.6 Expected supplies, production and deliveries for λ = 0.5: single batch shipping

• Expected production schedule


Ps b j wsjt ; t ∈ T (6.60)
s∈S j∈J

• Expected schedule of shipping of products from the producer to the distribution

centers
174 6 Integrated Selection of Supply Portfolio …

Expected Supplies
60000
from Region 3

40000
Parts

40000

0
1 2 3 4 5 6 7 8 9 10
Period

Expected Production
20000

15000
Products

10000

5000

0
1 2 3 4 5 6 7 8 9 10
Period

Expected Deliveries
25000 to DC2
to DC1
20000
Products

15000 to
DC3
10000
to
5000
to DC1 DC2

0
1 2 3 4 5 6 7 8 9 10 11
Period

Fig. 6.7 Expected supplies, production and deliveries for minimum cost (λ = 1): single batch
shipping

 
Ps b j ( wsjt )xkts ; k ∈ K , t ∈ T K , (6.61)
s∈S j∈Jk t ∈T :t <t

where T K = {mini∈I τi + 2, . . . , T + 1} is the set of shipping periods.

Note that the delivery schedule for the distribution center k is the shipping schedule
delayed by σk periods.
6.6 Computational Examples 175

The expected supply schedules shown in Figs. 6.5, 6.6 and 6.7 are marked with
the supplier region, while the expected delivery schedules indicate the distribution
center; DC1 (k = 1), DC2 (k = 2) and DC3 (k = 3).
As λ increases, i.e., the decision maker priority shifts from the maximum service
level to minimum cost and more parts are ordered from less reliable and lower cost
suppliers, the expected supply schedules and the corresponding production schedules
are more delayed as well as the delivery of products to the customers. Note that
despite constraint (6.11) for single batch shipping that ensures at most one shipment
of products to each distribution center for each disruption scenario, the expected
shipping schedule (6.61) may be split into more smaller size shipments, (cf. Figs. 6.6,
6.7 and 6.8).
For individual and multiple batch shipping, Figs. 6.8 and 6.9 show examples
of expected supply, production and delivery schedule for λ = 0.5. Comparison of
Fig. 6.6 for single batch shipping, with Fig. 6.8 and with Fig. 6.9 shows that the
expected supplies from different regions are more diversified, and the expected pro-
duction as well as the shipping schedule are distributed among more periods. The
results indicate that for individual and multiple batch shipping, where more ship-
ments are scheduled, supply, production and delivery schedules are more unlevelled.
On the other hand, comparison of the solution results shown in Tables 6.4, 6.5 and
6.6 indicate that individual and multiple batch shipping may lead to higher expected
service level and lower expected ordering, purchasing and shortage cost.
The solution results obtained for numerical examples modelled after realistic
problems in the electronics supply chain are in line with other approaches used in
the area of supply chain risk management. The main findings are listed below.
• The supply portfolios for different shipping methods are similar. For all shipping
methods, the service-oriented supply portfolio (λ ≤ 0.3) is more diversified than
the cost-oriented portfolio (λ ≥ 0.7).
• For the minimum cost objective the cheapest supplier is usually selected, while for
the maximum service level objective the most reliable and most expensive suppliers
are dominating in the supply portfolio.
• For the maximum service level objective, which is independent of any cost parame-
ters, the largest expected fraction of non-delayed customer demand simultaneously
leads to the largest expected fraction of rejected demand. This indicates that in
order to maximize expected service level, customer orders that cannot be fulfilled
by requested due dates are rejected.
• The more cost-oriented decision-making, the more delayed the expected supply,
production and distribution schedules.
• The individual and multiple batch shipping lead to higher expected service level
and lower expected ordering, purchasing, and shortage cost.
176 6 Integrated Selection of Supply Portfolio …

Expected Supplies
45000
from
Region 3
30000
Parts

from
Region 2
15000 from
Region 1

0
1 2 3 4 5 6 7 8 9 10
Period

Expected Production
20000

15000
Products

10000

5000

0
1 2 3 4 5 6 7 8 9 10
Period

Expected Deliveries
8000
to DC1
6000
Products

to DC2

4000

2000 to DC3

0
1 2 3 4 5 6 7 8 9 10 11 12 13
Period

Fig. 6.8 Expected supplies, production and deliveries for λ = 0.5: individual shipping

The computational experiments were performed using the AMPL programming

language and the XPRESS 27.01 solver on a MacBookPro laptop with Intel Core
i7 processor running at 2.8 GHz and with 16GB RAM. (For comparison, the com-
putational experiments have been also performed using CPLEX 12.6.2 and Gurobi
6.0.2 solvers, however they were outperformed by the XPRESS 27.01 solver.) The
MIP problem sizes (the numbers of variables, constraints and nonzero coefficients)
and CPU seconds required to find optimal solutions (or with GAP less than 1%)
for the examples are presented in Tables 6.4, 6.5 and 6.6. The solver was capable of
finding proven optimal solution for all examples for models SCS2_E and SCS3_E,
6.6 Computational Examples 177

Expected Supplies
40000 from
Region 3
30000
Parts

20000 from from

Region 1 Region 2
10000

0
1 2 3 4 5 6 7 8 9 10
Period

Expected Production
20000

15000
Products

10000

5000

0
1 2 3 4 5 6 7 8 9 10
Period

Expected Deliveries
5000
to DC1
to DC2
4000 to DC3
Products

3000

2000

1000

0
1 2 3 4 5 6 7 8 9 10 11 12 13
Period

Fig. 6.9 Expected supplies, production and deliveries for λ = 0.5: multiple batch shipping

with CPU time ranging from several seconds to several hours. Proven optimal solu-
tions were also found for most examples for single batch shipping model SCS1_E,
except for λ = 0.1, 0.9, where the solution with GAP less than 1% was found and
no improvements were observed within next 3600 CPU seconds. The smallest CPU
time was required for the individual shipping model SCS2_E. Since model SCS2 is
much simpler than SCS1_E and SCS3_E, the latter result is obvious.
178 6 Integrated Selection of Supply Portfolio …

6.6.2 Risk-Averse Decision-Making

In this subsection some computational examples are presented to illustrate the risk-
averse scheduling of supply, production and distribution using model SCS1_CV. In
the computational experiments data sets provided at the beginning of this section were
used. Table 6.7 shows solution results for the objective function (6.36) with λ = 0
(maximization of CVaR of service level), λ = 1 (minimization of CVaR of cost) and
λ = 0.5 (balanced trade-off between CVaR of cost and CVaR of service level). The
risk-averse supply portfolios and solution values of V a R c , CVaRc and V a R sl , CVaRsl
and the associated expected values E c and E sl , respectively of cost and service level
are presented for a subset of confidence levels α = 0.5, 0.75, 0.9, 0.95, 0.99.
As α increases, the CVaRc of cost and the associated expected cost, E c , increase,
while expected service level, E sl , decreases. Similarly, the CVaRsl of service level
decreases and the associated expected cost, E c , increases with confidence level α.
However, V a R sl , E sl and risk-averse supply portfolios are nearly independent of
confidence level, except for the highest α = 0.99. At the same time, CVaRsl and E sl
are very close to the corresponding V a R sl . For the CVaRc objective function and
increasing confidence level, a greater diversification of supply portfolios is observed,
with more demand shifted from the cheapest supplier 7 to more reliable and more
expensive suppliers 2, 3, 5, 6 and 9. A low-cost and most unreliable supplier 8 was
never selected (cf. Fig. 6.2). In contrast, for the CVaRsl objective function the supply
portfolios are diversified, even for small confidence levels.
Comparison of solution results for models SCS1_E and SCS1_CV (cf. Tables 6.4
and 6.7), demonstrates that the risk-neutral solutions for maximization of expected
service level (that is, for λ = 0) and minimization of expected cost (that is, for
λ = 1) are very close to the corresponding risk-averse solutions with confidence
level α = 0.5. Moreover, for the cost objective function, the corresponding risk-
neutral and the risk-averse supply portfolios are identical, while for the service level
objective they are very close each other.
The computational experiments were performed using the AMPL programming
language and the Gurobi 7.0.0 solver on a MacBookPro laptop with Intel Core
i7 processor running at 2.8 GHz and with 16GB RAM. The solver was incapable
of finding proven optimal solutions within 3600 CPU seconds. However, for all
examples with α = 0.5 and α = 0.75, the solution with GAP less than 1% was
found much earlier, while for the higher confidence levels, more than 3600 CPU
seconds were required to achieve GAP less than 1%.
6.6 Computational Examples 179

Table 6.7 Risk-averse solutions for model SCS1_CV

Confidence level α 0.50 0.75 0.90 0.95 0.99
λ = 0 (maximization of CVaRsl )
Var. = 125233, Bin. = 112723, Cons. = 85430, Nonz. = 1638087 (a)
CVaRsl 100% 78.58 78.40 77.84 76.89 68.83
V a R sl 100% 78.79 75.76
E sl 100% 78.69 75.88
Ec 16.54 16.56 16.72 16.64 17.40
Supply Portfolio: 1(21) 1(21) 1(21) 1(21) 1(21)
Supplier (% of total demand) (b) 2(21) 2(21) 2(21) 2(21) 2(21)
3(21) 3(21) 3(20) 3(20) 3(18)
4(8) 4(7) 4(7) 4(8) 4(8)
5(7) 5(7) 5(7) 5(7) 5(7)
6(7) 6(8) 6(8) 6(6) 6(8)
7(14) 7(11) 7(14) 7(13) 7(11)
8(3)
9(1) 9(4) 9(2) 9(4) 9(3)
λ = 1 (minimization of CVaRc )
CVaRc 5.66 8.56 12.78 13.82 16.10
V a Rc 2.77 2.77 11.25 12.32 15.30
Ec 4.22 4.22 10.17 11.27 13.05
sl
E 100% 48.35 48.35 15.57 11.26 5.75
Supply Portfolio:
Supplier (% of total demand) (b) 2(6) 2(26) 2(24)
3(21)
4(14) 4(19)
5(22) 5(18) 5(15)
6(21) 6(20) 6(15)
7(100) 7(100) 7(18) 7(17) 7(12)
9(19) 9(13)
λ = 0.5
CVaRc 5.66 8.74 12.88 13.89 16.17
V a Rc 2.77 3.12 10.79 12.47 15.36
Ec 4.22 4.53 10.73 11.72 14.89
CVaRsl 100% 45.18 42.36 55.40 72.28 59.06
V a R sl 100% 51.52 54.55 56.82 78.79 62.12
E sl 100% 48.35 51.50 57.75 78.46 67.85
(continued)
180 6 Integrated Selection of Supply Portfolio …

Table 6.7 (continued)

Confidence level α 0.50 0.75 0.90 0.95 0.99
Supply Portfolio:
Supplier (% of total demand) (b) 2(23) 2(22)
3(13) 3(20)
4(15) 4(17) 4(17)
5(20) 5(17)
6(5) 6(20) 6(17) 6(15)
7(100) 7(95) 7(16) 7(13) 7(13)
8(15)
9(14) 9(13)
(a)
Var. = number of variables, Bin. = number of binary variables,
Cons. = number of constraints, Nonz. = number of nonzero coefficients.

6.6.3 Expected Value - Based Decision-Making

For the example input data, the suppliers expected fulfillment rates, 1 − πi ; i ∈
I r , r ∈ R, are (0.993869, 0.992343, 0.989979, 0.959558, 0.955026, 0.965678,
0.938523, 0.908106, 0.916833), that is, supplier 1 is most reliable and supplier 8
most unreliable.

Table 6.8 Solution results for the expected value model ESCS1
λ 0 0.1 and 0.2 and 0.3 0.4 and 0.5 and 0.6 and 0.7 0.8 and 0.9 1
Var. = 265, Bin. = 233, Cons. = 157, Nonz. = 2249 (a)
c
Exp.Cost 17.88( E ) 9.33 7.65 4.49 4.39( E c )
Exp.Service Level (b)
sl
78.78( E ) 78.03 75.76 54.55 47.73( E sl )
Exp.Fulfilled Demand (c) 83.33 90.15 95.45 93.94 93.18
Suppliers Selected 1(97.6)
(% of total demand) 2(0.8) 2(30) 2(26)
5(0.8)
6(31) 6(31) 6(5)
7(0.8) 7(39) 7(43) 7(95) 7(100)
(a) Var. = number of variables, Bin. = number of binary variables,
Cons.
 = number of constraints, Nonz. = number of nonzero coefficients.
(b)
j∈J b j Y j /B × 100%.
(c)
 
j∈J t∈T b j W jt /B × 100%.

Table 6.8 presents a subset of nondominated solutions obtained for the expected
value problem ESCS1, and Fig. 6.10 shows supply, production and shipping sched-
6.6 Computational Examples 181

Fig. 6.10 Supplies, production and deliveries for the expected value problem
182 6 Integrated Selection of Supply Portfolio …

ules. Unlike the stochastic programming approach which accounts for all potential
disruption scenarios to optimize an expected performance of the supply chain, the
solution obtained using the deterministic approach is based on an aggregate infor-
mation on suppliers expected fulfillments.
The solution results are similar for both SCS1_E ESCS1 models (cf. Tables 6.4
and 6.8) and the corresponding optimal solution values are close to each other, which
indicates that the expected value problem can be used in practice, when stochas-
tic mixed integer programs are hard to solve. However, the stochastic wait-and-see
approach leads to a more diversified supply portfolio that will hedge against a variety
of scenarios. In particular, the service-oriented supply portfolio is more diversified
and may combine both high-cost, reliable suppliers and low-cost unreliable suppli-
ers, while the cost-oriented portfolio depends mainly on low-cost and less reliable
suppliers.
For the deterministic, expected value approach, most of nondominated supply
portfolios consist of three suppliers only: i = 2, 6, 7, while suppliers i = 1, 2, 5, 7
are selected only for the maximum service level objective (that is, for λ = 0). In
addition, the expected schedules are more delayed for the stochastic approach. Com-
parison of Figs. 6.5, 6.6 6.7 and 6.10 indicates that the expected schedules for model
SCS1_E, computed as expectations over all schedules for all potential disruption
scenarios, (6.59)–(6.61), are similar to the corresponding schedules determined by
model ESCS1. The main difference is a more delayed expected production and ship-
ping schedule for model SCS1_E when minimization of cost is considered (that is,
for λ = 1).
The computational experiments were performed using the AMPL programming
language and the Gurobi 7.0.0 solver on a MacBookPro laptop with Intel Core i7
processor running at 2.8 GHz and with 16GB RAM. The solver was capable of finding
proven optimal solution for all examples with CPU time ranging from fraction of a
second to several seconds. Since model ESCS1 deals with a single scenario only, the
low computational effort required is obvious.

6.7 Notes

There is a growing body of literature on deterministic models on integrated

production-distribution planning and scheduling, however, without the supply oper-
ations, which are mostly considered separately, e.g., Erenguc et al. (1999). For
example, Li et al. (2005) investigated an integrated scheduling of assembly and multi-
destination air-transportation in a consumer electronics supply chain. The problem
was divided into two sub-problems. The air transportation allocation was formulated
and solved using a MIP approach, and two heuristics were proposed for the assem-
bly scheduling problem. Lei et al. (2006) studied an integrated production, inventory
and distribution routing problem and proposed a MIP approach combined with a
heuristic routing algorithm to coordinate the production, inventory and transporta-
tion operations. A comprehensive review and classification of existing deterministic
6.7 Notes 183

models that integrate production and outbound distribution operations at the detailed
scheduling level was presented in Chen (2010). The variety of models were classified
by shipping and delivery methods. In particular, models with individual and imme-
diate delivery of each order and models with batch delivery and routing of orders to
different customers delivered together in the same shipment were considered. Such
models attempt to optimize detailed order-by-order production and delivery schedul-
ing jointly by taking into account relevant revenues, costs, and customer service lev-
els at the individual order level. In a more recent work on deterministic approaches,
Cakici et al. (2012) proposed a MIP approach for multi-objective scheduling of
customer orders in an integrated production and distribution system. The problem
objective was to optimize the trade-off between total weighted tardiness as a service
level measure and total distribution costs. Liu and Papageorgiou (2013) and Liu et al.
(2014) developed a multi objective MIP approach to address production, distribu-
tion and capacity planning of global supply chains considering cost, responsiveness
and customer service level simultaneously to achieve an equitably efficient or Pareto
optimal solution. An integrated production and distribution scheduling problem in
a make-to-order supply chain with limited production and distribution capacity was
considered by Viergutz and Knust (2014). The problem consists in finding a selection
of customers to be supplied such that the total satisfied demand is maximized. There
are also some reported studies on joint supplier selection and production and distri-
bution planning. Sawik (2009a) proposed a MIP approach for integrated scheduling
of material manufacturing, supply and product assembly in a customer-driven supply
chain. A monolithic approach, where the manufacturing, supply and assembly sched-
ules are determined simultaneously, was compared with a hierarchical approach, see
also Sawik (2006). Numerical examples modeled after a real-world scheduling in the
electronics supply chain were reported. A monolithic and a hierarchical approach
to multi-objective, integrated supply chain scheduling were also compared in Sawik
(2009b). A decomposition of the complex multi-objective production, manufacturing
and supply scheduling into a hierarchy of much simpler decision-making problems
was proposed and simple MIP formulations were provided. The objective functions
integrated both the total cost and the customer service level and the scheduling was
combined with the selection of part suppliers for each customer order and due date
setting for some orders. In Cui (2014) a MIP model was proposed for joint optimiza-
tion problem of production planning and supplier selection. The objective was to
maximize the manufacturers total profit subject to various operating constraints of
the supply chain. Gao et al. (2015) developed a MIP formulation and a heuristic with
a guaranteed worst-case bound for integrated production and distribution problem in
which orders are processed and delivered in batches with limited vehicle capacity.
Cheng et al. (2015) considered an integrated scheduling of production and distri-
bution to minimize total cost of production and distribution for the manufacturer.
A MIP model was developed and an improved ant colony method was proposed to
solve the production scheduling and the First-Fit-Decreasing heuristic used in the
bin-packing problem, for the distribution scheduling.
The material presented in this chapter is based on research reported in Sawik
(2016a), where bi-objective MIP models were proposed for an integrated selection of
184 6 Integrated Selection of Supply Portfolio …

supply portfolio and scheduling of production and distribution in the presence of sup-
ply chain disruption risks. The models incorporate supply-production, production-
distribution and supply-distribution coordinating constraints to efficiently coordi-
nate supply, production and distribution schedules. However, in Sawik (2016a) only
a risk-neutral trade-off between expected cost and expected service level has been
considered to optimize an overall performance of a supply chain. In this chapter, the
MIP approach and the risk-neutral models have been enhanced for the risk-averse
decision-making, using CVaR of cost and CVaR of service level as risk measures.
Moreover, the risk-neutral SMIP formulation based on the wait-and-see approach is
compared with a deterministic MIP model based on the expected value approach, in
which random parameters are replaced by their expected values. The expected value
problem is often used in practice as the related stochastic mixed integer program is
in general much harder to solve since it considers multiple scenarios, e.g., Durbach
and Stewart (2009), Maggioni and Wallace (2012).
The future research should concentrate on relaxations of the various simplified
assumptions used to formulate the problem (see, Sect. 6.2). For example, multi-
ple supplies of parts to the producer, e.g., partially disrupted supplies (e.g., Sawik
(2015b)), and multiple deliveries of finished products from distribution centers to
customers, along with an assignment of customers to distribution centers should
be considered. In addition, more shipping methods can be modelled such as batch
shipping with vehicle routing, methods with fixed shipping or delivery dates or with
shipping or delivery time windows, e.g., Chen et al. (2010).

Problems

6.1 Modify the SMIP models presented in this chapter for multiple part types with
subsets of part types required for each customer order and subsets of suppliers avail-
able for each part type.

6.2 Modify the SMIP models with single and multiple batch shipping of products
to distribution centers to account for a limited storage space for products waiting for
shipments.

6.3 Mean-risk scheduling in a supply chain

(a) Modify the SMIP models presented in this chapter to minimize expected cost and
CVaR of cost or maximize expected service level and CVaR of service level.
(b) How should the values of the optimized objective functions be scaled into the
interval [0,1] to avoid dimensional inconsistency among the two objectives and how
should the trade-off parameter be selected?
(c) How would you interpret the mean-risk supply chain schedules?

6.4 Mixed mean-risk scheduling in a supply chain

(a) Modify the SMIP models presented in this chapter to optimize expected cost and
CVaR of service level or optimize expected service level and CVaR of cost.
Problems 185

(b) How should the values of the optimized objective functions be scaled into the
interval [0,1] to avoid dimensional inconsistency among the two objectives and how
should the trade-off parameter be selected?
(c) How would you interpret the mixed mean-risk supply chain schedules?

6.5 Explain why the supply portfolios for different shipping methods of products to
distribution centers are similar and the service-oriented portfolios are more diversified
than the cost-oriented portfolios.
 Part III
Equitably Efficient Selection of Supply
Portfolio and Scheduling
Chapter 7
A Fair Decision-Making Under
Disruption Risks

7.1 Introduction

In global supply chains the optimization of material flows subject to unexpected

disruptive events, focuses on a variety of different optimality criteria. The most
commonly used criteria are minimization of cost and maximization of service level
that measures either fraction of customer orders or fraction of customer demand
satisfied on time. The cost and service level are in conflict and, in addition, the
decision makers often do not have preference to any objective, i.e., the two objectives
are equally important. Then, an equitably efficient solution should be generated, in
which the two normalized objective function values are as much close to each other
as possible. In this chapter, the two risk-neutral conflicting criteria: expected cost
and expected service level are fairly optimized to achieve an equitably efficient
supply portfolio and production schedule in the presence of supply chain disruption
risks. The supplies of parts are subject to independent random local and correlated
regional disruptions. The cost includes the cost of ordering, purchasing and shortage
of parts, while the service level is independent of any cost parameters. The two
alternative service level measures are compared: the expected order fulfillment rate
and the expected demand fulfillment rate (see, Chap. 5). In order to obtain an equitably
efficient solution to the combinatorial stochastic optimization problem, the Ordered
Weighted Averaging (OWA) aggregation of the two conflicting objective functions
is applied, e.g., Yager (1988). The equitably efficient solutions obtained for the
ordered weighted averaging aggregation of the two conflicting objective functions
will be compared with nondominated solutions obtained using the weighted-sum
aggregation approach.
The following time-indexed SMIP models are presented in this chapter:

© Springer International Publishing AG 2018 189

T. Sawik, Supply Chain Disruption Management Using Stochastic Mixed
Integer Programming, International Series in Operations Research
& Management Science 256, DOI 10.1007/978-3-319-58823-0_7
190 7 A Fair Decision-Making Under Disruption Risks

ESPS_E for risk-neutral selection of equitably efficient supply portfolio and

scheduling of customer orders to minimize expected cost and maximize
expected service level;
SPS_E(c) for risk-neutral selection of supply portfolio and scheduling of
customer orders to minimize expected cost;
SPS_E(sl) for risk-neutral selection of supply portfolio and scheduling of
customer orders to maximize expected service level;
WSPS_E for risk-neutral selection of supply portfolio and scheduling of
customer orders to minimize weighted-sum of normalized expected cost
and service level.

Numerical examples and some computational results are provided in Sect. 7.5, in
particular, comparison of OWA aggregation with weighted-sum aggregation of the
two objective functions is discussed.

7.2 The Lexicographic Minimax Optimization

and Equitable Aggregation

This section provides an overview of the lexicographic minimax optimization and

equitable aggregation.
Consider an optimization problem with m objective functions fk (x), k = 1, . . . m,
that are to be minimized. The problem can be formulated as follows:

min{f1 (x), f2 (x), . . . , fm (x) : x ∈ Q}, (7.1)

where f (x) = (f1 (x), f2 (x), . . . , fm (x)) is a vector-function that maps the decision
space X = R n into the criterion space Y = R m , Q ⊂ X denotes the feasible set, and
x ∈ X denotes the vector of decision variables. The elements of the criterion space
are referred to as outcome vectors.
Model (7.1) specifies that the objective is to minimize of all outcomes yk , k =
1, . . . m, however the solution concepts are defined by properties of the correspond-
ing preference model within the outcome space. If all the objective functions are
equally important to the decision makers, an equitable solution is sought, in which
all normalized objective function values should be as much close to each other as
possible.
The relation of equitable dominance can be expressed as a vector inequality on
the cumulative ordered outcomes. Let Θ : R m → R m be the ordering map such
that Θ(y) = (θ1 (y), θ2 (y), . . . θm (y)), where θ1 (y) ≥ θ2 (y) ≥ . . . ≥ θm (y)) and there
exists a permutation  of set {1, . . . , m} such that θk (y) = y (k) for k = 1, . . . m.
Apply to ordered outcomes Θ(y), a linear cumulative map that results in the  cumula-
tive ordering map Θ(y) = (θ 1 (y), θ 2 (y), . . . θ m (y)), defined as θ l (y) = lk=1 θk (y)
7.2 The Lexicographic Minimax Optimization and Equitable Aggregation 191

for l = 1, . . . , m. The coefficients of vector Θ(y) express: the largest outcome, the
total of the two largest outcomes,...,the total of the m outcomes, respectively.
The outcome vector y ∈ Y equitably dominates y ∈ Y , if and only if θl (y ) ≤
θ l (y ) for all l = 1, . . . , m where at least one strict inequality holds.
The following relationship between equitable efficiency and Pareto-optimality
can be provided (e.g., Kostreva et al. 2004).

Theorem 1 A feasible solution x ∈ Q is an equitably efficient solution of the multiple

criteria problem (7.1), iff it is a Pareto-optimal solution of the multiple criteria
problem with objectives Θ(f (x)):

min{(θ 1 (f (x)), θ 2 (f (x)), . . . θ m (f (x))) : x ∈ Q}. (7.2)

Note that minimization of the first objective in (7.2) corresponds to minimization

of the worst outcome, while minimization of the last (mth) objective in (2) aims at
minimizing the sum of outcomes.
The quantities θ l (y) can be modeled with simple auxiliary variables and con-
straints. For a given outcome vector y, the total of the l largest outcomes, θ l (y), may
be found by solving the following linear program (e.g., Ogryczak and Tamir 2003):


m
θ l (y) = min{lλ + δk : λ + δk ≥ yk , δk ≥ 0; k = 1, . . . m} (7.3)
k=1

where λ is an unrestricted variable while nonnegative variables δk represent, for

several outcome values yk , their upside deviations from the value of λ.
Using the linear program (7.3), problem (7.2) can be formulated as the following
multiple criteria optimization problem:


m 
m 
m
min{(λ1 + δk1 , 2λ2 + δk2 , . . . , mλm + δkm ) :
k=1 k=1 k=1
x ∈ Q, λl + δkl ≥ fk (x), δkl ≥ 0; k, l = 1, . . . m} (7.4)

In order to find an equitable efficient solution, the following lexicographic mini-

max problem can be solved

lex min{Θ(f (x)) : x ∈ Q}. (7.5)

Unlike in the standard minimax problem where only the worst objective value is min-
imized, in the lexicographic minimax problem (7.5), first the worst objective value
is minimized, then the second worst objective value is minimized (provided that the
worst one is as small as possible), the third worst objective value is minimized (pro-
vided that the two worst remain as small as possible), and so on. The lexicographic
192 7 A Fair Decision-Making Under Disruption Risks

minimax solution can be considered in some sense the “most equitable solution”
(Kostreva et al. 2004).
The multiple criteria problem (7.2) can be replaced with the minimization problem
of an equitably aggregated objective function. An equitable aggregation can be based
on the use of the cumulative ordered outcomes θ k (y), for example, the weighted sum
of θ k (y),
m
νk θ k (y), (7.6)
k=1

The above aggregation is the so-called ordered weighted averaging (OWA) aggrega-
tion, was introduced by Yager (1988). If the weights νk are positive, then applying
OWA aggregation to the multiple criteria problem (7.1) yields an equitably efficient
solution of the latter problem.
Note that in the OWA aggregation the weights are assigned to the ordered values
(i.e., to the largest value, the two largest values, etc.) rather than to the specific criteria,
like in the commonly used weighted sum aggregation.
By applying the OWA aggregation (7.6) to (7.4), the following OWA problem can
be formulated
m 
m
min{ νl (lλl + δkl ) : x ∈ Q, λl + δkl ≥ fk (x), δkl ≥ 0; k, l = 1, . . . m}. (7.7)
l=1 k=1

Liu and Papageorgiou (2013) considered the optimization problem (7.7) with
equal weights νl = 1 for all l = 1, . . . , m and m = 2 objective functions

2 
2
min{ (lλl + δkl ) : x ∈ Q, λl + δkl ≥ fk (x), δkl ≥ 0; k, l = 1, 2}, (7.8)
l=1 k=1

and in a recent paper Liu et al. (2014) proved the following new theorem.
Theorem 2 If there exists an optimal solution x ∗ ∈ Q of the optimization problem
(7.7) with equal weights νl = 1 for all l = 1, . . . , m, which also satisfies perfect
equity (i.e. f1 (x ∗ ) = f2 (x ∗ ) = . . . = fm (x ∗ ) ), then it is the optimal solution of the
lexicographic minimax problem (7.5).
In addition, the following theorem was introduced in Liu et al. (2014), according
to Theorem 1 (Kostreva et al. 2004).
Theorem 3 If x ∗ ∈ Q is an optimal solution of the optimization problem (7.7), with
equal weights νl = 1 for all l = 1, . . . , m, then it is an equitably efficient solution, as
well as a Pareto-optimal solution, of the multiobjective optimization problem (7.1).
The results of this section will be used to formulate and solve the fair, risk-neutral
optimization problem in this chapter and the fair, mean-risk (robust) optimization
problem in Chap. 8, for a bi-objective selection of supply portfolio and scheduling
of customer orders in the presence of supply chain disruption risks.
7.3 Problem Description 193

7.3 Problem Description

Consider a three-echelon customer-driven supply chain in which various types of

products are assembled by a single producer to meet customer orders, using the
same critical part type that can be manufactured and provided by many suppliers (for
notation used, see Table 7.1).

Table 7.1 Notation: selection of equitably efficient supply portfolio and scheduling
Indices
i = supplier, i ∈ I
j = customer order, j ∈ J
r = region, r ∈ R
s = disruption scenario, s ∈ S
t = planning period, t ∈ T
Input Parameters
aj = per unit requirement for parts of each product in customer order j
bj = size (number of products) of customer order j

A = j∈J aj bj , total demand for parts

B = j∈J bj , total demand for products
cj = per unit capacity consumption of producer for customer order j
Ct = capacity of producer in period t
dj = due date for customer order j
ei = fixed cost of ordering parts from supplier i
gj = per unit and per period penalty cost of delayed customer order j
hj = per unit penalty cost of unfulfilled customer order j
I r = subset of suppliers in region r
oi = per unit price of parts purchased from supplier i
pi = local disruption probability for supplier i
pr = regional disruption probability for all suppliers in region r
τi = delivery lead time from supplier i

Let I = {1, . . . , I} be the set of I suppliers, J = {1, . . . , J} the set of J customer

orders for products, and T = {1, . . . , T } the set of T planning periods (for notation
used, see Table 7.1).
Denote by bj and dj , respectively the size and the due date of customer order
j ∈ J, i.e., the number of units of ordered product type and the latest period of their
completion required to deliver the products to the customer by requested date. Let aj
be the unit requirement for the criticalpart of each product in customer order j ∈ J.
demand for all parts is A = j∈J aj bj and the total demand for all products
The total
is B = j∈J bj .
The customer orders are single-period orders such that each order must be com-
pleted in one planning period, e.g., Sawik (2011a). Assume that the producer has
194 7 A Fair Decision-Making Under Disruption Risks

limited time-varying capacity, and denote by Ct the producer capacity available in

planning period t ∈ T , and by cj the unit capacity consumption for each product in
customer order j ∈ J.
The orders for parts are assumed to be placed at the start of the planning horizon,
when all customer orders for products are known. Let oi be the unit purchasing price
of parts from supplier i ∈ I and denote by ei the fixed ordering cost of creating con-
tracts and maintaining relationships with supplier i ∈ I. Each supplier have sufficient
capacity to meet total demand for parts. The order preparation and transportation time
of a shipment from supplier i ∈ I to the producer is constant and equals to τi periods
so that the parts ordered from supplier i ∈ I are delivered in period τi and then can
be used for the assembly of products in period τi + 1, at the earliest.
Assume that the suppliers are located in a number of disjoint geographic
 regions
and denote by I r ⊆ I the subset of suppliers in region r ∈ R, where r∈R I r = I.
The supplies of parts are subject to random independent local disruptions of
each supplier individually, which may arise from equipment breakdowns, local labor
strike, fires, etc. Denote by pi the local disruption probability for supplier i.
In addition to independent local disruptions of each supplier, there are potential
correlated regional disasters that may result in disruption of all suppliers in the
same region simultaneously. For example, such regional disaster events may include
an earthquake, flooding, etc. Denote by pr the probability of correlated regional
disruption of all suppliers i ∈ I r in region r ∈ R. The regional disasters in each
region and the local disasters at each supplier are assumed to be independent events.
Let πi be the disruption probability of every supplier i ∈ I r , r ∈ R

πi = pr + (1 − pr )pi ; i ∈ I r , r ∈ R. (7.9)

Denote by S = {1, . . . , S} the index set of all disruption scenarios, where each
scenario s ∈ S is comprised of a unique subset Is ⊂ I of suppliers who deliver parts
without disruptions. All potential disruption scenarios will be considered, that is
S = 2I . For each scenario s ∈ S, the supplies from every supplier, i ∈ I \ Is , can
be disrupted either by a local or a regional disaster event. The probability Ps for
disruption scenario s ∈ S with the subset Is of non-disrupted suppliers, and with all
possible combinations of different disaster events considered, is (cf. Sect. 1.3)

Ps = Psr , (7.10)
r∈R

where Psr is the probability of realizing of disruption scenario s for suppliers in I r

   
(1 − pr ) i∈I r  Is (1 − pi ) i∈I r \Is pi if I r Is
= ∅
Psr =   (7.11)
pr + (1 − pr ) i∈I r pi if I r Is = ∅.

The producer can be charged with a contractual, order specific penalty cost for
delayed or unfulfilled customer orders, caused by the shortage of parts, that are
7.3 Problem Description 195

delivered late or not at all due to supply disruptions. Let gj and hj be, respectively
the per unit and per period penalty cost of delayed customer order j ∈ J and the per
unit total penalty cost of unfulfilled customer order j ∈ J.
The objective of the equitable optimization of a supply chain under disruption
risks is to allocate the total demand for parts among a subset of selected suppli-
ers and to schedule the customer orders for products over the planning horizon to
equitably minimize expected cost of ordering, purchasing and shortage of parts and
maximize expected customer service level, i.e., the fraction of customer orders (order
fulfillment rate) or customer demand (demand fulfillment rate) filled on or before
their due dates. The resulting equitably efficient supply portfolio (the allocation of
total demand for parts among the selected suppliers) is determined ahead of time
as well as the equitably efficient schedule of customer orders for every potential
disruption scenario.

7.4 Problem Formulation

In this section the time-indexed SMIP model ESPS_E is presented for selection of
equitably efficient supply portfolio and scheduling of customer orders to fairly min-
imize expected cost per product and maximize expected service level, the expected
fraction of customer orders filled on or before their due dates (i.e., expected order
fulfillment rate). The decision variables thatare jointly determined using the proposed
model are defined in Table 7.2.

Table 7.2 Variables: selection of equitably efficient supply portfolio and scheduling
First stage variables
ui = 1, if supplier i is selected; otherwise ui = 0 (supplier selection)
vi ∈ [0, 1], the fraction of total demand for parts ordered from supplier i (supply portfolio)
Auxiliary variables
λl unrestricted auxiliary variable, l = 1, 2
δkl ≥ 0, upside deviation of outcome value fk , k = 1, 2 from the value of λl , l = 1, 2
Second stage variables
wjts = 1, if under disruption scenario s customer order j is scheduled for period t; otherwise
wjts = 0 (production scheduling)

Let E c be the expected cost per product to be minimized and E sl , the expected
service level to be maximized
196 7 A Fair Decision-Making Under Disruption Risks
  
Ec = ( ei ui + Ps ( Aoi vi
i∈I s∈S i∈Is
   
+ gj bj (t − dj )wjts + hj bj (1 − wjts )))/B (7.12)
j∈J t∈T :t>dj j∈J t∈T
  
E sl = Ps wjts /J. (7.13)
j∈J t∈T :t≤dj s∈S

In order to avoid dimensional inconsistency among the two objectives, the values
ofc thec optimized objective functions are scaled into the interval [0,1]. Denote by f1 =
E −E c
c
E −E c
, the normalized expected cost per product (E c , E are the minimum and the
sl
E −E sl
maximum values of E c , respectively), and by f2 = sl , the normalized expected
E −E sl
sl
customer service level (E sl , E are the minimum and the maximum values of E sl ,
respectively).
The normalized objective functions f1 and f2 are defined below
    
f1 = ( ei ui + Ps ( Aoi vi + gj bj (t − dj )wjts
i∈I s∈S i∈Is j∈J t∈T :t>dj
  c
+ hj bj (1 − wjts )))/B − E c )/(E − E c ) (7.14)
j∈J t∈T
sl   
E − j∈J t∈T :t≤dj s∈S Ps wjts /J
f2 = sl
. (7.15)
(E − E ) sl

The SMIP model ESPS_E is formulated below. The model is based on the SMIP
formulations SPSm_E(c) and SPSm_E(sl) presented in Chap. 5. The objective func-
tion (7.16) subject to constraints (7.17) represent the so-called ordered weighted
averaging aggregation of the two conflicting criteria with equal weights assigned to
each criterion, see, OWA aggregation in Sect. 7.2. Applying OWA aggregation to
the bi-objective problem yields an equitably efficient solution to the problem. In the
model presented below λl are unrestricted variables, while nonnegative variables δkl
represent, for outcome values fk , their upside deviations from the value of λl (see,
Sect. 7.2).

ESPS_E: Risk-neutral selection of Equitably efficient Supply Portfolio and

Scheduling of customer orders to minimize expected cost and maximize
expected service level
Minimize


2 
2
(lλl + δkl ) (7.16)
l=1 k=1
7.4 Problem Formulation 197

subject to (7.14), (7.15) and

λl + δkl ≥ fk ; k, l = 1, 2 (7.17)

Demand allocation constraints:

– the total demand for parts must be fully allocated among the selected
suppliers,
– demand for parts cannot be assigned to non-selected suppliers,

vi = 1 (7.18)
i∈I
vi ≤ ui ; i ∈ I (7.19)

Order-to-period assignment constraints:

– for each disruption scenario s, each customer order j is either sched-
t∈T wjt = 1), or unscheduled and rejected
s
uled
 during the planning horizon (
( t∈T wjt = 0),
s

– for each disruption scenario s and each planning period t, the cumulative
demand for parts of all customer orders scheduled in periods 1 through t cannot
exceed the cumulative deliveries of parts in periods 1 through t − 1, from the
non-disrupted suppliers i ∈ Is ,
– for each disruption scenario s, the total requirement for parts of scheduled
customer orders is not greater than the total supplies from the non-disrupted
suppliers i ∈ Is ,


wjts ≤ 1; j ∈ J, s ∈ S (7.20)
t∈T
  
aj bj wjts  ≤ A vi ; t ∈ T , s ∈ S (7.21)
j∈J t  ∈T :t  ≤t i∈Is :τi ≤t−1
 
aj bj wjts ≤ A vi ; s ∈ S (7.22)
j∈J t∈T i∈Is

Producer capacity constraints:

– for any period t and each disruption scenario s, the total demand on
capacity of all customer orders scheduled in period t must not exceed the
producer capacity available in this period,

bj cj wjts ≤ Ct ; t ∈ T , s ∈ S (7.23)
j∈J

Non-negativity and integrality conditions:

198 7 A Fair Decision-Making Under Disruption Risks

δkl ≥ 0; k, l = 1, 2 (7.24)
ui ∈ {0, 1}; i ∈ I (7.25)
vi ∈ [0, 1]; i ∈ I (7.26)
wjts ∈ {0, 1}; j ∈ J, t ∈ T , s ∈ S. (7.27)

In the above model, each supplier is assumed to have sufficient capacity to meet
total demand for parts. Such an assumption allows the decision maker to select a
single sourcing solution, if such a portfolio is an equitably efficient supply portfolio.
However, the assumption can be easily relaxed to account for multiple capacitated
suppliers, see Sect. 7.4.2.

7.4.1 Minimum and Maximum Values of the Objective

Functions

In this subsection the minimum and maximum values for all objective functions are
calculated to determine the normalized values of the objective functions, f1 , (7.14), f2 ,
(7.15), that is, the values of the optimized objective functions scaled into the interval
[0,1]. Note that the cost and the service level objectives are in conflict. Therefore,
c
the minimum and maximum values of expected cost E c , E , and expected customer
sl
service level, E sl , E , are obtained by solving the following stochastic mixed integer
programs:

SPS_E(c): Risk-neutral selection of supply portfolio and scheduling of

customer orders to minimize expected cost per product
Minimize E c , (7.12),
subject to (7.18)–(7.23), (7.25)–(7.27).

SPS_E(sl): Risk-neutral selection of supply portfolio and scheduling of

customer orders to maximize expected service level
Maximize E sl , (7.13)
subject to (7.18)–(7.23), (7.25)–(7.27).

In problem SPS_E(c), E c is the minimized objective function, while E sl is not

considered. In problem SPS_E(sl), E sl is the maximized objective function, while
E c is not considered. Thus, by solving problem SPS_E(c), the minimum value E c of
E c and the minimum value E sl of E sl are determined. Similarly, by solving problem
7.4 Problem Formulation 199

sl c
SPS_E(sl), the maximum value E of E sl and the maximum value E of E c are
determined.
So far, in the proposed models the customer service level is measured by order
fulfillment rate, i.e., the number of customer orders fulfilled by their due dates, with
no account for the size of each customer order. For example, a high customer service
level can be achieved by fulfilling a large number of small-size orders, while leaving
the unfulfilled demand relatively high. To avoid such a solution, in particular when the
customer orders of different size are simultaneously considered, the service level can
be measured by demand fulfillment rate, i.e., the fraction of total customer demand
fulfilled by the requested due dates.
If the customer service level is defined as demand fulfillment rate, then E sl , (7.13)
and f2 , (7.15) should be replaced with the following formulae, (7.28) and (7.29),
respectively.
  
E sl = Ps bj wjts /B (7.28)
j∈J t∈T :t≤dj s∈S
sl   
E − j∈J t∈T :t≤dj s∈S Ps bj wjts /B
f2 = sl
, (7.29)
(E − E ) sl

sl
where E sl , E are the minimum and the maximum values of E sl , (7.28), respectively.
sl
The values of E sl , E can be determined using the models SPS_E(c) and SPS_E(sl).
In the computational examples presented in the next section, the two metrics of
the customer service level will be considered and compared against each other.

7.4.2 Model Limitations and Possible Enhancements

The proposed model has been developed to support decision-making in a make-to-

order environment under disruption risks. The model, however, has been formulated
under various simplified assumptions that may limit its practical usefulness. The
basic assumptions are listed below.
1. A single critical part type is required to fulfill all customer orders for products.
2. The orders for parts are placed at the start of the planning horizon, when all
customer orders for products are known.
3. Each supplier have sufficient capacity to meet the total demand for parts.
4. The order preparation time at each supplier is constant, independent of order size,
and all parts ordered from a supplier are delivered during a fixed transportation
time.
5. Transportation costs are not explicitly considered and the unit purchasing price
from each supplier is constant, independent of total volume or value of order for
parts, i.e., no quantity or business volume discounts are considered.
200 7 A Fair Decision-Making Under Disruption Risks

6. The customer orders are single-period orders such that each order must be com-
pleted in one planning period.
7. The penalty costs for delayed or unfulfilled customer demand are linear.
8. The inventory of parts and products are not considered.
9. The two conflicting objectives: reduction of expected cost and increase of
expected customer service level are equally important for the decision maker.
Some of the above assumptions can be easily relaxed, while the other needs a
more advanced model to be developed. Possible relaxations of the corresponding
assumptions and the model enhancements are listed below.
1. The model can be easily enhanced for multiple part types required to fulfill all
customer orders with different subsets of part types needed for different product
types and different subsets of capacitated suppliers capable of providing different
subsets of part types, e.g., Sawik (2013c).
2. A rolling planning horizon approach can be used to account for a dynamic arrival
and scheduling of customer orders as well as the corresponding supply portfolio.
In practice, however, the supply portfolio needs to be decided at the start of the
planning horizon, based on a forecast of the customer demand.
3. The model can be easily enhanced for multiple capacitated suppliers by the
addition of suppliers capacity constraints.
4. The more advanced model can be developed to consider order-dependent process-
ing and transportation times to better coordinate manufacturing and transportation
of parts and production of finished products.
5. The model can be easily enhanced to account for quantity or business volume
discounts (e.g., Sawik 2010) and the unit purchasing price from each supplier can
include unit transportation cost.
6. The model can be easily enhanced to account for large, multi-period customer
orders that cannot be completed in one period and must be split into single-period
portions to be processed in consecutive planning periods, e.g., Sawik (2011a).
7. The introduction of non-linear penalty costs may lead to a non-linear SMIP model.
The model, however, can be linearized in some cases, e.g., by using a piecewise
linear representation of the non-linear penalty cost.
8. The model can be enhanced to account for the output inventory of parts at sup-
pliers, the input inventory of parts at the producer and the output inventory of
products waiting for shipment to customers. The inventory balance constraints
can be added to the model and the producer inventory holding costs, to the cost-
based objective function.
9. The number of conflicting and equally important objectives can be increased, for
example by the addition of responsiveness (e.g., Liu and Papageorgiou 2013) as
another objective function.
7.5 Computational Examples 201

7.5 Computational Examples

In this section the proposed SMIP approach for selection of equitably efficient supply
portfolio and scheduling of customer orders in a supply chain under disruption risks
is compared with the weighted-sum approach and illustrated with computational
examples. The following parameters have been used for the example problems (cf.
Sect. 6.6):
• I = 9, J = 25, T = 10, and S = 2I = 512;
• R = {1, 2, 3}, and I 1 = {1, 2, 3}, I 2 = {4, 5, 6}, I 3 = {7, 8, 9};
• τi , the order preparation and shipping times from suppliers were 2, 3 and 4 time
periods, respectively for suppliers i ∈ I 1 , i ∈ I 2 and i ∈ I 3 ;
• aj ∈ {1, 2, 3}, bj ∈ {500, 1000, . . . , 5000}, cj ∈ {1, 2, 3}, dj ∈ {1 + mini∈I (τi ),
. . . , T };
• Ct , the capacity
 of producer in each period t, was integer drawn from
1000(2 j∈J bj cj /(T − maxi∈I τi ))U[0.75; 1.25]/1000 distribution, i.e., in
each period the producer capacity was from 75% to 125% of the double capacity
required to complete all customer orders during the planning horizon, after the
latest delivery of parts;
• ei ∈ {5000, 6000, . . . , 10000}, i ∈ I 1 , ei ∈ {10000, 11000, . . . , 15000}, i ∈ I 2
and ei ∈ {15000, 11000, . . . , 30000}, i ∈ I 3 ;
• oi , the unit price of parts purchased from supplier i, was uniformly distributed over
[11,16], [6,11] and [1,6], respectively for suppliers i ∈ I 1 , i ∈ I 2 and i ∈ I 3 ;
• gj = aj maxi∈I (oi )/350 , j ∈ J, i.e., the unit penalty cost per period of each
delayed customer order j was approximately 0.28% of the maximum unit price of
required parts;
• hj = 2aj maxi∈I (oi ) , j ∈ J, i.e., the unit penalty cost of each unfulfilled customer
order j was approximately twice as large as the maximum unit price of required
parts;
• pi , the local disruption probability was uniformly distributed over [0.005,0.01],
[0.01,0.05] and [0.05;0.10], respectively for suppliers i ∈ I 1 , i ∈ I 2 and i ∈ I 3 ;
• p1 = 0.001, p2 = 0.005 and p3 = 0.01.
The detailed data set was based on the example presented in Sawik (2013c), e.g.,:
unit requirements for parts, a = (2, 1, 3, 3, 1, 3, 2, 1, 2, 2, 2, 2, 3, 2, 1, 3, 2, 1, 3, 3,
2, 1, 1, 2, 1);
size of customer orders, b = (1, 2, 9, 7, 8, 5, 1, 7, 5, 4, 7, 4, 10, 6, 8, 1, 4, 2, 4, 8, 6,
3, 8, 7, 3) × 500,
(the resulting total demand for parts and products is A = 132500 and B = 66000,
respectively);
unit capacity consumption, c = (2, 1, 1, 2, 3, 3, 1, 3, 2, 1, 2, 1, 3, 1, 1, 3, 2, 3, 1, 1,
3, 2, 2, 1, 2);
producer available capacity, Ct = C = 38000, ∀t = 1, . . . , 10;
unit prices, o = (13, 12, 12, 8, 6, 6, 2, 5, 4);
local disruption probabilities, p = (0.00513571, 0.00666354, 0.00902974, 0.0356206,
202 7 A Fair Decision-Making Under Disruption Risks

0.040175, 0.0294692, 0.0519967, 0.0827215, 0.0739062);

and the corresponding disruption probabilities (7.9), π = (0.00613057, 0.00765688,
0.0100207, 0.0404425, 0.0449741, 0.0343219, 0.0614767, 0.0918943, 0.0831672).
Note that the constant producer
 capacity, C = 38000, allows for completing all
customer orders in at most  j∈J bj cj /C = 3.26 = 4 planning periods, that is,
in less than (T − maxi∈I τi ) = 6 periods remaining in the planning horizon after the
latest delivery of parts.
In the computational experiments all potential disruption scenarios, S = 2I =
512, and all possible combinations of local and regional disaster events were consid-
ered. Each scenario s ∈ S with the subset Is of non-disrupted suppliers is represented
by an I-dimensional 0-1 vector with 1, if i ∈ Is , i.e., if supplier i is not disrupted, and
0; otherwise. The corresponding disruption probability, Ps , for each scenario s ∈ S
was calculated using formulae (7.10) and (7.11).
The unit price per part oi and the disruption probability πi , (7.9), of each supplier
i ∈ I are shown in Fig. 6.2. The figure indicates that the most reliable (with the lowest
disruption probability, π1 = 0.00613057) is supplier 1, the least reliable (with the
highest disruption probability, π8 = 0.0918943) is supplier 8, the most expensive
(with the highest price per part, o1 = 13) is supplier 1, and the cheapest (with the
lowest price per part, o7 = 2) is supplier 7. Note that the geographic regions are
numbered in such a way that the unit prices are nonincreasing with r, while the fixed
ordering costs and the disruption probabilities are nondecreasing with r.
The solution results are presented in Table 7.3. In addition to the optimal absolute
and normalized solution values for the primary objective functions and the alloca-
tion of demand among the selected suppliers, Table 7.3 presents the expected values
of the associated objective function, i.e., the minimum expected service level, E sl ,
c
for model SPS_E(c) and the maximum expected cost per product, E , for model
SPS_E(sl). Table 7.3 indicates that for the cost-based objective (model SPS_E(c))
the cheapest supplier i = 7 is selected only, while for the customer service level
(model SPS_E(sl)), the total demand for parts is allocated among the three most
reliable and most expensive suppliers i = 1, 2, 3 for objective (7.13) and among
the two suppliers i = 1, 2 for objective (7.28). For the equitable solution (model
ESPS_E), the supply portfolio contains one reliable and expensive supplier i = 2
and two low-cost and unreliable suppliers i = 6, 7 for both (7.15) and (7.29), service
level objectives. 
As an illustrative example, Fig. 7.1 presents the demand for products, j∈J:dj =t bj ,
 
t ∈ T , and the expected production schedules, s∈S Ps j∈J bj wjts , t ∈ T for the opti-
mal cost, optimal customer service level and for the equitably efficient solution.
Figure 7.1 compares the expected production schedules for the two different metrics
of customer service level: (a) order fulfillment rate (i.e., fraction of customer orders
fulfilled on time), (7.13), and (b) demand fulfillment rate (i.e., fraction of customer
demand fulfilled on time), (7.28). While for the two different service level metrics,
the corresponding supply portfolios are very similar (cf. Table 7.3), and the expected
production schedules with respect to the two service level objectives are also very
similar, the corresponding schedules for the equitably efficient solutions are different.
7.5 Computational Examples 203

Table 7.3 Solution results

Model SPS_E(c): Var. = 100468, Bin. = 100459, Cons. = 21341, Nonz. = 765902 (a)
Expected Cost (E c ) 7.66
Suppliers Selected (% of total demand) 7(100)
Expected Service Level (E sl ) 67.60 (b) , 66.32 (c)
Model SPS_E(sl): Var. = 100468, Bin. = 100459, Cons. = 21341, Nonz. = 765902 (a)
sl
Expected Service Level (E ) 99.62 (b)
Suppliers Selected (% of total demand) 1(48), 2(31), 3(21)
c
Expected Cost (E ) 25.64
Model ESPS_E: Var. = 100476, Bin. = 100459, Cons. = 21347, Nonz. = 921880 (a)
Expected Cost 9.31
Expected Service Level 96.06 (b)
Normalized Expected Cost 0.098
Normalized Expected Service Level 0.111
Suppliers Selected (% of total demand) 2(6), 6(13), 7(81)
Model SPS_E(sl) with (7.13) replaced by (7.28)
sl
Expected Service Level (E ) 99.49 (c)
Suppliers Selected (% of total demand) 1(55), 2(45)
c
Expected Cost (E ) 25.61
Model ESPS_E with (7.15) replaced by (7.29)
Expected Cost 9.25
Expected Service Level 95.29 (c)
Normalized Expected Cost 0.088
Normalized Expected Service Level 0.127
Suppliers Selected (% of total demand) 2(6), 6(10), 7(84)
(a) Var. = number of variables, Bin. = number of binary variables,

Cons. = number of constraints, Nonz. = number of nonzero coefficients.

(b) (
  
t∈T :t≤dj wjt /J)100%.
s
s∈S Ps j∈J
(c) (
  
t∈T :t≤dj bj wjt /B)100%.
s
s∈S Ps j∈J

In general, the service level-based solution, when no cost components are included
in the objective function, better meets the customer demand, with the smallest fraction
of unfulfilled demand. The total customer demand is met with only a small fraction
of the expected unfulfilled demand: 0.0615 for the cost-based solution, 0.0055 for
the service level-based solution (a), 0.0051 for the service level-based solution (b),
0.0468 for the equitably efficient solution (a), and 0.0470 for the equitably efficient
solution (b). In addition, for the service level-based objective functions, the expected
production schedule approximately follows the customer demand pattern, while for
the minimum cost objective function the most unbalanced production schedule is
achieved.
204 7 A Fair Decision-Making Under Disruption Risks

Fig. 7.1 Expected production schedules: a order fulfillment rate; b demand fulfillment rate

In order to compare the solution results for the two different service level-
based objective functions (7.13) and (7.28), the computational experiments were
repeated for another example with more diversified size of customer orders, bj ∈
{500, 1000, . . . , 12000}, i.e., from 500 to 12000 products, for the same total demand
for parts and products, A = 132500 and B = 66000. The solution results are pre-
sented in Table 7.4, and Fig. 7.2 compares the expected production schedules for the
two different metrics of customer service level. In general, the results presented in
Table 7.4 are similar to those in Table 7.3, in particular the supply portfolios are very
similar. However, Fig. 7.2 shows that for the more diversified customer orders, the
less smoothed expected production schedules are obtained for the equitably efficient
solutions. On the other hand, Table 7.4 shows that the equitably efficient solutions
with a perfect equity were found (i.e., with identical values of normalized expected
cost and normalized expected customer service level), which indicates that the
obtained supply portfolios and schedules of customer orders are also the lexico-
graphic minimax optimal solutions as well as the Pareto-optimal solutions (see, Liu
and Papageorgiou 2013, Liu et al. 2014).
7.5 Computational Examples 205

Table 7.4 Solution results: diversified customer orders

Model SPS_E(c)
Expected Cost (E c ) 7.44
Suppliers Selected (% of total demand) 7(100)
Expected Service Level (E sl ) 71.32 (a) , 78.83 (b)
Model SPS_E(sl)
sl
Expected Service Level (E ) 99.77 (a)
Suppliers Selected (% of total demand) 1(42), 2(41), 3(17)
c
Expected Cost (E ) 25.15
Model ESPS_E
Expected Cost 9.38
Expected Service Level 97.48 (a)
Normalized Expected Cost 0.110
Normalized Expected Service Level 0.110
Suppliers Selected (% of total demand) 2(3), 6(22), 7(75)
Model SPS_E(sl) with (7.13) replaced by (7.28)
sl
Expected Service Level (E ) 99.47 (b)
Suppliers Selected (% of total demand) 1(65), 2(35)
c
Expected Cost (E ) 25.43
Model ESPS_E with (7.15) replaced by (7.29)
Expected Cost 10.44
Expected Service Level 95.99 (b)
Normalized Expected Cost 0.168
Normalized Expected Service Level 0.168
Suppliers Selected (% of total demand) 2(3), 5(5), 6(30), 7(62)
(a) (
   s /n)100%
P
s∈S s j∈J w
t∈T :t≤dj jt
(b) (
  
t∈T :t≤dj bj wjt /B)100%
s
s∈S Ps j∈J

The solution results demonstrate that for the minimum cost objective the cheapest
supplier is usually selected, for the maximum service level objective a subset of
most reliable and most expensive suppliers is usually chosen, whereas the equitably
efficient supply portfolio usually combines the two types of suppliers.
Weighted-Sum Approach
The equitably efficient solutions obtained using model ESPS_E have been com-
pared with the nondominated solutions obtained by minimizing the weighted-sum
aggregation of the two normalized objective functions, f1 , (7.14) and f2 , (7.29), i.e.,
the weighted-sum of expected cost per product and expected fraction of customer
demand fulfilled on time. The weighted-sum model WSPS_E is shown below and
the solution results for the example with similar and diversified customer orders are
presented in Tables 7.5 and 7.6, respectively.
206 7 A Fair Decision-Making Under Disruption Risks

Fig. 7.2 Expected production schedules for diversified customer orders: a order fulfillment rate;
b demand fulfillment rate

WSPS_E: Risk-neutral selection of supply portfolio and scheduling of

customer orders to minimize weighted-sum of normalized expected cost and
service level
Minimize

λf1 + (1 − λ)f2 , (7.30)

where 0 ≤ λ ≤ 1,
subject to (7.14), (7.18)–(7.23), (7.25)–(7.27), (7.29).

Tables 7.5 and 7.6 indicate that for λ = 1 (minimization of cost) the cheapest
supplier i = 7 is selected only, while for λ = 0 (maximization of customer service
level) the total demand for parts is allocated among the two most reliable and most
expensive suppliers i = 1, 2. As λ increases from 0 to 1, i.e., the decision maker
preference shifts from customer service level to cost, more demand is moved from
7.5 Computational Examples 207

Table 7.5 Nondominated solutions

λ 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Var. = 100470, Bin. = 100459, Cons. = 21343, Nonz. = 921868 (a)
Exp.Cost 25.61 19.40 11.93 9.25 9.25 8.88 8.88 8.88 8.88 8.01 7.66
Exp.Service Level(b) 99.48 98.57 96.86 95.29 95.29 94.73 94.73 94.73 94.73 82.69 66.32
Normalized Exp.Cost 1 0.654 0.238 0.088 0.088 0.068 0.068 0.068 0.068 0.020 0
Normalized Exp.Service Level 0 0.028 0.079 0.127 0.127 0.143 0.143 0.143 0.143 0.506 1
Suppliers Selected 1(55) 1(6)
(% of total demand) 2(45) 2(44) 2(10) 2(6) 2(6) 2(6) 2(6) 2(6) 2(6)
6(50) 6(40) 6(10) 6(10) 6(5) 6(5) 6(5) 6(5) 6(5)
7(50) 7(84) 7(84) 7(89) 7(89) 7(89) 7(89) 7(95) 7(100)
(a) Var. = number of variables, Bin. = number of binary variables,
Cons.
(b) (
= number
 ofconstraints, Nonz. = number of nonzero coefficients
t∈T :t≤dj bj wjt /B)100%
s
s∈S Ps j∈J

expensive and reliable suppliers to low-cost, unreliable suppliers. For the example
with similar customer orders, the subset of nondominated solutions contains (for
λ = 0.3, 0.4) the optimal solution obtained for model ESPS_E with constraint (7.21)
(cf., Tables 7.3 and 7.5). In contrast to the example with diversified customer orders
(cf., Tables 7.4 and 7.6). For diversified orders, Fig. 7.3 shows the nondominated
supply portfolios (the allocation of total demand for parts among selected suppliers)
for 11 levels of trade-off parameter λ. The subset of selected suppliers consists of
four suppliers i = 1, 2, 6, 7 of which suppliers i = 1, 2 are most reliable and suppliers
i = 6, 7 are the cheapest suppliers in region r = 2, 3, respectively.
For the example with diversified customer orders, the trade-off between the
expected cost and the expected customer service level is clearly shown in Fig. 7.4,
where the efficient frontier is presented. The results emphasize the effect of varying
service level/cost preference of the decision maker; the higher the trade-off parameter
λ, the more cost-oriented the decision-making.

Table 7.6 Nondominated solutions: diversified customer orders

λ 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Var. = 100470, Bin. = 100459, Cons. = 21343, Nonz. = 921868 (a)
Exp.Cost 25.43 24.81 12.95 11.00 8.61 8.61 8.31 8.31 8.31 8.31 7.44
Exp.Service Level(b) 99.47 99.42 97.07 96.30 94.90 94.90 94.45 94.45 94.45 94.45 78.83
Normalized Exp.Cost 1 0.9655 0.3062 0.1977 0.0650 0.0650 0.0486 0.0486 0.0486 0.0486 0
Normalized Exp.Service Level 0 0.0025 0.1169 0.1532 0.2226 0.2224 0.2439 0.2440 0.2442 0.2446 1
Suppliers Selected 1(65) 1(33)
(% of total demand) 2(35) 2(67) 2(13) 2(8) 2(4) 2(4) 2(3) 2(3) 2(3) 2(3)
6(53) 6(35) 6(9) 6(9) 6(5) 6(5) 6(5) 6(5)
7(34) 7(57) 7(87) 7(87) 7(92) 7(92) 7(92) 7(92) 7(100)
(a) Var. = number of variables, Bin. = number of binary variables,
Cons.
(b) (
= number
 ofconstraints, Nonz. = number of nonzero coefficients
t∈T :t≤dj bj wjt /B)100%
s
s∈S Ps j∈J
208 7 A Fair Decision-Making Under Disruption Risks

Supply Portfolios
100%

80%
Demand allocation
60%

40%

20%

0%
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Trade-off parameter lambda

Supplier1 Supplier2 Supplier6 Supplier7

Fig. 7.3 Nondominated supply portfolios: diversified customer orders

The computational experiments were performed using the AMPL programming

language and the CPLEX 12.5 solver on a MacBookPro laptop with Intel Core
i7 processor running at 2.8 GHz and with 16GB RAM. The solver was capable of
finding proven optimal solution for all examples with CPU time ranging from several
seconds to a few hours.

Pareto Frontier
100
lambda=0
Customer Service Level [%]

95

90

85

80
lambda=1

75
5 10 15 20 25 30
Cost per Product

Fig. 7.4 Pareto front for model WSPS_E: diversified customer orders

7.6 Notes

The research on equitably efficient multiple objective optimization of a supply chain

under disruption risks is rarely reported in the literature. In order to generate an
equitably efficient solution with respect to multiple, equally important objective
functions, the lexicographic minimax method, e.g., Kostreva et al. (2004), as a special
7.6 Notes 209

case of the ordered weighted averaging aggregation introduced Yager (1988), can be
applied. The lexicographic minimax problem can be transferred to a lexicographic
minimization problem, e.g., Erkut et al. (2008), Ogryczak et al. (2008). Recently
Liu and Papageorgiou (2013) developed an approach to transfer the lexicographic
minimax problem to a minimization optimization problem, instead of a lexicographic
minimization problem, which needs to solve a sequence of optimization problems
iteratively. The recent approach, however, is restricted to some special cases of a
multiple objective problem (Liu et al. 2014).
The material presented in this chapter is based on research reported by Sawik
(2014d, 2015a), where a SMIP approach was proposed for the integrated selection of
supply portfolio and scheduling of customer orders in a supply chain under disruption
risks. In order to equitably optimize expected cost and expected service level, the
ordered weighted averaging aggregation of the two conflicting objective functions
was applied. The equitably efficient solutions obtained for the ordered weighted
averaging aggregation were compared with nondominated solutions obtained using
the weighted-sum aggregation approach.
In the proposed model, each supplier is assumed to have sufficient capacity to
meet total demand for parts, which allows the decision maker to select a single
sourcing type of a supply portfolio, if it is an equitably efficient solution. In the future
research, however, that assumption can be easily relaxed to account for multiple
capacitated suppliers. Furthermore, the other assumptions can also be relaxed to
develop a more advanced model (for possible model relaxations and enhancements,
see Sect. 7.4.2). In particular, the future research should focus on equitably efficient
decision-making with respect to more equally important and conflicting objectives
such as responsiveness, robustness, competitiveness, etc.

Problems

7.1 Modify the SMIP models presented in this chapter for multiple part types with
subsets of part types required for each customer order and subsets of capacitated
suppliers available for each part type.

7.2 Enhance the SMIP models presented in this chapter for multi-period customer
orders that cannot be completed in one period and must be split into single-period
portions to be processed in consecutive planning periods (see, Sawik 2011a).

7.3 Limited storage space for parts and products

Enhance the SMIP models presented in this chapter:
(a) for the limited input inventory of parts at the producer;
(b) for the limited output inventory of products at the producer.
Formulate the inventory balance constraints that should be added to the SMIP models.
210 7 A Fair Decision-Making Under Disruption Risks

7.4 In addition to cost and service level, what are the other conflicting and equally
important objective functions that can be considered when selecting supply portfolio
and scheduling customer orders in the presence of supply chain disruption risks?

7.5 How would you select a nondominated solution obtained using the weighted-
sum approach that is as close as possible to the equitably efficient solution?
Chapter 8
A Robust Decision-Making Under
Disruption Risks

8.1 Introduction

In this chapter we look for an equitably efficient solution with respect to both average-
case and worst-case performance measures of a supply chain. Such an equitably
efficient average and worst-case solution, or equivalently equitably efficient risk-
neutral and risk-averse solution will be called a fair mean-risk solution. The solution
will equitably focus on the two objective functions: the expected value (average-case
performance measure) and the expected worst-case value (worst-case performance
measure), i.e., Conditional Value-at-Risk of the selected optimality criterion, cost or
service level. The fair mean-risk decision-making aims at achieving the normalized
expected and expected worst-case values of the selected objective function as much
close to each other as possible, that is, the decision-making aims at equalizing the
distance to optimality both under business-as-usual and under worst-case conditions.
The mean-risk fairness reflects the decision makers common requirement to maintain
an equally good performance of a supply chain under varying operating conditions,
which is close to the idea of robustness of a supply chain. Therefore, the mean-risk
fairness, i.e., the equitably efficient performance of a supply chain in the average-case
as well as in the worst-case, in this chapter will be called robustness.
In a make-to-order environment, the choice between high-cost, reliable suppliers
and low-cost, unreliable suppliers has a direct impact on both cost and service level.
The more reliable and expensive suppliers selected, the higher the service level and
the purchasing costs, and the lower penalties for delayed or unfulfilled customer
orders. While the risk-neutral solution for business-as-usual conditions does not
account for high impact disruptions and hence may imply poor results when such
disruptions occur, the risk-averse solution that focuses on potential worst-case losses
may not perform well under business-as-usual conditions. It is often most desirable
to equitably consider these conflicting objectives, that is, the expected value and the
expected worst-case value of cost or customer service level. The models presented

© Springer International Publishing AG 2018 211

T. Sawik, Supply Chain Disruption Management Using Stochastic Mixed
Integer Programming, International Series in Operations Research
& Management Science 256, DOI 10.1007/978-3-319-58823-0_8
212 8 A Robust Decision-Making Under Disruption Risks

in this chapter provide such a mean-risk fairness in decision-making to equitably

optimize average and worst-case performance of a customer-driven supply chain.
The following time-indexed SMIP models are presented in this chapter:

RSPS_ECV(c) for selection of robust supply portfolio and scheduling of cus-

tomer orders to equitably optimize expected cost and CVaR of cost;
RSPS_ECV(sl) for selection of robust supply portfolio and scheduling of cus-
tomer orders to equitably optimize expected service level and CVaR of
service level;
SPS_E(c) for risk-neutral selection of supply portfolio and scheduling of cus-
tomer orders to minimize expected cost;
SPS_E(sl) for risk-neutral selection of supply portfolio and scheduling of cus-
tomer orders to maximize expected service level;
SPS_CV(c) for risk-averse selection of supply portfolio and scheduling of
customer orders to minimize CVaR of cost;
SPS_CV(sl) for risk-averse selection of supply portfolio and scheduling of
customer orders to maximize CVaR of service level;
SPS_E(c,α), enhanced model SPS_E(c) for computing the maximum CVaR
of cost, for a given confidence level α;
SPS_E(sl,α), enhanced model SPS_E(sl) for computing the minimum CVaR
of service level, for a given confidence level α;
SPS_ECV(c) for mean-risk minimization of expected cost and CVaR of cost;
SPS_ECV(sl) for mean-risk maximization of expected service level and CVaR
of service level.

Numerical examples and some computational results, in particular comparison of

the mean-risk with the fair mean-risk approach, are provided in Sect. 8.4.

8.2 Problem Description

In a supply chain under consideration various types of products are assembled over
a planning horizon by a single producer to meet customer demand, using the same
critical part type that can be manufactured and provided by different suppliers. For
detailed problem description see Sect. 7.3, and the notation used is introduced in
Table 8.1.
The objective of the equitably efficient optimization of average and worst-case
performance of a supply chain under disruption risks is to allocate the total demand
for parts among a subset of selected suppliers and to schedule the customer orders for
products over the planning horizon to equitably minimize expected cost and expected
worst-case cost or equitably maximize expected service level and expected worst-
case service level. The cost includes the cost of ordering and purchasing of parts plus
penalty cost of delayed and unfulfilled customer orders due to the parts shortages,
8.2 Problem Description 213

Table 8.1 Notation: selection of robust supply portfolio and scheduling

Indices
i = supplier, i ∈ I
j = customer order, j ∈ J
r = region, r ∈ R
s = disruption scenario, s ∈ S
t = planning period, t ∈ T
Input Parameters
aj = per unit requirement for parts of each product in customer order j
bj = size (number of products) of customer order j

A = j∈J a j b j , total demand for parts

B = j∈J b j , total demand for products
cj = per unit capacity consumption of producer for customer order j
Ct = capacity of producer in period t
dj = due date for customer order j
ei = fixed cost of ordering parts from supplier i
gj = per unit and per period penalty cost of delayed customer order j
hj = per unit penalty cost of unfulfilled customer order j
Ir = subset of suppliers in region r
oi = per unit price of parts purchased from supplier i
pi = local disruption probability for supplier i
pr = regional disruption probability for all suppliers in region r
τi = delivery lead time from supplier i

while the customer service level is a performance measure independent of any cost
parameters, defined as the fraction of customer orders filled on or before their due
dates (i.e., order fulfillment rate). The equitable solution means an equitably effi-
cient risk-neutral and risk-averse solution. In this chapter such an equitably efficient
solution will be called a robust solution. The robust solution is capable of equitably
optimizing the performance of a supply chain in the average-case as well as in the
worst-case. The robust solution (the supply portfolio and the schedule of customer
orders) aims at achieving the normalized expected and expected worst-case values
of the selected objective function as much close to each other as possible. To this
end, the ordered weighted averaging aggregation of the expected value and condi-
tional value-at-risk of the selected objective function will be applied. The resulting
robust supply portfolio (the allocation of total demand for parts among the selected
suppliers) is determined ahead of time as well as the equitably efficient schedule of
customer orders for every potential disruption scenario.
214 8 A Robust Decision-Making Under Disruption Risks

8.3 Problem Formulation

In this section the two time-indexed SMIP models RSPS_ECV(c) and RSPS_ECV
(sl) are proposed for selection of robust supply portfolio and customer order schedul-
ing to equitably optimize average and worst-case performance of a supply chain in the
presence of disruption risks. The models are based on the bi-objective optimization
problem formulation (7.8). The objective of model RSPS_ECV(c) is to equitably
minimize expected cost per product and expected worst-case cost per product and
the objective of model RSPS_ECV(sl) is to equitably maximize expected service
level and expected worst-case service level. The problem variables are introduced in
Table 8.2.

Table 8.2 Variables: selection of robust supply portfolio and scheduling

First stage variables
ui = 1, if supplier i is selected; otherwise u i = 0 (supplier selection)
vi ∈ [0, 1], the fraction of total demand for parts ordered from supplier i (supply portfolio)
Auxiliary variables
λl unrestricted auxiliary variable, l = 1, 2
δkl ≥ 0, upside deviation of outcome value f k , k = 1, 2 from the value of λl , l = 1, 2
Second stage variables
wsjt = 1, if under disruption scenario s customer order j is scheduled for period t; otherwise
wsjt = 0 (production scheduling)
Auxiliary variables
VaRc Cost-at-Risk, the targeted cost such that for a given confidence level α, for 100α% of
the scenarios, the outcome is below VaRc
VaRsl Service-at-Risk, the targeted service level such that for a given confidence level α, for
100α% of the scenarios, the outcome is above VaRsl
Cs ≥ 0, the tail cost for scenario s, i.e., the amount by which costs in scenario s exceed VaRc
Ss ≥ 0, the tail service level for scenario s, i.e., the amount by which VaRsl exceeds service
level in scenario s
zs = 1, if for scenario s, cost per product is not less than VaRc ; otherwise z s = 0
ζs = 1, if for scenario s, service level is not less than VaRsl ; otherwise ζs = 0

To control the risk of supply disruptions, the two percentile measures of risks
will be applied: Value-at-Risk (VaR) and Conditional Value-at-Risk (CVaR). In the
selection of supply portfolio and scheduling of customer orders under disruption
risks, the decision maker controls the risk of high losses due to supply disruptions
by choosing the confidence level α. The greater the confidence level α, the more
risk aversive is the decision maker and the smaller percent of the highest cost (or the
lowest service level) outcomes is focused on.
8.3 Problem Formulation 215

8.3.1 Equitable Minimization of Average and Worst-Case

Costs

Let VaRc be the targeted cost such that for a given confidence level α, for 100α% of
the scenarios, the outcome is below VaRc , and let CVaRc be the expected cost in the
worst 100(1 − α)% of the cases with the cost above VaRc . Define Cs as the tail cost
for scenario s, where tail cost is defined as the amount by which costs in scenario s
exceed VaRc . The risk-averse supply portfolio and the production schedule will be
optimized by calculating VaRc and minimizing CVaRc simultaneously.
Let E c be the minimized expected cost per product and C V a R c , the minimized
expected worst-case cost per product
  
Ec = ( ei u i + Ps ( Aoi vi
i∈I s∈S i∈Is
   
+ g j b j (t − d j )wsjt + h j b j (1 − wsjt )))/B (8.1)
j∈J t∈T :t>d j j∈J t∈T

C V a R c = V a R c + (1 − α)−1 Ps Cs . (8.2)
s∈S

In order to avoid dimensional inconsistency among various objectives, the values

of the optimized objective functions are scaled into the interval [0,1]. Denote by
f 1c = EE c −E
c c c
−E c
, the normalized expected cost per product (E c , E are the minimum and
C V a R c −C V a R c
the maximum values of E c , respectively), and by f 2c = c
C V a R −C V a R c
, the normalized
c
expected worst-case cost per product (C V a R c , C V a R are the minimum and the
maximum values of C V a R c for a given confidence level α, respectively).
The normalized cost objective functions f 1c and f 2c are defined below
    
f 1c = ( ei u i + Ps ( Aoi vi + g j b j (t − d j )wsjt
i∈I s∈S i∈Is j∈J t∈T :t>d j
  c
+ h j b j (1 − wsjt )))/B − E c )/(E − E c ) (8.3)
j∈J t∈T
 c
f 2c = (V a R c + (1 − α)−1 Ps Cs − C V a R c )/(C V a R − C V a R c ) (8.4)
s∈S

The SMIP model RSPS_ECV(c) for selection of robust supply portfolio and
customer order scheduling to equitably minimize expected and expected worst-case
cost per product is formulated below.
216 8 A Robust Decision-Making Under Disruption Risks

RSPS_ECV(c): Selection of Robust Supply Portfolio and Scheduling of

customer orders to equitably optimize expected cost and CVaR of cost
Minimize


2 
2
(lλl + δkl ) (8.5)
l=1 k=1

subject to (8.3), (8.4) and

λl + δkl ≥ f kc ; k, l = 1, 2 (8.6)

Demand allocation constraints:

– the total demand for parts must be fully allocated among the suppliers,
– demand for parts cannot be assigned to non-selected suppliers,

vi = 1 (8.7)
i∈I
vi ≤ u i ; i ∈ I (8.8)

Order-to-period assignment constraints:

– for each disruption scenario s, each customer order j is either sched-
t∈T w jt = 1), or unscheduled and rejected
s
uled
 during the planning horizon (
( t∈T w jt = 0),
s

– for each disruption scenario s and each planning period t, the cumulative
demand for parts of all customer orders scheduled in periods 1 through t cannot
exceed the cumulative deliveries of parts in periods 1 through t − 1, from the
non-disrupted suppliers i ∈ Is ,
– for each disruption scenario s, the total requirement for parts of scheduled
customer orders is not greater than the total supplies from the non-disrupted
suppliers i ∈ Is ,

wsjt ≤ 1; j ∈ J, s ∈ S (8.9)
t∈T
  
a j b j wsjτ ≤ A vi ; t ∈ T, s ∈ S (8.10)
j∈J τ ∈T :τ ≤t i∈Is :τi ≤t−1
 
a j b j wsjt ≤ A vi ; s ∈ S (8.11)
j∈J t∈T i∈Is

Producer capacity constraints:

– for any period t and each disruption scenario s, the total demand on
capacity of all customer orders scheduled in period t must not exceed the
8.3 Problem Formulation 217

producer capacity available in this period,


b j c j wsjt ≤ Ct ; t ∈ T, s ∈ S (8.12)
j∈J

Risk constraints:
– the tail cost for scenario s is defined as the nonnegative amount by which
cost in scenario s exceeds VaRc ,
 
Cs ≥ ei u i /B + Aoi vi /B
i∈I i∈Is
   
+ g j b j (t − d j )wsjt /B + h j b j (1 − wsjt )/B
j∈J t∈T :t>d j j∈J t∈T

−V a R ; s ∈ S
c
(8.13)

Non-negativity and integrality conditions:

δkl ≥ 0; k, l = 1, 2 (8.14)
u i ∈ {0, 1}; i ∈ I (8.15)
vi ∈ [0, 1]; i ∈ I (8.16)
wsjt ∈ {0, 1}; j ∈ J, t ∈ T, s ∈ S (8.17)
Cs ≥ 0; s ∈ S. (8.18)

8.3.2 Equitable Maximization of Average and Worst-Case

Service Level

In the next model, VaRsl is the targeted customer service level such that for a given
confidence level α, for 100α% of the scenarios, the outcome is above VaRsl , while
CVaRsl is the expected service level in the worst 100(1 − α)% of the cases with
the service level below VaRsl . Let Ss be the tail service level for scenario s, where
tail service level is defined as the amount by which VaRsl exceeds service level in
scenario s. The aim of the following SMIP model is to maximize average customer
service level as well as to reduce the outcomes below the service level represented
by VaRsl , and by this to equitably optimize average and worst-case customer service
level. The solutions equitably maximize the expected and the expected worst-case
fraction of customer orders filled on or before their due dates.
218 8 A Robust Decision-Making Under Disruption Risks

Let E sl be the maximized expected service level and C V a R sl , the maximized

expected worst-case service level
  
E sl = Ps wsjt /J (8.19)
j∈J t∈T :t≤d j s∈S

C V a R sl = V a R sl − (1 − α)−1 Ps Ss . (8.20)
s∈S

sl sl
E −E sl
Denote by f 1sl = sl , the normalized expected service level (E sl , E are the
E −E sl
sl
C V a R −C V a R sl
minimum and the maximum values of E sl , respectively), and by f 2sl = sl ,
C V a R −C V a R sl
sl
the normalized expected worst-case service level (C V a R sl , C V a R are the mini-
mum and the maximum values of C V a R sl for a given confidence level α, respec-
tively).
The normalized service level objective functions f 1sl and f 2sl are defined below

sl   
E − j∈J t∈T :t≤d j s∈S Ps wsjt /J
f 1sl = sl
(8.21)
(E − E ) sl

sl 
C V a R − (V a R sl − (1 − α)−1 Ps Ss )
f 2sl = sl
s∈S
(8.22)
CV aR − CV aR sl

The SMIP model RSPS_ECV(sl) for selection of robust supply portfolio and
customer order scheduling to equitably maximize expected and expected worst-case
service level is formulated below.

RSPS_ECV(sl): Selection of robust supply portfolio and scheduling of

customer orders to equitably optimize expected service level and CVaR of
service level
Minimize (8.5)
subject to (8.6)–(8.12), (8.14)–(8.17), (8.21), (8.22) and
Risk constraints:
– the tail service level for scenario s is defined as the nonnegative amount
by which VaRsl exceeds service level in scenario s,
 
Ss ≥ V a R sl − wsjt /J ; s ∈ S (8.23)
j∈J t∈T :t≤d j

Ss ≥ 0. (8.24)

Note that the objective functions (8.19) and (8.20) (or normalized objective func-
tions, (8.21) and (8.22)) do not directly account for any cost components. As a result,
the optimal solution to RSPS_ECV(sl) that aims at equitable optimization of average
8.3 Problem Formulation 219

and worst-case service level depends mainly on the distribution of disruption prob-
abilities among the suppliers. However, the optimization of service level implicitly
reduces the penalty costs of delayed and unfulfilled customer orders, that represent
an important part of the total cost structure.

8.3.3 Minimum and Maximum Values of the Objective

Functions

In this subsection the minimum and maximum values for all objective functions are
calculated to determine the normalized values of all objective functions, f 1c (8.3), f 2c
(8.4), f 1sl (8.21) and f 2sl (8.22), that is, the values of the optimized objective func-
tions scaled into the interval [0,1]. Note that the cost and the service level objectives
are in conflict. Similarly, the expected and the expected worst-case values of a given
objective function are also in conflict.

Minimum and Maximum Values of Expected Values

c
The minimum and maximum values of expected cost E c , E , and expected ser-
sl
vice level, E sl , E , are obtained by solving the following stochastic mixed integer
programs:

SPS_E(c): Risk-neutral selection of supply portfolio and scheduling of

customer orders to minimize expected cost
Minimize E c , (8.1),
subject to (8.7)–(8.12), (8.15)–(8.17).

SPS_E(sl): Risk-neutral selection of supply portfolio and scheduling of

customer orders to maximize expected service level
Maximize E sl , (8.19)
subject to (8.7)–(8.12), (8.15)–(8.17).

In problem SPS_E(c), E c is the minimized objective function, while E sl is not

considered. In problem SPS_E(sl), E sl is the maximized objective function, while
E c is not considered. Thus, by solving problem SPS_E(c), the minimum value E c of
E c and the minimum value E sl of E sl are determined. Similarly, by solving problem
sl c
SPS_E(sl), the maximum value E of E sl and the maximum value E of E c are
determined.
220 8 A Robust Decision-Making Under Disruption Risks

Minimum and Maximum Values of CVaR

The minimum value of expected worst-case cost, C V a R c , and the maximum value
sl
of expected worst-case service level, C V a R , for a given confidence level α, are
obtained as the optimal solution of problem SPS_CV(c) and SPS_CV(sl), respec-
tively. The SMIP models SPS_CV(c) and SPS_CV(sl) are shown below.

SPS_CV(c): Risk-averse selection of supply portfolio and scheduling of

customer orders to minimize CVaR of cost
Minimize (8.2)
subject to (8.7)–(8.13), (8.15)–(8.18).

SPS_CV(sl): Risk-averse selection of supply portfolio and scheduling of

customer orders to maximize CVaR of service level
Maximize (8.20)
subject to (8.7)–(8.12), (8.15)–(8.17), (8.23), (8.24).

Since the expected and the expected worst-case values of a given objective function
c
are in conflict, the maximum value of expected worst-case cost C V a R , and the
sl
minimum value of expected worst-case service level, C V a R for a given confidence
level α, are associated with the optimal solutions of SMIP problem SPS_E(c) and
SPS_E(sl), respectively. In order to compute the associated values of CVaR, the
stochastic mixed integer programs are enhanced as shown below.

SPS_E(c,α)
Minimize (8.1),
subject to (8.2), (8.7)–(8.13), (8.15)–(8.18) and
 
zs ≥ ( ei u i /B + Aoi vi /B
i∈I i∈Is
   
+ g j b j (t − d j )wsjt /B + h j b j (1 − wsjt )/B
j∈J t∈T :t>d j j∈J t∈T

−V a R )/Cmax ; s ∈ S
c
(8.25)
 
zs ≤ 1 + ( ei u i /B + Aoi vi /B
i∈I i∈Is
8.3 Problem Formulation 221

   
+ g j b j (t − d j )wsjt /B + h j b j (1 − wsjt )/B
j∈J t∈T :t>d j j∈J t∈T

−V a R )/Cmax ; s ∈ S
c
(8.26)

Ps z s ≤ 1 − α (8.27)
s∈S
z s ∈ {0, 1}; s ∈ S, (8.28)

where (8.25) and (8.26) determine scenarios s with the cost per product not less
than VaRc , and (8.27) ensures that the total probability of all such scenarios is
not greater than 1 − α.
Cmax is an upper bound on cost per product, and the additional binary variable
is defined as follows: z s = 1, if for scenario s, cost per product,
     
ei u i /B + Aoi vi /B + g j b j (t − d j )wsjt /B + h j b j (1 − wsjt )/B
i∈I i∈Is j∈J t∈T :t>d j j∈J t∈T

is not less than VaRc ; otherwise z s = 0.

SPS_E(sl,α)
Maximize (8.19)
subject to (8.7)–(8.12), (8.15)–(8.17), (8.20), (8.23), (8.24) and
 
ζs ≥ wsjt /J − V a R sl ; s ∈ S (8.29)
j∈J t∈T :t≤d j
 
ζs ≤ 1 + wsjt /J − V a R sl ; s ∈ S (8.30)
j∈J t∈T :t≤d j

Ps ζs ≥ α (8.31)
s∈S
ζs ∈ {0, 1}; s ∈ S, (8.32)

where (8.29) and (8.30) determine scenarios s with the customer service level
not less than VaRsl , and (8.31) ensures that the total probability of all such
scenarios is not less than the confidence level α. The additional binary vari-
able
 is defined as follows: ζs = 1, if for scenario s, customer service level,
t∈T :t≤d j w jt /J , is not less than VaR ; otherwise ζs = 0.
s sl
j∈J
222 8 A Robust Decision-Making Under Disruption Risks

For a given confidence level α, the maximum value of expected worst-case cost,
c
C V a R , is obtained by solving problem SPS_CV(c), and the minimum value of
expected worst-case service level, C V a R sl by solving problem SPS_CV(sl).

8.4 Computational Examples

In this section some computational examples are presented to illustrate possible

applications of the proposed SMIP approach for the robust supplier selection, order
quantity allocation and customer orders scheduling in a supply chain under disruption
risks. The following parameters have been used for the example problems:
• I = 9, J = 25, K = 3, R = 3, T = 10, and S = 2 I = 512;
• K = {1, 2, 3}, and J1 = {1, . . . , 10}, J2 = {11, . . . , 20}, J3 = {21, . . . , 25};
• R = {1, 2, 3}, and I 1 = {1, 2, 3}, I 2 = {4, 5, 6}, I 3 = {7, 8, 9};
• a j ∈ {1, 2, 3}, ∀ j = 1, . . . , 25;
• b j ∈ {500, 1000, . . . , 5000} ∀ j = 1, . . . , 25;
• c j ∈ {1, 2, 3}, ∀ j = 1, . . . , 25;
• Ct = C = 38000, ∀t = 1, . . . , 10;
• e = (8, 7, 10, 13, 14, 11, 19, 17, 19) × 1000;
• g j = 1, h j = 26 ∀ j = 1, . . . , 25;
• o = (13, 12, 12, 8, 6, 6, 2, 5, 4);
• τi , was 2, 3 and 4 time periods, respectively for suppliers i ∈ I 1 , i ∈ I 2 and i ∈ I 3 ;
• d j ∈ {1 + mini∈I (τi ), . . . , T } ∀ j = 1, . . . , 25;
• pi , the local disruption probability was uniformly distributed over [0.005,0.01],
[0.01,0.05] and [0.05;0.10], respectively for suppliers i ∈ I 1 , i ∈ I 2 and i ∈ I 3 ;
• pr , the regional disruption probability was 0.001, 0.005 and 0.01, respectively for
region r = 1, r = 2 and r = 3;
• α, the confidence level, was equal to 0.50, 0.75, 0.90, 0.95 or 0.99.
The computational experiments were performed for the same replication of the
above input data set. All potential disruption scenarios were considered and to cal-
culate the corresponding disruption probabilities, formulae (7.10) and (7.11) were
applied. For all test examples, the resulting total demand for parts and products is
A = 132500 and B = 66000, respectively. The unit price per part oi and the disrup-
tion probability πi , (7.9), of each supplier i ∈ I are shown in Fig. 6.1.
The solution results for the risk-neutral, risk-averse and robust decision-making
are presented in Tables 8.3, 8.4 and 8.5, respectively. The minimum and maximum
values of all considered objective functions are presented in Table 8.6. In addition to
the optimal absolute and normalized solution values for the primary objective func-
tions and the allocation of demand among the selected suppliers, Table 8.5 presents
also the expected values of the associated objective function, i.e., the expected ser-
vice level for model RSPS_ECV(c) and the expected cost per product for model
RSPS_ECV(sl).
8.4 Computational Examples 223

Table 8.3 Risk-neutral solutions

Model SPS_E(c): Var.=100468, Bin.=100459, Cons.=21341, Nonz.=765902 (a)
Expected Cost 7.66
Suppliers Selected(% of total demand) 7(100)
Model SPS_E(sl): Var.=100468, Bin.=100459, Cons.=21341, Nonz.=765902 (a)
Expected Service Level(%) 99.62
Suppliers Selected(% of total demand) 1(48), 2(31), 3(21)
(a)
Var.=number of variables, Bin.=number of binary variables,
Cons.=number of constraints, Nonz.= number of nonzero coefficients

For the risk-neutral minimization of cost, Table 8.3 indicates that the cheapest
supplier i = 7 is selected only, while for the risk-neutral maximization of service
level, the total demand is allocated among the three most reliable and most expensive
suppliers i = 1, 2, 3.
For the risk-averse minimization of cost, Table 8.4 indicates that for the lowest
confidence levels α = 0.5, 0.75, the cheapest supplier i = 7 is selected only, while
for a higher α, more suppliers are selected, including the reliable but expensive
suppliers: i = 2 for α = 0.95 and i = 2, 3 for α = 0.99.

Table 8.4 Risk-averse solutions

Confidence level α 0.50 0.75 0.90 0.95 0.99
Model SPS_CV(c): Var.=100981, Bin.=100459, Cons.=21853, Nonz.=874288 (a)

CVaRc 10.60 16.47 23.75 26.66 30.90

Suppliers Selected(% of total demand) 7(100) 7(100) 4(23) 2(30) 2(22)
5(21) 4(16) 3(22)
6(21) 5(19) 5(15)
7(17) 6(19) 6(16)
9(18) 7(16) 7(12)
9(13)
Model SPS_CV(sl): Var.=100981, Bin.=100459, Cons.=21853, Nonz.=822422 (a)
CVaRsl (%) 99.25 98.49 96.23 92.46 85.73
Suppliers Selected(% of total demand) 1(48) 1(48) 1(48) 1(48) 1(21)
2(31) 2(31) 2(31) 2(31) 2(21)
3(21) 3(21) 3(21) 3(21) 3(14)
4(7)
5(7)
6(8)
7(8)
8(7)
9(7)
(a)
Var.=number of variables, Bin.=number of binary variables,
Cons.=number of constraints, Nonz.= number of nonzero coefficients
224 8 A Robust Decision-Making Under Disruption Risks

For the risk-averse maximization of service level, where the supplier selection is
independent of any cost parameters and the solution mainly depends on the distrib-
ution of disruption probabilities, Table 8.4 demonstrates that the most reliable (and
most expensive) suppliers i = 1, 2, 3 are selected for all confidence levels, except
for the highest confidence level, α = 0.99, for which all nine suppliers are selected.
Table 8.5 indicates that for minimization of cost and the low confidence lev-
els, α = 0.5, 0.75, the robust supply portfolio is identical with the risk-neutral
portfolio. Similarly, for maximization of service level and the confidence levels,
α = 0.5, 0.75, 0.9, 0.95, the robust supply portfolio is identical with the risk-neutral
portfolio, while for α = 0.99, the robust supply portfolio is similar to the risk-averse
portfolio. For the service level objective function, the robust solutions with a perfect
equity were found for all confidence levels, except for the highest α = 0.99, which
indicates that the obtained robust solutions are also the lexicographic minimax opti-

Table 8.5 Robust solutions

Confidence level α 0.50 0.75 0.90 0.95 0.99
Model RSPS_ECV(c): Var.=100989, Bin.=100459, Cons.=21859, Nonz.=975283 (a)
Ec 7.66 7.66 10.67 11.75 11.04
CVaR c 10.60 16.47 28.55 29.05 35.44
VaRc 4.73 4.73 18.51 23.70 30.83
Normalized E c 0 0 0.167 0.227 0.188
Normalized CVaRc 0 0 0.167 0.093 0.210
Suppliers Selected(% of total demand) 7(100) 7(100) 6(26) 6(37) 6(54)
7(50) 7(30) 7(46)
9(24) 9(33)
E sl (b) 67.46 67.56 85.37 84.79 87.18
Model RSPS_ECV(sl): Var.=100989, Bin.=100459, Cons.=21859, Nonz.=878445 (a)
E sl (b) 99.62 99.62 99.62 99.62 98.91
CVaRsl (%) 99.25 98.49 96.23 92.46 85.61
VaRsl (%) 100 100 100 100 92.00
Normalized E sl 0 0 0 0 0.022
Normalized CVaRsl 0 0 0 0 0.002
Suppliers Selected(% of total demand) 1(48) 1(48) 1(48) 1(48) 1(22)
2(31) 2(31) 2(31) 2(31) 2(22)
3(21) 3(21) 3(21) 3(21) 3(19)
4(11)
6(10)
7(9)
9(7)
Ec 25.64 25.64 25.64 25.64 21.84
(a) Var.=number of variables, Bin.=number of binary variables,
Cons.=number
 of constraints,
 Nonz.= number of nonzero coefficients.
(b) (
t∈T :t≤d j w jt /J )100%
s
s∈S Ps j∈J
8.4 Computational Examples 225

Table 8.6 Minimum and maximum values of objective functions

Expected values
Ec 7.66
c
E 25.64
(associated with Maximum Expected Service Level)
sl
E (%) 99.62
E sl (%) 67.60
(associated with Minimum Expected Cost)
Expected worst-case values, CVaR
Confidence level α 0.50 0.75 0.90 0.95 0.99
C V a Rc 10.60 16.47 23.75 26.66 30.91
c
CV aR 52.48 for all confidence levels
(associated with Minimum Expected Cost)
sl
C V a R (%) 99.25 98.49 96.23 92.46 85.73
C V a R sl (%) 67.86 67.72 71.26 74.43 39.58
(associated with Maximum Expected Service Level)

mal solutions (see, Theorem 2 in Sect. 7.2). As a result, the obtained expected service
level and the expected worst-case service level are equally very close to their best
available values. However, the associated expected cost is much higher than its best
value (recall that the service level-based solution is independent of any cost parame-
ters). For the cost-based objective function, the solutions with perfect equity were
found only for α = 0.5, 0.75, 0.9.
For the cost-based objective function, Fig. 8.1 compares the distribution of cost
per product for the risk-neutral, risk-averse and robust decision-making for the two
confidence levels α = 0.75 and α = 0.99. The probability mass functions are con-
centrated in a few points, which is typical for the scenario-based optimization under
uncertainty, where the probability measure is concentrated in finitely many points.
For α = 0.75, the risk-averse and robust solution are identical with the risk-neutral
solution, and so are the corresponding probability mass functions. A large probability
atom 0.06148 is concentrated at the highest cost of 52.49. For α = 0.99, different
solutions are obtained for different types of the decision-making. For the risk-averse
solution, the total probability measure of costs between 24 and 29 is 0.9977, while
the probability of costs greater than 40 is 0.0023. For the robust solution, the largest
probability measure of 0.906 is concentrated at the lowest cost of 8.95 and the prob-
ability of the highest cost, 52.65, is only 0.002.
For the service level-based objective function, Fig. 8.2 compares the distribution
of customer service level for the risk-neutral, risk-averse and robust decision-making,
for the two confidence levels α = 0.75 and α = 0.99. The solution results for the three
types of decision making are identical for α = 0.75 and very similar for α = 0.99,
with the most reliable and most expensive suppliers i = 1, 2, 3 predominating the
226 8 A Robust Decision-Making Under Disruption Risks

Cost for Minimum Cost Objective Functions

1

0.8
Probability

0.6

0.4

0.2

0
10 20 30 40 50 60 70 80 90 100
Cost per Product

Risk-neutralsolution Risk-aversesolution(alpha=0.75) Robustsolution(alpha=0.75)

Cost for Minimum Cost Objective Functions

1

0.8
Probability

0.6

0.4

0.2

0
10 20 30 40 50 60 70 80 90 100
Cost per Product

Risk-neutralsolution Risk-aversesolution(alpha=0.99) Robustsolution(alpha=0.99)

Fig. 8.1 Distribution of cost per product for α = 0.75 and α = 0.99

supply portfolio. As a result, the probability mass functions presented in Fig. 8.2 are
identical or very similar, respectively.
 As an illustrative example, Fig. 8.3 presents the demand  for
 products,
j∈J :d j =t b j , t ∈ T , and the expected production schedules, j∈J b j w jt ,
s
s∈S Ps
t ∈ T for the optimal cost- and service level-based, risk-neutral, risk-averse and
robust solutions with the confidence level α = 0.99. The total customer demand is
met with a small fraction of the expected rejected demand (from 0.0038 for the
risk-neutral, service level-based solution up to 0.2407 for the risk-averse, cost-based
8.4 Computational Examples 227

Service Level for Maximum Service Level Objective Functions

1

0.8
Probability

0.6

0.4

0.2

0
0 10 20 30 40 50 60 70 80 90 100
Service Level [%]

Risk-neutral solution -Risk-averse solution (alpha=0.75) -Robust solution (alpha=0.75)

Service Level for Maximum Service Level Objective Functions

1

0.8
Probability

0.6

0.4

0.2

0
0 10 20 30 40 50 60 70 80 90 100
Service Level [%]

Risk-neutral solution Risk-averse solution (alpha=0.99) Robust solution (alpha=0.99)

Fig. 8.2 Distribution of service level for α = 0.75 and α = 0.99

solution). In general, the service level-based solution, when no cost components are
included in the objective function, better meets the customer demand, with a smaller
fraction of unfulfilled demand. For the service level objective, the expected produc-
tion approximately follows the demand pattern. In addition, the expected production
schedules for the risk-neutral and the robust decision-making are very close to each
other, which indicates that in the average-case, the robust solution is nearly as good
as the risk-neutral solution.
Finally, Figs. 8.4 and 8.5 present the expected worst-case production schedules for
the optimal risk-averse and the robust solutions with the confidence level α = 0.99,
228 8 A Robust Decision-Making Under Disruption Risks

Expected Production for Minimum Cost

20000

16000
Production

12000

8000

4000

0
1 2 3 4 5 6 7 8 9 10
Period
Risk-neutral solution -Risk-averse solution (alpha=0.99)
-Robust solution (alpha=0.99) Demand

Expected Production for Maximum Service Level

14000

12000

10000
Production

8000

6000

4000

2000

0
1 2 3 4 5 6 7 8 9 10
Period
Risk-neutral solution Risk-averse solution (alpha=0.99)
Robust solution (alpha=0.99) Demand

Fig. 8.3 Expected production schedules for α = 0.99

for maximization of service level (Fig. 8.4),

 
s∈S:Ss >0 (Ps j∈J b j wsjt )
 , t ∈ T,
s∈S:Ss >0 Ps

and for minimization of cost (Fig. 8.5),

 
s∈S:Cs >0 (Ps j∈J b j wsjt )
 , t ∈ T,
s∈S:Cs >0 Ps
8.4 Computational Examples 229

Expected Worst-Case Production Schedule

for Maximum Service Level
14000

12000
Production/Demand

10000

8000

6000

4000

2000

0
1 2 3 4 5 6 7 8 9 10
Period

-Risk-averse solution(alpha=0.99) -Robust solution (alpha=0.99) Demand

Fig. 8.4 Expected worst-case production schedules for maximum service level, α = 0.99

respectively.
The risk-averse and the robust expected worst-case schedules for maximization of
service level are very close to each other (cf. similar supply portfolios in Tables 8.4
and 8.5), which indicates that in the worst-case the robust solution is nearly as good
as the risk-averse solution. The expected worst-case fractions of fulfilled customer
demand for products are similar and close to 75%.
For the minimization of cost, however, the expected worst-case schedules for the
risk-averse and the robust solutions are different (cf. different supply portfolios in
Tables 8.4 and 8.5). For the risk-averse solution, above 55% of the customer demand
is fulfilled in the expected worst-case. However, the expected worst-case production
for the robust solution is virtually negligible (see, Fig. 8.5(a)). The corresponding
expected worst-case fraction of fulfilled customer demand for products is nearly
equal to zero. Such a result may be due to the too low unit penalty cost,
(h j = 2a j maxi∈I (oi ), j ∈ J ), for unfulfilled customer demand, with respect to
the purchasing cost of required parts, i.e., the unit penalty cost is approximately twice
as large as the maximum unit price of required parts. The resulting robust supply
portfolio is based on the low cost, unreliable suppliers i = 6, 7, similarly to the
risk-neutral portfolio. In contrast, the pure risk-averse portfolio contains the most
reliable suppliers i = 2, 3.
In order to emphasize the impact of higher penalty costs, the computational experi-
ments were repeated with a double unit penalty cost, h j = 4a j maxi∈I (oi ), j ∈ J ,
i.e., with the unit penalty cost for unfulfilled customer demand, approximately four
times as large as the maximum unit price of required parts. The solution results
for the minimization of cost with a double penalty are presented in Table 8.7. The
table shows that for the low confidence levels, α = 0.5, 0.75, the risk-averse and
230 8 A Robust Decision-Making Under Disruption Risks

Expected Worst-Case Production Schedule

for Minimum Cost (a)
14000

12000
Production/Demand

10000

8000

6000

4000

2000

0
1 2 3 4 5 6 7 8 9 10
Period

-Risk-averse solution (alpha=0.99) Demand

Expected Worst-Case Production Schedule

for Minimum Cost (b)
14000

12000
Production/Demand

10000

8000

6000

4000

2000

0
1 2 3 4 5 6 7 8 9 10
Period
Risk-averse solution (alpha=0.99) Robust solution (alpha=0.99) Demand

Fig. 8.5 Expected worst-case production schedules for minimum cost, α = 0.99: a h j =
2a j maxi∈I (oi ), b h j = 4a j maxi∈I (oi )

the robust supply portfolios are identical with the risk-neutral portfolio that assigns
total demand for parts to a single, cheapest supplier i = 7. For a higher confidence
level, the risk-averse portfolio contains a single, reliable supplier (i = 2, for α = 0.9,
and i = 1, for α = 0.95), whereas the robust portfolio contains the reliable supplier,
i = 2, only for the highest confidence level, α = 0.99. This simultaneously leads to
8.4 Computational Examples 231

the highest expected service level, 94.03%. Table 8.7 shows that for most confidence
levels, the robust solution with a perfect equity was obtained, which indicates that
the obtained robust solutions are also the lexicographic minimax optimal solutions
(see, Theorem 2 in Sect. 7.2).
The expected worst-case schedules for a double unit penalty cost,
h j = 4a j maxi∈I (oi ), j ∈ J , is shown in Fig. 8.5(b). The obtained expected worst-
case production for the robust solution is no longer negligible, however the production
is still smaller than that for a pure risk-averse solution. Now, the robust supply
portfolio has been enforced with a reliable supplier i = 2, while the pure risk-averse
portfolio contains all nine suppliers (see, Table 8.7). For the robust solution only 18%
of the total customer demand is fulfilled in the expected worst-case, while for the risk-
averse solution, over 60% of the customer demand is fulfilled. On the other hand, the
expected percentage of fulfilled customer demand is over 95% for the robust solution
and only 75% for the risk-averse solution. Thus, the risk-averse solution outperforms
the robust solution in the worst-case, and vice-versa the robust solution outperforms
the risk-averse solution in the average-case. The above results demonstrate that the
robust solution for the cost-based objective functions is strongly dependent on the
cost parameters.
The equitably efficient solutions obtained using models RSPS_ECV(c) and
RSPS_ECV(sl) have been compared with the nondominated solutions minimizing
the weighted-sum of the normalized objective functions, respectively f 1c , f 2c and
f 1sl , f 2sl , with equal weights of each objective. In the literature this type of trade-off
model (with a varying trade-off parameter) is known as the mean-risk model (e.g.,
Ogryczak and Ruszczynski 2002) with weighted-sum objective consisting of the
expected value and the CVaR as a risk measure. The weighted-sum programs (with
equal weights of each objective function), SPS_ECV(c), (8.33), and SPS_ECV(sl),
(8.34), are presented below and the solution results are shown in Table 8.8.

SPS_ECV(c) (8.33)

Minimize f 1c + f 2c
subject to (8.3), (8.4), (8.7)–(8.13), (8.15)–(8.18).

SPS_ECV(sl) (8.34)

Maximize f 1sl + f 2sl

subject to (8.7)–(8.12), (8.15)–(8.17), (8.21)–(8.24).

Table 8.8 indicates that the mean-risk solutions are close to the robust solutions
presented in Tables 8.5 and 8.7, while the perfect equity is very rarely achieved. In
particular, for the service level objective function the robust and the mean-risk solu-
232 8 A Robust Decision-Making Under Disruption Risks

Table 8.7 Solution results for a double unit penalty cost, h j = 4a j maxi∈I (oi )
Risk-neutral solution: model SPS_E(c)
Ec 10.87
Suppliers Selected(% of total demand) 7(100)
Confidence level α 0.50 0.75 0.90 0.95 0.99
Risk-averse solutions: model SPS_CV(c)
CVaRc 17.02 24.99 30.35 35.82 46.97
Suppliers Selected(% of total demand) 7(100) 7(100) 2(100) 1(100) 1(19)
2(19)
3(20)
4(6)
5(7)
6(9)
7(9)
8(5)
9(6)
Minimum and maximum values of objective functions.
Ec 10.87
c
E 26.03
C V a Rc 17.02 24.99 30.35 35.82 46.97
c
CV aR 104.68 for all confidence levels
Robust solutions: model RSPS_ECV(c)
Ec 10.87 10.87 14.35 14.41 14.92
CVaRc 17.02 29.31 47.38 51.56 62.38
VaRc 4.73 4.73 37.85 46.17 52.88
Normalized E c 0 0 0.23 0.23 0.27
Normalized CVaRc 0 0.05 0.23 0.23 0.27
Suppliers Selected(% of total demand) 7(100) 7(100) 5(30) 5(25) 2(13)
6(31) 6(39) 6(45)
7(39) 7(36) 7(42)
E sl (b) 67.38 67.46 86.55 87.12 94.03
(b) (
   s /J )100%
P
s∈S s j∈J w
t∈T :t≤d j jt

tions are nearly identical for most confidence levels. However, for the cost objective
function, the robust solutions based on the ordered weighted averaging aggregation
approach, in most cases are different from the nondominated solutions obtained using
the mean-risk approach.
The robust solution that equitably optimizes both the average and the worst-case
performance of a supply chain has been compared with the risk-neutral solution
that focuses on the average performance only, and with the risk-averse solution
that focuses on the worst-case performance only. The computational experiments
demonstrate that the robust solution may be outperformed by an optimal risk-neutral
8.4 Computational Examples 233

Table 8.8 Mean-risk solutions

Confidence level α 0.50 0.75 0.90 0.95 0.99
Model SPS_ECV(c) (8.33)
Ec 7.67 7.67 12.74 12.74 11.04
CVaRc 10.61 16.49 24.17 27.36 35.44
VaRc 4.73 4.73 20.97 21.00 30.83
Normalized E c 0.0005 0.0004 0.2827 0.2828 0.1880
Normalized CVaRc 0.0002 0.0005 0.0146 0.0269 0.2100
Suppliers Selected(% of total demand) 7(100) 7(100) 5(27) 5(27) 6(54)
6(27) 6(27) 7(46)
7(22) 7(22)
9(24) 9(24)
E sl (%) 67.01 67.80 84.32 85.15 87.18
Model SPS_ECV(c) (8.33) for a double unit penalty cost, h j = 4a j maxi∈I (oi ).
Ec 10.87 10.87 15.53 14.73 13.37
CVaRc 17.02 29.31 41.13 48.70 66.94
VaRc 4.73 4.73 33.74 41.98 56.76
Normalized E c 0 0 0.307 0.255 0.165
Normalized CVaRc 0 0.05 0.145 0.187 0.346
Suppliers Selected(% of total demand) 7(100) 7(100) 5(25) 5(34)
6(26) 6(34) 6(52)
7(24) 7(32) 7(48)
9(25)
E sl (%) 67.05 67.50 85.06 86.59 87.13
Model SPS_ECV(sl) (8.34)
E sl (%) 99.62 99.62 99.62 99.62 99.07
CVaRsl (%) 99.25 98.49 96.20 92.40 84.92
VaRsl (%) 100 100 100 100 92.00
Normalized E sl 0 0 0 0 0.0172
Normalized CVaRsl 0 0 0.001 0.003 0.0176
Suppliers Selected(% of total demand) 1(48) 1(48) 1(48) 1(48) 1(21)
2(31) 2(31) 2(31) 2(31) 2(21)
3(21) 3(21) 3(21) 3(21) 3(18)
4(10)
5(10)
6(10)
7(10)
Ec 25.64 25.64 26.03 25.64 21.44
234 8 A Robust Decision-Making Under Disruption Risks

solution in the average-case and by an optimal risk-averse solution in the worst-case.

However, the robust solution outperforms the risk-averse solution in the average-case
and the risk-neutral solution in the worst-case.
The following basic insights can be derived from the computational experiments.
• The robust solutions are frequently also the lexicographic minimax optimal solu-
tions as well as the Pareto-optimal solutions.
For both cost-based and service-based objective functions, the solutions with a
perfect equity are frequently found (see, Theorems 2 and 3 in Sect. 7.2).
• In average-case the robust solution may perform as well as the risk-neutral solu-
tion.
For both cost-based and service-based objective functions, the robust solution for
a low confidence level is identical with the risk-neutral solution, i.e., the supply
portfolio is based on low-cost, unreliable suppliers and on high-cost, reliable sup-
pliers, respectively. Moreover, the expected production schedules for the robust
and for the risk-neutral solutions are very close to each other.
• In worst-case the robust solution for the maximum service level objective is nearly
as good as the risk-averse solution.
For the maximum service level objective and a high confidence level, the robust
solution is very close to the risk-averse solution. Moreover, the expected worst-
case schedules for the robust and for the risk-averse solutions are very close to
each other.
• For the minimum cost objective and a high confidence level, the robust solution is
strongly dependent on the cost parameters.
For example, if unit penalty cost for unfulfilled demand for products is comparable
with purchasing cost of required parts, the robust solution for a high confidence
level may more resemble the risk-neutral solution, while for a high penalty cost
the robust solution may be closer to the risk-averse solution.
• The mean-risk solutions are rarely the lexicographic minimax optimal solutions.
The mean-risk solutions that minimize the sum of the normalized expected value
and the normalized CVaR are close to the robust solutions, while the perfect equity
is very rarely achieved. For the service level objective function the robust and the
mean-risk solutions are nearly identical for most confidence levels. However, for
the cost objective function, the robust solution and the mean-risk solutions are
different in most cases.
• The service level and the cost objective functions are in conflict.
The best service level-based solutions may perform poorly with respect to the
cost-based metrics, and vice versa.
The computational experiments were performed using the AMPL programming
language and the CPLEX 12.5 solver on a MacBookPro laptop with Intel Core
i7 processor running at 2.8 GHz and with 16 GB RAM. The solver was capable
of finding proven optimal solutions for all examples with CPU time ranging from
several seconds for the risk-neutral solutions to several hours for the robust or mean-
risk solutions and cost-based objectives.
8.5 Notes 235

8.5 Notes

In the literature on supply chain risk management, robustness implies that the supply
chain is strong enough to be unaffected by disturbances, while retaining its original
structure, e.g., Klibi et al. (2010), Gabrel et al. (2014). In contrast to resiliency (see,
Chap. 4) that implies that the supply chain needs to adapt (reconfigure) its structure
to survive and grow in the face of change and uncertainty (e.g., Fiksel 2006; Klibi
and Martel 2012).
The major contribution of this chapter is that it proposes a simple approach for
the robust decision-making associated with supplies of parts and deliveries of fin-
ished products in a customer-driven supply chain under disruption risks. Here, the
robustness is defined as the mean-risk fairness and refers to an equitably efficient per-
formance of a supply chain in the average-case as well as in the worst-case. The fair
mean-risk decision-making aims at equalizing the distance to optimality both under
business-as-usual and under worst-case conditions, which reflects the decision mak-
ers common requirement to maintain an equally good performance of a supply chain
under varying operating conditions. The robust decision-making equitably focuses
on the two objective functions: the expected value and the expected worst-case value
(i.e., Conditional Value-at-Risk) of the selected criterion, cost or service level.
The material presented in this chapter is based on results presented in Sawik
(2014c, 2016c), where SMIP models were developed. The models were the enhance-
ments of formulations proposed in Sawik (2013c, 2014a, 2014b, see also Chap. 5)
for a single-objective decision-making. In contrast, the stochastic optimization prob-
lems considered in this chapter are formulated as bi-objective SMIP models with
the two conflicting objective functions: expected cost and expected worst-case cost
(Conditional Cost-at-Risk) or expected service level and expected worst-case ser-
vice level (Conditional Service-at-Risk). In order to obtain an equitably efficient
solution to the combinatorial stochastic optimization problem, the ordered weighted
averaging aggregation (Yager 1988) of the expected and the expected worst-case
value of the selected objective function has been applied. The equitable optimization
of the supply chain network under disruption risks and the associated coordinated
scheduling of the disrupted material flows are rarely considered. However, another
type of a trade-off model is well-known in the literature: the mean-risk model (e.g.,
Ogryczak and Ruszczynski 2002). The mean-risk model is formulated as the opti-
mization of weighted-sum objective consisting of the expected value and the CVaR
as a risk measure. The mean-risk approach aims at balancing the expected value
with the risk tolerance, however, the mean-risk model is not designed to achieve an
equitably efficient solution.
236 8 A Robust Decision-Making Under Disruption Risks

Problems

8.1 Modify the SMIP models presented in this chapter for multiple part types with
subsets of part types required for each customer order and subsets of suppliers avail-
able for each part type.

8.2 How would you determine an upper bound, Cmax , on the cost per product for
model SPS_E(c,α)?

8.3 Mixed mean-risk decision-making

(a) Modify models SPS_ECV(sl) and SPS_ECV(c) to optimize expected cost and
CVaR of service level and optimize expected service level and CVaR of cost, respec-
tively.
(b) How should the values of the optimized objective functions be scaled into the
interval [0,1] to avoid dimensional inconsistency among the two objectives and how
should the trade-off parameter be selected?
(c) How would you interpret the mixed mean-risk solutions?

8.4 Explain why the robust supply portfolios for the cost-based objective is strongly
dependent on the cost parameters and which parameters are the most influential?

8.5 Explain why the expected worst-case production for the robust solution in
Fig. 8.5(a) is nearly negligible?
 Part IV
Selection of Primary and Recovery
Portfolios and Scheduling
Chapter 9
Selection of Primary and Recovery Supply
Portfolios and Scheduling

9.1 Introduction

In this chapter the portfolio approach presented in previous chapters for the selection
of primary suppliers and order quantity allocation to mitigate the impact of disruption
risks is enhanced also for the recovery process, i.e., for the selection of both primary
and recovery suppliers and order quantity allocation to mitigate the impact of dis-
ruption risks and optimize the recovery process. Unlike most of reported research
on the supply chain risk management which focuses on the risk mitigation decisions
taken prior to a disruption, this chapter combines decisions made before, during and
after the disruption. The two decision-making approaches will be considered: inte-
grated approach with the perfect information about the future disruption scenarios,
and hierarchical approach with no such information available ahead of time. In the
integrated approach, which accounts for all potential disruption scenarios, the pri-
mary supply portfolio that will hedge against all scenarios is determined along with
the recovery supply portfolio and production schedule of finished products for each
scenario, to minimize expected cost or CVaR of cost and maximize expected service
level or CVaR of service level over all scenarios. In the hierarchical approach first
the primary supply portfolio is selected, and then, when a primary supplier is hit
by a disruption, the recovery supply portfolio is selected to optimize the process of
recovery from the disruption.
The following time-indexed SMIP and MIP models are presented in this chapter:

Support_E for risk-neutral selection of primary and recovery supply port-

folios and production scheduling;
SupportMP_E model Support_E for multiple part types and product types;
PSupport for selection of primary supply portfolio and production schedul-
ing under deterministic conditions;

© Springer International Publishing AG 2018 239

T. Sawik, Supply Chain Disruption Management Using Stochastic Mixed
Integer Programming, International Series in Operations Research
& Management Science 256, DOI 10.1007/978-3-319-58823-0_9
240 9 Selection of Primary and Recovery Supply Portfolios and Scheduling

RSupport(s) for selection of recovery supply portfolio and production

scheduling, for predetermined primary supply portfolio and the realized
disruption scenario;
RSupport_E model Support_E for predetermined primary supply portfo-
lio;
Support_CV(c) for risk-averse selection of primary and recovery supply
portfolios and production scheduling to minimize CVaR of cost;
Support_CV(sl) for risk-averse selection of primary and recovery supply
portfolios and production scheduling to maximize CVaR of service level.

Numerical examples and computational results are reported in Sects. 9.5.1 and
9.5.2, respectively for risk-neutral and risk-averse decision-making.
In the next chapter, the portfolio approach will be further enhanced to simultane-
ously select supply and demand portfolios, when a disruption impacts both a primary
supplier of parts and the buyer’s firm primary assembly plant. Then, in addition to
determining the primary and recovery supply portfolios, the firm may also choose to
move production to alternate (recovery) plants along with transshipment of parts from
the impacted primary assembly plant to the recovery plants. The resulting allocation
of unfulfilled demand for products among recovery assembly plants will determine
a recovery, demand or capacity portfolio.

9.2 Problem Description

Consider a supply chain in which a single producer of one product type, assembles
products to meet customer demand, using a critical part type that can be manufactured
and provided by several suppliers.
Let I = {1, . . . , I} be the set of I suppliers, T = {1, . . . , T } the set of T planning
periods, and let dt be the demand for products in period t ∈ T (for notation, see
Table 9.1).
The orders for parts are assumed to be placed at the beginning of the planning
horizon, and under normal conditions the parts ordered from supplier i are delivered
in period τi , where τi represents total of manufacturing lead time and transportation
time. If production does not meet the demand, the producer is charged with penalty
cost for unfulfilled demand for products.
The suppliers of parts are located in R geographic regions, subject to potential
regional disasters that may result in complete shutdown of all suppliers in the same
region simultaneously. In addition to correlated regional disruptions, each supplier
i ∈ I is subject to random independent local disruptions of different levels, l ∈ Li =
{0, . . . , Li }, where the disruption level refers to the fraction of an order that can be
delivered (supplier fulfillment rate). Level l = 0 represents complete shutdown of a
supplier, i.e., no order delivery, while level l = Li represents normal conditions with
no disruption, i.e., full order delivery. The fraction of an order that can be delivered by
supplier i under disruption level l is described by the associated supplier fulfillment
rate, γil
9.2 Problem Description 241

Table 9.1 Notation: selection of primary and recovery supply portfolios and scheduling
Indices
i = supplier, i ∈ I
l = disruption level, l ∈ Li , i ∈ I
r = region, r ∈ R
s = disruption scenario, s ∈ S
t = planning period, t ∈ T
Input Parameters
c = per period capacity of producer
C = total available capacity of producer
dt = demand for products in period t
D = total demand for parts/products
ei = fixed ordering cost of creating contracts and maintaining relationships with
supplier i
g = per unit penalty cost of unfulfilled demand for products
oi = per unit price of parts purchased from supplier i
pil = probability of disruption level l for supplier i
pr = regional disruption probability for region r
ts = start time period of disruptive event s
γil = fraction of an order delivered by supplier i under disruption level l (supplier
fulfillment rate)
τi = delivery lead time from supplier i
ρis = firm’s portion of supplier i cost-to-recover from disruption under scenario s
θis = time-to-recover of supplier i from disruption under scenario s
CTR(i, l) = firm’s portion of cost-to-recover for supplier i hit by disruption at level l
TTR(i, l) = time-to-recover for supplier i hit by disruption at level l

⎧
⎨= 0 if l = 0
γil ∈ (0, 1) if l = 1, . . . , Li − 1 (9.1)
⎩
=1 if l = Li .

Denote by S = {1, . . . , S} the index set of all disruption scenarios, and by Ps the
probability of disruption scenario s ∈ S. Each scenario s ∈ S can be represented by
an integer-valued vector λs = {λ1s , . . . , λIs }, where λis ∈ Li is the disruption level
of an order delivery from supplier i ∈ I under scenario s ∈ S. When all potential
disruption scenarios are considered, then S = i∈I (Li + 1).
For each scenario s ∈ S, the supplies from every supplier can be disrupted either
by a local or a regional disaster event. Denote by Is ⊂ I the subset of non-shutdown
suppliers, who can deliver parts under scenario s. The probability Ps for disruption
scenario s ∈ S with the subset Is of non-shutdown suppliers is (cf. Sect. 1.3)

Ps = Psr .
r∈R
242 9 Selection of Primary and Recovery Supply Portfolios and Scheduling

Psr is the probability of realizing of disruption scenario s for suppliers in I r (cf.

Sect. 1.4)
  
(1 − pr ) i∈I r
(pi,λis ) if I r  Is = ∅
Psr =
pr + (1 − pr ) i∈I r pi0 if I r Is = ∅,

where pi,λis is the probability of occurrence the disruption at level l = λis of an order
delivery from supplier i under scenario s.
A disruptive event under scenario s ∈ S is assumed to occur in period ts . When
supplier i is hit by disruption at level l, its recovery process to normal conditions
starts in period ts + 1 and takes TTR(i, l) time periods (Time-To-Recover), so that the
supplier i recovers to its full pre-disruption capacity in period t = ts + TTR(i, l) + 1.
The cost of recovery process can be shared between the supplier of parts and the firm
(producer of products). Let CTR(i, l) be the firm’s portion of Cost-To-Recover. Note
that CTR(i, L) = TTR(i, L) = 0, i.e., no recovery is required for a non-disrupted
supplier.
For each supplier i, denote by θis and ρis , respectively time-to-recover and firm’s
portion of cost-to recover from disruption under scenario s

θis = TTR(i, l); i ∈ I, s ∈ S : l = λis (9.2)

ρis = CTR(i, l); i ∈ I, s ∈ S : l = λis . (9.3)

The following assumptions are made to formulate the problem.

• Each supplier has sufficient capacity to meet total demand for parts.
• A single recovery mode is considered for each supplier and each disruption level.
• Disruption start times are constant parameters.
• Time-to recover and the associated cost-to-recover are constant parameters that
represent recovery of a disrupted supplier to its full capacity.
• The buyer firm may participate in supplier’s cost-to-recover.
• Disruption to primary supplier under scenario s ∈ S occurs in period ts and recovery
process starts in period ts + 1, so that the disrupted supplier i ∈ I returns to its full
capacity in period t = ts + θis + 1.
• Multiple disruptions, one after the other in a series, during the recovery process
are not considered.
• A recovery supplier can be a disrupted primary supplier after its recovery to full
capacity or a new supplier.
• Transition time required for switching to recovery supplier different from the
primary suppliers is negligible.
• Each product requires one unit of a critical part type.
• A penalty cost is charged for the demand for products unfulfilled by the end of the
planning horizon.
Many of the above assumptions can be easily relaxed. For example, the assumption
that each supplier has sufficient capacity to meet total demand for parts, which allows
for a single sourcing, often met in practice. Some of them, however, would require
9.2 Problem Description 243

development of more advanced models to be relaxed, e.g., random disruption start

times, durations and suppliers cost-to-recover.
Notice that no recovery supply portfolio needs to be selected if a disruptive event
occurs after the latest delivery lead time from primary suppliers, i.e., when ts >
maxi∈I (τi ).
In the sequel two decision-making approaches will be considered: an integrated
approach with the perfect information about the future disruption scenarios, and
a hierarchical approach with no such information available ahead of time. In the
integrated approach, which accounts for all potential disruption scenarios, decisions
are made prior to a disruption. The primary supply portfolio that will hedge against
all scenarios is determined along with the recovery supply portfolio and production
schedule for each scenario. The problem objective is to minimize expected cost
or maximize expected service level over all scenarios, for risk-neutral models, and
minimize CVaR of cost or maximize CVaR of service level for risk-averse models.
In the hierarchical approach, first prior to a disruption, the primary supply portfolio
is determined to minimize cost or maximize service level, and then, when a primary
supplier is hit by a disruption, the recovery supply portfolio is selected to optimize
the process of recovery from the disruption.

9.3 Models for Risk-Neutral Decision-Making

9.3.1 Integrated Selection of Primary and Recovery

Supply Portfolios

In this subsection a time-indexed SMIP model Support_E is presented for the inte-
grated risk-neutral decision making in the presence of supply chain disruption risks.
The objective of the integrated decision making is to jointly:
• determine the primary supply portfolio, i.e., to allocate the total demand for parts
among a subset of selected primary suppliers,
• determine the recovery supply portfolio for each disruption scenario, i.e., to allo-
cate the unfulfilled demand for parts among a subset of selected recovery suppliers,
when primary suppliers are disrupted,
• schedule production over the planning horizon for each disruption scenario,
to optimize the expected cost. The cost includes ordering and purchasing cost of
parts, recovery cost of disrupted suppliers and penalty cost for unfulfilled demand
for products.
The variables used to formulate the SMIP model are defined in Table 9.2.
The primary supply portfolio, (v1 , . . . , vI ), where i∈I vi = 1 and 0 ≤ vi ≤ 1, i ∈
I, is the initial allocation of total demand for parts among primary
suppliers.
244 9 Selection of Primary and Recovery Supply Portfolios and Scheduling

Table 9.2 Variables: selection of primary and recovery supply portfolios and scheduling
First stage variables
ui = 1, if supplier i is selected as a primary supplier; otherwise ui = 0 (primary supplier
selection)
vi ∈ [0, 1], the fraction of total demand for parts ordered from primary supplier i (primary
supply portfolio)
Second stage variables
Ui s = 1, if supplier i is selected as a recovery supplier under scenario s; otherwise Uis = 0
(recovery supplier selection)
Vis ∈ [0, 1], the fraction of total demand for parts ordered from recovery supplier i under
scenario s (recovery supply portfolio)
xts ≥ 0, production in period t under scenario s (production scheduling)
Auxiliary variables
qis = 1, if ui = Uis = 1; otherwise qis = 0 (elimination of double fixed ordering costs)
VaRc Cost-at-Risk, the targeted cost such that for a given confidence level α, for 100α% of
the scenarios, the outcome is below VaRc
VaR sl Service-at-Risk, the targeted service level such that for a given confidence level α, for
100α% of the scenarios, the outcome is above VaRsl
Cs ≥ 0, the tail cost for scenario s, i.e., the amount by which costs in scenario s exceed VaRc
Ss ≥ 0, the tail service level for scenario s, i.e., the amount by which VaRsl exceeds service
level in scenario s

The recovery supply portfolio for each disruption scenario s, (V1s , . . . , VIs ),
where i∈I (γi,λis vi + Vis ) = 1 and 0 ≤ Vis ≤ 1, i ∈ I, is the allocation among recov-
ery suppliers of unfulfilled demand for parts caused by supply disruptions under
scenario s.
Notice that the actual quantity ordered from
each supplier i can be determined by
multiplying total demand for parts, (D = t∈T dt ), by fractional variable vi or Vis ,
i.e., by the fraction of total demand allotted to supplier i.
In addition, the following auxiliary binary variable, qis , is introduced in model
Support_E: qis = 1, if ui = Uis = 1; otherwise qis = 0.
This variable eliminates double charging with fixed ordering cost ei of each sup-
plier i, who is selected both as primary and recovery supplier.
Let E c be the expected cost per product to be minimized

Ec = Ps ( ei (ui + Uis − qis )/D

s∈S i∈I

+ ρis Uis /D + oi (γi,λis vi + Vis )) + g(1 − E sl ), (9.4)

i∈I i∈I

where λis is disruption level of supplier i under scenario s, and γi,λis is the corre-
sponding fulfillment rate, i.e., the fraction of an order delivered by supplier i under
disruption scenario s.
9.3 Models for Risk-Neutral Decision-Making 245

E sl = Ps xts /D (9.5)
s∈S t∈T

is the expected service level, i.e., the expected fraction of the total fulfilled demand
for products (expected demand fulfillment rate).
The expected cost per product, E c , (9.4), constitutes of expected ordering cost per
product,

i∈I ei (ui + Ui − qi )/D = i∈I ei ui /D +
s s
s∈S Ps s∈S Ps i∈I ei
(Uis − qis )/D,
firm’s
expected portion of suppliers cost-to-recover from disruptions, per product,
i∈I is Ui /D,
ρ s
P
s∈S s
expected
purchasing cost per product
for delivered parts,
P
s∈S s o (γ v
i∈I i i,λis i + V i
s
) = o Γ v
i∈I i i i + s∈S Ps
s
i∈I oi Vi ,
and expected
penalty per product for the unfulfilled demand for products,
g(1 − s∈S t∈T Ps xts /D) = g(1 − E sl ),
where Γi is the expected fulfillment rate of supplier i

Γi = Ps γi,λis ; i ∈ I.
s∈S

Support_E: Selection of primary and recovery Supply portfolios and

scheduling of production
Minimize (9.4)
subject to
Primary supply portfolio selection constraints
- the total demand for parts must be fully allocated among the selected
primary suppliers,
- demand for parts cannot be assigned to non-selected primary suppliers,

vi = 1 (9.6)
i∈I
vi ≤ ui ; i ∈ I (9.7)

Recovery supply portfolio selection constraints

- the unfulfilled demand for parts caused by supply disruptions under sce-
nario s must be fully allocated among the selected recovery suppliers,
- the unfulfilled demand for parts caused by supply disruptions under sce-
nario s cannot be assigned to non-selected recovery suppliers,
- each supplier selected to both primary and recovery portfolio is charged
exactly once with fixed ordering cost in the objective function (9.4),

(γi,λis vi + Vis ) = 1; s ∈ S (9.8)

i∈I
246 9 Selection of Primary and Recovery Supply Portfolios and Scheduling

Vis ≤ Uis ; i ∈ I, s ∈ S (9.9)

qis ≤ (ui + Uis )/2; i ∈ I, s ∈ S (9.10)

Supply-production coordinating constraints

- for each disruption scenario s and each period t, the cumulative demand for
parts of production scheduled in periods 1 through t cannot exceed the initial
inventory of parts and the cumulative deliveries by period t − 1 (delivery in
period τi ≤ t − 1 from each primary supplier i and in period ts + θis + τi ≤
t − 1 from each recovery supplier i),

xts /D ≤
t  ∈T :t  ≤t

V0 + γi,λis vi + Vis ; t ∈ T , s ∈ S, (9.11)

i∈I:τi ≤t−1 i∈I:ts +θis +τi ≤t−1

where DV0 is the initial inventory of parts

Production capacity constraints
- for each disruption scenario s and each period t, production cannot exceed
the producer capacity,

xts ≤ c; t ∈ T , s ∈ S (9.12)

Non-negativity and integrality conditions

ui ∈ {0, 1}; i ∈ I (9.13)

vi ∈ [0, 1]; i ∈ I (9.14)
Ui ∈ {0, 1}; i ∈ I, s ∈ S
s
(9.15)
Vis ∈ [0, 1]; i ∈ I, s ∈ S (9.16)
qis ∈ {0, 1}; i ∈ I, s ∈ S (9.17)
xts ≥ 0; t ∈ T , s ∈ S. (9.18)

A simple upper bound on the expected service level (9.5) is derived below.

Proposition 9.1
E sl ≤ min{1, C/D}, (9.19)

min{DV0 , cτmin } + c(T − τmin ), is the total available capacity of pro-

where C =
ducer, D = t∈T dt , is the total demand for products, and τmin = mini∈I (τi ) is the
minimum delivery lead time from suppliers.

Proof Supply-production coordinating constraints (9.11) and production capacity

constraints (9.12) imply that
9.3 Models for Risk-Neutral Decision-Making 247

xts ≤ min{DV0 , cτmin }; s ∈ S

t∈T :t≤τmin

xts ≤ c(T − τmin ); s ∈ S.

t∈T :t>τmin

Thus

E sl = Ps xts /D ≤ Ps C/D = C/D.

s∈S t∈T s∈S

Since E sl cannot be greater than 1, its upper bound is 1, if C/D > 1.

Notice that worst-case disruption scenarios can be identified by the lowest service
level, SL. When there is no initial inventory of parts (i.e., V0 = 0), SL can be calculated
as below.
SL = (T − max{ts + min{θis + τi }})c/D. (9.20)
s∈S i∈I

Model Support_E is a deterministic equivalent mixed integer program of a two-

stage stochastic mixed integer program with recourse. The primary portfolio selection
variables, ui , vi , are referred to as first-stage decisions, and the recovery portfolio
selection variables, Uis , Vis and production scheduling variables, xts , are referred to
as recourse or second-stage decisions (cf. Table 9.2). Unlike the first-stage decisions,
the latter variables are dependent on disruption scenario s ∈ S. Model Support_E
illustrates the wait-and-see approach. Basically, this approach is based on the per-
fect information about the future. In contrast to the two-stage stochastic program
with recourse, in the wait-and-see approach, both the first stage and the second stage
decisions are made simultaneously only when disruption scenario is known. Model
Support_E is capable of simultaneously determining both stage variables to min-
imize expected cost over all disruption scenarios. When this problem is solved a
recommendation is obtained for selection of primary supply portfolio (ui , vi ) that
will hedge against a variety of disruption scenarios in which fulfillment rates of
certain suppliers are not sufficient to satisfy demand for parts. The recovery supply
portfolio selection and production scheduling (Uis , Vis and xts ) are decisions that will
be implemented in the future, when scenario s ∈ S is finally realized.
Even though the recovery portfolio selection, Uis , Vis , and production scheduling,
s
xt , decisions are associated with each disruption scenario, they can be viewed as
recourse decisions during disruption: each scenario s ∈ S is associated with time
period of disruption occurrence, ts , as well as disruption duration, θis , (i.e., time-to-
recover from disruption of each supplier i). Moreover, a recovery supplier selection
decision is interrelated with production scheduling during supply disruption.
248 9 Selection of Primary and Recovery Supply Portfolios and Scheduling

9.3.2 Multiple Part Types and Product Types

In this subsection an enhancement of model Support_E is described for multiple

types of parts and products. Let H and K be, respectively, the set of part types and
set of product types, and denote by ahk , h ∈ H, k ∈ K the number of parts type h
required to produce one unit of product type k.
If we denote by Dk the total demand for products type k, then D = k∈K Dk is
the total demand for all products and Ah = k∈K ahk Dk is the total demand for parts
type h.
Now, the portfolio decision variables vi , Vis , i ∈ I, s ∈ S are replaced by vih ,
Vih ; i ∈ I, h ∈ H, s ∈ S, defined as fractions of total demand Ah for parts type
s

h, ordered from primary supplier i, recovery supplier i under disruption scenario

s, respectively. In addition, the production scheduling variable xts , t ∈ T , s ∈ S is
replaced by xkts , k ∈ K, t ∈ T , s ∈ S - the production of product type k in period t
under disruption scenario s.
Model SupportMP_E for multiple part and product types is presented below.

SupportMP
Minimize

Ec = Ps ( ei (ui + Uis − qis )

s∈S i∈I

+ ρis Uis + oih Ah (γi,λis vih + Vihs )

i∈I i∈I h∈H

+ gk (Dk − xkts ))/D,

k∈K t∈T

where oih is unit purchasing price of part type h from supplier i, and gk is unit
penalty cost of unfulfilled demand for product type k,
subject to

vih = 1; h ∈ H
i∈I
vih ≤ ui ; i ∈ I, h ∈ H

(γi,λis vih + Vihs ) = 1; h ∈ H, s ∈ S

i∈I
Vihs ≤ Uis ; i ∈ I, h ∈ H, s ∈ S
qis ≤ (ui + Uis )/2; i ∈ I, s ∈ S

ahk xkts  /Ah ≤ Vh0 + γi,λis vih

k∈K t  ∈T :t  ≤t i∈I:τi ≤t−1
9.3 Models for Risk-Neutral Decision-Making 249

+ Vihs ; h ∈ H, t ∈ T , s ∈ S
i∈I:ts +θis +τi ≤t−1

xkts ≤ c; t ∈ T , s ∈ S
k∈K
ui ∈ {0, 1}; i ∈ I
vih ∈ [0, 1]; i ∈ I, h ∈ H
Uis ∈ {0, 1}; i ∈ I, s ∈ S
Vihs ∈ [0, 1]; i ∈ I, h ∈ H, s ∈ S
qis ∈ {0, 1}; i ∈ I, s ∈ S
xkts ≥ 0; k ∈ K, t ∈ T , s ∈ S,

where Ah Vh0 is the initial inventory of parts type h.

Notice that the supply-production coordinating constraints (9.11) are also bill-of-
material constraints in model SupportMP_E.
The above model can be further enhanced, for example by introducing subsets
Ih ⊂ I of suppliers for each part type h, unit capacity consumption for each product
type k in the left-hand side of production capacity constraints, etc.

9.3.3 Hierarchical Selection of Primary and Recovery

Supply Portfolios

In this subsection two deterministic MIP models PSupport and RSupport(s) are
presented for the hierarchical decision making in the presence of supply chain dis-
ruption risks. The two-stage decision making is described below (Fig. 9.1).
1. Selection of primary supply portfolio for deterministic environment.
The primary suppliers are determined ahead of time with no disruption scenarios
considered, using deterministic MIP model PSupport.
2. Selection of recovery supply portfolio, after disruption of primary suppliers.
The recovery suppliers are determined after a primary supplier was hit by a dis-
ruption to optimize the process of recovery from the disruption, using MIP model
RSupport(s).
The hierarchical approach is a purely top-down approach.
In the deterministic MIP model PSupport, stochastic variable, xts , (9.18), defined
in model Support for each disruption scenario s ∈ S has been replaced by its deter-
ministic equivalent Xt .
250 9 Selection of Primary and Recovery Supply Portfolios and Scheduling

Fig. 9.1 Hierarchical

selection of supply portfolios

PSupport: Primary supply portfolio selection and production scheduling

Minimize

Pc = (ei ui /D + oi vi ) + g(1 − Xt /D) (9.21)

i∈I t∈T

subject to

vi = 1 (9.22)
i∈I
vi ≤ ui ; i ∈ I (9.23)

Xt  /D ≤ V0 + vi ; t ∈ T (9.24)
t  ∈T :t  ≤t i∈I:τi ≤t−1
Xt ≤ c; t ∈ T (9.25)
ui ∈ {0, 1}; i ∈ I (9.26)
vi ∈ [0, 1]; i ∈ I (9.27)
Xt ≥ 0; t ∈ T . (9.28)

The solution to PSupport is primary supply portfolio: ui∗ , vi∗ ; i ∈ I, and the
associated production schedule, Xt∗ .
9.3 Models for Risk-Neutral Decision-Making 251

A simple lower bound on the objective function Pc , (9.21), of the top level problem
PSupport is derived below.
Proposition 9.2

Pc ≥ min(ei /D + oi ) + g max{0, 1 − C/D}. (9.29)

i∈I

Proof Primary supply portfolio selection constraints (9.22), (9.23) and Proposition
9.2, (9.19), imply that

Pc = (ei ui /D + oi vi ) + g(1 − Xt /D) ≥

i∈I t∈T

(ei /D + oi )vi + g max{0, 1 − C/D} ≥

i∈I
min(ei /D + oi ) + g max{0, 1 − C/D}.
i∈I

In the deterministic MIP model RSupport(s), stochastic variables, Uis , Vis , xts ,
(9.15), (9.16), (9.18), defined in model Support_E for each disruption scenario s ∈ S
have been replaced by their deterministic equivalents, Uis̃ , Vis̃ , xts̃ for the realized
disruption scenario s = s̃.

RSupport(s): Recovery supply portfolio selection and production scheduling

for predetermined primary supply portfolio and the realized disruption
scenario
Minimize

Rs̃c = (ei (1 − ui∗ )Uis̃ /D + ρis̃ Uis̃ /D + oi Vis̃ ) + g(1 − xts̃ /D)
i∈I t∈T

+ (ei ui∗ /D + oi γi,λis̃ vi∗ ) (9.30)

i∈I

subject to

Vis̃ = 1 − γi,λis̃ vi∗ (9.31)

i∈I i∈I
Vis̃ ≤ Uis̃ ; i ∈ I (9.32)

xts̃ /D − Vis̃ ≤ V0 + γi,λis̃ vi∗ ; t ∈ T (9.33)

t  ∈T :t  ≤t i∈I:ts̃ +θis̃ +τi ≤t−1 i∈I:τi ≤t−1

xts̃ ≤ c; t ∈ T (9.34)
Uis̃ ∈ {0, 1}; i ∈ I (9.35)
Vis̃ ∈ [0, 1]; i ∈ I (9.36)
xts̃ ≥ 0; t ∈ T . (9.37)
252 9 Selection of Primary and Recovery Supply Portfolios and Scheduling

Since the primary supply portfolio has been predetermined, the last summation
term in Eq. (9.30) as well as the right-hand sides of Eqs. (9.31) and (9.33) are constant.
The solution to RSupport(s) is the recovery supply portfolio, Uis̃ , Vis̃ ; i ∈ I, and
the production schedule, xts̃ ; t ∈ T , for the realized disruption scenario s = s̃:
Notice that no recovery supply portfolios, Vis = 0, ∀i ∈ I, will be determined for
scenarios s with non disrupted primary suppliers, i.e., for scenarios s ∈ S such that
λis = Li if vi∗ > 0. Since γi,Li = 1, i∈I γi,Li vi∗ = 1, and hence the right-hand side of
(9.31) is equal to zero.
Model Support_E for the predetermined primary supply portfolio is separable
with respect to disruption scenarios s ∈ S, since the objective function (9.4) is additive
and separable with respect to s and all constraints (9.8)–(9.12) are also separable with
respect to s. Thus, given the primary supply portfolio, the recovery supply portfolios,
Uis , Vis ; i ∈ I, and production schedules, xts , can be found simultaneously for all
potential disruption scenarios s ∈ S, by solving
RSupport_E = Support_E for predetermined primary supply portfolio: ui =
ui∗ , vi = vi∗ ; i ∈ I.

RSupport_E: Recovery supply portfolio selection and production scheduling

for predetermined primary supply portfolio
Minimize (9.4)
subject to (9.8)–(9.12), (9.15)–(9.18) and

ui = ui∗ ; i ∈ I (9.38)
vi = vi∗ ; i ∈ I. (9.39)

9.4 Models for Risk-Averse Decision-Making

In this section the two time-indexed SMIP models Support_CV(c) and Sup-
port_CV(sl) are proposed for the integrated, risk-averse selection of primary and
recovery supply portfolios and production scheduling to optimize, respectively CVaR
of cost and CVaR of service level under disruption risks. The models are based on
the risk-neutral model Support_E.
Let VaRc be the Value-at Risk of cost per product, i.e., the targeted cost such
that for a given confidence level α, for 100α% of disruption scenarios, the outcome
is below VaRc . Accordingly, let CVaRc be Conditional Value-at-Risk of cost per
product, i.e., the expected cost in the worst 100(1 − α)% of the scenarios with the
cost above VaRc .
9.4 Models for Risk-Averse Decision-Making 253

CV aRc = V aRc + (1 − α)−1 Ps Cs . (9.40)

s∈S

The risk-averse primary and recovery supply portfolio and the production schedule
will be optimized by calculating VaRc and minimizing CVaRc simultaneously. Model
Support_CV(c) is presented below.

Support_CV(c): Selection of primary and recovery supply portfolios and

scheduling of production to minimize CVaR of cost
Minimize (9.40)
subject to (9.6)–(9.18) and
Risk constraints:
- the tail cost for scenario s is defined as the nonnegative amount by which
cost in scenario s exceeds VaRc ,

Cs ≥ ei (ui + Uis − qis )/D + ρis Uis /D

i∈I i∈I

+ oi (γi,λis vi + Vis ) + g(1 − xts /D) − V aRc ; s ∈ S (9.41)

i∈I t∈T
Cs ≥ 0; s ∈ S, (9.42)

where Cs is the tail cost for scenario s.

In the next model, VaRsl is the Value-at Risk of service level, i.e., the targeted
service level such that for a given confidence level α, for 100α% of disruption
scenarios, the outcome is above VaRsl , while CVaRsl is the Conditional Value-at-
Risk of service level, i.e., the expected service level in the worst 100(1 − α)% of
scenarios with the service level below VaRsl .

CV aRsl = V aRsl − (1 − α)−1 Ps Ss . (9.43)

s∈S

The risk-averse primary and recovery supply portfolio and the production sched-
ule will be optimized by calculating VaRsl and maximizing CVaRsl simultaneously.
Model Support_CV(sl) is presented below.

Support_CV(sl): Selection of primary and recovery supply portfolios and

scheduling of production to maximize CVaR of service level
Maximize (9.43)
subject to (9.6)–(9.18) and
Risk constraints:
254 9 Selection of Primary and Recovery Supply Portfolios and Scheduling

- the tail service level for scenario s is defined as the nonnegative amount
by which VaRsl exceeds service level in scenario s,

Ss ≥ V aRsl − xts /D; s ∈ S (9.44)

t∈T
Ss ≥ 0; s ∈ S, (9.45)

where Ss is the tail service level for scenario s.

9.5 Computational Examples

In this section some computational examples are presented to illustrate the proposed
portfolio approach for selection of primary and recovery supply portfolios. The
problem of joint selection of primary and recovery supply portfolios and produc-
tion scheduling under disruption risks considered here is different from the existing
literature in many different ways. Although the input data for the examples are
hypothetical, their relations to each other are real and in part they have been taken
from real case studies. In particular, the case studies of Toyota supply chain disrup-
tion and recovery after the Great East Japan earthquake and tsunami of March 11,
2011 (Fujimoto and Park 2013; Park et al. 2013; MacKenzie et al. 2014; Matsuo
2015) and a case study of Thailands floods in 2011 (Haraguchi and Lall 2015) have
been analyzed. The following parameters have been selected for the computational
examples.
I = 4 suppliers, Li = 3 partial disruption levels for all i ∈ I, R = 2 geographic
regions, T = 30 planning periods.
I 1 = {1, 2}, I 2 = {3, 4}.
Delivery lead times from suppliers: τ = (2, 2, 4, 4).
The initial inventories of parts: V0 = 0.
Customer demand: dt = 10000 for all t ∈ T,
and total demand for parts/products: D = t∈T dt = 300000.
Fixed ordering costs for suppliers: e = (8000, 6000, 12000, 13000).
Unit penalties for unfulfilled demand:
g ∈ {1, 10, 100, 1000, 10000, 100000, ∞}, where g = ∞ denotes maximization
of service level.
Unit purchasing prices from suppliers: o = (14, 12, 8, 9).
Local disruption levels and the associated supplier fulfillment rates (the percentage
of an order that can be delivered) are shown below.
Li = L = {0, 1, 2, 3} for all i ∈ I, where l = 0, complete shutdown, γi0 = 0 ∀i ∈
I, i.e., 0% of an order delivered; l = 1, major disruption, γi1 ∈ [0.01, 0.50] ∀i ∈ I 1
and γi1 ∈ [0.01, 0.30] ∀i ∈ I 2 , i.e., 1% to 50% and 1% to 30% of an order deliv-
ered, respectively; l = 2, minor disruption, γi2 ∈ [0, 51, 0.99] ∀i ∈ I 1 and γi2 ∈
9.5 Computational Examples 255

Fig. 9.2 Suppliers

[0, 31, 0.99] ∀i ∈ I 2 , i.e., 51% to 99% and 31% to 99% of an order delivered, respec-
tively; l = L = 3, no disruption, γi3 = 1 ∀i ∈ I, i.e., 100% of an order delivered.
The total number of all potential scenarios is S = (L + 1)I = 44 = 256 scenarios,
where each scenario s ∈ S is represented by vector λs = {λ1s , . . . , λ4s }, where λis ∈
L, i ∈ I, see Table 9.3.
The probability of realizing of disruption scenario s ∈ S for suppliers in region
r = 1, 2 is calculated as follows (see, Sect. 9.2)
⎧  
⎨ (1 −

pr )( i∈I r :λis =0 0.1(1 − pi3 ))( i∈I r :λis =1 0.3(1 − pi3 ))

Psr = ×( i∈I r :λis =2 0.6(1 − pi3 ))( i∈I r :λis =3 pi3 ) if λ >0
⎩ r  i∈I r is
p + (1 − p ) i∈I r 0.1(1 − pi3 )
r if i∈I r λis = 0,

and the probability for disruption scenario s ∈ S is given by Ps = Ps1 Ps2 .

In the above formula, pi3 denotes the local probability of non-disruptive operation
(level l = 3) for supplier i and pr denotes the regional disruption probability for all
256 9 Selection of Primary and Recovery Supply Portfolios and Scheduling

suppliers in region r. The probability pi3 was uniformly distributed over [0.89,0.99]
and [0.79,0.89], respectively for suppliers i ∈ I 1 , and i ∈ I 2 .
Given local non disruption probabilities, pi3 , i ∈ I, the probabilities for the
remaining local disruption levels l = 0, 1, 2 were calculated as follows:
probability of complete shutdown (level l = 0), pi0 = 0.1(1 − pi3 );
probability of major disruption (level l = 1), pi1 = 0.3(1 − pi3 );
probability of minor disruption (level l = 2), pi2 = 0.6(1 − pi3 ) for all suppliers
i ∈ I.
Notice that, pi0 ≤ pi1 ≤ pi2 ≤ pi3 , i.e., the probability of disruption level of a
supplier increases with the level such that the smallest probability was assigned to
complete shutdown (level 0) and the largest to non-disruptive operation (level 3).
The regional disruption probabilities are p1 = 0.001 and p2 = 0.01.
The regional disruption probabilities are chosen to be lower than local shutdown
probabilities of suppliers in that region. In addition, the two regional disruption prob-
abilities are significantly different to represent two geographic regions differently
exposed to disruptive events.
Cost-to-recover and time-to-recover are defined below.
CTR(i, l)=if l = 0 then 100000ei ; if l = 1 then 10000ei ; if l = 2 then 1000ei
∀i ∈ I,
TTR(i, l)=if l = 0 then 12; if l = 1 then 10; if l = 2 then 8 ∀i ∈ I.
Figure 9.2 presents basic characteristics of all suppliers: purchasing price, oi , i ∈
shutdown,r p + (1 − pr )pi0 , i ∈ I , r ∈ R, and expected
r r r
I, probability of complete
fulfillment rate, Γi = l=1,2,3 (1 − p )pil γil , i ∈ I , r ∈ R.
The three levels of producer per period capacity will be considered:
c ∈ {5000, 10000, 15000},
and total available producer capacity, C = c(T − τmin ), respectively:
C ∈ {140000, 280000, 420000}.
The corresponding capacity-to-demand ratio is C/D = 0.47, 0.93 and 1.4.

9.5.1 Risk-Neutral Decision-Making

Scenarios with a Common Disruption Start Time

In this subsection each disruption s ∈ S to primary suppliers is assumed to occur in
the same period ts = 1, before the earliest delivery lead time, mini∈I (τi ) = 2, and the
recovery process starts in period t = 2 so that the disrupted supplier i returns to its
full capacity in period t = TTR(i, l) + 1.
The optimal solutions for both the integrated and the hierarchical approach are
summarized in Tables 9.4, 9.5, and 9.6, respectively for capacity-to-demand ratio,
C/D = 0.47, 0.93 and 1.4.
For different values of unit penalty cost g, the tables show:
• optimal primary supply portfolio, (v1 , v2 , v3 .v4 ), indicating percent of total demand
for parts ordered from each supplier;
9.5 Computational Examples 257

Table 9.3 Multi-level disruption scenarios: λ1s , λ2s , λ3s , λ4s ; s = 1, . . . , 256

s i=1 2 3 4 s i=1 2 3 4 s i=1 2 3 4 s i=1 2 3 4

1 0 0 0 0 65 1 0 0 0 129 2 0 0 0 193 3 0 0 0
2 0 0 0 1 66 1 0 0 1 130 2 0 0 1 194 3 0 0 1
3 0 0 0 2 67 1 0 0 2 131 2 0 0 2 195 3 0 0 2
4 0 0 0 3 68 1 0 0 3 132 2 0 0 3 196 3 0 0 3
5 0 0 1 0 69 1 0 1 0 133 2 0 1 0 197 3 0 1 0
6 0 0 1 1 70 1 0 1 1 134 2 0 1 1 198 3 0 1 1
7 0 0 1 2 71 1 0 1 2 135 2 0 1 2 199 3 0 1 2
8 0 0 1 3 72 1 0 1 3 136 2 0 1 3 200 3 0 1 3
9 0 0 2 0 73 1 0 2 0 137 2 0 2 0 201 3 0 2 0
10 0 0 2 1 74 1 0 2 1 138 2 0 2 1 202 3 0 2 1
11 0 0 2 2 75 1 0 2 2 139 2 0 2 2 203 3 0 2 2
12 0 0 2 3 76 1 0 2 3 140 2 0 2 3 204 3 0 2 3
13 0 0 3 0 77 1 0 3 0 141 2 0 3 0 205 3 0 3 0
14 0 0 3 1 78 1 0 3 1 142 2 0 3 1 206 3 0 3 1
15 0 0 3 2 79 1 0 3 2 143 2 0 3 2 207 3 0 3 2
16 0 0 3 3 80 1 0 3 3 144 2 0 3 3 208 3 0 3 3
17 0 1 0 0 81 1 1 0 0 145 2 1 0 0 209 3 1 0 0
18 0 1 0 1 82 1 1 0 1 146 2 1 0 1 210 3 1 0 1
19 0 1 0 2 83 1 1 0 2 147 2 1 0 2 211 3 1 0 2
20 0 1 0 3 84 1 1 0 3 148 2 1 0 3 212 3 1 0 3
21 0 1 1 0 85 1 1 1 0 149 2 1 1 0 213 3 1 1 0
22 0 1 1 1 86 1 1 1 1 150 2 1 1 1 214 3 1 1 1
23 0 1 1 2 87 1 1 1 2 151 2 1 1 2 215 3 1 1 2
24 0 1 1 3 88 1 1 1 3 152 2 1 1 3 216 3 1 1 3
25 0 1 2 0 89 1 1 2 0 153 2 1 2 0 217 3 1 2 0
26 0 1 2 1 90 1 1 2 1 154 2 1 2 1 218 3 1 2 1
27 0 1 2 2 91 1 1 2 2 155 2 1 2 2 219 3 1 2 2
28 0 1 2 3 92 1 1 2 3 156 2 1 2 3 220 3 1 2 3
29 0 1 3 0 93 1 1 3 0 157 2 1 3 0 221 3 1 3 0
30 0 1 3 1 94 1 1 3 1 158 2 1 3 1 222 3 1 3 1
31 0 1 3 2 95 1 1 3 2 159 2 1 3 2 223 3 1 3 2
32 0 1 3 3 96 1 1 3 3 160 2 1 3 3 224 3 1 3 3
33 0 2 0 0 97 1 2 0 0 161 2 2 0 0 225 3 2 0 0
34 0 2 0 1 98 1 2 0 1 162 2 2 0 1 226 3 2 0 1
35 0 2 0 2 99 1 2 0 2 163 2 2 0 2 227 3 2 0 2
36 0 2 0 3 100 1 2 0 3 164 2 2 0 3 228 3 2 0 3
37 0 2 1 0 101 1 2 1 0 165 2 2 1 0 229 3 2 1 0
38 0 2 1 1 102 1 2 1 1 166 2 2 1 1 230 3 2 1 1
39 0 2 1 2 103 1 2 1 2 167 2 2 1 2 231 3 2 1 2
40 0 2 1 3 104 1 2 1 3 168 2 2 1 3 232 3 2 1 3
41 0 2 2 0 105 1 2 2 0 169 2 2 2 0 233 3 2 2 0
42 0 2 2 1 106 1 2 2 1 170 2 2 2 1 234 3 2 2 1
43 0 2 2 2 107 1 2 2 2 171 2 2 2 2 235 3 2 2 2
44 0 2 2 3 108 1 2 2 3 172 2 2 2 3 236 3 2 2 3
45 0 2 3 0 109 1 2 3 0 173 2 2 3 0 237 3 2 3 0
46 0 2 3 1 110 1 2 3 1 174 2 2 3 1 238 3 2 3 1
47 0 2 3 2 111 1 2 3 2 175 2 2 3 2 239 3 2 3 2
48 0 2 3 3 112 1 2 3 3 176 2 2 3 3 240 3 2 3 3
49 0 3 0 0 113 1 3 0 0 177 2 3 0 0 241 3 3 0 0
50 0 3 0 1 114 1 3 0 1 178 2 3 0 1 242 3 3 0 1
51 0 3 0 2 115 1 3 0 2 179 2 3 0 2 243 3 3 0 2
52 0 3 0 3 116 1 3 0 3 180 2 3 0 3 244 3 3 0 3
53 0 3 1 0 117 1 3 1 0 181 2 3 1 0 245 3 3 1 0
54 0 3 1 1 118 1 3 1 1 182 2 3 1 1 246 3 3 1 1
55 0 3 1 2 119 1 3 1 2 183 2 3 1 2 247 3 3 1 2
56 0 3 1 3 120 1 3 1 3 184 2 3 1 3 248 3 3 1 3
57 0 3 2 0 121 1 3 2 0 185 2 3 2 0 249 3 3 2 0
58 0 3 2 1 122 1 3 2 1 186 2 3 2 1 250 3 3 2 1
59 0 3 2 2 123 1 3 2 2 187 2 3 2 2 251 3 3 2 2
60 0 3 2 3 124 1 3 2 3 188 2 3 2 3 252 3 3 2 3
61 0 3 3 0 125 1 3 3 0 189 2 3 3 0 253 3 3 3 0
62 0 3 3 1 126 1 3 3 1 190 2 3 3 1 254 3 3 3 1
63 0 3 3 2 127 1 3 3 2 191 2 3 3 2 255 3 3 3 2
64 0 3 3 3 128 1 3 3 3 192 2 3 3 3 256 3 3 3 3
258 9 Selection of Primary and Recovery Supply Portfolios and Scheduling

Table 9.4 Solution results for a common disruption start time: C/D = 0.47
g 1 10 102 103 104 105 ∞ (a)
Integrated approach: model Support_E
Var. = 10205, Bin. = 2052, Cons. = 10462, Nonz. = 10205 (e)
Exp.Cost E c , (9.4) 8.80 13.72 61.76 542 5343 53337 –
Exp.Service E sl × 100%, (9.5) 43 47 47 47 47 47 47
Primary Portfolio: Supplier(% of total demand) (b) 1(0) 1(0) 1(2) 1(4) 1(9) 1(31)
2(0) 2(3) 2(2) 2(4) 2(10) 2(19)
3(100) 3(97) 3(96) 3(92) 3(79) 3(26)
4(0) 4(0) 4(0) 4(0) 4(2) 4(24)
Exp.Recovery Portfolio: Supplier(% of total demand) (c) 1(0.18) 1(0.27) 1(0.22) 1(0.17) 1(0.16) 1(1.82)
2(1.86) 2(1.81) 2(1.82) 2(1.71) 2(1.51) 2(1.46)
3(0) 3(0.04) 3(0.06) 3(0.25) 3(0.7) 3(1.25)
4(7.21) 4(6.95) 4(6.93) 4(6.63) 4(5.80) 4(1.43)
Hierarchical approach: models PSupport and RSupport
PSupport: Var. = 36, Bin. = 4, Cons. = 37, Nonz. = 534
Primary Portfolio: Supplier(% of total demand) (b) 1(0) 1(0) 1(0)
2(0) 2(3) 2(100)
3(100) 3(97) 3(0)
4(0) 4(0) 4(0)
Cost Pc , (9.21) 8.61 13.53 61.53 542 5342 53342 –
Service (d) 43 47 47 47 47 47 47
RSupport_E: Var. = 8050, Bin. = 768, Cons. = 6250, Nonz. = 8050
Exp.Recovery Portfolio: Supplier(% of total demand) (c) 1(0.18) 1(0.22) 1(0.83)
2(1.86) 2(1.73) 2(3.14)
3(0) 3(0.20) 3(0.07)
4(7.21) 4(6.75) 4(0.04)
Exp.Cost E c , (9.4) 8.80 13.84 61.85 542 5344 53365 –
Exp.Service E sl × 100%, (9.5) 43 47 47 47 47 47 47
(a) Maximization of service level
(b) 1(v 1 × 100), 2(v2 × 100), 3(v3 × 100%), 4(v4 × 100)
(c) 1( Ps V1s × 100), 2( s∈S Ps V2s × 100), 3( s∈S Ps V3s × 100), 4( s∈S Ps V4s × 100)
(d)
s∈S
t∈T Xt /D × 100
(e) Var. = number of variables, Bin. = number of binary variables, Cons. = number of constraints,

Nonz. = number of nonzero coefficients

• optimal
expected recovery supply portfolio, ( s∈S Ps V1s , s∈S Ps V2s , s∈S Ps V3s ,
s∈S Ps V4 ), indicating expected percent of total demand for parts ordered from
s

each supplier.
For the integrated approach and the bottom level problem RSupport_E of the
hierarchical approach, expected cost E c , (9.4), is shown along with the associated
expected service level E sl , (9.5). Similarly, for the top level problem PSupport
c
approach, cost P , (9.21), is shown along with the associated
of the hierarchical
service level, t∈T Xt /D. In addition, Table 9.4 presents the size of each MIP model,
Support_E, PSupport and RSupport_E. The results indicate that for most examples
the expected service level E sl , (9.5), and the service level for PSupport problem,
9.5 Computational Examples 259

Table 9.5 Solution results for a common disruption start time: C/D = 0.93
g 1 10 102 103 104 105 ∞ (a)
Integrated approach: model Support_E
Exp.Cost E c , (9.4) 8.36 9.19 15.26 75.53 677 6685 –
Exp.Service E sl × 100%, (9.5) 87 93 93 93 93 93 93
Primary Portfolio: Supplier(% of total demand) (b) 1(0) 1(0) 1(3) 1(8) 1(18) 1(27)
2(0) 2(7) 2(4) 2(9) 2(20) 2(25)
3(100) 3(93) 3(93) 3(83) 3(57) 3(7)
4(0) 4(0) 4(0) 4(0) 4(5) 4(41)
Exp.Recovery Portfolio: Supplier(% of total demand) (c) 1(0.18) 1(0.36) 1(0.25) 1(0.16) 1(0.13) 1(1.99)
2(1.86) 2(1.77) 2(1.78) 2(1.55) 2(1.16) 2(1.99)
3(0) 3(0.07) 3(0.13) 3(0.49) 3(1.44) 3(1.37)
4(7.21) 4(6.70) 4(6.65) 4(6.06) 4(4.26) 4(0.40)
Hierarchical approach: models PSupport and RSupport_E
Primary Portfolio: Supplier(% of total demand) (b) 1(0) 1(0) 1(0)
2(0) 2(7) 2(100)
3(100) 3(93) 3(0)
4(0) 4(0) 4(0)
Cost Pc , (9.21) 8.17 9 15 75 675 6675 –
Service (d) 87 93 93
Exp. Recovery Portfolio: Supplier(% of total demand) (c) 1(0.18) 1(0.36) 1(1.18)
2(1.86) 2(1.77) 2(1.27)
3(0) 3(0.07) 3(1.55)
4(7.21) 4(6.70) 4(0.07)
Exp.Cost E c , (9.4) 8.36 9.19 15.26 76.02 684 6759 –
Exp.Service E sl × 100%, (9.5) 87 93 93
(a) Maximization of service level
(b) 1(v 1 × 100), 2(v2 × 100), 3(v3 × 100), 4(v4 ×
100)
(c) 1(
s∈S Ps V1 × 100), 2( s∈S Ps V2 × 100), 3( s∈S Ps V3 × 100), 4( s∈S Ps V4 × 100)
s s s s
(d)

t∈T Xt /D × 100

Xt /D, attain their upper bounds C/D, (9.19), whereas the cost Pc , (9.21),
t∈T
remains at its lower bound 8.04, (9.29), only for C/D = 1.4.
Tables 9.4 and 9.5 indicate that for C/D < 1 and small unit penalty cost g, the
cheapest supply portfolio is selected and the achieved service level is less than C/D.
The same solution results are obtained for both the integrated and the hierarchi-
cal approach. As g increases to reduce the unfulfilled demand, more expensive and
diversified supply portfolios are selected and the highest service level C/D, (9.19),
is attained. In particular, for integrated approach the more diversified primary supply
portfolios are selected for g > 1 to hedge against all disruption scenarios. In con-
trast, for hierarchical approach the results for g > 1 are independent of g. When the
objective is to maximize service level (g → ∞), the hierarchical approach selects
the most reliable supplier as the only primary supplier. In addition, Table 9.6 demon-
strates that for C/D > 1 and both approaches, the primary portfolio and the expected
260 9 Selection of Primary and Recovery Supply Portfolios and Scheduling

Table 9.6 Solution results for a common disruption start time: C/D = 1.4
g 1 10 102 103 104 105 ∞ (a)
Integrated approach: model Support_E
Exp.Cost E c , (9.4) 8.23 8.23 8.23 8.24 8.31 9.05 –
Exp.Service E sl × 100%, (9.5) 100 100
Primary Portfolio: Supplier(% of total demand) (b) 1(0) 1(0) 1(40)
2(0) 2(6) 2(10)
3(100) 3(94) 3(26)
4(0) 4(0) 4(24)
Exp.Recovery Portfolio: Supplier(% of total demand) (c) 1(0.18) 1(0.17) 1(0.43)
2(1.86) 2(1.74) 2(2.10)
3(0) 3(0.22) 3(1.73)
4(7.21) 4(6.78) 4(1.62)
Hierarchical approach: models PSupport and RSupport_E:
Primary Portfolio: Supplier(% of total demand) (b) 1(0) 1(100)
2(0) 2(0)
3(100) 3(0)
4(0) 4(0)
Cost Pc , (9.21) 8.04 –
Service (d) 100 100
Exp.Recovery Portfolio: Supplier(% of total demand) (c) 1(0.18) 1(1.45)
2(1.86) 2(1.04)
3(0) 3(0.13)
4(7.21) 4(0.08)
Exp.Cost E c , (9.4) 8.23 8.23 8.23 8.24 8.31 9.05 –
Exp.Service E sl × 100%, (9.5) 100 100
(a) Maximization of service level
(b) 1(v 1 × 100), 2(v2 × 100), 3(v3 × 100), 4(v4 ×
100)
(c) 1(
s∈S Ps V1 × 100), 2( s∈S Ps V2 × 100), 3( s∈S Ps V3 × 100), 4( s∈S Ps V4 × 100)
s s s s
(d)

t∈T Xt /D × 100

recovery portfolio are independent of g, since for C/D = 1.4, the total demand for
products is fulfilled for all g and hence no penalty cost appears.
Table 9.7 shows an example of optimal primary supply portfolio v1 , v2 , v3 , v4
and optimal recovery supply portfolios V1s , V2s , V3s , V4s , ∀s ∈ S for C/D = 1.4 and
unit penalty g = 1, 10, 100, 1000, 10000. Both the integrated approach and the hier-
archical approach yield the same solution results (cf. Table 9.6). The table shows
the optimal primary supply portfolio, v1 = v2 = 0, v3 = 1, v4 = 0, and the associ-
ated cost Pc , (9.21), as well as the optimal recovery supply portfolios and costs Rsc ,
(9.30) for each disruption scenarios s ∈ S. The results demonstrate that no recovery
supply portfolio is selected, i.e., V1s = V2s = V3s = V4s = 0, for scenarios s ∈ S with
undisrupted primary supplier, i = 3, i.e., for {s ∈ S : λ3s = 3} = {13 − 16, 29 −
32, 45 − 48, 61 − 64, 77 − 80, 93 − 96, 109 − 112, 125 − 128, 141 − 144, 157 −
160, 173 − 176, 189 − 192, 205 − 208, 221 − 224, 237 − 240, 253 − 256}.
9.5 Computational Examples 261

Table 9.7 Optimal recovery supply portfolios for a common disruption start time, C/D = 1.4,
g = 1, 10, 102 , 103 , 104 : integrated approach/hierarchical approach†
Scenario: s Optimal recovery supply portfolio: V1s V2s V3s V4s Optimal expected cost: Rsc , (9.30)
s i=1 2 3 4 Cost s i=1 2 3 4 Cost
1 0 1 0 0 2037.06 65 1 0 0 0 295.73
2 0 0 0 1 467.42 66 1 0 0 0 295.73
3 0 0 0 1 67.42 67 0 0 0 1 67.42
4 0 0 0 1 9.08 68 0 0 0 1 9.08
5 0 0 0.9 0 422.72 69 0.9 0 0 0 284.8
6 0 0 0.9 0 422.72 70 0.9 0 0 0 284.8
7 0 0 0 0.9 56.99 71 0 0 0 0.9 56.99
8 0 0 0 0.9 8.98 72 0 0 0 0.9 8.98
9 0 0 0.44 0 48.04 73 0 0 0.44 0 48.04
10 0 0 0.44 0 48.04 74 0 0 0.44 0 48.04
11 0 0 0.44 0 48.04 75 0 0 0.44 0 48.04
12 0 0 0 0.44 8.53 76 0 0 0 0.44 8.53
13 0 0 0 0 8.04 77 0 0 0 0 8.04
14 0 0 0 0 8.04 78 0 0 0 0 8.04
15 0 0 0 0 8.04 79 0 0 0 0 8.04
16 0 0 0 0 8.04 80 0 0 0 0 8.04
17 0 1 0 0 227.06 81 0 1 0 0 227.06
18 0 1 0 0 227.06 82 0 1 0 0 227.06
19 0 0 0 1 67.42 83 0 0 0 1 67.42
20 0 0 0 1 9.08 84 0 0 0 1 9.08
21 0 0.9 0 0 216.33 85 0 0.9 0 0 216.33
22 0 0.9 0 0 216.33 86 0 0.9 0 0 216.33
23 0 0 0 0.9 56.99 87 0 0 0 0.9 56.99
24 0 0 0 0.9 8.98 88 0 0 0 0.9 8.98
25 0 0 0.44 0 48.04 89 0 0 0.44 0 48.04
26 0 0 0.44 0 48.04 90 0 0 0.44 0 48.04
27 0 0 0.44 0 48.04 91 0 0 0.44 0 48.04
28 0 0 0 0.44 8.53 92 0 0 0 0.44 8.53
29 0 0 0 0 8.04 93 0 0 0 0 8.04
30 0 0 0 0 8.04 94 0 0 0 0 8.04
31 0 0 0 0 8.04 95 0 0 0 0 8.04
32 0 0 0 0 8.04 96 0 0 0 0 8.04
33 0 1 0 0 37.06 97 0 1 0 0 37.06
34 0 1 0 0 37.06 98 0 1 0 0 37.06
35 0 1 0 0 37.06 99 0 1 0 0 37.06
36 0 0 0 1 9.08 100 0 0 0 1 9.08
37 0 0.9 0 0 31.65 101 0 0.9 0 0 31.65
38 0 0.9 0 0 31.65 102 0 0.9 0 0 31.65
39 0 0.9 0 0 31.65 103 0 0.9 0 0 31.65
40 0 0 0 0.9 8.98 104 0 0 0 0.9 8.98
41 0 0.44 0 0 29.83 105 0 0.44 0 0 29.83
42 0 0.44 0 0 29.83 106 0 0.44 0 0 29.83
43 0 0.44 0 0 29.83 107 0 0.44 0 0 29.83
44 0 0 0 0.44 8.53 108 0 0 0 0.44 8.53
45 0 0 0 0 8.04 109 0 0 0 0 8.04
46 0 0 0 0 8.04 110 0 0 0 0 8.04
47 0 0 0 0 8.04 111 0 0 0 0 8.04
48 0 0 0 0 8.04 112 0 0 0 0 8.04
49 0 1 0 0 12.06 113 0 1 0 0 12.06
50 0 1 0 0 12.06 114 0 1 0 0 12.06
51 0 1 0 0 12.06 115 0 1 0 0 12.06
52 0 0 0 1 9.08 116 0 0 0 1 9.08
53 0 0.9 0 0 11.65 117 0 0.9 0 0 11.65
54 0 0.9 0 0 11.65 118 0 0.9 0 0 11.65
55 0 0.9 0 0 11.65 119 0 0.9 0 0 11.65
56 0 0 0 0.9 8.98 120 0 0 0 0.9 8.98
57 0 0.44 0 0 9.83 121 0 0.44 0 0 9.83
58 0 0.44 0 0 9.83 122 0 0.44 0 0 9.83
59 0 0.44 0 0 9.83 123 0 0.44 0 0 9.83
60 0 0 0 0.44 8.53 124 0 0 0 0.44 8.53
61 0 0 0 0 8.04 125 0 0 0 0 8.04
62 0 0 0 0 8.04 126 0 0 0 0 8.04
63 0 0 0 0 8.04 127 0 0 0 0 8.04
64 0 0 0 0 8.04 128 0 0 0 0 8.04

(continued)
262 9 Selection of Primary and Recovery Supply Portfolios and Scheduling

Table 9.7 (continued)

Scenario: s Optimal recovery supply portfolio: V1s V2s V3s V4s Optimal expected cost: Rsc , (9.30)
s i=1 2 3 4 Cost s i=1 2 3 4 Cost
129 1 0 0 0 45.73 193 1 0 0 0 14.07
130 1 0 0 0 45.73 194 1 0 0 0 14.07
131 1 0 0 0 45.73 195 1 0 0 0 14.07
132 0 0 0 1 9.08 196 0 0 0 1 9.08
133 0.9 0 0 0 40.11 197 0.9 0 0 0 13.45
134 0.9 0 0 0 40.11 198 0.9 0 0 0 13.45
135 0.9 0 0 0 40.11 199 0.9 0 0 0 13.45
136 0 0 0 0.9 8.98 200 0 0 0 0.9 8.98
137 0.44 0 0 0 37.39 201 0.44 0 0 0 10.72
138 0.44 0 0 0 37.39 202 0.44 0 0 0 10.72
139 0.44 0 0 0 37.39 203 0.44 0 0 0 10.72
140 0 0 0 0.44 8.53 204 0 0 0 0.44 8.53
141 0 0 0 0 8.04 205 0 0 0 0 8.04
142 0 0 0 0 8.04 206 0 0 0 0 8.04
143 0 0 0 0 8.04 207 0 0 0 0 8.04
144 0 0 0 0 8.04 208 0 0 0 0 8.04
145 1 0 0 0 45.73 209 1 0 0 0 14.07
146 1 0 0 0 45.73 210 1 0 0 0 14.07
147 1 0 0 0 45.73 211 1 0 0 0 14.07
148 0 0 0 1 9.08 212 0 0 0 1 9.08
149 0.9 0 0 0 40.11 213 0.9 0 0 0 13.45
150 0.9 0 0 0 40.11 214 0.9 0 0 0 13.45
151 0.9 0 0 0 40.11 215 0.9 0 0 0 13.45
152 0 0 0 0.9 8.98 216 0 0 0 0.9 8.98
153 0.44 0 0 0 37.39 217 0.44 0 0 0 10.72
154 0.44 0 0 0 37.39 218 0.44 0 0 0 10.72
155 0.44 0 0 0 37.39 219 0.44 0 0 0 10.72
156 0 0 0 0.44 8.53 220 0 0 0 0.44 8.53
157 0 0 0 0 8.04 221 0 0 0 0 8.04
158 0 0 0 0 8.04 222 0 0 0 0 8.04
159 0 0 0 0 8.04 223 0 0 0 0 8.04
160 0 0 0 0 8.04 224 0 0 0 0 8.04
161 0 1 0 0 37.06 225 1 0 0 0 14.07
162 0 1 0 0 37.06 226 1 0 0 0 14.07
163 0 1 0 0 37.06 227 1 0 0 0 14.07
164 0 0 0 1 9.08 228 0 0 0 1 9.08
165 0 0.9 0 0 31.65 229 0.9 0 0 0 13.45
166 0 0.9 0 0 31.65 230 0.9 0 0 0 13.45
167 0 0.9 0 0 31.65 231 0.9 0 0 0 13.45
168 0 0 0 0.9 8.98 232 0 0 0 0.9 8.98
169 0 0.44 0 0 29.83 233 0.44 0 0 0 10.72
170 0 0.44 0 0 29.83 234 0.44 0 0 0 10.72
171 0 0.44 0 0 29.83 235 0.44 0 0 0 10.72
172 0 0 0 0.44 8.53 236 0 0 0 0.44 8.53
173 0 0 0 0 8.04 237 0 0 0 0 8.04
174 0 0 0 0 8.04 238 0 0 0 0 8.04
175 0 0 0 0 8.04 239 0 0 0 0 8.04
176 0 0 0 0 8.04 240 0 0 0 0 8.04
177 0 1 0 0 12.06 241 0 1 0 0 12.06
178 0 1 0 0 12.06 242 0 1 0 0 12.06
179 0 1 0 0 12.06 243 0 1 0 0 12.06
180 0 0 0 1 9.08 244 0 0 0 1 9.08
181 0 0.9 0 0 11.65 245 0 0.9 0 0 11.65
182 0 0.9 0 0 11.65 246 0 0.9 0 0 11.65
183 0 0.9 0 0 11.65 247 0 0.9 0 0 11.65
184 0 0 0 0.9 8.98 248 0 0 0 0.9 8.98
185 0 0.44 0 0 9.83 249 0 0.44 0 0 9.83
186 0 0.44 0 0 9.83 250 0 0.44 0 0 9.83
187 0 0.44 0 0 9.83 251 0 0.44 0 0 9.83
188 0 0 0 0.44 8.53 252 0 0 0 0.44 8.53
189 0 0 0 0 8.04 253 0 0 0 0 8.04
190 0 0 0 0 8.04 254 0 0 0 0 8.04
191 0 0 0 0 8.04 255 0 0 0 0 8.04
192 0 0 0 0 8.04 256 0 0 0 0 8.04

† Optimal primary supply portfolio: v

1 = v2 = 0, v3 = 1, v4 = 0, optimal expected cost: 8.04, (9.21)
9.5 Computational Examples 263

Table 9.8 Solution results for different disruption start times: C/D = 0.93
g 1 10 102 103 104 105 ∞ (a)
Integrated approach: model Support_E
Var. = 30596, Bin. = 6148, Cons. = 31365, Nonz. = 451299 (e)
Exp.Cost E c , (9.4) 8.37 9.20 15.43 75.73 677 6686 –
Exp.Service E sl × 100%, (9.5) 87 93 93 93 93 93 93
Primary portfolio: supplier(% of total demand) (b) 1(0) 1(0) 1(0) 1(7) 1(8) 1(17) 1(33)
2(0) 2(7) 2(8) 2(7) 2(9) 2(20) 2(31)
3(100) 3(93) 3(92) 3(86) 3(80) 3(51) 3(10)
4(0) 4(0) 4(0) 4(0) 4(3) 4(12) 4(26)
Exp. recovery portfolio: supplier(% of total demand) (c) 1(0.18) 1(0.20) 1(0.19) 1(0.17) 1(0.16) 1(1.13) 1(2.90)
2(1.86) 2(1.81) 2(1.76) 2(1.64) 2(1.55) 2(1.16) 2(7.70)
3(0) 3(0.22) 3(0.29) 3(0.39) 3(0.7) 3(1.84) 3(10.39)
4(7.21) 4(6.67) 4(6.57) 4(6.27) 4(5.82) 4(3.80) 4(5.22)
Hierarchical approach: models PSupport and RSupport_E
PSupport: Var. = 36, Bin. = 4, Cons. = 37, Nonz. = 534
Primary Portfolio: Supplier(% of total demand) (b) 1(0) 1(0) 1(0)
2(0) 2(7) 2(100)
3(100) 3(93) 3(0)
4(0) 4(0) 4(0)
Cost Pc , (21) 8.17 9 15 75 675 6675 –
Service (d) 87 93 93 93 93 93 93
RSupport_E: Var. = 23846, Bin. = 2304, Cons. = 15838, Nonz. = 203756
Exp.Recovery Portfolio: Supplier(% of total demand) (c) 1(0.18) 1(0.20) 1(1.17)
2(1.86) 2(1.81) 2(1.27)
3(0) 3(0.22) 3(1.55)
4(7.21) 4(6.67) 4(0.07)
Exp.Cost E c , (9.4) 8.37 9.20 15.48 78.26 706 6985 –
Exp.Service E sl × 100%, (9.5) 87 93 93 93 93 93 93
(a) Maximization of service level
(b) 1(v 1 × 100), 2(v2 × 100), 3(v3 × 100), 4(v4 × 100)
(c) 1( Ps V1s × 100), 2( s∈S Ps V2s × 100), 3( s∈S Ps V3s × 100), 4( s∈S Ps V4s × 100)
(d)
s∈S
t∈T Xt /D × 100
(e) Var. = number of variables, Bin. = number of binary variables, Cons. = number of constraints,

Nonz. = number of nonzero coefficients

Scenarios with Different Disruption Start Times

In real-life cases, supply disruptions may occur at any time and for any supplier. In this
subsection, disruption scenario is defined as a combination of disruptive event and
its start time. The start time ts of each disruptive event s ∈ S cannot be greater than
the maximum delivery lead time, maxi∈I (τi ) = 4. In the computational examples,
ts ∈ {1, 2, 3}, and the total number of all potential scenarios to be considered is 256
× 3 = 768, i.e., S = {1, . . . , 768}. Now, each disruption scenario s ∈ S is represented
by vector λs = {λ1s , . . . , λ4s }, (see Table 9.3), where λs = λs+256 = λs+512 ; s ≤ 256,
and its start time ts = 1 for s ≤ 256, ts = 2 for 257 ≤ s ≤ 512 and ts = 3 for 513 ≤
s ≤ 768.
264 9 Selection of Primary and Recovery Supply Portfolios and Scheduling

The probability Ps of realizing each disruption scenario s ∈ S is calculated as

follows:
Ps = β1 Ps1 Ps2 for s ≤ 256,
Ps = β2 Ps−256
1 2
Ps−256 for 257 ≤ s ≤ 512,
Ps = β3 Ps−512 Ps−512
1 2
for 513 ≤ s ≤ 768,
where probabilities Psr ; r = 1, 2, s ≤ 256 are defined at the beginning of this
section, and β1 , β2 and β3 are nonnegative constants such that: β1 + β2 + β3 = 1.
Notice that for β1 = 1, Ps reduces to probability of realizing scenario s with a
common disruption start time ts = 1, from Sect. 9.5.1.
As illustrative examples, the optimal solutions for β1 = 0.1, β2 = 0.2, β3 = 0.7,
and capacity-to-demand ratio, C/D = 0.93, are summarized in Tables 9.8, for both
integrated and hierarchical approach. In contrast to examples in Sect. 9.5.1, now the
probabilities for disruptions with later start times are higher.
Comparison of solution results in Tables 9.5 and 9.8 indicate for the two types of
disruption scenarios, nearly identical optimal values of expected cost are achieved for
the integrated approach and slightly worse for the hierarchical approach. However,
the optimal supply portfolios for the corresponding penalty, g, for unfulfilled demand
for products, are not identical (except for the smallest, g = 1). This indicates that for
both types of scenarios the impact of disruption risks can be similarly mitigated by
best selection of primary and recovery portfolios, using the developed approach.

Table 9.9 Worst-case scenarios: C/D = 0.93, g = 100

Disruption scenario Recovery portfolio (V1s , V2s , V3s , V4s ) Cost (a) Service level (b)
(λ1s , λ2s , λ3s , λ4s ), ts
Integrated approach: primary portfolio (v1 , v2 , v3 , v4 ) = (0,0.08,0.92,0)
Hierarchical approach: primary portfolio (v1 , v2 , v3 , v4 ) = (0,0.07,0.93,0)
(0,0,0,0), 1 (0,1,0,0) 2062 50
(0,0,0,0), 2 (0,1,0,0) 2065 47
(0,0,0,0), 3 (0,1,0,0) 2069 43 (c)
(0,0,0,1), 1 (0,0,0,1) 492 50
(0,0,0,1), 2 (0,0,0,1) 496 47
(0,0,0,1), 3 (0,0,0,1) 499 43 (c)
(1,0,0,0), 3 (1,0,0,0) 331 50
(1,0,0,1), 3 (1,0,0,0) 331 50
(0,0,0,2), 3 (0,0,0,1) 102 50
(1,0,0,2), 3 (0,0,0,1) 102 50
(a)

e (u + Ui − qi )/D + i∈I ρis Ui /D + i∈I oi (γi,λis vi + Vi ) + g(1 − t∈T xts /D)
s s s s
(b)
i∈I i s i
t∈T xt /D × 100
(c) SL × 100, (9.20)
9.5 Computational Examples 265

9.5.1.1 Best-Case and Worst-Case Analysis

In order to provide more insights, best-case and worst-case scenarios were identified
such that service level associated with optimal solution is not less than capacity-
to-demand ratio C/D and not greater than some fraction of C/D, respectively. For
example, the total number of best-case scenarios for C/D = 0.93 and unit penalty
g = 100, with service level t∈T xts /D = 0.93 is 298 and 192, respectively for inte-
grated and hierarchical approach. The corresponding cost per product is ranging,
respectively from 15 to 36.69 and from 15 to 20.65. The best-case scenarios for
hierarchical approach are a subset of those for the integrated approach. In contrast,
the number of worst-case scenarios is much lower. Table 9.9 presents all worst-
case scenarios for unit penalty g = 100, with service level t∈T xts /D ≤ 0.5. For
both integrated and hierarchical approach the same set of 10 worst-case scenar-
ios were identified. The table shows disruption scenario, (λ1s , λ2s , λ3s , λ4s ) along
with its start time, ts , and the corresponding optimal recovery supply portfo-
lio, (V s
1 2 3 4s, V s
, V s
, V s
), cost per product,
i∈I ei (ui + Uis − qis )/D
+ s i∈I ρis Ui /D +
s

i∈I oi (γi,λis vi + Vi ) + g(1 − t∈T xt /D), and service level t∈T xt /D. Notice that
s

the lowest service level, SL = 0.43, (9.20), is attained for two scenarios only.
Table 9.9 indicates that for worst-case scenarios all primary suppliers are shutdown
as well as the other suppliers are hit by disruptions. When all suppliers were shutdown,
then a single sourcing recovery portfolio is selected with one of primary suppliers
chosen as recovery supplier. Otherwise, a single recovery supplier is selected from
among those less severely disrupted.
Similar results were obtained for examples with other ratios C/D and unit penalty
g as well as for scenarios with a common disruption start time.
Overall, the main results of computational study for the risk-neutral models are
in line with other research and indicate that:
• for both cost and service level objective function, the integrated decision-making
selects a more diversified primary supply portfolio, that will hedge against all
potential disruption scenarios,
• the primary supply portfolio for the hierarchical approach is made up of cheapest
suppliers or a single, most reliable primary supplier only, to minimize expected
cost or maximize expected service level, respectively,
• a single sourcing recovery supply portfolio is usually selected when all primary
suppliers are shutdown by disruption.
The computational experiments were performed using the AMPL programming
language and the Gurobi 7.0.0 solver on a MacBookPro laptop with Intel Core i7
processor running at 2.8 GHz and with 16GB RAM. The portfolio approach leads to
SMIP formulations with a strong LP relaxation and proves to be computationally very
efficient. The solver was capable of finding proven optimal solution for all examples
with CPU time ranging from fraction of a second for MIP models PSupport and
RSupport(s) to a few seconds for SMIP models Support_E and RSupport_E.
266 9 Selection of Primary and Recovery Supply Portfolios and Scheduling

9.5.2 Risk-Averse Decision-Making

In this subsection some computational examples are presented to illustrate the risk-
averse selection of primary and recovery supply portfolios to minimize CVaR of
cost or maximize CVaR of service level. In the computational experiments data
sets provided at the beginning of this section were used with producer per period
capacity fixed to c = 10000 and hence C/D = 0.93. In addition, for minimization
of CVaR of cost, unit penalty was fixed to g = 100. The solution results are shown
in Tables 9.10 and 9.11, respectively for scenarios with a common disruption start

Table 9.10 Risk-averse solutions for a common disruption start time: C/D = 0.93
Confidence level α 0.50 0.75 0.90 0.95 0.99
Model Support_CV(c), g = 100:
Var. = 10462, Bin. = 2052, Cons. = 10718, Nonz. = 163986 (a)
CV aRc 15.50 15.92 16.93 18.03 22.37
V aRc 15.05 15.09 15.57 16.09 18.82
Ec 15.28 15.30 15.71 15.65 17.70
E sl 100% 93.29 93.31 92.92 93.13 93.21
Primary Portfolio: Supplier(% of total demand) (b) 1(2) 1(3) 1(3) 1(3) 1(4)
2(5) 2(4) 2(5) 2(7) 2(64)
3(93) 3(93) 3(92) 3(90) 3(31)
4(0) 4(0) 4(0) 4(0) 4(0)
Exp.Recovery Portfolio: Supplier(% of total demand) (c) 1(0.34) 1(0.39) 1(0.31) 1(0.76) 1(0.29)
2(1.81) 2(1.85) 2(1.78) 2(1.82) 2(2.34)
3(0.03) 3(0.02) 3(0.07) 3(0.14) 3(1.67)
4(6.69) 4(6.69) 4(6.64) 4(5.93) 4(1.39)
Model Support_CV(sl):
Var. = 10462, Bin. = 2052, Cons. = 10718, Nonz. = 159122 (a)
CV aRsl 100% 93.32 93.30 93.25 93.17 92.53
V aRsl 100% 93.33
E sl 100% 93.32
Ec 19.04 19.06 19.06 19.02 19.06
Primary Portfolio: Supplier(% of total demand) (b) 1(27)
2(25)
3(7)
4(41)
Exp.Recovery Portfolio: Supplier(% of total demand) (c) 1(1.81) 1(1.14) 1(1.14) 1(1.80) 1(1.52)
2(0.86) 2(1.55) 2(1.55) 2(0.86) 2(1.16)
3(2.00) 3(2.41) 3(2.41) 3(2.54) 3(2.25)
4(1.07) 4(0.64) 4(0.64) 4(0.53) 4(0.81)
(a) Var. = number of variables, Bin. = number of binary variables, Cons. = number of constraints,
Nonz. = number of nonzero coefficients
(b) 1(v × 100), 2(v × 100), 3(v × 100), 4(v × 100)
(c) 1(

1 2 3 4
s∈S Ps V1 × 100), 2( s∈S Ps V2 × 100), 3( s∈S Ps V3 × 100), 4( s∈S Ps V4 × 100)
s s s s
9.5 Computational Examples 267

Table 9.11 Risk-averse solutions for different disruption start times: C/D = 0.93
Confidence level α 0.50 0.75 0.90 0.95 0.99
Model Support_CV(c), g = 100:
Var. = 31365, Bin. = 6148, Cons. = 32133, Nonz. = 488799 (a)
CV aRc 15.68 16.11 17.07 18.13 22.49
V aRc 15.25 15.28 15.81 16.22 18.90
Ec 15.47 15.49 15.64 15.80 17.43
E sl 100% 93.28 93.29 93.29 93.22 93.08
Primary Portfolio: Supplier(% of total demand) (b) 1(5) 1(4) 1(6) 1(6) 1(8)
2(5) 2(6) 2(6) 2(7) 2(46)
3(90) 3(90) 3(83) 3(80) 3(46)
4(0) 4(0) 4(5) 4(7) 4(0)
Exp.Recovery Portfolio: Supplier(% of total demand) (c) 1(0.21) 1(0.18) 1(0.31) 1(0.45) 1(0.62)
2(1.76) 2(1.71) 2(1.80) 2(2.45) 2(1.19)
3(0.23) 3(0.13) 3(0.26) 3(0.38) 3(1.09)
4(6.46) 4(6.60) 4(6.12) 4(5.13) 4(2.44)
Model Support_CV(sl):
Var. = 31365, Bin. = 6148, Cons. = 32133, Nonz. = 474207 (a)
CV aRsl 100% 93.32 93.30 93.25 93.17 92.50
V aRsl 100% 93.33
E sl 100% 93.32
Ec 20.52 20.48 20.50 20.45 20.47
Primary portfolio: supplier(% of total demand) (b) 1(33)
2(31)
3(10)
4(26)
Exp. recovery portfolio: supplier(% of total demand) (c) 1(2.23) 1(2.35) 1(2.26) 1(1.79) 1(2.31)
2(0.89) 2(0.81) 2(0.75) 2(0.59) 2(0.54)
3(1.27) 3(1.16) 3(1.30) 3(1.77) 3(0.97)
4(0.84) 4(0.90) 4(0.92) 4(1.07) 4(1.41)
(a) Var. = number of variables, Bin. = number of binary variables, Cons. = number of constraints,
Nonz. = number of nonzero coefficients
(b) 1(v × 100), 2(v × 100), 3(v × 100), 4(v × 100)
(c) 1(

1 2 3 4
s∈S Ps V1 × 100), 2( s∈S Ps V2 × 100), 3( s∈S Ps V3 × 100), 4( s∈S Ps V4 × 100)
s s s s

time and with different disruption start times. For the risk-averse solutions, VaR,
CVaR and the associated expected values of cost and service level are presented for a
subset of confidence levels α = 0.5, 0.75, 0.9, 0.95, 0.99. The risk-averse portfolio
selected using model Support_CV(sl) is more diversified than the risk-averse port-
folio determined by model Support_CV(c). As α increases, CV aRc of cost increases
and CV aRsl of service level decreases. At the same, a greater diversification of the
primary supply portfolio is observed for model SupportCV(c), with more demand
shifted from the cheapest supplier 3 to more reliable supplier 2. For model Sup-
268 9 Selection of Primary and Recovery Supply Portfolios and Scheduling

port_CV(sl), however, V aRsl , E sl and primary supply portfolio are independent of

α. CV aRsl and E sl are very close to V aRsl , while the latter is equal to C/D = 0.93.
Comparison of solution results in Tables 9.10 and 9.11, for scenarios with a common
and different disruption start times indicate that while the latter scenarios lead to
higher values of cost, the corresponding values of service level are nearly identical.
Comparison of solution results for models Support_E and Support_CV(c), Sup-
port_CV(sl) (cf. Tables 9.5, 9.10 and 9.8, 9.11), demonstrates that the risk-neutral
solutions for minimization of expected cost are very close to the corresponding
risk-averse solutions with confidence level α = 0.5. Moreover, for the service level
objective, the corresponding risk-neutral and the risk-averse primary supply portfo-
lios are identical.
All the above results indicate that for the service level objective and the example
problem parameters, the impact of the worst-case disruption scenarios can be fully
mitigated by the optimal selection of primary and recovery supply portfolios.
The computational experiments were performed using the AMPL programming
language and the Gurobi 7.0.0 solver on a MacBookPro laptop with Intel Core i7
processor running at 2.8 GHz and with 16GB RAM. Proven optimal solutions were
found for all examples with CPU time ranging from fraction of a second for model
Support_CV(sl) to a few seconds for model Support_CV(c).

9.6 Notes

The low-probability and high-impact disruptions of material flows in global supply

chains and the resulting losses may threaten financial state of firms. For example,
the disruptive events that occurred in 2011 in the automotive and electronics supply
chains (the Great East Japan earthquake and tsunami in March and Thailand’s floods
in October) resulted in huge losses of major automakers and electronics manufactur-
ers, e.g., Park et al. (2013), Haraguchi and Lall (2015). In order to minimize losses
caused by the shortage of material supplies, customer companies (firms) apply differ-
ent disruption management strategies, such as maintaining inventory, buying from an
alternate supplier or helping a primary supplier recover more quickly. When the latter
strategy is applied, the firm participates in supplier’s recovery process after disrup-
tion to reduce recovery time, which means that the firm may participate in suppliers
cost-to-recover. The real-world examples presented in Sect. 1.1 illustrate a typical
disruption management strategy. Whenever a primary supplier is hit by disruption,
the customer company needs to choose whether to support recovery of disrupted pri-
mary supplier or select an alternate (recovery) supplier, non-disrupted or disrupted
less severely than the primary supplier. The recovery suppliers are selected in such
a way that the recovery process is optimized with respect to recovery time and cost.
However, the literature on mitigation the impact of disruption risks and optimization
of a recovery process in supply chains is limited (e.g., Tomlin 2006; Ruiz-Torres
et al. 2013; MacKenzie et al. 2014; Zeng and Xia 2015; Hamdi et al. 2015; Ivanov
et al. 2016). A recent literature review on Operations Research/Management Science
9.6 Notes 269

models for supply chain disruptions was presented by Snyder et al. (2016). They dis-
cussed 180 scholarly works on the topic, organized into six categories: evaluating
supply disruptions; strategic decisions; sourcing decisions; contracts and incentives;
inventory; and facility location. Paul et al. (2016a) reviewed literature on manag-
ing risk and disruption in production-inventory systems and supply chains. They
considered four categories of disruptions: disruption in production, disruption in
supply, disruption in transportation, and fluctuation in demand. The authors focused
on reviewing the mathematical models and the solution approaches used in solving
the models using both hypothetical and real-world problem scenarios.
In practice, the probability and magnitude of low-probability and high-impact
disruptive events is difficult and often impossible to be estimated. The classical risk
assessment and mitigation methods that require such knowledge at an early stage
of the risk analysis may fail to prepare supply chains for such disruptive events. A
novel risk-exposure model for analyzing operational-disruption risk with no need to
estimate the probability of any specific disruptive events, was proposed by Simchi-
Levi et al. (2015). The model is capable of assessing the impact of a disruption
originating anywhere in a supply chain. The model has been applied by Ford Motor
Company to identify risk exposures, evaluate risk mitigation actions, and develop
optimal contingency plans in the automotive supply chain.
In another stream of research Paul et al. (2014a, b, 2015a, b) developed mathe-
matical models and solution algorithms for disruption management in production–
inventory systems, under single or multiple disruptions. The authors considered back
orders, lost sales and/or outsourcing options as recovery strategies.
The material presented in this chapter is based on results achieved by Sawik
(2017), where a computationally efficient portfolio approach developed in Sawik
(2011a, b, 2013a, b, c, d) was enhanced for simultaneous or sequential selection of
primary and recovery supply portfolios under local and regional disruptions risks.
However, in Sawik (2017) only a risk-neutral, expected cost (or expected service
level) objective has been considered to optimize an overall performance of a supply
chain. Minimizing expected cost (or maximizing expected service level) under dis-
ruption risks may sometimes be impractical in the long run, especially when large
losses could threaten financial state of firms. Then, a downside risk measure such
as expected worst-case cost, may be more appropriate. In this chapter, the portfolio
approach and the risk-neutral models developed by Sawik (2017) have been enhanced
to assess the effect of risk-averse decision making on the selection of primary and
recovery supply portfolios, using CVaR as a risk measure.
In the literature on supply uncertainty, the supply is either subject to complete
disruptions or yield uncertainty. Yield uncertainty occurs when the quantity of supply
delivered is a random variable, modeled as either a random additive or multiplicative
quantity, whereas disruptions occur when supply is subject to partial or complete
failure. Typically disruptions are modeled as events which occur randomly and may
have a random length. Schmitt and Snyder (2012) considered inventory systems
subject to both supply disruptions and yield uncertainty. They compared single-
period versus multi-period models and showed that the former can lead to selecting
the wrong strategy for mitigating supply risk.
270 9 Selection of Primary and Recovery Supply Portfolios and Scheduling

The future research should concentrate on relaxations of the various simplified

assumptions used to formulate the problem. For example, the assumption that each
supplier has sufficient capacity to meet total demand for parts that allows for a
single sourcing, can be easily relaxed. The developed models can be enhanced for
finite capacity suppliers. A single recovery mode defined by constant recovery time,
TTR(i, l) and the associated recovery cost, CTR(i, l) for each supplier i and each
disruption level l, can be replaced by a number of available recovery modes, each
represented by different value of time, TTR(i, l) and the associated cost, CTR(i, l).
Then, a new recovery mode selection variable can enhance the problem formulation.
In a more general setting, both disruption start time, recovery time and recovery
cost can be modeled as random parameters, e.g., Schmitt (2011), Schmitt and Singh
(2012), Paul et al. (2016c).

Problems

9.1 In the SMIP models presented in this chapter introduce delay penalty when
demand dt for period t is not met by that period.

9.2 In the SMIP models presented in this chapter introduce transition time required
for switching to recovery supplier different from the primary supplier.

9.3 Enhance the SMIP models presented in this chapter for the limited output inven-
tory of products at the producer. Formulate the inventory balance constraints that
should be added to the models.

9.4 Formulate the mean-risk models Support_ECV(c) and Support_ECV(sl) to

minimize expected cost and CVaR of cost or maximize expected service level and
CVaR of service level, respectively.

9.5 Enhance model Support_E for multiple recovery modes defined by different
recovery time, TTR(i, l) and the associated recovery cost, CTR(i, l) for each supplier
i and each disruption level l.
Chapter 10
Selection of Primary and Recovery Supply
and Demand Portfolios and Scheduling

10.1 Introduction

In this chapter the portfolio approach proposed in Chap. 9 for the selection of pri-
mary and recovery suppliers and order quantity allocation to mitigate the impact
of disruption risks is enhanced also for the recovery process of the firm’s assembly
plants for finished products. Unlike most of reported research on supply chain dis-
ruption management a disruptive event is assumed to impact both a primary supplier
of parts and the buyer’s firm primary assembly plant. Then the firm may choose
alternate (recovery) suppliers and move production to alternate (recovery) plants
along with transshipment of parts from the impacted primary plant to the recov-
ery plants. The resulting allocation of unfulfilled demand for parts among recovery
suppliers and unfulfilled demand for products among recovery assembly plants deter-
mines recovery supply and demand portfolio, respectively. The enhanced portfolio
approach and SMIP formulations with an embedded network flow problem devel-
oped in this chapter are capable of selecting primary suppliers, the decision to be
implemented before a disruption and of selecting recovery suppliers and recovery
assembly plants, the decision to be implemented during and after the disruption. The
supply and demand portfolios are determined along with production scheduling in
assembly plants. Multi-level disruptions of suppliers and assembly plants will be
considered. The objective of selection primary and recovery suppliers and assem-
bly plants and allocation of order quantity for parts and demand for products is to
mitigate the impact of disruption risks and optimize the recovery process. The two
decision making approaches will be considered: an integrated approach with the per-
fect information about the future disruption scenarios, and a hierarchical approach
with no such information available ahead of time. In the integrated approach, which
accounts for all potential disruption scenarios, the primary supply portfolio that will
hedge against all scenarios is determined along with the recovery supply and demand
portfolios and production schedule of finished products for each scenario, to mini-
mize expected cost (or CVaR of cost) or maximize expected service level (or CVaR

© Springer International Publishing AG 2018 271

T. Sawik, Supply Chain Disruption Management Using Stochastic Mixed
Integer Programming, International Series in Operations Research
& Management Science 256, DOI 10.1007/978-3-319-58823-0_10
272 10 Selection of Primary and Recovery Supply …

of service level) over all scenarios. In the hierarchical approach first the primary
supply portfolio is selected, and then, when a primary supplier or primary assembly
plant is hit by a disruption, the recovery supply and demand portfolios are selected
to optimize the process of recovery from the disruption.
The following time-indexed SMIP and MIP models are presented in this chapter:

DSupport_E for risk-neutral selection of primary and recovery supply and

demand portfolios and production scheduling to minimize expected cost;
DSupportMP_E model DSupport_E for multiple part types and product
types;
PSupport for selection of primary supply portfolio and production scheduling
to minimize cost under deterministic conditions;
RDSupport(s) for selection of recovery supply and demand portfolios and
production scheduling to minimize cost, for predetermined primary port-
folios and the realized disruption scenario;
RDSupport_E model DSupport_E for predetermined primary portfolios.
DSupport_CV(c) for risk-averse selection of primary and recovery supply
and demand portfolios and production scheduling to minimize CVaR of
cost;
DSupport_CV(sl) for risk-averse selection of primary and recovery supply
and demand portfolios and production scheduling to maximize CVaR of
service level;
RDSupport_CV(c) model DSupport_CV(c) for predetermined primary port-
folios;
RDSupport_CV(sl) model DSupport_CV(sl) for predetermined primary
portfolios.

Numerical examples, computational results, best-case and worst-case analysis,

and some comparison of the two approaches for the selection of primary and recovery
supply and demand portfolios are provided in Sects. 10.5.1 and 10.5.2, respectively
for risk-neutral and risk-averse decision-making.

10.2 Problem Description

In this section the problem of simultaneous selection of primary and recovery supply
and demand portfolios is considered.
Consider a supply chain in which a single producer of one product type, assembles
products in several assembly plants to meet customer demand, using a critical part
type that can be manufactured and provided by several suppliers (for notation used,
see Table 10.1).
Let I = {1, . . . , I} be the set of I suppliers, J = {1, . . . , J} be the set of J
assembly plants, T = {1, . . . , T } the set of T planning periods, and denote by D the
10.2 Problem Description 273

Table 10.1 Notation: selection of primary and recovery supply and demand portfolios and
scheduling
Indices
i = supplier, i ∈ I
j = assembly plant, j ∈ J
l = disruption level, l ∈ Li , i ∈ I, l ∈ L j , j ∈ J
r = region, r ∈ R
s = disruption scenario, s ∈ S
t = planning period, t ∈ T
Input Parameters
cj = per period capacity of plant j
D = total demand for products
εj = additional per unit production cost at recovery plant j > 1
ei = fixed ordering cost for supplier i
ϕj = fixed production setup cost at recovery plant j > 1
ψj = per unit transshipment cost from primary plant j = 1 to plant j > 1
g = per unit penalty cost for unfulfilled demand for products
oi = per unit price of parts purchased from supplier i
pil = probability of disruption level l for supplier i
πjl = probability of disruption level l for plant j
pr = regional disruption probability for region r
ts = start time period of disruptive event s
γil = fraction of an order delivered by supplier i under disruption level l (supplier
fulfillment rate)
δjl = fraction of capacity of plant j available under disruption level l
τij = delivery lead time from supplier i to plant j
σj = transshipment time from primary plant j = 1 to plant j > 1
θis = time-to-recover of supplier i from disruption under scenario s
ϑjs = time-to-recover of plant j from disruption under scenario s
ρis = cost-to-recover of supplier i from disruption under scenario s
ρjs = cost-to-recover of plant j from disruption under scenario s

total demand for products. Let j = 1 be the primary plant, where the total demand
for products, D, is initially assigned.
The suppliers of parts and assembly plants are located in R geographic regions,
subject to potential regional disasters that may result in complete shutdown of all
suppliers and plants in the same region simultaneously. Denote by I r and J r , respec-
tively the subsets of suppliers and plants in region r ∈ R, and by pr , the regional
disruption probability for region r.
In addition to correlated regional disruptions, each supplier i ∈ I is subject to
independent local disruptions of different levels, l ∈ Li = {0, . . . , Li }, where dis-
ruption level refers to the fraction of an order that can be delivered, see, Sect. 9.2.
274 10 Selection of Primary and Recovery Supply …

Level l = 0 represents complete shutdown of a supplier, i.e., no order delivery, while

level l = Li represents normal conditions with no disruption, i.e., full order delivery.
Denote by pil , the probability of disruption level l for supplier i, and by γil , the frac-
tion of an order that can be delivered by supplier i under disruption level l (supplier
fulfillment rate),

⎧
⎨= 0 if l = 0
γil ∈ (0, 1) if l = 1, . . . , Li − 1 (10.1)
⎩
=1 if l = Li .

Similarly to suppliers, each plant j ∈ J is subject to random local disruptions

of different levels, l ∈ L j = {0, . . . , L j }, where disruption level refers to available
fraction of full capacity, cj , available per period under normal conditions. Level l = 0
represents complete shutdown of an assembly plant, while level l = L j represents
normal conditions, i.e., full capacity, cj , available. Denote by πjl , the probability of
disruption level l for plant j, and by δjl , the fraction of available capacity of plant j
under disruption level l.
⎧
⎨= 0 if l = 0
δjl ∈ (0, 1) if l = 1, . . . , L j − 1 (10.2)
⎩
=1 if l = L j .
 
The total number of all potential scenarios is S = i∈I (Li + 1) j∈J
(L j + 1). Each scenario s ∈ S is represented by an (I + J)- dimensional vector
λs = {λ1s , . . . , λIs , λI+1,s , . . . , λI+J,s }, where λis ∈ Li is the disruption level of
supplier i ∈ I and λI+j,s ∈ L j is the disruption level of plant j ∈ J, under scenario
s ∈ S. 
The probability Ps of disruption scenario s ∈ S is Ps = r∈R Psr , where Psr is the
probability of realizing disruption scenario s in region r
⎧    
⎪ (1 − pr )( i∈I r l∈Li :λis =l pil )( j∈J r l∈Lj :λI+j,s =l πjl ),
⎪
⎨  
Psr =
if i∈I r λis + j∈J r λI+j,s 
>0
(10.3)
⎪
⎪ p r
+ (1 − p r
)( p )( j∈J r πj0 ),
⎩  i∈I r i0
if i∈I r λis + j∈J r λI+j,s = 0.

When supplier i is hit by disruption at level l, its recovery process to normal

conditions takes TTR(i, l) time periods (Time-To-Recover) and let CTR(i, l) be the
firm’s portion of Cost-To-Recover. For each supplier i, denote by θis and ρis , respec-
tively time-to-recover and firm’s portion of cost-to recover from disruption under
scenario s

θis = TTR(i, l); i ∈ I, s ∈ S : l = λis (10.4)

ρis = CTR(i, l); i ∈ I, s ∈ S : l = λis . (10.5)
10.2 Problem Description 275

Similarly, when plant j is hit by disruption at level l, its recovery process to normal
conditions takes PRT (j, l) time periods (Plant Recovery Time) and cost PRC(j, l)
(Plant Recovery Cost). For each plant j, denote by ϑjs and ρjs , respectively time-to-
recover and cost-to-recover from disruption under scenario s

ϑjs = PRT (j, l); j ∈ J, s ∈ S : l = λI+j,s (10.6)

ρjs = PRC(j, l); j ∈ J, s ∈ S : l = λI+j,s . (10.7)

The orders for parts are assumed to be placed at the beginning of the planning
horizon, and under normal conditions the parts ordered from supplier i are delivered
to assembly plant j in period τij , where τij is the total of manufacturing lead time and
transportation time. Denote by σj the transshipment time from primary plant j = 1
to plant j.
The firm who moves production to an alternate assembly plant j ∈ J incurs a fixed
cost ϕj and encounters additional per unit cost of production εj , and per unit cost, ψj ,
of transshipment of parts from the primary plant, where ϕ1 = 0, ε1 = 0 and ψ1 = 0.
A recovery plant can be a disrupted primary plant j = 1 with reduced capacity
during recovery process and then with its full capacity or a new plant, non-disrupted
or disrupted less severely than the primary plant.
The following assumptions are made to formulate the problem.
• Each supplier has sufficient capacity to meet total demand for parts.
• A single disruptive event is assumed to occur over the entire planning horizon.
Multiple disruptions, one after the other in a series, during the recovery process
are not considered.
• Partial recovery of a disrupted supplier to its partial capacity as well as partial
recovery of a disrupted assembly plant to its partial capacity are not considered.
• Time-to recover and the associated cost-to-recover are constant parameters that
represent recovery of a disrupted supplier or a disrupted assembly plant to its full
capacity.
• A single recovery mode is considered for each supplier, each assembly plant and
each disruption level.
• Disruption to primary supplier (primary assembly plant) under scenario s ∈ S
occurs in period ts and recovery process starts in period ts + 1, so that the disrupted
supplier i ∈ I (disrupted plant j ∈ J) returns to its full capacity in period t = ts +θis
(t = ts + ϑjs ).
• A recovery supplier can be a disrupted primary supplier after its recovery to full
capacity or a new supplier.
• A recovery assembly plant can be a disrupted primary assembly plant during and
after its recovery to full capacity or a new assembly plant.
• Transshipment of parts from the impacted primary assembly plant to recovery
plants starts along with the primary plant recovery process.
• Transition time required for switching to recovery supplier different from the
primary suppliers is negligible.
• The buyer firm participates in supplier’s cost-to-recover.
276 10 Selection of Primary and Recovery Supply …

Table 10.2 Variables: selection of primary and recovery supply and demand portfolios and
scheduling
First stage variables
ui = 1, if supplier i is selected as a primary supplier; otherwise ui = 0 (primary
supplier selection)
vi ∈ [0, 1], the fraction of total demand for parts ordered from primary supplier i to
primary plant j = 1 (primary supply portfolio)
Second stage variables
Uis = 1, if supplier i is selected as a recovery supplier under disruption scenario s;
otherwise Uis = 0 (recovery supplier selection)
Vijs ∈ [0, 1], the fraction of total demand for parts ordered from recovery supplier i to
recovery plant j, under disruption scenario s (recovery supply portfolio)
wjs ∈ [0, 1], the fraction of total demand for parts, transshipped from the primary
plant j = 1 to recovery plant j, under disruption scenario s, where w1s represents
the parts that remain in the primary plant j = 1 (transshipment variables)
xjts ≥ 0, production in plant j in period t under disruption scenario s (production scheduling)
yjs = 1, if assembly plant j is selected as a recovery plant under disruption scenario s;
otherwise yjs = 0 (recovery plant selection)
s
zj ∈ [0, 1] the fraction of total demand for products to be completed by recovery plant
j under disruption scenario s (recovery demand portfolio).
Auxiliary variables
qis = 1, if ui = Uis = 1; otherwise qis = 0 (elimination of double fixed ordering costs)
VaRc Cost-at-Risk, the targeted cost such that for a given confidence level α,
for 100α% of the scenarios, the outcome is below VaRc
VaRsl Service-at-Risk, the targeted service level such that for a given confidence level α,
for 100α% of the scenarios, the outcome is above VaRsl
Cs ≥ 0, the tail cost for scenario s, i.e., the amount by which costs in scenario s exceed
VaRc
Ss ≥ 0, the tail service level for scenario s, i.e., the amount by which VaRsl exceeds
service level in scenario s

• A penalty cost is charged for the demand for products unfulfilled by the end of the
planning horizon.

10.3 Models for Risk-Neutral Decision-Making

10.3.1 Integrated Selection of Primary and Recovery Supply

and Demand Portfolios

In this subsection a SMIP model DSupport_E is presented for the integrated selection
of supply and demand portfolios under disruption risks.
The following primary and recovery portfolios are jointly determined using the
proposed model (for definitions of first and second stage variables, see Table 10.2):
10.3 Models for Risk-Neutral Decision-Making 277

• Primary supply portfolio: vi ∈ [0, 1] - the fraction of total demand for parts ordered
from primary supplier i to be delivered to primary plant j = 1.
• Recovery supply portfolio: Vijs ∈ [0, 1] - the fraction of total demand for parts
ordered from recovery supplier i to recovery assembly plant j, under disruption
scenario s.
• Recovery demand portfolio: zjs ∈ [0, 1] is the fraction of total demand for products
to be completed by recovery plant j under disruption scenario s.
Since the total demand for products has been assigned to the single primary assembly
plant j = 1 and all parts from the primary suppliers are ordered for that plant, the
primary demand portfolio is not considered.
The primary supply portfolio, (v1 , . . . , vI ), which determines supply of parts to the
primary assembly plant j = 1 is selected along with the recovery supply portfolio for
each disruption scenario. The recovery supply portfolio for scenario s is determined
by supplies of  recovery suppliers, i ∈ I, to recovery assembly plants,
parts, DVijs , from
j ∈ J, where i∈I (γis vi + j∈J Vijs ) = 1 and 0 ≤ Vijs ≤ 1, i ∈ I, j ∈ J. The selection
of recovery supply and demand portfolios may be combined with transshipment of
parts from the primary plant to recovery plants.
Let E c be the expected cost per product to be minimized, and E sl , the expected
service level, i.e., the expected fraction of the total fulfilled demand for products (i.e.,
the expected demand fulfillment rate) to be maximized.
  
Ec = Ps (ei (ui + Uis − qis )/D + ρis Uis /D + oi (γis vi + Vijs ))
s∈S i∈I j∈J
  
+ Ps (ψj wjs + (ϕj + ρjs )yjs /D + εj xjts /D) + g(1 − E sl ), (10.8)
s∈S j∈J t∈T


E sl = Ps xjts /D. (10.9)
s∈S j∈J t∈T

The auxiliary variable, qis , is introduced to eliminate double charging with fixed
ordering cost ei of each supplier i, who is selected both as primary and recovery
supplier: qis = 1, if ui = Uis = 1; otherwise qis = 0.
The expected cost per product, E c , (10.8), constitutes of different fixed and
variable cost per product. The fixed cost per product includes expected cost: of
ordering parts fromprimary suppliers  and  from recoverys suppliers different from
primary suppliers, i∈I ei ui /D + s∈S i∈I Ps ei (Uis − q i )/D, 
of recovery process
for impacted suppliers selected as recovery suppliers, s∈S i∈I Ps ρis Uis /D, of
moving
  production from primary assembly plant to recovery assembly plants,
s∈S j∈J P s ϕj y s
/D,
j and  of recovery process for impacted assembly plants selected
as recovery plants, s∈S j∈J Ps ρjs yjs /D. The variable cost per product includes
cost:of purchasing
 parts for partially fulfilled deliveries from primary  suppli-

ers,
 s∈S i∈I P s o i γ i
s
vi , of purchasing parts from recovery suppliers, s∈S i∈I
s
j∈J P s oi Vij , of transshipment of parts from primary plant to recovery plants,
j∈J Ps ψj wj , of additional production cost in recovery plants different from
s
s∈S
278 10 Selection of Primary and Recovery Supply …

  
t∈T εj xjt /D, and of penalty for unfulfilled demand,
s
primary
plant,
  s∈S j∈J
g(1 − s∈S j∈J t∈T Ps xjt /D).
s

DSupport_E: Risk-neutral selection of primary and recovery Demand and

Supply portfolios and production scheduling
Minimize (10.8)
subject to
Primary supply portfolio selection constraints
- the total demand for parts must be fully allocated among the selected
primary suppliers (i.e., the primary supply portfolio must be selected),
- demand for parts cannot be assigned to non-selected primary suppliers,

vi = 1 (10.10)
i∈I
vi ≤ ui ; i ∈ I (10.11)

Recovery supply and demand portfolio selection constraints

- the unfulfilled demand for parts must be fully allocated among the selected
recovery suppliers (i.e., the recovery supply portfolio must be selected),
- the unfulfilled demand for parts cannot be assigned to non-selected recov-
ery suppliers,
- the unfulfilled demand for products must be fully allocated among the
selected plants (i.e., the recovery demand portfolio must be selected),
- the unfulfilled demand for products cannot be assigned to non-selected
recovery plants,
- the supply and demand portfolio flow conservation constraints ensure that
for each recovery plant, the recovery supplies and transshipment of parts are
in balance with the demand for products to be fulfilled by that plant,
- the balance constraints for parts ensure that the total transshipment of
parts from primary plant is equal to the initial inventory of parts and partially
fulfilled supplies from primary suppliers less the usage of parts for production
before a disruptive event,
- each supplier selected to both primary and recovery portfolio is charged
exactly once with fixed ordering cost in the objective function,
 
(γis vi + Vijs ) = 1 − V0 ; s ∈ S (10.12)
i∈I j∈J
Vijs ≤ Uis ; i ∈ I, j ∈ J, s ∈ S (10.13)
 
s
x1t /D + zjs = 1; s ∈ S (10.14)
t∈T :t<ts j∈J
zjs ≤ yjs ; j ∈ J, s ∈ S (10.15)
10.3 Models for Risk-Neutral Decision-Making 279

Vijs + wjs = zjs ; j ∈ J, s ∈ S (10.16)
i∈I
  
wjs = V0 + γis vi − s
x1t /D; s ∈ S (10.17)
j∈J i∈I t∈T :t<ts
qis ≤ (ui + Uis )/2; i ∈ I, s ∈ S, (10.18)

where DV0 is the initial inventory of parts at primary assembly plant j = 1.

Production capacity constraints
- before a disruptive event, the production at primary plant in every period
cannot exceed its capacity,
- after a disruptive event, the production at each selected recovery plant in
every period cannot exceed the plant available capacity,
- the total production at each recovery plant cannot exceed the assigned
portion of total demand for products,
s
x1t ≤ c1 ; t ∈ T , s ∈ S : t < ts (10.19)
xjts ≤ cjts yjs ;
j ∈ J, t ∈ T , s ∈ S : t ≥ ts (10.20)

xjts /D ≤ zjs ; j ∈ J, s ∈ S (10.21)
t∈T :t≥ts

Supply-transshipment-production coordinating constraints

- for each disruption scenario s and each period t, the cumulative demand
for parts of production in primary assembly plant j = 1, scheduled in periods
1 through t cannot exceed the initial inventory of parts and the cumulative
deliveries by period t − 1 (delivery in period τi1 ≤ t − 1 from each primary
supplier i and in period ts + θis + τi1 ≤ t − 1 from each recovery supplier i),
less transhipment of parts to recovery assembly plants j > 1,
- for each disruption scenario s and each period t ≤ σj , the cumulative
demand for parts of production scheduled in periods 1 through t in recovery
assembly plant j > 1, cannot exceed the cumulative deliveries by period t − 1
(delivery in period ts + θis + τij ≤ t − 1 from each recovery supplier i, that can
be used for production in period ts + θis + τij + 1, at the earliest),
- for each disruption scenario s and each period t ≥ σj + 1, the cumulative
demand for parts of production scheduled in periods 1 through t in recovery
assembly plant j > 1, cannot exceed the cumulative deliveries by period t − 1
(delivery in period ts + θis + τij ≤ t − 1 from each recovery supplier i, that
can be used for production in period ts + θis + τij + 1, at the earliest), plus
transhipment of parts from primary assembly plant j = 1,

 /D ≤
s
x1t
t  ∈T :t  ≤t
  
V0 + γis vi + Vijs − wjs ;
i∈I:τi1 ≤t−1 i∈I:ts +θis +τi1 ≤t−1 j∈J:j>1
280 10 Selection of Primary and Recovery Supply …

t ∈ T, s ∈ S (10.22)

xjts  /D ≤
t  ∈T :t  ≤t

Vijs ;
i∈I:ts +θis +τij ≤t−1

j ∈ J, t ∈ T , s ∈ S : j > 1, t < σj + 1 (10.23)


xjts  /D ≤
t  ∈T :t  ≤t

Vijs + wjs ;
i∈I:ts +θis +τij ≤t−1
j ∈ J, t ∈ T , s ∈ S : j > 1, t ≥ σj + 1 (10.24)

Non-negativity and integrality conditions

qis ∈ {0, 1}; i ∈ I, s ∈ S (10.25)

ui ∈ {0, 1}; i ∈ I (10.26)
vi ∈ [0, 1]; i ∈ I (10.27)
Uis ∈ {0, 1}; i ∈ I, s ∈ S (10.28)
Vijs ∈ [0, 1]; i ∈ I, j ∈ J, s ∈ S (10.29)
wjs ∈ [0, 1]; j ∈ J, s ∈ S (10.30)
xjts ≥ 0; j ∈ J, t ∈ T , s ∈ S (10.31)
yjs ∈ {0, 1}; j ∈ J, s ∈ S (10.32)
zjs ∈ [0, 1]; j ∈ J, s ∈ S, (10.33)

where cjts is the capacity in period t of plant j under disruption scenario s

δjs cj , if ts ≤ t ≤ ts + ϑjs − 1
cjts = (10.34)
cj , if t ≤ ts − 1, t ≥ ts + ϑjs ,

where δjs = δj,λI+j,s is the fraction of capacity available at assembly plant j

under scenario s, and λI+j,s is disruption level of plant j under scenario s.

The supply and demand portfolio selection constraints form an embedded net-
work flow problem. In particular, Eqs. (10.16) and (10.17) are flow conservation
constraints for each node j (assembly plant) and node j = 1 (primary assembly
plant), respectively. Notice that vi represents flow of parts from supplier i (source
node) to primary plant j = 1 (sink/transshipment node) and wj represents flow of
parts from primary plant j = 1 (transshipment node) to plant j (sink node).
10.3 Models for Risk-Neutral Decision-Making 281

Model DSupport_E is a deterministic equivalent mixed integer program of

a two-stage stochastic mixed integer program with recourse and illustrates the
wait-and-see approach. The primary supply portfolio selection variables, ui , vi , are
referred to as first-stage decisions, and the recovery supply portfolio and recovery
demand portfolio selection variables, Uis , Vijs and yjs , zjs , as well as transshipment
and production scheduling variables, wjs and xjts , are referred to as recourse or second-
stage decisions (see, Table 10.2). Unlike the first-stage decisions, the latter variables
are dependent on disruption scenario s ∈ S.
Worst-case service level

Proposition 10.1
If there is no initial inventory of parts (i.e., V0 = 0), the lowest service level, SL,
can be calculated as below.

SL = max{cj [T − max{ts + max(min{θis + τij }, ϑjs , σj )]}/D. (10.35)

j∈J s∈S i∈I

Proof.
The lowest service level is associated with worst-case disruption scenario s ∈ S
for which primary supplier and primary assembly plant are both hit by disruption at
time ts and then the plant j with maximum capacity available after its full recovery is
selected as a recovery plant. The maximum available capacity of the recovery plant
is based on the number of periods remaining for production after its full recovery
(after ts + ϑjs periods), after the earliest delivery of parts by a recovery supplier (i.e.,
after ts + mini∈I {θis + τij } periods) and after transshippment of parts from the primary
plant (i.e., after ts + σj periods), whichever occurs later.
In order to strengthen MIP model DSupport_E the following constraint can be
added   
SL ≤ s
x1t /D + xjts /D ≤ 1; s ∈ S, (10.36)
t∈T :t<ts j∈J t∈T :t≥ts

where the right-hand side of (10.36) is implied by Eqs. (10.14) and (10.21).

10.3.2 Multiple Part Types and Product Types

In this subsection an enhancement of model DSupport_E is described for multiple

types of parts and products. Let H and K be, respectively, the set of part types and
the set of product types, and denote by ahk , h ∈ H, k ∈ K the number of parts type
h required to produce one unit of product type k (for notation used and definition of
problem variables, see Table 10.3). 
If we denote by Dk the total demand forproducts type k, then D = k∈K Dk is
the total demand for all products and Ah = k∈K ahk Dk is the total demand for parts
type h.
282 10 Selection of Primary and Recovery Supply …

Table 10.3 Notation and variables: multiple part types and product types
Indices
h = part type, h ∈ H
k = product type, k ∈ K
Input Parameters
ahk = Unit requirement for part type h of product type k

Ah = k∈K ahk Dk total demand for parts type h
Dk = Total demand for products type k
εjk = Additional per unit production cost for product type k at recovery plant j > 1
gk = Per unit penalty cost for unfulfilled demand for products type k
oih = Per unit price of parts type h purchased from supplier i
First stage variables
vih ∈ [0, 1], the fraction of total demand for parts type h ordered from primary supplier i to
primary plant j = 1 (primary supply portfolio)
Second stage variables
Vijhs ∈ [0, 1], the fraction of total demand for parts type h ordered from recovery supplier i to
recovery plant j, under disruption scenario s (recovery supply portfolio)
wjhs ∈ [0, 1], the fraction of total demand for parts type h, transshipped from the primary plant
j = 1 to recovery plant j, under disruption scenario s, where w1h s represents the parts type

h that remain in the primary plant j = 1 (transshipment variables)

s
xjkt ≥ 0, production in plant j of product type k in period t under disruption scenario s (production
scheduling)
s
zjk ∈ [0, 1] the fraction of total demand for products type k to be completed by recovery plant
j under disruption scenario s (recovery demand portfolio).

Now, the portfolio decision variables vi , Vijs , i ∈ I, j ∈ J, s ∈ S are replaced by

vih , Vijh
s
; h ∈ H, i ∈ I, j ∈ J, s ∈ S, defined as fractions of total demand Ah for parts
type h, ordered from primary supplier i, recovery supplier i for recovery plant j under
s
disruption scenario s, respectively. Accordingly, the new transshipment variable wjh
denotes the fraction of total demand Ah for parts type h transshipped from the primary
plant j = 1 to recovery plant j, under scenario s. In addition, the production scheduling
variable xjts , j ∈ J, t ∈ T , s ∈ S is replaced by xjkt
s
, j ∈ J, k ∈ K, t ∈ T , s ∈ S - the
production in plant j of product type k in period t under disruption scenario s, and
zjs , j ∈ J, s ∈ S by zjks , j ∈ J, k ∈ K, s ∈ S - the fraction of demand for products
type k to be completed in recovery plant j under disruption scenario s.
Model DSupportMP_E for multiple part and product types is presented below.

DSupportMP
Minimize
   
Ec = Ps (ei (ui + Uis − qis ) + ρis Uis + oih Ah (γis vih + s
Vijh ))/D
s∈S i∈I h∈H j∈J
10.3 Models for Risk-Neutral Decision-Making 283
  
+ Ps ( ψj Ah wjh
s
+ (ϕj + ρjs )yjs + εjk xjkt
s
)/D
s∈S j∈J h∈H k∈K t∈T
  
+ gk (1 − Ps s
xjkt /Dk ),
k∈K s∈S j∈J t∈T

subject to

vih = 1; h ∈ H
i∈I
vih ≤ ui ; h ∈ H, i ∈ I
 
(γis vih + s ) = 1 − V ; h ∈ H, s ∈ S
Vijh h0
i∈I j∈J
s ≤ U s ; h ∈ H, i ∈ I, j ∈ J, s ∈ S
Vijh i
 
s /D +
x1kt s = 1; k ∈ K, s ∈ S
zjk
k
t∈T :t<ts j∈J
s ≤ ys ; j ∈ J, k ∈ K, s ∈ S
zjk j
 
s + ws =
Vijh s /A ; h ∈ H, j ∈ J, s ∈ S
ahk Dk zjk
jh h
i∈I k∈K
   
s =V +
wjh γis vih − s /A ; h ∈ H, s ∈ S
ahk x1kt
h0 h
j∈J i∈I k∈K t∈T :t<ts
qis ≤ (ui + Uis )/2; i ∈ I, s ∈ S

s ≤ c ; t ∈ T, s ∈ S : t < t
x1kt 1 s
k∈K

s ≤ cs ys ; j ∈ J, t ∈ T , s ∈ S : t ≥ t
xjkt jt j s
k∈K

s /D ≤ zs ; j ∈ J, k ∈ K, s ∈ S
xjkt k jk
t∈T :t≥ts
    
s /A ≤ V +
ahk x1kt γis vih + s −
Vijh s ;
wjh
 h h0
k∈K t  ∈T :t  ≤t i∈I:τi1 ≤t−1 i∈I:ts +θis +τi1 ≤t−1 j∈J:j>1
h ∈ H, t ∈ T , s ∈ S
  
s /A ≤
ahk xjkt s ;
Vijh
 h
k∈K t  ∈T :t  ≤t i∈I:ts +θis +τij ≤t−1
h ∈ H, j ∈ J, t ∈ T , s ∈ S : j > 1, t < σj + 1
  
s /A ≤
ahk xjkt s + ws ;
Vijh
 h jh
k∈K t  ∈T :t  ≤t i∈I:ts +θis +τij ≤t−1
h ∈ H, j ∈ J, t ∈ T , s ∈ S : j > 1, t ≥ σj + 1
qis ∈ {0, 1}; i ∈ I, s ∈ S
ui ∈ {0, 1}; i ∈ I
vih ∈ [0, 1]; i ∈ I, h ∈ H
Uis ∈ {0, 1}; i ∈ I, s ∈ S
s
Vijh ∈ [0, 1]; i ∈ I, j ∈ J, h ∈ H, s ∈ S
284 10 Selection of Primary and Recovery Supply …

s
wjh ∈ [0, 1]; j ∈ J, h ∈ H, s ∈ S
xjkt ≥ 0; j ∈ J, k ∈ K, t ∈ T , s ∈ S
s

yjs ∈ {0, 1}; j ∈ J, s ∈ S

zjks ∈ [0, 1]; j ∈ J, k ∈ K, s ∈ S,

where oih is unit purchasing price of part type h from supplier i, εjk is the
additional per unit production cost for product type k at recovery plant j > 1,
gk is unit penalty cost of unfulfilled demand for product type k, and Ah Vh0 is
the initial inventory of parts type h at primary assembly plant.

Notice that problem DSupportMP_E is a multi-portfolio selection problem, since

supply and demand portfolios are simultaneously selected for each part type and
each product type. Now, the supply and demand portfolio selection constraints
form an embedded multicommodity network flow problem. Moreover, the supply
and demand portfolio balance constraints (10.16) and the supply-transshipment-
production coordinating constraints (10.22)–(10.24) are also bill-of-material con-
straints in model DSupportMP_E.
The demand fulfillment constraints for each product type k ∈ K,
 
s
x1kt /Dk + zjks = 1; k ∈ K, s ∈ S,
t∈T :t<ts j∈J

imply the following requirement constraints for each part type h ∈ H,

  
s
ahk x1kt /Ah + ahk Dk zjks /Ah = 1; h ∈ H, s ∈ S.
k∈K t∈T :t<ts j∈J k∈K

The latter constraints can be obtained from the previous ones by multiplying both
sides by ahk Dk and summing over all k ∈ K, and then replacing the right-hand side,

k∈K ahk Dk , by Ah . The requirement constraints for part types can be added to model
DSupportMP_E to strengthen MIP formulation.
The above model can be further enhanced, for example by introducing subsets
Ih ⊂ I of suppliers for each part type h, unit capacity consumption for each product
type k and each plant j in the left-hand side of production capacity constraints, etc.

10.3.3 Hierarchical Selection of Primary and Recovery

Supply and Demand Portfolios

In this subsection two deterministic MIP models PSupport and RDSupport(s) are
presented for the hierarchical decision making in the presence of supply chain dis-
ruption risks. The two-stage decision making is described below (Fig. 10.1).
10.3 Models for Risk-Neutral Decision-Making 285

Fig. 10.1 Hierarchical

selection of supply and
demand portfolios

1. Selection of primary supply portfolio for deterministic environment.

The primary suppliers are determined ahead of time with no disruption scenarios
considered, using deterministic MIP model PSupport. The primary assembly
plant is plant j = 1, by default.
2. Selection of recovery supply and demand portfolios, after disruption of a primary
supplier and/or the primary assembly plant.
The recovery portfolios are determined to optimize the process of recovery from
the disruption, using MIP model RDSupport(s).
s
In the deterministic MIP model PSupport, stochastic variable, x1t , (10.31),
defined in model DSupport_E for each disruption scenario s ∈ S has been replaced
by its deterministic equivalent Xt .

PSupport: Primary supply portfolio selection and production scheduling

Minimize
 
Pc = (ei ui /D + oi vi ) + g(1 − Xt /D) (10.37)
i∈I t∈T

subject to

vi = 1 (10.38)
i∈I
vi ≤ ui ; i ∈ I (10.39)
286 10 Selection of Primary and Recovery Supply …

 
Xt  /D ≤ V0 + vi ; t ∈ T (10.40)
t  ∈T :t  ≤t i∈I:τi1 ≤t−1
Xt ≤ c1 ; t ∈ T (10.41)
ui ∈ {0, 1}; i ∈ I (10.42)
vi ∈ [0, 1]; i ∈ I (10.43)
Xt ≥ 0; t ∈ T . (10.44)

The solution to PSupport is primary supply portfolio: ui∗ , vi∗ ; i ∈ I, and the
associated production schedule, Xt∗ .
A disruptive event under scenario s ∈ S is assumed to occur in period ts and it may
impact a primary supplier (as well as the other suppliers) and the primary assembly
plant j = 1 (as well as the other assembly plants). Then, the recovery process starts
in period ts + 1 so that the disrupted supplier i ∈ I returns to its full capacity in period
t = ts + θis and disrupted plant j ∈ J, in period t = ts + ϑjs .
Denote by Ds and V0s respectively, the unfulfilled demand for products in period
ts and the inventory of parts at plant j = 1 in period ts , expressed by the fraction of
parts required to complete the unfulfilled demand Ds

Ds = D − s
x1t ; s∈S (10.45)
t∈T :t<ts
 
V0s = D(V0 + γis vi∗ )/Ds − s
x1t /Ds ; s ∈ S, (10.46)
i∈I:τi ≤ts t∈T :t<ts

where vi∗ , i ∈ I, is a predetermined primary supply portfolio.

If V0s > 0, then selection of recovery portfolios may be combined with transship-
ment of parts from the primary plant to recovery plants.
When a disruption to primary assembly plant occurs after the latest delivery lead
time from primary suppliers of parts, ts > maxi∈I (τi1 ui∗ ), i.e., after all required parts
are delivered to the primary plant, then all parts can be transshipped to recovery
plants and no recovery supply portfolio needs to be selected.
In model RDSupport(s) presented below, the recovery supply and demand port-
folios are selected whenever a disruption occurs to a primary supplier or the primary
assembly plant j = 1, given a primary supply portfolio and the realized disruption
scenario s = s̃.

RDSupport(s): Recovery supply and demand portfolio selection and

production scheduling for predetermined primary portfolios and the realized
disruption scenario
Minimize
10.3 Models for Risk-Neutral Decision-Making 287

   
Rs̃c = ei (1 − ui∗ )Uis̃ /Ds̃ + ρis̃ Uis̃ /Ds̃ + (ψj wjs̃ + oi Vijs̃ )
i∈I i∈I j∈J i∈I
 
+ ((ϕj + ρjs̃ )yjs̃ + (εj − g) xjts̃ )/Ds̃ + g
j∈J t∈T s̃
 
+ ei ui∗ /Ds̃ + oi γis vi∗ (10.47)
i∈I i∈I

subject to

Vijs̃ = 1 − V0s̃ (10.48)
i∈I j∈J

Vijs̃ ≤ Uis̃ ; i ∈ I, j ∈ J (10.49)


zjs̃ = 1 (10.50)
j∈J

zjs̃ ≤ yjs̃ ; j ∈ J (10.51)


Vijs̃ + wjs̃ = zjs̃ ; j∈J (10.52)
i∈I

wjs̃ = V0s̃ (10.53)
j∈J
  
 /D ≤ V0 + Vi1s̃ − wjs̃ ; t ∈ T s̃
s̃ s̃ s̃
x1t (10.54)
t  ∈T s̃ :t  ≤t i∈I:ts̃ +θis̃ +τi1 ≤t−1 j∈J:j>1
 
xjts̃  /Ds̃ ≤ Vijs̃ ;
t  ∈T s̃ :t  ≤t i∈I:ts̃ +θis̃ +τij ≤t−1

j ∈ J, t ∈ T s̃ : j > 1, t < ts̃ + 1 + σj (10.55)

 
xjts̃  /Ds̃ ≤ Vijs̃ + wjs̃ ;
t  ∈T s̃ :t  ≤t i∈I:ts̃ +θis̃ +τij ≤t−1

j ∈ J, t ∈ T s̃ : j > 1, t ≥ ts̃ + 1 + σj (10.56)


xjts̃ /Ds̃ ≤ zjs̃ ; j ∈ J (10.57)
t∈T s̃

xjts̃ ≤ cjts̃ yjs̃ ; j ∈ J, t ∈ T s̃ (10.58)

Uis̃ ∈ {0, 1}; i ∈ I (10.59)
Vijs̃ ∈ [0, 1]; i ∈ I, j ∈ J (10.60)
wjs̃ ∈ [0, 1]; j ∈ J (10.61)
xjts̃ ≥ 0; j ∈ J, t ∈ T s̃ (10.62)
yjs̃ ∈ {0, 1}; j ∈ J (10.63)
zjs̃ ∈ [0, 1]; j ∈ J, (10.64)

where T s = {ts + 1, . . . , T } and

cjts̃ is the capacity in period t of plant j under disruption scenario s̃

δjs̃ cj , if ts̃ ≤ t ≤ ts̃ + ϑjs̃ − 1
cjts̃ = (10.65)
cj , if t ≥ ts̃ + ϑjs̃ .
288 10 Selection of Primary and Recovery Supply …

The objective function (10.47) constitutes of different fixed and variable cost per
product (cf. Eq.(10.8)), and the last two constant components of (10.47) represent
per product costs incurred by predetermined primary supply portfolio.
Notice that model DSupport_E for the predetermined primary supply portfolio
is separable with respect to disruption scenarios s ∈ S, since the objective function
(10.8) is additive and separable with respect to s as well as all constraints (10.12) -
(10.24) are separable with respect to s. Thus, for a given primary supply portfolio and
start time ts of disruption s, the recovery supply portfolios, Uis , Vijs ; i ∈ I, j ∈ J, the
recovery demand portfolios, yjs , zjs ; j ∈ J, the transshipments of parts, wjs , j ∈ J,
and the production schedules, xjts , j ∈ J, can be found simultaneously for all potential
disruption scenarios s ∈ S, by solving
RDSupport_E = DSupport_E for predetermined primary supply portfolio: ui =
ui∗ , vi = vi∗ ; i ∈ I.

RDSupport_E: Risk-neutral selection of recovery supply and demand

portfolios and production scheduling for predetermined primary portfolios
Minimize (10.8)
subject to (10.12)–(10.25), (10.28)–(10.34) and

ui = ui∗ ; i ∈ I (10.66)
vi = vi∗ ; i ∈ I. (10.67)

Worst-case cost per product

A simple upper bound on cost per product, Rsc , for scenario s, of model RDSup-
port_E for predetermined primary supply and demand portfolios is derived below.
Proposition 10.2

Rsc ≤ Rc ; s ∈ S, (10.68)

where
  
Rsc = ei (1 − ui∗ )Uis /D + ρis Uis /D + (ϕj + ρjs )yjs /D
i∈I i∈I j∈J
    
+ (ψj wjs + oi Vijs ) + εj xjts /D + g(1 − xjts /D)
j∈J i∈I j∈J t∈T t∈T
 
+ ei ui∗ /D + oi γis vi∗ , (10.69)
i∈I i∈I
10.3 Models for Risk-Neutral Decision-Making 289

and Rc is the worst-case cost per product associated with worst-case disruption sce-
nario with respect to cost

Rc = ei /D + g(1 − SL) + ψmax + εmax
i∈I

+ max{ ρis ui∗ /D + max(ϕj + ρjs )/D
s∈S j∈J
i∈I
 
+(omax − ψmin )(1 − γis vi∗ ) + oi γis vi∗ }. (10.70)
i∈I i∈I

Proof. 
Without loss of generality, assume that V0 = 0 and t∈T :t<ts x1t
s
/D = 0. Recov-
ery supply and demand portfolio selection constraints (10.12), (10.14), (10.16) and
minimization of cost objective function, imply
    
ρis Uis /D + (ϕj + ρjs )yjs /D + ψj wjs + oi Vijs + oi γis vi∗ =
i∈I j∈J j∈J i∈I j∈J i∈I
     
ρis Uis /D + (ϕj + ρjs )yjs /D + ψj (zjs − Vijs ) + oi Vijs + oi γis vi∗ =
i∈I j∈J j∈J i∈I i∈I j∈J i∈I
     
ρis Uis /D + (ϕj + ρjs )yjs /D + ψj zjs − ψj Vijs + oi Vijs + oi γis vi∗ ≤
i∈I j∈J j∈J i∈I j∈J i∈I j∈J i∈I
   
ρis Uis /D + (ϕj + ρjs )yjs /D + ψmax + (omax − ψmin ) Vijs + oi γis vi∗ ≤
i∈I j∈J i∈I j∈J i∈I
  
ψmax + max{ ρis ui∗ /D + max(ϕj + ρjs )/D + (omax − ψmin )(1 − γis vi∗ ) + oi γis vi∗ },
s∈S j∈J
i∈I i∈I i∈I

where omax = maxi∈I oi , ψmax = maxj∈J ψj , ψmin = minj∈J ψj .

In addition, Eqs. (10.14), (10.21) and (10.36) imply
    
ei (1 − ui∗ )Uis /D + ei ui∗ /D + εj xjts /D + g(1 − xjts /D) ≤
i∈I i∈I j∈J t∈T t∈T
 
ei /D + εmax xjts /D + g(1 − SL) ≤
i∈I j∈J t∈T
 
ei /D + εmax zjs + g(1 − SL) ≤
i∈I j∈J

ei /D + εmax + g(1 − SL),
i∈I

where εmax = maxj∈J εj .

By summing up the final parts of the above two expressions, formula (10.70) for
Rc is achieved.
290 10 Selection of Primary and Recovery Supply …

10.4 Models for Risk-Averse Decision-Making

10.4.1 Integrated Approach

In this subsection the two time-indexed SMIP models DSupport_CV(c) and DSup-
port_CV(sl) are proposed for the integrated, risk-averse selection of primary and
recovery supply and demand portfolios and production scheduling to optimize,
respectively expected worst-case cost and expected worst-case service level under
disruption risks. The models are based on the risk-neutral model DSupport_E.
When the worst-case cost is focused on, the risk-averse primary and recovery
supply and demand portfolios and the production schedule will be optimized by
calculating VaRc and minimizing CVaRc simultaneously. Model DSupport_CV(c)
is presented below.

DSupport_CV(c): Risk-averse selection of primary and recovery supply and

demand portfolios and scheduling of production to minimize CVaR of cost
Minimize

CV aRc = V aRc + (1 − α)−1 Ps Cs (10.71)
s∈S

subject to (10.10)–(10.34) and

Risk constraints:
• the tail cost for scenario s is defined as the nonnegative amount by which
cost in scenario s exceeds VaRc ,
 
Cs ≥ (ei (ui + Uis − qis )/D + ρis Uis /D + oi (γis vi + Vijs ))
i∈I j∈J
 
+ (ψj wjs + (ϕj + ρjs )yjs /D + εj xjts /D)
j∈J t∈T

+g(1 − xjts /D) − V aRc ; s ∈ S (10.72)
j∈J t∈T
Cs ≥ 0; s ∈ S, (10.73)

where Cs is the tail cost for scenario s.

If worst-case service level is to be maximized, the risk-averse primary and recov-

ery supply portfolio and the production schedule will be optimized by calculating
VaRsl and maximizing CVaRsl simultaneously. Model DSupport_CV(sl) is pre-
sented below.
10.4 Models for Risk-Averse Decision-Making 291

DSupport_CV(sl): Risk-averse selection of primary and recovery supply and

demand portfolios and scheduling of production to maximize CVaR of service
level
Maximize

CV aRsl = V aRsl − (1 − α)−1 Ps Ss (10.74)
s∈S

subject to (10.10)–(10.34) and

Risk constraints:

• the tail service level for scenario s is defined as the nonnegative amount by
which VaRsl exceeds service level in scenario s,

Ss ≥ V aRsl − xjts /D; s ∈ S (10.75)
j∈J t∈T
Ss ≥ 0; s ∈ S, (10.76)

where Ss is the tail service level for scenario s.

10.4.2 Hierarchical Approach

In this subsection the two time-indexed SMIP models RDSupport_CV(c) and

RDSupport_CV(sl) are presented for the hierarchical, risk-averse selection of recov-
ery supply and demand portfolios and production scheduling to optimize, respec-
tively expected worst-case cost and expected worst-case service level, given primary
supply portfolio.
The models are based on the risk-averse decision-making models,
DSupport_CV(c) and DSupport_CV(sl), respectively:
• RDSupport_CV(c) = DSupport_CV(c) for predetermined primary supply port-
folio: ui = ui∗ , vi = vi∗ ; i ∈ I;
• RDSupport_CV(sl) = DSupport_CV(sl) for predetermined primary supply port-
folio: ui = ui∗ , vi = vi∗ ; i ∈ I.
For model RDSupport_CV(c), the primary supply portfolio, ui∗ , vi∗ ; i ∈ I, is
determined using model PSupport with cost-based objective function (10.37). For
model RDSupport_CV(sl), however, the cost-based objective function (10.37) is
replaced by the following service-based objective function

Maximize Xt /D. (10.77)
t∈T
292 10 Selection of Primary and Recovery Supply …

RDSupport_CV(c): Risk-averse selection of recovery supply and demand

portfolios and production scheduling to minimize CVaR of cost for
predetermined primary portfolios.
Minimize (10.71)
subject to (10.10)–(10.34), (10.66), (10.67), (10.72), (10.73).

RDSupport_CV(sl): Risk-averse selection of recovery supply and demand

portfolios and production scheduling to minimize CVaR of service level for
predetermined primary portfolios.
Minimize (10.74)
subject to (10.10)–(10.34), (10.66), (10.67), (10.75), (10.76).

10.5 Computational Examples

In this section some computational examples are presented to illustrate the proposed
portfolio approach for selection of primary and recovery supply and demand port-
folios and production scheduling. The input data for the examples are hypothetical,
however their relations to each other are real and in part they have been taken from
a real case study (e.g., Fujimoto and Park 2013, Park et al. 2013, MacKenzie et al.
2014, Matsuo 2015, Haraguchi and Lall 2015). The basic input parameters are taken
from the computational examples presented in Sect. 9.5.
I = 4 suppliers, Li = 3, i.e., four disruption levels for each supplier i ∈ I,
J = 2 assembly plants, L j = 1, i.e., two (all-or-nothing) disruption levels for each
plant j ∈ J,
R = 2 geographic regions, T = 30 planning periods.
I 1 = {1, 2}, I 2 = {3, 4}, J 1 = {1, 2}, i.e., two suppliers and two plants are
located in region r = 1, and two suppliers in region r = 2.
The initial inventories of parts: V0 = 0.
Total demand for parts/products: D = 300000.
Delivery lead times from suppliers: τ1j = τ2j = 2, τ3j = τ4j = 4, ∀j ∈ J.
Fixed ordering costs for suppliers: e = (8000, 6000, 12000, 13000).
Unit purchasing prices from suppliers: o = (14, 12, 8, 9).
Plant capacity: c1 = 10000, c2 = 5000.
Production costs: ε2 = 1, ϕ2 = 100.
Transshipment cost and time: ψ2 = 0.1, σ2 = 2.
Unit penalties for unfulfilled demand:
10.5 Computational Examples 293

g ∈ {1, 10, 100, 1000, 10000, 100000, ∞}, where g = ∞ denotes maximization
of service level.
Supplier cost-to-recover and time-to-recover are defined below.
CTR(i, l)=if l = 0 then 100000ei ; if l = 1 then 10000ei ; if l = 2 then 1000ei
∀i ∈ I,
TTR(i, l)=if l = 0 then 12; if l = 1 then 10; if l = 2 then 8 ∀i ∈ I.
Plant recovery cost, PRC(j, 0) and recovery time, PRT (j, 0), are
PRC(1, 0) = PRC(2, 0) = 10000 and PRT (1, 0) = 10, PRT (2, 0) = 5.
Local disruption levels of suppliers and the associated fulfillment rates are shown
below.
Li = L = {0, 1, 2, 3} for all i ∈ I, where l = 0, complete shutdown,
γi0 = 0 ∀i ∈ I, i.e., 0% of an order delivered; l = 1, major disruption,
γi1 ∈ [0.01, 0.50] ∀i ∈ I 1 and γi1 ∈ [0.01, 0.30] ∀i ∈ I 2 , i.e., 1–50% and 1–30% of
an order delivered, respectively; l = 2, minor disruption, γi2 ∈ [0, 51, 0.99] ∀i ∈ I 1
and γi2 ∈ [0, 31, 0.99] ∀i ∈ I 2 , i.e., 51–99% and 31–99% of an order delivered,
respectively; l = L = 3, no disruption, γi3 = 1 ∀i ∈ I, i.e., 100% of an order
delivered.
The total number of all potential scenarios is S = (L + 1)I (L 1 + 1)(L 2 + 1) =
(4 )(22 ) = 1024 scenarios. Each scenario s ∈ S is represented by a 6-dimensional
4

vector λs = (λ1s , . . . , λ4s , λ5s , λ6s ), where λis ∈ Li , i ∈ I, and λ5s , λ6s ∈ {0, 1},
where λ5s and λ6s represent disruption pattern under scenario s of assembly plant j =
1 and j = 2, respectively. The selected supply and production disruption scenarios
are presented in Tables 10.5 and 10.6.
The local probability of suppliers non-disruptive operation (level l = 3), pi3 , was
uniformly distributed over [0.89, 0.99] and [0.79, 0.89], respectively for suppliers
i ∈ I 1 , and i ∈ I 2 .
Given local non disruption probabilities of suppliers, pi3 , i ∈ I, the probabilities
for the remaining local disruption levels l = 0, 1, 2 were calculated as follows:
probability of complete shutdown (level l = 0), pi0 = 0.1(1 − pi3 );
probability of major disruption (level l = 1), pi1 = 0.3(1 − pi3 );
probability of minor disruption (level l = 2), pi2 = 0.6(1 − pi3 ) for all suppliers
i ∈ I.
The local probabilities of non-disruptive operation of assembly plants were π1 =
0.85 and π2 = 0.95, i.e., the primary assembly plant j = 1 was modeled to be less
reliable to emphasize the impact of its disruption.
The regional disruption probabilities are p1 = 0.001 and p2 = 0.01.
The probability Ps1 of realizing disruption scenario s ∈ S for suppliers and assem-
bly plants in region r = 1, and Ps2 of realizing disruption scenario s ∈ S for suppliers
in region r = 2 are calculated as follows
294 10 Selection of Primary and Recovery Supply …
⎧  
⎪ (1 −
⎪
⎪  p1 )( i∈I 1 :λis =0 0.1(1 − pi3 ))( i∈I 1 :λis =1 0.3(1 − pi3 ))
⎪
⎪ ×( i∈I 1 :λis =2 0.6(1 − pi3  ))( i∈I 1 :λis =3 pi3 )
⎪
⎨ ×(
⎪
(1 − π ))( j∈J:λI+j,s =1 πj )
j∈J:λI+j,s =0
j
Ps1 =
⎪
⎪ if

i∈I 1 {5,6} λis > 0

⎪
⎪ 
⎪
⎪ p1 + (1 − p1 ) i∈I 1 0.1(1 − pi3 ) j∈J (1 − πj )
⎪
⎩ if 
i∈I 1 {5,6} λis = 0.

⎧  
⎨ (1 −

p2 )( i∈I 2 :λis =0 0.1(1 − pi3 ))( i∈I 2 :λis =1 0.3(1 − pi3 ))
 
Ps2 = ×( i∈I 2 :λis =2 0.6(1 − pi3 ))( i∈I 2 :λis =3 pi3 ) if λ >0
⎩ 2  i∈I 2 is
p + (1 − p ) i∈I 2 0.1(1 − pi3 )
2 if i∈I 2 λis = 0.

The probability for disruption scenario s ∈ S is given by Ps = Ps1 Ps2 .

10.5.1 Risk-Neutral Decision-Making

In the computational examples presented in this subsection each disruption s ∈ S

to primary suppliers (and possibly other suppliers) and/or to primary plant j = 1
(and possibly other plants) is assumed to occur in the same period ts = 1, before
the earliest delivery lead time, mini∈I (τi1 ) = 2, and the recovery process starts in
period t = 2 so that the disrupted supplier i and disrupted plant j return to its full
capacity in period t = θis + 1 and t = ϑjs + 1, respectively. The assumption of a
common disruption start time of all disruptive events will be relaxed in Sect. 10.5.1.1
to provide more insights by best-case and worst-case analysis, and in Sect. 10.5.2 to
illustrate the risk-averse decision-making.
The solution results for both the integrated and the hierarchical approach are
summarized in Table 10.4. The table shows primary supply portfolio and expected
recovery supply portfolio and recovery demand portfolio for different values of unit
penalty cost g. Notice that for each disruption scenario, the recovery supply portfolio
fulfills the remaining demand for parts, unmet by partially disrupted primary suppliers
(see, (10.12) and Table 10.5). Thus, the numbers representing the expected fraction,
over all scenarios, of total demand to be fulfilled by recovery suppliers do not need
to add up to 100%. In contrast to expected recovery demand portfolio, that sums
up to 100%, since all demand for products is assigned to a single primary plant that
is shutdown before any production can be started, if a disruptive event occurs at time
ts = 1 (see, (10.14) and Table 10.5). For the integrated approach, DSupport_E, and
the bottom level problem RDSupport of the hierarchical approach, expected cost E c ,
(10.8), is shown along with the associated expected service level E sl , (10.9). Similarly,
for the top level problem PSupport of the hierarchical
 approach, cost Pc , (10.37),
is shown along with the associated service level, t∈T Xt /D. In addition, Table 10.4
presents the size of each MIP model, DSupport_E, PSupport and RDSupport_E.
10.5 Computational Examples 295

As g increases to reduce the unfulfilled demand, more expensive and diversified

supply portfolios are selected and a higher service level is attained. In particular,
for integrated approach more diversified primary supply portfolios are selected for
g > 1 to hedge against all disruption scenarios. In contrast, for hierarchical approach
the results for g > 1 are independent of g. When the objective is to maximize service
level (g → ∞), the hierarchical approach selects the most reliable supplier (i = 2)
as the only primary supplier. Table 10.4 demonstrates that for low values of g, the two
approaches yield identical solution values. This is due to negligible penalty cost for
unfulfilled demand for products with respect to other cost components. However, as
g increases, the integrated approach outperforms the purely deterministic, top-down
hierarchical approach. Both, expected cost over all scenarios is lower and expected
service level is higher.
Overall, the computational results demonstrate that:

• the integrated approach selects for both objectives a more diversified primary
supply portfolio that will hedge against all potential disruption scenarios,
• the hierarchical approach selects the primary supply portfolio that is made up of
cheapest suppliers or a single, most reliable supplier only, to minimize expected
cost or maximize expected service level, respectively.

For the integrated approach (model DSupport_E) and selected supply and pro-
duction disruption scenarios s ∈ S, Table 10.5 shows examples of recovery supply
and demand portfolios (V11 s
+ V12
s
, V21
s
+ V22
s
, V31
s
+ V32
s
, V41
s
+ V42s
, z1s , z2s ) along with
the associated cost,
 
(ei (ui + Uis − qis )/D + ρis Uis /D + oi (γis vi + Vijs ))
i∈I j∈J
  
+ (ψj wjs + (ϕj + ρjs )yjs /D + εj xjts /D) + g(1 − xjts /D),
j∈J t∈T j∈J t∈T

and service level,


xjts /D,
j∈J t∈T

for a unit penalty g = 100. The primary supply portfolio for the example is (see,
Table 10.4): v1 = 0, v2 = 0.08, v3 = 0.74, v4 = 0.18. The results indicate that the
cost is ranging from 8.57 for scenario s = 845 to 2095.44 for scenario s = 1 with all
suppliers and assembly plants completely shutdown. The corresponding service level
is ranging from 100% to 50%. Notice that the objective of model DSupport_E is
to select primary supply portfolio to hedge against all potential disruption scenarios
and to select recovery supply and demand portfolios for each scenario to minimize
expected cost over all scenarios.
296 10 Selection of Primary and Recovery Supply …

Table 10.4 Supply and demand portfolios

The corresponding results for the hierarchical approach (models PSupport and
RDSupport_E) are presented in Table 10.6. The primary supply portfolio obtained to
minimize purchasing cost for parts and penalty for unfulfilled demand for products
in a deterministic operating environment (model PSupport) is (see, Table 10.4):
v1 = 0, v2 = 0.07, v3 = 0.93, v4 = 0. Then, given the primary supply and
demand portfolios and the realized disruption scenario, the recovery supply and
demand portfolios are selected to minimize total cost, including cost of recovery
from the disruption. Now, the cost is ranging from 9.05 for scenario s = 844 to
10.5 Computational Examples 297

Table 10.5 Supply and demand recovery portfolios, (V11 s + Vs ,Vs + Vs ,Vs + Vs ,Vs +
12 21 22 31 32 41
V42 , z1 , z2 ) and associated cost and service, for selected supply and production disruption scenarios
s s s

s ∈ S: integrated approach, g = 100

Disruption scenario Supply portfolio Demand portfolio Solution values
s λ1s λ2s λ3s λ4s λ5s λ6s i=1 i=2 i=3 i=4 j=1 j=2 Cost Service %
1 0 0 0 0 0 0 0 1 0 0 1 0 2095.44 50
7 0 0 1 2 0 0 0 0 0 0.81 1 0 119.03 66.67
17 0 1 0 0 0 0 0 0.97 0 0 1 0 285.92 59.51
18 0 1 0 1 0 0 0 0.94 0 0 1 0 282.62 62.72
19 0 1 0 2 0 0 0 0 0 0.86 1 0 119.19 66.67
26 0 1 2 1 0 0 0 0 0.53 0 0.67 0.33 120.53 95
28 0 1 2 3 0 0 0 0 0 0.38 1 0 75.44 66.67
35 0 2 0 2 0 0 0 0.82 0 0 0.67 0.33 100.76 98.33
36 0 2 0 3 0 0 0 0 0 0.76 1 0 75.97 66.67
42 0 2 2 1 0 0 0 0.49 0 0 0.67 0.33 99.35 98.33
44 0 2 2 3 0 0 0 0 0 0.34 1 0 75.56 66.67
49 0 3 0 0 0 0 0 0.92 0 0 0.67 0.33 79.44 100
50 0 3 0 1 0 0 0 0.88 0 0 1 0 78.67 66.67
321 1 0 0 0 0 1 1 0 0 0 0.57 0.43 329.75 85
322 1 0 0 1 0 1 0.1 0 0 0.87 0.6 0.4 755.37 88.21
323 1 0 0 2 0 1 0 0 0 0.89 0.72 0.28 91.39 95
324 1 0 0 3 0 1 0 0 0 0.82 0.67 0.33 43.1 100
325 1 0 1 0 0 1 0.92 0 0 0 0.64 0.36 321.67 92.62
326 1 0 1 1 0 1 0.89 0 0 0 0.72 0.28 319.13 95
328 1 0 1 3 0 1 0 0 0 0.75 0.67 0.33 43.03 100
329 1 0 2 0 0 1 0 0 0.59 0 0.72 0.28 87.05 95
332 1 0 2 3 0 1 0 0 0 0.4 1 0.67 0.33 42.69 100
339 1 1 0 2 0 1 0 0 0 0.86 0.7 2 0.28 91.47 95
340 1 1 0 3 0 1 0 0 0 0.79 0.67 0.33 43.19 100
341 1 1 1 0 0 1 0 0.9 0 0 0.67 0.33 250.75 95
344 1 1 1 3 0 1 0 0 0 0.72 0.67 0.33 43.11 100
353 1 2 0 0 0 1 0 0.93 0 0 0.68 0.32 67.75 98.33
356 1 2 0 3 0 1 0 0 0 0.76 0.67 0.33 43.3 100
357 1 2 1 0 0 1 0 0.86 0 0 0.67 0.33 67.45 98.33
360 1 2 1 3 0 1 0 0 0 0.68 0.67 0.33 43.23 100
361 1 2 2 0 0 1 0 0.52 0 0 0.68 0.32 66.11 98.33
381 1 3 3 0 0 1 0 0 0.18 0 0.82 0.18 57.9 84.39
382 1 3 3 1 0 1 0 0 0.15 0 0.85 0.15 61.1 81.18
383 1 3 3 2 0 1 0 0 0.07 0 0.93 0.07 68.95 73.33
384 1 3 3 3 0 1 0 0 0 0 1 0 75.29 66.67
640 1 3 3 3 1 0 0 0 0 0 1 0 15.29 93.33
(continued)
298 10 Selection of Primary and Recovery Supply …

Table 10.5 (continued)

Disruption scenario Supply portfolio Demand portfolio Solution values
s λ1s λ2s λ3s λ4s λ5s λ6s i=1 i=2 i=3 i=4 j=1 j=2 Cost Service %
646 2 0 1 1 1 0 0.89 0 0 0 1 0 66.03 74.15
647 2 0 1 2 1 0 0.81 0 0 0 0.68 0.32 73.77 100
648 2 0 1 3 1 0 0 0 0 0.75 1 0 22.36 86.67
657 2 1 0 0 1 0 0.97 0 0 0 1 0 74.56 66.18
658 2 1 0 1 1 0 0.94 0 0 0 1 0 71.19 69.39
659 2 1 0 2 1 0 0.86 0 0 0 1 0 62.95 77.24
660 2 1 0 3 1 0 0 0 0 0.79 1 0 19.67 89.51
661 2 1 1 0 1 0 0.9 0 0 0 1 0 66.49 73.8
662 2 1 1 1 1 0 0.86 0 0 0 1 0 63.12 77
663 2 1 1 2 1 0 0.78 0 0 0 1 0 54.88 84.85
664 2 1 1 3 1 0 0 0 0 0.72 1 0 19.6 89.51
673 2 2 0 0 1 0 0 0.93 0 0 1 0 62.1 70
676 2 2 0 3 1 0 0 0 0 0.76 1 0 15.97 93.33
677 2 2 1 0 1 0 0 0.86 0 0 1 0 54.18 77.62
844 1 0 2 3 1 1 0 0 0 0.41 0.87 0.13 9.16 100
845 1 0 3 0 1 1 0 0 0.26 0 0.87 0.13 8.57 100
961 3 0 0 0 1 1 1 0 0 0 0.9 0.1 14.56 100
976 3 0 3 3 1 1 0 0 0.08 0 0.92 0.08 13.55 95.14
977 3 1 0 0 1 1 0.97 0 0 0 0.93 0.07 14.48 100
992 3 1 3 3 1 1 0 0 0.06 0 0.94 0.06 13.64 95.14
1009 3 3 0 0 1 1 0 0.92 0 0 0.93 0.07 12.5 100
1024 3 3 3 3 1 1 0 0 0 0 1 0 15.29 93.33

2095.39 for scenario s = 1 with complete shutdown of all suppliers and assembly
plants. The corresponding service level is ranging from 100% to 50%.
s
Observe that no recovery supply or demand portfolio is selected, i.e., V11 + V12
s
=
V21 + V22 = V31 + V32 = V41 + V42 = 0, z1 = 1, z2 = 0, for scenarios s ∈ S with
s s s s s s s s

non disrupted all primary suppliers or primary assembly plant, respectively, e.g., for
scenarios s = 640, 1024.

10.5.1.1 Best-Case and Worst-Case Analysis

In order to provide more insights, in this subsection best-case and worst-case dis-
ruption scenarios with respect to cost and service level were identified. Since supply
disruptions can happen at any time and for any supplier and assembly plant, from
now on a disruption scenario is defined as a combination of disruptive event and its
start time. The start time ts of each disruptive event s ∈ S is assumed to be not greater
10.5 Computational Examples 299

Table 10.6 Supply and demand recovery portfolios, (V11 s + Vs ,Vs + Vs ,Vs + Vs ,Vs +
12 21 22 31 32 41
V42 , z1 , z2 ) and associated cost and service, for selected supply and production disruption scenarios
s s s

s ∈ S: hierarchical approach, g = 100

Disruption scenario Supply portfolio Demand portfolio Solution values
s λ1s λ2s λ3s λ4s λ5s λ6s i=1 i=2 i=3 i=4 j=1 j=2 Cost Service %
1 0 0 0 0 0 0 0 1 0 0 1 0 2095.39 50
7 0 0 1 2 0 0 0 0 0 0.9 1 0 119.38 66.3
17 0 1 0 0 0 0 0 0.98 0 0 1 0 286.49 58.91
19 0 1 0 2 0 0 0 0 0 0.98 1 0 126.93 58.91
26 0 1 2 1 0 0 0 0 0.46 0 1 0 114.82 66.67
28 0 1 2 3 0 0 0 0 0 0.46 1 0 75.32 66.67
35 0 2 0 2 0 0 0 0.95 0 0 1 0 98.73 66.67
36 0 2 0 3 0 0 0 0 0 0.95 1 0 75.93 66.67
42 0 2 2 1 0 0 0 0.43 0 0 1 0 96.65 66.67
44 0 2 2 3 0 0 0 0 0 0.43 1 0 75.41 66.67
49 0 3 0 0 0 0 0 0.93 0 0 1 0 78.73 66.67
50 0 3 0 1 0 0 0 0.93 0 0 1 0 78.73 66.67
321 1 0 0 0 0 1 1 0 0 0 0.57 0.43 329.7 85
322 1 0 0 1 0 1 1 0 0 0 0.72 0.28 329.7 85
323 1 0 0 2 0 1 0 0 0 1 0.57 0.43 101.39 85
324 1 0 0 3 0 1 0 0 0 1 0.67 0.33 43.1 100
325 1 0 1 0 0 1 0.9 0 0 0 0.66 0.34 319.49 94.63
326 1 0 1 1 0 1 0.9 0 0 0 0.66 0.34 319.49 94.63
328 1 0 1 3 0 1 0 0 0 0.9 0.67 0.33 43.01 100
329 1 0 2 0 0 1 0 0 0.48 0 0.72 0.28 87.01 95
332 1 0 2 3 0 1 0 0 0 0.48 0.67 0.33 42.58 100
339 1 1 0 2 0 1 0 0 0 0.98 0.72 0.28 99.21 87.24
340 1 1 0 3 0 1 0 0 0 0.98 0.67 0.33 43.17 100
341 1 1 1 0 0 1 0 0.88 0 0 0.67 0.33 250.62 95
344 1 1 1 3 0 1 0 0 0 0.88 0.67 0.33 43.07 100
353 1 2 0 0 0 1 0 0.95 0 0 0.68 0.32 67.71 98.33
356 1 2 0 3 0 1 0 0 0 0.95 0.67 0.33 43.26 100
357 1 2 1 0 0 1 0 0.85 0 0 0.68 0.32 67.32 98.33
360 1 2 1 3 0 1 0 0 0 0.85 0.67 0.33 43.16 100
361 1 2 2 0 0 1 0 0.43 0 0 0.67 0.33 65.63 98.33
381 1 3 3 0 0 1 0 0 0 0 1 0 74.99 66.67
382 1 3 3 1 0 1 0 0 0 0 1 0 74.99 66.67
383 1 3 3 2 0 1 0 0 0 0 1 0 74.99 66.67
384 1 3 3 3 0 1 0 0 0 0 1 0 74.99 66.67
(continued)
300 10 Selection of Primary and Recovery Supply …

Table 10.6 (continued)

Disruption scenario Supply portfolio Demand portfolio Solution values
s λ1s λ2s λ3s λ4s λ5s λ6s i=1 i=2 i=3 i=4 j=1 j=2 Cost Service %
640 1 3 3 3 1 0 0 0 0 0 1 0 14.99 93.33
646 2 0 1 1 1 0 0.9 0 0 0 1 0 67.21 72.96
647 2 0 1 2 1 0 0.9 0 0 0 1 0 67.21 72.96
648 2 0 1 3 1 0 0 0 0 0.9 1 0 22.34 86.67
657 2 1 0 0 1 0 0.98 0 0 0 1 0 75.14 65.57
658 2 1 0 1 1 0 0.98 0 0 0 1 0 75.14 65.57
659 2 1 0 2 1 0 0.98 0 0 0 1 0 75.14 65.57
660 2 1 0 3 1 0 0 0 0 0.98 1 0 23.6 85.57
661 2 1 1 0 1 0 0.88 0 0 0 1 0 64.93 75.2
662 2 1 1 1 1 0 0.88 0 0 0 1 0 64.93 75.2
663 2 1 1 2 1 0 0.88 0 0 0 1 0 64.93 75.2
664 2 1 1 3 1 0 0 0 0 0.88 1 0 20.17 88.91
673 2 2 0 0 1 0 0 0.95 0 0 1 0 63.48 68.58
676 2 2 0 3 1 0 0 0 0 0.95 1 0 20.68 88.58
677 2 2 1 0 1 0 0 0.85 0 0 1 0 53.47 78.21
844 1 0 2 3 1 1 0 0 0 0.48 0.87 0.13 9.05 100
845 1 0 3 0 1 1 0 0 0.07 0 0.93 0.07 15.13 93.33
961 3 0 0 0 1 1 1 0 0 0 0.9 0.1 14.52 100
976 3 0 3 3 1 1 0 0 0.07 0 0.93 0.07 15.13 93.33
977 3 1 0 0 1 1 0.98 0 0 0 0.92 0.08 14.45 100
992 3 1 3 3 1 1 0 0 0.04 0 0.96 0.04 15.19 93.33
1009 3 3 0 0 1 1 0 0.93 0 0 0.93 0.07 12.46 100
1024 3 3 3 3 1 1 0 0 0 0 1 0 14.99 93.33

than the maximum delivery lead time, maxi∈I,j∈J (τij ) = 4. Thus ts ∈ {1, 2, 3, 4},
and the total number of all potential scenarios to be considered is 1024×4=4096,
i.e., S = {1, . . . , 4096}. Now, each disruption scenario s ∈ S is represented by
vector λs = (λ1s , . . . , λ6s ), where λs = λs+1024 = λs+2048 = λs+3072 ; s ≤ 1024,
and its start time ts = 1 for s ≤ 1024, ts = 2 for 1025 ≤ s ≤ 2048, ts = 3 for
2049 ≤ s ≤ 3072 and ts = 4 for 3073 ≤ s ≤ 4096.
The probability Ps of realizing each disruption scenario s ∈ S is calculated as
follows:
Ps = β1 Ps1 Ps2 for s ≤ 1024,
Ps = β2 Ps−1024
1 2
Ps−1024 for 1025 ≤ s ≤ 2048,
Ps = β3 Ps−2048 Ps−2048 for 2049 ≤ s ≤ 3072,
1 2

Ps = β4 Ps−3072
1 2
Ps−3072 for 3073 ≤ s ≤ 4096,
where probabilities Psr ; r = 1, 2, s ≤ 1024 are defined at the beginning of this
section, and β1 , β2 , β3 and β4 are nonnegative constants such that: β1 +β2 +β3 +β4 =
1. In the computational examples, optimal solutions were determined for β1 = 0.1,
10.5 Computational Examples 301

Table 10.7 Worst-case disruption scenarios: g = 100, integrated approach

Disruption scenario Recovery portfolio Cost (a) Service Level (b)
(λ1s , λ2s , λ3s , λ4s , λ5s , λ6s ), ts s + V s , V s + V s , V s + V s , V s + V s , zs , zs )
(V11 12 21 22 31 32 41 42 1 2
Primary supply portfolio (v1 , v2 , v3 , v4 )= (0,0.04,0.78,0.18)
Worst-case disruption scenarios with respect to cost
(0,0,0,0,0,0), 1 (0,1,0,0,1,0) 2095 50
(0,0,0,0,0,1), 1 (0,1,0,0,0.5,0.5) 2071 75
(0,0,0,0,1,0), 1 (0,1,0,0,0.4,0.6) 2081 65
(0,0,0,0,1,1), 1 (0,1,0,0,0.75,0.25) 2038 75
(0,0,0,0,0,0), 2 (0,1,0,0,1,0) 2099 47
(0,0,0,0,0,1), 2 (0,1,0,0,0.77,0.23) 2076 70
(0,0,0,0,1,0), 2 (0,1,0,0,1,0) 2065 47
(0,0,0,0,1,1), 2 (0,1,0,0,0.47,0.53) 2043 70
(0,0,0,0,0,0), 3 (0,1,0,0,1,0) 2102 43
(0,0,0,0,0,1), 3 (0,1,0,0,0.78,0.22) 2081 65
(0,0,0,0,1,0), 3 (0,1,0,0,1,0) 2069 43
(0,0,0,0,1,1), 3 (0,1,0,0,0.43,0.57) 2048 65
(0,0,0,0,0,0), 4 (0,1,0,0,1,0) 2105 40(c)
(0,0,0,0,0,1), 4 (0,1,0,0,0.4,0.6) 2086 60
(0,0,0,0,1,0), 4 (0,1,0,0,1,0) 2072 40(c)
(0,0,0,0,1,1), 4 (0,1,0,0,0.8,0.2) 2052 60
Selected worst-case disruption scenarios with respect to service level
(0,0,0,0,0,0), 4 (0,1,0,0,1,0) 2105 40(c)
(0,0,0,0,1,0), 4 (0,1,0,0,1,0) 2072 40(c)
(a)
  
(e (u + Uis − qis )/D + ρis Uis /D + oi (γis vi + j∈J Vijs )) + j∈J (ψj wjs + (ϕj + ρjs )yjs /D
i∈I i s i  
+ εj t∈T xjt /D) + g(1 − j∈J t∈T xjts /D).
(b)
 
t∈T xjt /D × 100.
s
j∈J
(c) SL × 100, (10.35).

β2 = 0.2, β3 = 0.3, β4 = 0.4, i.e., disruptive events with later start times are
modelled to be more likely.
In order to identify the cost best-case and worst-case disruption scenarios, the
cost per product associated with optimal solution of DSupport_E for the integrated
approach (or PSupport and RDSupport_E for the hierarchical approach) is chosen
to be not greater than 10 and not less than 2000, respectively. For the examples with
unit penalty g = 100, the total number of cost best-case scenarios is 384 and 171,
respectively
  for the integrated and the hierarchical approach. The maximum service
level, j∈J t∈T xjts /D = 1, was achieved for all best-case scenarios with respect
to cost, which clearly shows that for the maximum service level, no penalty cost for
unfulfilled customer demand is incurred. The cost best-case disruption scenarios for
the hierarchical approach are a subset of those for the integrated approach.
In contrast, the number of cost worst-case disruption scenarios is much smaller,
16 and 53 scenarios, respectively for the integrated and the hierarchical approach.
302 10 Selection of Primary and Recovery Supply …

Table 10.8 Selected worst-case disruption scenarios with respect to cost and service level: g = 100,
hierarchical approach
Disruption scenario Recovery portfolio Cost (a) Service Level (b)
(λ1s , λ2s , λ3s , λ4s , λ5s , λ6s ), ts s + V s , V s + V s , V s + V s , V s + V s , zs , zs )
(V11 12 21 22 31 32 41 42 1 2
Primary supply portfolio: (v1 , v2 , v3 , v4 )=(0,0.07,0.93,0)
(0,1,1,0,0,0), 3 (0,0,0,0.88,0.98,0.02) 4469 40
(1,0,0,2,0,0), 3 (0,0.03,0.97,0,1,0) 6102 40
(1,0,1,0,0,0), 3 (0,0,0,0.90,0.38,0.62) 4469 40
(2,0,0,1,0,0), 3 (0,0,0.97,0.03,1,0) 4534 40
(0,1,1,0,0,0), 4 (0,0,0,0.88,0.98,0.02) 4469 40
(1,0,1,0,0,0), 4 (0,0,0,0.90,0.98,0.02) 4469 40
(a)
  
(e (u + Uis − qis )/D + ρis Uis /D + oi (γis vi + j∈J Vijs )) + j∈J (ψj wjs + (ϕj + ρjs )yjs /D
i∈I i s i  
+ εj t∈T xjt /D) + g(1 − j∈J t∈T xjts /D).
(b) SL × 100, (10.35).

The corresponding service level is ranging, respectively from 0.40 to 0.75 and from
0.40 to 0.65. The cost worst-case disruption scenarios for the integrated approach
are a subset of those for the hierarchical approach.
In order to identify the service level best-case and worst-case disruption scenarios,
the service level associated with optimal solution of DSupport_E for the integrated
approach (or PSupport and RDSupport_E for the hierarchical approach) is chosen
to be not less than 1 and not greater than SL = 0.40, (10.35), respectively. For the
examples with unit penalty g = 100, the total number of best-case scenarios with
service level j∈J t∈T xjts /D = 1 is 1290 and 720, respectively for the integrated
and the hierarchical approach. The corresponding cost per product is ranging, respec-
tively from 8.57 to 102.60 and from 9.05 to 81.44. The best-case scenarios for the
hierarchical approach are a subset of those for the integrated approach. In contrast,
the number of service level worst-case scenarios for the integrated approach is much
smaller, 71 scenarios, while for the hierarchical approach there are 573 scenarios.
The corresponding cost per product is ranging, respectively from 124 to 2105 and
from 68 to 6102. The worst-case scenarios for the integrated approach are a subset
of those for the hierarchical approach.
For the examples with unit penalty g = 100, Tables 10.7 and 10.8 present worst-
case disruption scenarios with respect to cost and service level, respectively for the
integrated and the hierarchical approach. For the integrated approach, Table 10.7
presents all 16 worst-case scenarios with cost per product
 not  less than 2000, and a
subset of two worst-case scenarios with service level j∈J t∈T xjts /D = SL = 0.4
and cost per product not less than 2000. For the hierarchical approach, Table 10.8
presents a subset of 6 worst-case scenarios with respect to cost such that cost per
product is not less than 0.7Rc (for the example problems, R = 6108, (10.70)) and
c

with respect to service level such that service level is j∈J t∈T xjts /D = SL = 0.4.
Tables 10.7 and 10.8 show disruption scenario, (λ1s , λ2s , λ3s , λ4s , λ5s , λ6s ), along
with its start time, ts , and the corresponding optimal recovery supply and demand
10.5 Computational Examples 303

portfolio, (V11
s
+ V12
s
, V21
s
+ V22
s
, V31
s
+ V32
s
, V41
s
+ V42
s
, z1s , z2s ), cost per product,
 
(ei (ui + Uis − qis )/D + ρis Uis /D + oi (γis vi + Vijs ))
i∈I j∈J
 
+ (ψj wjs + (ϕj + ρjs )yjs /D + εj xjts /D)
j∈J t∈T

+g(1 − xjts /D),
j∈J t∈T

 
and service level j∈J t∈T xjts /D.
The results in Table 10.7 demonstrate that for the worst-case scenarios with respect
to cost, all primary suppliers are completely shutdown and then the one with shorter
delivery lead time is selected as a single sourcing recovery portfolio. When both
assembly plants are shutdown, the primary plant is selected as a single recovery plant.
Otherwise, the recovery demand portfolio may include two plants. The results also
indicate that for a given disruption pattern, (λ1s , λ2s , λ3s , λ4s , λ5s , λ6s ), both cost per
product and service level are deteriorating with disruption start time, ts . For example
(see Table 10.7), for disruption pattern (0, 0, 0, 0, 0, 0), cost per product is increasing,
2095, 2099, 2102, 2105, and service level is decreasing, 50%, 47%, 43%, 40%,
respectively with start times, ts = 1, 2, 3, 4. For the hierarchical approach, Table 10.8
shows that for the worst-case scenario with respect to both service level and cost, all
assembly plants are shutdown, while primary suppliers are fully or partially disrupted.
The highest cost per product, 6102, is connected with scenario, (1, 0, 0, 2, 0, 0) and
start time ts = 3, where the primary suppliers were first completely shutdown and
then recovered and selected as recovery suppliers. The recovery supply portfolio
is very close to the primary supply portfolio and the recovery demand portfolio is
identical with the primary demand portfolio.
The computational experiments for the risk-neutral decision-making demonstrate
that:
• when all primary suppliers are completely shutdown, a single sourcing recovery
supply portfolio is usually selected,
• if all assembly plants are shutdown, the integrated approach selects the primary
plant as a single recovery plant, whereas the hierarchical approach may choose
multiple recovery plants,
• both cost per product and service level are deteriorating with disruption start time,
• the best-case and worst-case disruption scenarios for the hierarchical approach
are, respectively subsets and supersets of the corresponding scenarios for the
integrated approach.
The computational experiments were performed using the AMPL programming
language and the Gurobi 7.0.0 solver on a MacBookPro laptop with Intel Core i7
processor running at 2.8 GHz and with 16GB RAM. The solver was capable of finding
proven optimal solution for all examples with CPU time ranging from fraction of a
304 10 Selection of Primary and Recovery Supply …

second for deterministic models PSupport and RDSupport(s) to a few seconds for
stochastic models DSupport_E and RDSupport_E.

10.5.2 Risk-Averse Decision-Making

In this subsection some computational examples are presented to illustrate the risk-
averse selection of primary and recovery supply and demand portfolios to minimize
CVaR of cost or maximize CVaR of service level. In the computational experiments
the data sets provided at the beginning of this section were used, except for disruption
scenarios that are described in Sect. 10.5.1.1.
In addition, for minimization of CVaR of cost, unit penalty was fixed to g = 100.
In order to eliminate recovery cost of impacted suppliers that are not selected to
recovery supply portfolio, additional constraint was introduced into the models to
assign at least 5% of the total demand of parts to each recovery supplier under each
scenario, that is, 
Vijs ≥ 0.05Uis ; i ∈ I, s ∈ S.
j∈J

The solution results for the integrated approach are summarized in Tables 10.9 and
10.10, respectively for models DSupport_CV(c) and DSupport_CV(sl). The tables
show primary supply portfolio and expected recovery supply and demand portfolios
for different values of confidence level α = 0.5, 0.75, 0.9, 0.95, 0.99. The solution
values, CV aRc and CV aRsl , are presented along with the associated expected values
of cost, E c , and service level, E sl . While CV aRc and V aRc increase, and CV aRsl
and V aRsl decrease with the confidence level α, the associated expected values, E c
and E sl , are not varying monotonously.
In addition, Tables 10.9 and 10.10 show best-case and worst-case scenarios with
respect to cost, such that cost per product associated with optimal solution is not
greater than 10 and not less than 2000, respectively. The tables also show best-case
and worst-case scenarios with respect to service such that service level is equal to 1
and to SL = 0.40, (10.35), respectively. Tables 10.9 and 10.10, clearly indicate that
the best-case scenarios with respect to cost are associated with the maximum service
level 100%, i.e., with no penalty cost for unfulfilled demand. On the other hand,
however, the best case scenarios with respect to service, for which service level also
attains 100%, can be associated with a higher cost, e.g., due to required recovery
processes. The results indicate that the best case scenarios with respect to cost are a
subset of the best case scenarios with respect to service.
For model DSupport_CV(sl), where the objective function does not account for
any cost parameters, the worst-case scenarios with respect to service are a subset of
the worst-case scenarios with respect to cost, see Table 10.10.
In addition, Tables 10.9 and 10.10 show the worst-case scenarios with respect
to both cost and service level such that cost per product is not less than 2000 and
service level is equal to SL = 0.40. These scenarios are a subset of maximum cost
10.5 Computational Examples 305

Table 10.9 Risk-averse solutions for model DSupport_CV(c), g = 100

Confidence level α 0.50 0.75 0.90 0.95 0.99
Var. = 274413, Bin. = 40964, Cons. = 491761, Nonz. = 4451648 (a)
CV aRc 33.98 52.07 81.92 82.84 84.86
V aRc 15.66 16.03 77.84 82.33 82.33
Ec 24.56 24.80 39.76 33.14 34.93
E sl 100% 88.72 89.91 76.74 81.77 79.86
Primary Supply Portfolio: Supplier(% of total demand) (b) 2(23) 2(23) 2(24) 2(23) 2(23)
3(77) 3(45) 3(63) 3(77) 3(77)
4(32) 4(13)
Exp.Recovery Supply Portfolio: Supplier(% of total demand) (c) 1(0.98) 1(1.51) 1(1.61) 1(3.34) 1(1.97)
2(1.68) 2(2.70) 2(1.61) 2(1.46) 2(2.56)
3(0.67) 3(1.59) 3(2.04) 3(0.48) 3(0.56)
4(4.04) 4(1.26) 4(1.98) 4(2.09) 4(2.27)
Exp.Recovery Demand Portfolio: Supplier(% of total demand) (d) 1(95.80) 1(94.65) 1(97.53) 1(98.03) 1(97.65)
2(2.93) 2(4.12) 2(2.21) 2(1.46) 2(1.86)
Best-Case and Worst-Case Scenarios
Number of Cost Best-Case Scenarios (Cost≤ 10) 15 112 1 0 4
Minimum service level % 100 100 100 100
Maximum service level % 100 100 100 100
Number of Service Best-Case Scenarios (Service = 100%) 706 970 63 26 39
Minimum cost 8.53 8.57 10 33.25 8.90
Maximum cost 212 212 212 82.73 212
Number of Cost Worst-Case Scenarios (Cost≥ 2000) 18 16 16 16 16
Minimum service % 40
Maximum service % 75
Number of Service Worst-Case Scenarios (Service=SL) 73 2 787 392 389
Minimum cost 124 2072 68 68 68
Maximum cost 4337 2105 2105 2105 2105
Number of Cost and Service Worst-Case Scenarios 4 2 2 2 3
(Cost≥ 2000 and Service = SL)
Maximum Cost 4337 2105 2105 2105 2105
(a) Var. = number of variables, Bin. = number of binary variables,
Cons.=number of constraints, Nonz. = number of nonzero coefficients.
(b) 1(v × 100), 2(v × 100), 3(v × 100), 4(v × 100).
(c) 1(
1 2 3 4  
j∈J,s∈S Ps V1j ×100), 2( j∈J,s∈S Ps V2j ×100), 3( j∈J,s∈S Ps V3j ×100), 4( j∈J,s∈S Ps V4j ×
s s s s

100).  
(d) 1(
s∈S Ps z1 × 100), 2( s∈S Ps z2 × 100).
s s

worst-case scenarios with respect to service. Such worst-case scenarios for model
DSupport_CV(c) are presented in more details in Table 10.11.
Table 10.11 indicates that the worst-case scenarios with respect to both cost and
service level are nearly identical for all confidence level, α, except for the lowest
α = 0.5. Under the worst-case scenario, all suppliers and assembly plants are either
completely shutdown or hit by major disruption. When all primary suppliers are
shutdown, only one of them is selected for recovery to reduce recovery cost. In
306 10 Selection of Primary and Recovery Supply …

Table 10.10 Risk-averse solutions for model DSupport_CV(sl)

Confidence level α 0.50 0.75 0.90 0.95 0.99
Var. = 274413, Bin. = 40964, Cons. = 491761, Nonz. = 4451648 (a)
CV aRsl 100% 85.53 77.73 61.32 60 60
V aRsl 100% 93.33 93.33 66.67 60 60
E sl 100% 89.82 89.74 77.50 75.29 75.44
Ec 127 95 137 149 177
Primary Supply Portfolio: Supplier(% of total demand) (b) 1(34) 1(26) 1(26)
2(23) 2(23) 2(31) 2(23) 2(23)
3(37) 3(37) 3(15) 3(22) 3(22)
4(40) 4(40) 4(20) 4(29) 4(29)
Exp.Recovery Supply Portfolio: Supplier(% of total demand) (c) 1(2.19) 1(2.30) 1(1.86) 1(1.67) 1(1.21)
2(1.88) 2(1.33) 2(0.96) 2(1.25) 2(1.08)
3(1.82) 3(1.62) 3(1.09) 3(1.76) 3(2.22)
4(1.11) 4(1.75) 4(0.75) 4(0.79) 4(0.96)
Exp.Recovery Demand Portfolio: Supplier(% of total demand) (d) 1(94.13) 1(94.16) 1(96.10) 1(95.63) 1(95.98)
2(4.68) 2(4.61) 2(3.51) 2(4.03) 2(3.58)
Best-Case and Worst-Case Scenarios
Number of Cost Best-Case Scenarios (Cost≤ 10) 12 3 1 0 0
Minimum service level % 100 100 100
Maximum service level % 100 100 100
Number of Service Best-Case Scenarios (Service = 100%) 312 262 22 23 19
Minimum cost 8.71 9.27 9.71 11.04 46
Maximum cost 8676 7044 6379 7044 4478
Number of Cost Worst-Case Scenarios (Cost≥ 2000) 1602 1589 1425 1642 1987
Minimum service % 40
Maximum service % 100
Number of Service Worst-Case Scenarios (Service = SL) 367 472 581 478 260
Minimum cost 103 102 126 126 159
Maximum cost 10436 8469 7139 8470 7104
Number of Cost and Service Worst-Case Scenarios 192 260 281 257 183
(Cost≥ 2000 and Service = SL)
Maximum Cost 10436 8469 7139 8470 7104
(a) Var. = number of variables, Bin. = number of binary variables,
Cons. = number of constraints, Nonz. = number of nonzero coefficients.
(b) 1(v × 100), 2(v × 100), 3(v × 100), 4(v × 100).
(c) 1(
1 2 3 4  
j∈J,s∈S Ps V1j ×100), 2( j∈J,s∈S Ps V2j ×100), 3( j∈J,s∈S Ps V3j ×100), 4( j∈J,s∈S Ps V4j ×
s s s s

100).  
(d) 1(
s∈S Ps z1 × 100), 2( s∈S Ps z2 × 100).
s s
10.5 Computational Examples 307

Table 10.11 Worst-case scenarios with respect to cost and service level: model DSupport_CV(c),
g = 100
Disruption scenario Recovery portfolio Cost (a) Service Level (b)
(λ1s , λ2s , λ3s , λ4s , λ5s , λ6s ), ts (V11s + Vs ,Vs + Vs ,Vs + Vs ,Vs +
12 21 22 31 32 41
V42 , z1 , z2s )
s s

Confidence level α = 0.5, primary suppliers i = 2, 3

(0,1,0,1,0,0), 1 (0,0.32,0.6,0,0.4,0.6) 4337 40(c)
(1,0,0,1,0,0), 1 (0,0,1,0,0.18,0.82) 4135 40
(0,0,0,0,0,0), 4 (0,1,0,0,1,0) 2105 40
(0,0,0,0,1,0), 4 (0,1,0,0,1,0) 2072 40
Confidence level α = 0.75, 0.90, primary suppliers i = 2, 3, 4
(0,0,0,0,0,0), 4 (0,1,0,0,1,0) 2105 40
(0,0,0,0,1,0), 4 (0,1,0,0,1,0) 2072 40
Confidence level α = 0.95, primary suppliers i = 2, 3
(0,0,0,0,0,0), 4 (0,1,0,0,1,0) 2105 40
(0,0,0,0,1,0), 4 (0,1,0,0,1,0) 2072 40
Confidence level α = 0.99, primary suppliers i = 2, 3
(0,0,0,0,0,1), 1 (0,1,0,0,0.4,0.6) 2105 40
(0,0,0,0,0,0), 4 (0,1,0,0,1,0) 2105 40
(0,0,0,0,1,0), 4 (0,1,0,0,1,0) 2072 40
(a)
  
(e (u + Uis − qis )/D + ρis Uis /D + oi (γis vi + j∈J Vijs )) + j∈J (ψj wjs + (ϕj + ρjs )yjs /D
i∈I i s i  
+ εj t∈T xjt /D) + g(1 − j∈J t∈T xjts /D).
(b)
 
t∈T xjt /D × 100.
s
j∈J
(c) SL × 100, (10.35).

contrast to selection of recovery plants, where both shutdown plants may be selected
for recovery. When both plants are selected as recovery plants, a larger fraction of
demand for products is assigned to plant j = 2. While per period capacity, c2 = 5000
of plant j = 2 is half of the capacity, c1 = 10000 of plant j = 1, its recovery time
PRT (2, 0) = 5 is half of the recovery time PRT (1, 0) = 10. In the examples,
however,
 the largest portion of total cost is recovery cost of impacted suppliers,
i∈I ρis Ui /D.
s

Tables 10.12 and 10.13 provide risk-averse solutions for the hierarchical approach
and models RDSupport_CV(c) and RDSupport_CV(sl), respectively, with the pri-
mary supply portfolio determined using model PSupport. A general comparison of
solution results in Tables 10.9 and 10.12 and in Tables 10.10 and 10.13 demonstrates
that both the integrated and the hierarchical approach lead to similar risk-averse solu-
tions. The main differences are less diversified primary supply portfolios determined
by the deterministic upper level model PSupport: a dual sourcing for cost-based
objective function (10.37) and a single sourcing for service-based objective function
(10.77).
The computational experiments for the risk-averse decision-making demonstrate
that:
• the integrated decision-making selects a more diversified primary supply portfolio
to hedge against all potential disruption scenarios,
308 10 Selection of Primary and Recovery Supply …

Table 10.12 Risk-averse solutions for model RDSupport_CV(c), g = 100

Primary supply portfolio: (v1 , v2 , v3 , v4 ) = (0,0.07,0.93,0)
Confidence level α 0.50 0.75 0.90 0.95 0.99
Var. = 247105, Bin. = 27456, Cons. = 310590, Nonz. = 3050044 (a)
CV aRc 33.43 51.84 81.51 82.18 84.22
V aRc 15.01 15.01 78.34 81.67 81.67
Ec 24.00 24.08 32.16 32.19 32.63
E sl 100% 88.83 89.61 82.46 83.07 81.54
Exp.Recovery Supply Portfolio: Supplier(% of total demand) (b) 1(0.32) 1(0.37) 1(1.44) 1(2.18) 1(1.64)
2(2.04) 2(2.56) 2(2.44) 2(2.07) 2(3.11)
3(0.38) 3(2.32) 3(0.87) 3(1.48) 3(0.46)
4(0.13) 4(4.93) 4(3.67) 4(2.36) 4(2.88)
Exp.Recovery Demand Portfolio: Supplier(% of total demand) (c) 1(97.56) 1(95.50) 1(98.01) 1(96.41) 1(97.06)
2(1.15) 2(3.31) 2(0.84) 2(2.40) 2(1.84)
Best-Case and Worst-Case Scenarios
Number of Cost Best-Case Scenarios (Cost≤ 10) 11 5 1 3 3
Minimum service level % 100
Maximum service level % 100
Number of Service Best-Case Scenarios (Service = 100%) 555 744 146 121 106
Minimum cost 9.19 9.32 9.48 9.72 9.62
Maximum cost 82.05 210 87 86.54 125
Number of Cost Worst-Case Scenarios (Cost≥ 2000) 33 16 16 17 16
Minimum service % 40
Maximum service % 47 75 75 95 75
Number of Service Worst-Case Scenarios (Service = SL) 1038 9 670 541 578
Minimum cost 68 146 68 68 68
Maximum cost 4135 2105 2105 2105 2105
Number of Cost and Service Worst-Case Scenarios 31 2 2 2 2
(Cost≥ 2000 and Service = SL)
Maximum Cost 4135 2105 2105 2105 2105
(a) Var. = number of variables, Bin. = number of binary variables,
Cons. = number
 of constraints, Nonz. = number
 of nonzero coefficients. 
(b) 1( j∈J,s∈S Ps V1js × 100), 2( j∈J,s∈S Ps V2js × 100), 3( j∈J,s∈S Ps V3js ×

100), 4( j∈J,s∈S Ps V4js × 100).
(c) 1(
 
s∈S Ps z1 × 100), 2( s∈S Ps z2 × 100).
s s

• when all primary suppliers are completely shutdown, a single sourcing recovery
supply portfolio is usually selected,
• multiple recovery plants may be selected even if all assembly plants are shutdown,
in particular for a low confidence level,
• the best-case and worst-case disruption scenarios for the hierarchical approach
are, respectively subsets and supersets of the corresponding scenarios for the
integrated approach.
Notice that very similar findings were identified for the risk-neutral decision-
making in Sect. 10.5.1.
10.5 Computational Examples 309

Table 10.13 Risk-averse solutions for model RDSupport_CV(sl)

Primary supply portfolio: (v1 , v2 , v3 , v4 ) = (0,1,0,0)
Confidence level α 0.50 0.75 0.90 0.95 0.99
Var. = 210553, Bin. = 19456, Cons. = 254876, Nonz. = 2338786 (a)
CV aRsl 100% 84.32 75.36 60.15 60 60
V aRsl 100% 93.33 93.33 63.33 60 60
E sl 100% 88.83 88.84 86.23 85.93 85.95
Ec 62.34 62.40 61.38 62.64 61.57
Exp.Recovery Supply Portfolio: Supplier(% of total demand) (b) 1(1.44) 1(1.43) 1(1.24) 1(0.80) 1(1.41)
2(2.04) 2(2.04) 2(1.89) 2(2.14) 2(1.54)
3(0.38) 3(0.38) 3(0.42) 3(0.52) 3(0.59)
4(0.13) 4(0.13) 4(0.43) 4(0.53) 4(0.44)
Exp.Recovery Demand Portfolio: Supplier(% of total demand) (c) 1(97.56) 1(97.55) 1(96.73) 1(96.75) 1(97.27)
2(1.15) 2(1.16) 2(2.04) 2(2.02) 2(1.50)
Best-Case and Worst-Case Scenarios
Number of Cost Best-Case Scenarios (Cost≤ 10) 0
Minimum service level %
Maximum service level %
Number of Service Best-Case Scenarios (Service = 100%) 2 2 9 3 7
Minimum cost 2037 2037 44 12 474
Maximum cost 2071 2071 4478 6376 6342
Number of Cost Worst-Case Scenarios (Cost≥ 2000) 1652 1703 1374 1595 1604
Minimum service % 40
Maximum service % 100
Number of Service Worst-Case Scenarios (Service = SL) 1848 1711 538 458 257
Minimum cost 99 99 101 134 167
Maximum cost 6773 6773 8469 8469 7104
Number of Cost and Service Worst-Case Scenarios 1158 1135 338 344 196
(Cost≥ 2000 and Service=SL)
Maximum Cost 6773 6773 8469 8469 7104
(a) Var. = number of variables, Bin. = number of binary variables,
Cons. = number
 of constraints, Nonz. = number
 of nonzero coefficients. 
(b) 1( j∈J,s∈S Ps V1js × 100), 2( j∈J,s∈S Ps V2js × 100), 3( j∈J,s∈S Ps V3js ×

100), 4( j∈J,s∈S Ps V4js × 100).
(c) 1(
 
s∈S Ps z1 × 100), 2( s∈S Ps z2 × 100).
s s

Overall, the results of computational experiments indicate that the proposed port-
folio approach and developed SMIP models with an embedded network flow problem
is a flexible and efficient tool for supply chain disruption management. The approach
leads to SMIP formulations with a strong LP relaxation and has proven to be com-
putationally very efficient. CPU time required to find proven optimal solutions for
all examples, using commercially available software for MIP, was ranging from
fraction of a second for model RDSupport_CV(sl) to less than an hour for model
DSupport_CV(c).
310 10 Selection of Primary and Recovery Supply …

10.6 Notes

The literature on mitigation the impact of disruption risks and optimization of a

recovery process in supply chains is limited, e.g., Gurnani et al. (2012). For exam-
ple, mitigation and contingency/recovery actions were studied by Tomlin (2006)
in a dual-sourcing setting, one unreliable supplier and another reliable and more
expensive. A buyer that suffers a supply shortage can buy from a more expensive
alternate supplier or produce less, and its decision depends on its inventory. If reli-
able supplier has volume flexibility, contingent rerouting by temporarily increasing
its production may prove to be an effective way to speed up recovery process. The
author established that, along with cost, percentage of supplier uptime, disruption
length, capacity, and flexibility, play an important role in determining a buyers dis-
ruption management strategy. MacKenzie et al. (2014) proposed a model for severe
disruptions in which a disruption simultaneously impacts several suppliers, while
the suppliers’ customers (firms) may face supply shortages. The model incorporated
decisions made by both suppliers and firms during the disruption of random duration.
The decisions may include whether or not suppliers move production to an alternate
facility, hold parts inventory; a firm purchases parts from alternate suppliers that
are not impacted, helps a primary supplier recover more quickly or holds finished
products inventory. The supplier and the firm optimal decisions are expressed in
terms of model parameters, e.g., the cost of each strategy, the chances of losing
business, and the probability of suppliers recovery. The impact of a disruption was
evaluated based on the optimal response of an existing network once that disruption
has occurred. The authors derived threshold parameters to support decisions whether
and when a supplier moves production to an alternate facility and how much the firm
should produce during the disruption to trade off between two conflicting objectives:
maximizing profit and service level. They applied model to a simulation based on
the Great East Japan earthquake and tsunami in March 2011. Ivanov et al. (2016)
analyzed seven proactive supply chain structures and computed recovery policies to
re-direct material flows in the case of two disruption scenarios. They assessed the
performance impact of the duration of disruptions and the costs of recovery for both
service level and costs with the help of a supply chain (re)planning model contain-
ing elements of system dynamics and linear programming. A review of literature
on Operations Research/Management Science models for supply chain disruptions
was presented by Snyder et al. (2016). They discussed 180 scholarly works on the
topic, organised into six categories: evaluating supply disruptions; strategic deci-
sions; sourcing decisions; contracts and incentives; inventory; and facility location.
The future research should concentrate on relaxations of the various simplified
assumptions used to formulate the problem (see, Sect. 10.2). For example, a sin-
gle recovery mode defined by constant recovery times, TTR(i, l), PRT (j, l), and the
associated recovery costs, CTR(i, l), PRC(j, l), for each supplier i, each assembly
plant j and each disruption level l, can be replaced by a number of available recov-
ery modes, each represented by different values of recovery time and the associated
recovery cost. Then, introduction of new recovery mode selection variables for each
10.6 Notes 311

supplier and each assembly plant can enhance the problem formulation. In addition,
the inventory of parts at the impacted primary assembly plant may be non available
for transshipment or the transshipment itself can be non available. Then, the recovery
supply portfolio should be selected, even after full delivery of all required parts to the
impacted primary assembly plant. In a more general setting, both recovery time and
recovery cost can be modeled as random parameters, e.g., Schmitt (2011),Schmitt
and Singh (2012).

Problems

10.1 In the SMIP models presented in this chapter introduce delay penalty when
demand dt for period t is not met by that period.

10.2 Enhance the SMIP models presented in this chapter for the case in which
the inventory of parts at the impacted primary assembly plant is not available for
transshipment.

10.3 Formulate the mean-risk models DSupport_ECV(c) and DSupport_ECV(sl)

to minimize expected cost and CVaR of cost or maximize expected service level and
CVaR of service level, respectively.

10.4 Enhance model DSupport_E for multiple recovery modes defined by differ-
ent recovery time, TTR(i, l), PRT (j, l), and the associated recovery cost, CTR(i, l),
PRC(j, l), for each supplier i, each assembly plant j and each disruption level l.

10.5 Explain why the best-case and worst-case disruption scenarios for the hierar-
chical approach are, respectively subsets and supersets of the corresponding scenarios
for the integrated approach.
 Part V
Information Flow Disruption Management
Chapter 11
Selection of Cybersecurity Safequards
Portfolio

11.1 Introduction

Nowadays information technology (IT) becomes one of the core elements of global
supply chains that are increasingly at risk of a disruption of their information systems.
IT may improve a supply chain robustness and resilience, e.g., Pereira (2009). For
example, an important determinant of supply chain competitiveness is supply chain
visibility defined as “the extent to which actors within a supply chain have access to or
share information which they consider as key or useful to their operations and which
they consider will be of mutual benefit” (Lee et al. 2014). The supply chain visibility
is accomplished through inter organizational information systems, which should be
very well protected against cyber-attacks. This chapter deals with the selection of
cybersecurity safeguards portfolio to prevent and mitigate the impact of information
flow disruptions in a supply chain. The objective of IT security planning in supply
chains is to protect information flows against disruptions and the supply chain assets
against a compromise in the area of confidentiality, integrity or availability, where
asset types may include systems and applications, networks, end-user systems, and
off line media and devices. The various actions developed to prevent intrusions or to
mitigate the impact of successful breaches and information flow disruptions are called
security safeguards, controls or countermeasures. In view of the variety of methods
used by attackers to infiltrate a supply chain IT infrastructure and disrupt operations,
a wide range of different countermeasures are developed. Some countermeasures
are used to limit physical access to an IT infrastructure, (e.g., key entry systems,
retinal or finger print scans), other block access or protect privacy over networks
(e.g., firewalls, data encryption, or virus and spyware scanners), while additional
countermeasures are designed to permit recovery after a successful intrusion. A
common practice is information redundancy. Firms back up supplier data, customer
data, bill-of-material data, etc. However, backup of hard drives may not be sufficient
for a supply chain to continue its operations, more often backup of the system itself

© Springer International Publishing AG 2018 315

T. Sawik, Supply Chain Disruption Management Using Stochastic Mixed
Integer Programming, International Series in Operations Research
& Management Science 256, DOI 10.1007/978-3-319-58823-0_11
316 11 Selection of Cybersecurity Safequards Portfolio

is required. For example, if a widespread power outage severely impacts business

continuity by compromising information systems.
In practice, even the most sophisticated countermeasures cannot be expected to
completely block attacks as new attack profiles proliferate and the time and cost
required to adapt to them the implemented countermeasures are not negligible. The
potential threats, however, cannot be ignored and without countermeasures the losses
resulting from many more breaches and disruptions of information flows could easily
excess the costs of countermeasure implementations. In this chapter a scenario-based
SMIP approach with CVaR as a risk measure is proposed for the decision-making.
Given a set of potential threats and a set of available countermeasures, the decision
maker needs to decide which countermeasure to implement under limited budget to
minimize potential losses from successful cyber-attacks. The selection of counter-
measures is based on their effectiveness of blocking different threats, implementation
costs and probability of potential attack scenarios. The risk-neutral, risk-averse mod-
els and, in particular, mean-risk trade-off model provide the decision maker with a
simple tool for balancing expected and worst-case losses and for shaping of the result-
ing cost distribution through the selection of optimal subset of countermeasures for
implementation, i.e., the selection of optimal cybersecurity safeguards portfolio.
The following SMIP models are presented in this chapter:

NCP_E for risk-neutral selection of cybersecurity safeguards portfolio to

minimize expected loss (nonlinear model);
CP_E for risk-neutral selection of cybersecurity safeguards portfolio to
minimize expected loss (linear model);
CP_CV for risk-averse selection of cybersecurity safeguards portfolio to
minimize CVaR of loss;
CP_EB for risk-neutral selection of cybersecurity safeguards portfolio to
minimize expected loss and required budget;
CP_CVB for risk-averse selection of cybersecurity safeguards portfolio to
minimize CVaR of loss and required budget;
CP_EBCV for mean-risk selection of cybersecurity safeguards portfolio to
optimize trade-off between expected loss plus required budget and CVaR
of loss.

In the computational experiments described in Sect. 11.6, an application of the

developed models is illustrated with numerical examples.

11.2 Problem Description

Let I = {1, . . . m} be the set of m threats and J = {1, . . . n} the set of n countermea-
sures (for notation used, see Table 11.1). Denote by pi the probability of threat i, i.e.,
attack episode of threat i occurs with probability pi , or not at all with probability
11.2 Problem Description 317

Table 11.1 Notation: selection of cybersecurity safeguards portfolio

Indices
i = threat, i ∈ I
j = countermeasure, j ∈ J
s = attack scenario, s ∈ S
l = countermeasure implementation level, l ∈ L = {0, 1} (l = 0 - off, l = 1 - on)
Input Parameters
ai = loss caused by a successful attack episode of threat i
B = available budget for countermeasures
c j = cost of countermeasure j
Is = subset of threats in scenario s
pi = probability of threat i
qi jl = max{1 − l, ri j } - proportion of threats i that survive if countermeasure j is implemented
at level l (qi j0 = 1 and qi j1 = ri j )
ri j = proportion of threats i that survive if countermeasure j is implemented
α = confidence level

(1 − pi ). Let Ps be the probability that attack scenario s is realized, where each

scenario s ∈ S is comprised of a unique subset Is ⊂ I of threats that appear in the
cyberattack, and S = {1, . . . , S} is the index set of all scenarios (note that there are
a total of S = 2m potential attack scenarios). The probability of attack scenario s in
the presence of independent threat events is
 
Ps = pi · (1 − pi ). (11.1)
i∈Is i ∈I
/ s

Let ri j ∈ [0, 1] be the proportion of threats i that survive if countermeasure j is

implemented, where ri j = 0 indicates that countermeasure j totally prevents suc-
cessful attacks of threat i, whereas ri j = 1 denotes that countermeasure j is totally
incapable of mitigating threat i.
Note, that even the most sophisticated countermeasures, in practice are rarely
expected to completely block attacks, and hence the survival proportions ri j are
rarely equal to 0.
The blocking effectiveness of each countermeasure is assumed to be independent
whether or not it is used alone or together with other countermeasures, and the
proportion of successful attacks of threats type i that survive all countermeasures in
the subset JS ⊆J of selected countermeasures is a multiplication of proportions
ri j , j ∈ JS. Let j∈JS ri j be the proportion of threats i that survive if subset JS of
countermeasures is implemented. Note that the expected proportion of successful 
attacks of threat type i for the subset JS of selected countermeasures is pi j∈JS ri j .
Denote by c j the cost of implementing countermeasure j, and by ai the cost of
(loss from) a successful attack episode of threat i. The available budget B for selected
318 11 Selection of Cybersecurity Safequards Portfolio

countermeasures implementation is limited. The subset of selected countermeasures

JS ⊆ J must satisfy the available budget constraint, j∈JS c j ≤ B, i.e., the total
expenditures on the selected countermeasures cannot exceed the available budget B.
Furthermore, the selected countermeasures are assumed to be feasible with respect
to various additional constraints, e.g., they cannot contain mutually exclusive coun-
termeasures or they must contain all countermeasures contingent on each other, etc.
The decision maker needs to decide which countermeasures to select to minimize
losses from surviving occurrences of threats under limited budget for countermea-
sures implementation.

11.3 Models for Risk-Neutral Decision-Making

In this section SMIP models are proposed for a risk-neutral selection of optimal
cybersecurity safeguards portfolio, i.e., the risk-neutral selection of optimal sub-
set of countermeasures for implementation. The problem variables are defined in
Table 11.2.
Let L = {0, 1} be the set of implementation levels of each countermeasure, where
l = 0 denotes that a particular countermeasure is not selected for implementation,
otherwise l = 1. The decision whether or not to select a particular countermeasure
will be represented by a binary variable u jl , j ∈ J, l ∈ L, where u jl = 1, if counter-
measure j is implemented at level l, otherwise u jl = 0, that is, countermeasure j is
selected for implementation if u j1 = 1 and u j0 = 0, otherwise u j1 = 0 and u j0 = 1.
The above definition implies that each countermeasure j is selected at exactly one
level, i.e., l∈L u jl = 1.
Denote by qi jl = max{1 − l, ri j } the proportion of threats i that survive if coun-
termeasure j is implemented at level l, that is qi j0 = 1 and qi j1 = ri j .

11.3.1 Nonlinear Model for Risk-Neutral Decision-Making

For a risk-neutral decision-making the overall quality of the selected countermeasure

portfolio can be measured by the expected cost of losses from successful attacks.
The proportion of successful attacks of threats type i that survive all selected
countermeasures is a multiplication of individual proportions of the selected coun-
termeasures and can be expressed as

( qi jl u jl ).
j∈J l∈L

As a result, the expected loss from successful attacks is given by a nonlinear

formula
11.3 Models for Risk-Neutral Decision-Making 319
 
Ps ai ( ( qi jl u jl )). (11.2)
s∈S i∈Is j∈J l∈L

Table 11.2 Variables: selection of cybersecurity safeguards portfolio

First stage variables
u jl = 1, if countermeasure j is implemented at level l, otherwise u jl = 0 (countermeasure
selection)
vi jl = proportion of surviving occurrences of threat i to be addressed by countermeasure j at
level l
wi j = proportion of surviving occurrences of threat i that passed countermeasures 1 through j
Wi = proportion of successful attacks of type i
Auxiliary variables
VaR Loss-at-Risk, the targeted loss such that for a given confidence level α, for 100α% of the
scenarios, the outcome is below VaR
Cs ≥ 0, the tail loss for attack scenario s, i.e., the amount by which losses in scenario s exceed
VaR

The 0–1 nonlinear stochastic programming program NCP_E for risk-neutral

selection of optimal subset of security safeguards is formulated below.

NCP_E: Nonlinear model for risk-neutral selection of Cybersecurity

safeguards Portfolio to minimize expected loss
Minimize (11.2)
subject to
1. Countermeasure selection constraints:
– each countermeasure is selected at exactly one level (i.e., implemented or
not implemented),
– the expenditures on selected countermeasures cannot exceed available
budget,


u jl = 1; j ∈ J (11.3)
l∈L

c j u j1 ≤ B (11.4)
j∈J

3. Integrality conditions:

u jl ∈ {0, 1}; j ∈ J, l ∈ L . (11.5)

320 11 Selection of Cybersecurity Safequards Portfolio

The 0–1 nonlinear stochastic program NCP_E is computationally hard for solv-
ing, even for small size instances of the problem. The next subsection describes a
recursive procedure that is capable of computing the nonlinear objective function
(11.2) using a set of linear equations.

11.3.2 Linear Model for Risk-Neutral Decision-Making

The nonlinear objective function (11.2) can be replaced with a formula


Ec = Ps ai Wi , (11.6)
s∈S i∈Is

where

Wi = ( qi jl u jl ); i ∈ I, (11.7)
j∈J l∈L

is the proportion of successful attacks of type i.

In order to compute Wi for each threat i, a recursive procedure is proposed below.
Denote by
 
wi j = ( qikl u kl ); i ∈ I, j ∈ J, (11.8)
k∈J :k≤ j l∈L

the proportion of surviving occurrences of threat i that passed countermeasures 1

through j. Note that (cf. (11.7), (11.8))

Wi = win ; i ∈ I. (11.9)

For each threat i ∈ I and countermeasure j ∈ J , wi j can be calculated recursively

as follows.
The initial condition that explicitly prescribes the first term is

wi1 = qi1l u 1l ; i ∈ I. (11.10)
l∈L

The recurrence formula by means of which the remaining terms are determined
inductively is

wi j = ( qi jl u jl )wi, j−1 , i ∈ I, j ∈ J. (11.11)
l∈L
11.3 Models for Risk-Neutral Decision-Making 321

In order to eliminate nonlinear terms in the right-hand side of Eq. (11.11), define
an auxiliary variable

vi jl = u jl wi j−1 ; i ∈ I, j ∈ J, l ∈ L , (11.12)

where wi0 = 1 for all i, i.e., all threat events of each type i are capable of attacking
supply chain IT infrastructure. Note that vi jl represents the proportion of surviving
occurences of threat i to be addressed by countermeasure j at level l
Substituting vi jl = u jl wi j−1 into Eq. (11.11) yields

wi j = qi jl vi jl ; i ∈ I, j ∈ J, (11.13)
l∈L

and, in particular, for j = n,


win = qinl vinl ; i ∈ I, (11.14)
l∈L

where win = Wi for all i ∈ I , (11.9).

On the other hand, replacing j with j + 1 in Eq. (11.12) gives

vi, j+1,l = u j+1,l wi j ; i ∈ I, j ∈ J : j < n, (11.15)

 Summing both sides of Eq. (11.15) on l for each i and j, and using the fact that
l∈L u j+1,l = 1; j ∈ J , (11.3), yields


vi, j+1,l = wi j ; i ∈ I, j ∈ J : j < n. (11.16)
l∈L

Comparison of Eqs. (11.13) and (11.16) produces to the following relation

 
qi jl vi jl = vi, j+1,l ; i ∈ I, j ∈ J : j < n, (11.17)
l∈L l∈L

that is, for each threat i, the proportion of occurrences which survive countermeasure
j are addressed by countermeasure j + 1.
Finally, setting j = 1 and summing both sides of Eq. (11.12) on l, using the fact
that wi0 = 1 for all i and l∈L u 1l = 1, (11.3), yields

vi1l = 1; i ∈ I, (11.18)
l∈L

that is, all threat events of each type i are addressed by countermeasure j = 1.
The above procedure eliminates all variables wi j for each i, except for win . Sum-
marizing, the proportion of successful attacks Wi = win for each threat i can be
322 11 Selection of Cybersecurity Safequards Portfolio

calculated recursively, using Eqs. (11.18), (11.17) and (11.14) with win replaced by
Wi .
Note that the linearizing procedure would not be possible if the implementation
level l ∈ L = {0, 1} of each countermeasure j ∈ J was not introduced and the
variable u jl was replaced with a simple binary selection variable u j ∈ {0, 1}, denoting
whether or not countermeasure j is selected. In the latter case the proportion of threats
i that survive whether or not countermeasure j is selected, is max{1 − u j , ri j }, i.e.,
the survival proportions would depend nonlinearly on the selection variables u j . In
contrast to constant survival proportions qi jl = max{1 − l, ri j } for variables u jl , the
survival proportions for variables u j could not be expressed by constant coefficients
in the constraints.
The linearized version CP_E of model NCP_E is shown below (see also, network
flow model in Deane et al. 2009; Rakes et al. 2012).

CP_E: Risk-neutral selection of cybersecurity safeguards portfolio to

minimize expected loss
Minimize (11.6)
subject to
1. Countermeasure selection constraints: (11.3), (11.4)
2. Surviving threats balance constraints:


vi1l = 1; i ∈ I (11.19)
l∈L
 
qi jl vi j1 = vi j+1l ; i ∈ I, j ∈ J : j < n (11.20)
l∈L l∈L

qinl vinl = Wi ; i ∈ I (11.21)
l∈L
vi jl ≤ u jl ; i ∈ I, j ∈ J, l ∈ L (11.22)

3. Non-negativity and integrality conditions:

u jl ∈ {0, 1}; j ∈ J, l ∈ L (11.23)

vi jl ≥ 0; i ∈ I, j ∈ J, l ∈ L (11.24)
Wi ≥ 0; i ∈ I. (11.25)

Constraint (11.22) (cf. (11.12)) ensures that threats can only by addressed by the
implemented countermeasures.
Note that the expected proportion of successful attacks of threat type i is pi Wi .
Model CP_E will be used to compare the risk-neutral results with those obtained
by applying a risk aversive decision-making model described in the next section.
11.4 Model for Risk-Averse Decision-Making 323

11.4 Model for Risk-Averse Decision-Making

In this section a risk-averse selection of optimal cybersecurity safeguards portfolio

is considered with the two popular in financial engineering percentile measures of
risk, VaR and CVaR, applied to control the risk of high losses.
When selecting the optimal risk-averse countermeasure portfolio, the decision
maker controls the risk of high losses caused by operational disruptions by choosing
the confidence level α. The greater the confidence level α, the more risk aversive is
the decision maker and the smaller percent of the highest loss outcomes is focused
on. The risk aversive decision maker wants to minimize the expected worst-case
losses exceeding VaR, by minimizing CVaR, given available budget B for selected
countermeasures.
Define Cs as the tail loss for scenario s, where tail loss is defined as the amount by
which losses in scenario s exceed VaR. The cybersecurity safeguards portfolio will be
optimized by calculating VaR and minimizing CVaR simultaneously. By measuring
CVaR, the magnitude of the tail loss is considered to achieve a more accurate estimate
of the risks of minimizing loss. In the proposed model, CVaR is represented by
an auxiliary function (11.26) introduced by Rockafellar and Uryasev (2000). The
SMIP model CP_CV for selection of risk-averse cybersecurity safeguards portfolio
to reduce the risk of high losses is formulated below.

CP_CV: Risk-averse selection of cybersecurity safeguards portfolio to

minimize CVaR of loss
Minimize 
C V a R = V a R + (1 − α)−1 Ps Cs (11.26)
s∈S

subject to
1. Countermeasure selection constraints: (11.3)–(11.4)
2. Surviving threats balance constraints: (11.19)–(11.22)
2. Risk constraints:
– the tail loss for scenario s is defined as the nonnegative amount by which
losses in scenario s exceed VaR,

Cs ≥ ai Wi − V a R; s ∈ S (11.27)
i∈Is

3. Non-negativity and integrality conditions: (11.23)–(11.25).

Cs ≥ 0; s ∈ S. (11.28)

Models CP_E and CP_CV can be enhanced for simultaneous optimization of

the expenditures on countermeasures and the loss from successful attacks. Then, the
324 11 Selection of Cybersecurity Safequards Portfolio

budget constraints (11.4) should

 be removed from the models and the fixed cost of
selected countermeasures, j∈J c j u j1 , added to the objective function.
Define the fixed cost of selected countermeasures as the required budget, B(u),

B(u) = c j u j1 . (11.29)
j∈J

The enhanced models CP_EB and CP_CVB for simultaneous optimization of

fixed cost of countermeasures and variable loss from successful attacks are presented
below.

CP_EB: Risk-neutral selection of cybersecurity safeguards portfolio to

minimize expected loss and required budget
Minimize
E c + B(u) (11.30)

subject to (11.3), (11.6), (11.19)–(11.25), (11.29).

CP_CVB: Risk-averse selection of cybersecurity safeguards portfolio to

minimize CVaR of loss and required budget
Minimize
C V a R + B(u) (11.31)

subject to (11.3), (11.19)–(11.29).

11.5 Models for Mean-Risk Decision-Making

In the single objective approach the countermeasure portfolio is selected by mini-

mizing either the expected loss (plus the required budget) or the expected worst-case
loss (plus the required budget). Since the probability of worst-case loss outcomes
is usually very low, the expected cost function that aims at optimizing an average
performance of IT security system, virtually neglects the worst-case loss. In contrast,
CVaR that aims at optimizing worst-case performance, focuses on the low probabil-
ity, high loss outcomes, and as the confidence level α increases a more risk-averse
decision-making focuses on a smaller set of the highest loss outcomes.
In this section the two conflicting objectives are considered simultaneously, and
a bi-objective selection of countermeasures is presented aimed at minimizing both
objective functions to balance the required budget and expected cost with the risk
tolerance. This trade-off model, known as the mean-risk model, is formulated as the
optimization of a convex combination of the expected loss and the CVaR of loss as
a risk measure.
11.5 Models for Mean-Risk Decision-Making 325

The nondominated solution set of the bi-objective countermeasure portfolio can

be found by the parameterization on λ the weighted-sum program CP_EBCV. The
scalarizing mixed integer program is based on CP_CVB model with the addition of
objective (11.6) of model CP_EB.

CP_EBCV: Mean-risk selection of cybersecurity safeguards portfolio to

optimize trade-off between expected loss plus required budget and CVaR of
loss
Minimize
λ(E c + B(u)) + (1 − λ)C V a R (11.32)

where 0 ≤ λ ≤ 1,
subject to (11.3), (11.6), (11.19)–(11.29).

Now, the decision maker controls both the risk of high losses from successful
cyber-attacks by choosing the confidence level α as well as the trade-off between
expected and worst-case losses by choosing the trade-off parameter λ. Using the
latter parameter, the level of risk allowed into the solution can be controlled. In
particular, a decision maker may decide to minimize expected loss while requiring
the percentile worst-case loss to be not greater than some parameter. In other words,
model CP_EBCV with λ = 1 may include an upper bound on CVaR. However,
when α increases so that the worst possible outcomes are considered only, imposing
a strict bound on CVaR may result in an infeasible solution, e.g., Chahara and Taaffe
(2009).

11.6 Computational Examples

In this section some computational examples are presented to illustrate possible

applications of the proposed portfolio approach for the selection of countermeasures
to mitigate the impact of information flow disruptions caused by IT security incidents.
The following parameters have been used for the example problems:
• m = n, the number of threats and the number of countermeasures, were equal to
10, and the corresponding number S = 2m of potential attack scenarios, was equal
to 1024;
• ai , loss from a successful attack of each threat i (in $1000): a1 = 24, a2 = 122,
a3 = 350, a4 = 5, a5 = 250, a6 = 20, a7 = 20, a8 = 25, a9 = 30, a10 =
10000;
• c j , cost of each countermeasure j (in $1000): c1 = 40, c2 = 28, c3 = 80, c4 =
24, c5 = 70, c6 = 50, c7 = 40, c8 = 45, c9 = 50, c10 = 80;
• pi , probability of each threat i: p1 = 0.35, p2 = 0.25, p3 = 0.15, p4 = 0.25,
p5 = 0.20, p6 = 0.25, p7 = 0.50, p8 = 0.35, p9 = 0.40, p10 = 0.003;
• ri j , surviving proportions for each threat i and countermeasure j:
326 11 Selection of Cybersecurity Safequards Portfolio

i/j 1 2 3 4 5 6 7 8 9 10
1 0.01 0.5 1 1 1 1 1 1 1 1
2 1 0.04 1 1 0.6 1 1 1 1 1
3 1 1 0.01 0.8 1 0.8 0.9 0.95 0.9 1
4 1 1 1 0.25 1 0.8 0.9 0.95 0.9 0.8
5 1 0.5 1 0.8 0.02 0.8 0.9 0.95 0.9 1
6 1 0.6 1 1 1 0.1 0.6 0.65 0.6 1
7 1 0.5 1 1 1 0.5 0.15 0.2 0.25 1
8 1 0.5 1 1 1 0.55 0.2 0.25 0.3 1
9 1 0.5 1 1 1 0.5 0.15 0.2 0.35 1
10 1 1 1 1 1 1 1 1 1 0.2;

• α, the confidence level, was equal to 0.50, 0.75, 0.90, 0.95 or 0.99;
• 
B, available budget for countermeasures (in $1000) was equal to 150,300 or
j∈J c j = 507.

The above data set is similar to the one presented in Rakes et al. (2012), which was
based on the threat set reported on IT security forum EndpointSecurity.org. Note that
threat i = 10 is a high-impact and low-probability threat with the loss from successful
attack equal to a10 = $10M and the probability of occurrence equal to p10 = 0.003.
Such a serious and rarely occurring event might be a compromise of some important
business data. The remaining threats are more typical such as viruses, data damages,
or information disclosures, occurring more frequently and resulting in lower loss
from successful attacks. The off-diagonal surviving proportions ri j , i = j, less than
one, indicate that some countermeasures work against their primary threat as well as
against other threats. For example, countermeasures j = 2, 6, 7, 8, 9 may represent
strong security policies focused on particular threats and simultaneously showing
benefits across the other threats. In contrast, countermeasures j = 1, 3 reduce the
survival proportions of a single threat only.
For the risk-neutral models CP_E and CP_EB, solution results are shown in
Table 11.3, and for the risk-averse models CP_CV and CP_CVB with different con-
fidence levels, in Tables 11.4 and 11.5. In addition to VaR and CVaR, for comparison
with risk-neutral solutions, Tables 11.4 and 11.5 show the corresponding expected
loss for the optimal risk-averse countermeasure portfolios. In the tables all costs are
expressed in thousands of dollars.
For the risk-neutral models, Table 11.3 indicates that when CP_E model is applied,
the greater the available budget, the more countermeasures are selected to minimize
expected loss. If, however, model CP_EB is used, a single countermeasure is selected
only to minimize expenditures on countermeasures and expected loss from successful
attacks. For CP_E model, Fig. 11.1 shows the probability mass function of optimal
cost of losses for the available budget B = 150, with the tail of cost distribution
presented in a separate chart. Similarly, for CP_EB model with a variable budget,
the probability mass function of optimal cost is shown in Fig. 11.2, where the tail of
cost distribution is also presented in a separate chart.
For the risk-averse models, Tables 11.4 and 11.5 demonstrate that the number of
selected countermeasures increases with the available budget, which indicates that
11.6 Computational Examples 327

the impact of disruption risks is mitigated by providing a more secure environment.

When α increases, a more risk-averse decision-making focuses on a smaller set of
outcomes, however the number of selected countermeasures is not increasing. For
a limited budget, when α increases, the more expensive countermeasures against
high-impact, low probability threats are more frequently selected. As a result the
total number of all selected countermeasures decreases with α. The latter result is
partly due to the parameter settings for which the countermeasure implementation
costs do not always dominate the losses from successful cyber-threats. For example,
costs of countermeasures 1, 4, 6, 7, 8 and 9 are higher than the losses from a single
successful attack of the corresponding threats.
Note that for SP_CV model, VaR becomes smaller than expected loss for all
budget levels B, when α = 0.50 and α = 0.75, and for all confidence levels α, when
the highest budget level B =$507,000 allows for the selection of all countermeasures,
(see, Table 11.4).
Table 11.3 Solution results for risk-neutral models CP_E and CP_EB
Model SP_E CP_EB
Budget B 150 300 507 28
Expected cost 63.842 17.079 7.589 132.545
Selected 2, 3, 7 2, 3, 5, 7, 10 1–10 2
countermeasures

Table 11.4 Solutions results for risk-averse model CP_CV

Confidence level α 0.50 0.75
Budget B 150 300 507 150 300 507
CVaR 121.130 29.154 14.839 224.294 44.849 27.798
VaR 13.500 10.128 1.078 23.780 16.428 2.965
Expected cost 63.842 17.079 7.589 63.842 17.079 7.589
Selected countermeasures 2, 3, 7 2, 3, 5, 7, 10 1–10 2, 3, 7 2, 3, 5, 7, 10 1–10

Table 11.4 (continued)

Confidence level α 0.90 0.95 0.99
Budget B 150 300 507 150 300 507 150 300 507
CVaR 393.775 84.185 64.597 478.204 145.744 124.985 921.449 624.627 604.858
VaR 302.500 21.450 3.769 318.880 24.178 4.663 414.500 29.028 5.882
Expected cost 92.045 17.079 7.589 92.045 17.079 7.589 92.045 17.079 7.589
Selected countermeasures 2, 4, 10 2, 3, 5, 7, 10 1–10 2, 4, 10 2, 3, 5, 7, 10 1–10 2, 4, 10 2, 3, 5, 7, 10 1–10

Table 11.5 Solutions results for risk-averse model CP_CVB

Confidence level α 0.50 0.75 0.90 0.95 0.99
CVaR 144.008 67.089 109.323 172.808 652.214
VaR 29.380 34.500 42.928 49.500 59.000
Expected cost 80.570 31.332 31.332 31.332 31.332
Required budget 108 258 258 258 258
Selected countermeasures 2, 3 2, 3, 5, 10 2, 3, 5, 10 2, 3, 5, 10 2, 3, 5, 10
328 11 Selection of Cybersecurity Safequards Portfolio

In the computational experiments with risk-averse models, the confidence level

α is set at five levels of 0.5, 0.75, 0.90, 0.95, and 0.99, which means that focus is
on minimizing the highest 50%, 25%, 10%, 5%, and 1% of all scenario outcomes,
i.e., losses from successful attacks. For SP_CV model, Fig. 11.3 shows the proba-
bility mass function of optimal cost of losses for the available budget B = 150 and
confidence level α = 0.99, with the tail of loss distribution presented in a separate
chart. For SP_CVB model with a variable budget and confidence level α = 0.99 the
probability mass function of optimal cost is shown in Fig. 11.4, where the tail of cost
distribution is also presented in a separate chart. Figures 11.1, 11.2, 11.3 and 11.4
indicate that the probability measure is concentrated in finitely many points, which is
typical for the scenario-based optimization under uncertainty, (e.g., in Fig. 11.1, the
probability measure is concentrated in four points). Moreover, the tail of cost distri-
bution shows that the high loss probability is very low. As a result the expected loss
for the optimal risk-neutral portfolio and CVaR for the optimal risk-averse portfolio
are much lower than the corresponding worst-case loss outcomes.

Fig. 11.1 Probability mass function for risk-neutral model CP_E with B =$150,000: Expected
loss = $63,842
11.6 Computational Examples 329

Fig. 11.2 Probability mass function for risk-neutral model CP_EB: Required budget = $28,000,
Expected cost = $132,545

Comparison of tails of the cost distributions for the corresponding risk-neutral

and risk-averse solutions (i.e., Figs. 11.1 vs. 11.3 and 11.2 vs. 11.4) indicates that the
optimal risk-averse countermeasure portfolio positively shapes the cost distribution,
i.e., significantly reduces the worst-case cost outcomes. For example, the worst-
case cost of $10.6M for the risk-neutral portfolio determined using CP_EB model
is reduced to $2.06 M, when the risk-averse portfolio is applied, determined using
CP_CVB model.
Table 11.6 Nondominated solutions for mean-risk model CP_ECV: α = 0.9
λ 0.01 0.10 0.25 0.50 0.75 0.90 0.99
CVaR 65.049 68.145 84.185 109.323 218.193 633.842 710.691
Expected cost 7.908 9.718 17.079 31.332 56.320 116.108 132.545
Required budget 457 388 298 258 188 52 28
Selected 1–8, 10 1–3, 5–7, 10 2, 3, 5, 7, 10 2, 3, 5, 10 2, 3, 10 2, 4 2
countermeasures
330 11 Selection of Cybersecurity Safequards Portfolio

Fig. 11.3 Probability mass function for risk-averse model CP_CV with B =$150,000: α = 0.99,
CVaR = $921,449, VaR = $414,500, Expected cost = $92,045

For the bi-objective approach, the subsets of nondominated solutions were com-
puted by parameterization on λ ∈ {0.01, 0.10, 0.25, 0.50, 0.75, 0.90, 0.99} the
weighted-sum program CP_EBCV. The results obtained for the confidence level
α = 0.9 are presented in Table 11.6. The trade-off between the required budget and
expected cost, and CVaR is clearly shown in Fig. 11.5, where the convex efficient
frontier of the mean-risk model with α = 0.9 is presented. The results emphasize the
effect of varying cost/risk preference of the decision maker. The higher the weight λ
for the required budget and expected cost, the smaller the number of selected counter-
measures. When λ increases from 0.01 to 0.99, the size of the optimal countermeasure
portfolio decreases from nine to one selected countermeasure. At the same time the
required budget and expected cost decreases from $464,908 to $160,545, while CVaR
increases from $65,049 to $710,691.
Note that the nondominated solutions of the weighted-sum program CP_EBCV
with λ = 1 and λ = 0 are identical with the optimal solutions to single objec-
tive models CP_EB (Table 11.3) and CP_CV with unlimited (the highest) budget
B =$507,000 (Table 11.4), respectively.
11.6 Computational Examples 331

Fig. 11.4 Probability mass function for risk-averse model CP_CVB: α = 0.99, CVaR = $652,214,
VaR = $59,000, Required budget = $258,000, Expected cost = $31,332

The computational experiments were performed using the AMPL programming

language and Gurobi 5.0.1 solver on a laptop MacBookPro 6.2 with Intel Core i7
processor running at 2.66 GHz and with 8 GB RAM. The Gurobi solver was capable
of finding proven optimal solutions within CPU seconds for all examples.
Examples of the proposed mixed integer programs size for different numbers
m = n of threats and countermeasures are shown in Table 11.7. The size of the
risk-neutral model CP_EB and the risk-averse model CP_CVB is represented by
the total number of variables, Var., number of binary variables, Bin., and number
of constraints, Cons. The CPU time in seconds required to prove optimality of the
solution was ranging from fraction of a second to several hundred seconds. Note that
the number of variables and constraints in the risk-averse model CP_CVB grows
exponentially in the number m of threats, when all potential attack scenarios are
considered. As a result, CPU time increases rapidly when the number of threats
increases from m = 10 to m = 20.
332 11 Selection of Cybersecurity Safequards Portfolio

Fig. 11.5 Pareto front for mean-risk model CP_EBCV: α = 0.9

Table 11.7 Examples of mixed integer program size and CPU time
m 2m(a) Var.(b) Bin. Cons. CPU(c)
Risk-neutral model CP_EB
10 1024 231 20 321 <1
15 32768 496 30 766 <1
20 1 048 576 861 40 1241 <1
Risk-averse model CP_CVB
10 1024 1256 20 1345 <1
15 32768 33265 30 33474 5
20 1 048 576 1 049 438 40 1 049 817 416
(a) m = number of threats (number of countermeasures),

2m = number of attack scenarios.

(b) Var. = number of variables, Bin. = number of binary variables,

Cons. = number of constraints.

(c) CPU seconds for proving optimality on a MacBookPro 6.2, Intel Core i7, 2.66 GHz,

RAM 8 GB/Gurobi 5.0.1

• The higher the budget available and the higher the confidence level, the more
risk-oriented is the countermeasure portfolio selected. In most cases the number
of selected countermeasures increases with the available budget, and for a limited
budget decreases with the confidence level.
• For a limited budget and lower confidence level α, the expensive countermeasures
against low-probability, high-impact threats are rarely selected, and as a result
the total number of all selected countermeasures is usually greater than that for a
higher α, when the more expensive countermeasures against high-impact threats
are chosen.
11.6 Computational Examples 333

• The trade-off mean-risk model provides the decision maker with a simple tool for
balancing expected and worst-case losses and for shaping of the resulting cost
distribution through the selection of optimal countermeasure portfolio. The deci-
sion maker is capable of controlling both the risk of high losses from successful
cyber-attacks by choosing the confidence level α as well as the trade-off between
expected and worst-case losses by choosing the trade-off parameter λ that rep-
resents cost versus risk preference. Using the latter parameter, the level of risk
allowed into the solution can be controlled.

The computational experiments prove that for a limited number of attack scenarios
considered, the optimal risk-averse portfolio can be found within CPU seconds, using
the Gurobi MIP solver.

11.7 Notes

The problem of a right choice of countermeasures for implementation under limited

budget is not an easy task. The choice depends not only on reliable data on potential
cyber-threats and losses but also requires a good security risk planning tool. The
National Institute of Standards and Technology classifies information security con-
trols (NIST SP800-53) into the three categories: technical, operational, and manage-
ment controls (e.g., Viduto et al. 2012), in which the last two categories can become
as critical as the first one. For example, operational countermeasures include mon-
itoring and logging procedures, business continuity/incident response procedures,
backup and recovery procedures, while management countermeasures include peri-
odic employee training, periodic testing and review procedures, network auditing,
and protection for all sensitive informational assets.
Most of IT security planning models in the literature are qualitative. For example,
Bojanc and Jerman-Blazic (2008) introduced methods for identification of the assets,
the threats, the vulnerabilities of the ICT systems and proposed a procedure that
enables selection of the optimal investment of the necessary security technology
based on the quantification of the values of the protected systems. In Egan (2005) a
checklist in table form was developed to help decision maker planning a coverage
strategy. Chen et al. (2011) discussed current research findings in enterprise risk and
security management using web mining techniques.
A study by Gordon and Loeb (2002) uses risk analysis to suggest an optimal budget
for a risk-neutral decision maker. In their approach, they compared the loss caused
by security incidents to the investment required to reduce the related vulnerability.
Based on two general classes of security breach functions, they state that the amount
to invest is considerably lower than the expected loss caused by an incident. In fact,
they find that the amount to invest in security never exceeds 37% of the expected
loss and in most cases will be substantially less. However, these observations only
hold true if the security breach functions meet the condition of decreasing marginal
returns in case of security investments. Hausken (2006) examined four additional
334 11 Selection of Cybersecurity Safequards Portfolio

types of security breach functions with different shapes and found that the amount
to invest is no longer limited by 37% and different investment strategies should be
applied in each case. Wang et al. (2008) proposed a more detailed analysis which
makes use of security incident data and value-at-risk to support decision-making.
In contrast to qualitative approaches, the literature on quantitative methods for
selection of countermeasures to block or mitigate security attacks is very limited.
Based on NIST SP800-30 guidelines, Viduto et al. (2012) developed a risk assess-
ment and optimization model to satisfy organizational security needs in a cost-
effective manner. The security countermeasure selection problem was formulated
as a multi-objective optimization problem, where variables such as financial cost
and risk may affect the final solution. A tailored multi-objective tabu search-based
heuristic approach was constructed to solve the proposed multi-objective optimiza-
tion problem and asses the qualities of its solutions with respect to optimal ones.
Deane et al. (2009) developed a linear generalized network flow model that quan-
tifies IT security risk in the supply chain. It was shown how to find solutions for
optimal risk reduction under several definitions of optimality: minimizing upstream
risk, minimizing downstream risk, and minimizing global supply chain risk. Then,
following the mathematical models proposed by Deane et al. (2009), an integer pro-
gramming model was developed by Rakes et al. (2012), for optimally choosing a
subset of countermeasures to block or mitigate security attacks in the presence of a
given threat level profile. The model was used to examine the two different types of
scenarios: under expected threat levels and under worst-case levels. The authors illus-
trated the tradeoffs in optimal security planning when expected threats are used to
parameterize the model versus worst-case values for threat outcomes. To demonstrate
the trade-off which occurs if decision makers divert budgets away from planning for
ordinary risk in an effort to mitigate the effects of potential high-impact outcomes,
budget-dependent risk curves were developed. Schilling and Werners (2016) pro-
posed a combinatorial optimization model to efficiently select security safeguards in
order to protect IT infrastructures and systems. The approach is designed to provide
decision support for an organization as a whole or separately for specific systems.
The material presented in this chapter is based on research reported in Sawik
(2013d), where portfolio approach with CVaR as a risk measure was proposed for
the selection of countermeasures in IT security planning to prevent and mitigate
the impact of cyber-attacks. A critical issue that need to be considered before any
practical application of the proposed models is attempted, however, is the estima-
tion of probabilities and the resulting losses associated with each type of threats
and countermeasures. In practice, threat likelihood estimates are provided by secu-
rity experts (e.g., Ryan et al. 2012) and complete distributional information is not
available. However, the proposed scenario-based approach does not require such a
complete information to be available and only assumes independence of different
threat events. In many cases they are independent of each other, but in principle
might have a joint probability distribution. The future research should also consider
selection of a countermeasure portfolio under the less restrictive assumptions on a
joint probability of different cyber-threats. However, relaxation of the assumption
of independent threat events may significantly complicate the problem of building
11.7 Notes 335

potential attack scenarios. For example, modeling of attack scenarios with correlated
threat events represented by Bernoulli random variables may require constructing of
the corresponding correlated probability distribution, i.e., a correlated binomial dis-
tribution instead of the binomial distribution used in this chapter.

Problems

11.1 Design attack scenarios for correlated threat events.

11.2 Enhance the SMIP models presented in this chapter for multiple implementa-
tion levels of countermeasures, where different levels are associated with different
implementation costs and different effectiveness of blocking different threats.

11.3 Modify the SMIP models presented in this chapter for selection of cybersecurity
portfolio such that cannot contain mutually exclusive countermeasures or such that
must contain subsets of countermeasures contingent on each other.

11.4 Relax the assumption that the blocking effectiveness of each countermeasure is
independent whether or not it is used alone or together with other countermeasures.
What would be the implication of that relaxation on the developed models?

11.5 Identify the best-case and the worst-case attack scenarios such that the CVaR
of loss is not greater and not less than than a fixed threshold loss, respectively.
References

Ahmadjian, C., Lincoln, J.: Keiretsu, governance, and learning: Case studies in change from the
Japanese automotive industry. Org. Sci. 12(6), 683–701 (2001)
Aissaoui, N., Haouari, M., Hassini, E.: Supplier selection and order lot sizing modeling: a review.
Comput. Oper. Res. 34, 3516–3540 (2007)
Altendorfer, K., Jodlbauer, H.: An analytical model for service level and tardiness in a single machine
MTO production system. Int. J. Prod. Res. 49(7), 1827–1850 (2011)
Arshinder, K.A., Deshmukh, S.G.: Supply chain coordination: perspectives, empirical studies and
research directions. Int. J. Prod. Econ. 115, 316–335 (2008)
Bard, J.F., Nananukul, N.: The integrated production-inventory-distribution-routing problem. J.
Sched. 12(3), 257–280 (2009)
Basnet, C., Leung, J.M.Y.: Inventory lot-sizing with supplier selection. Comput. Oper. Res. 32,
1–14 (2005)
Berger, P.D., Gerstenfeld, A., Zeng, A.Z.: How many suppliers are best? A decision-analysis
approach. Omega Int. J. Manag. Sci. 32(1), 9–15 (2004)
Berger, P.D., Zeng, A.Z.: Single versus multiple sourcing in the presence of risks. J. Oper. Res. Soc.
57(3), 250–61 (2006)
Bilgen, B., Celebi, Y.: Integrated production scheduling and distribution planning in dairy supply
chain by hybrid modelling. Ann. Oper. Res. 211, 55–82 (2013)
Bijulal, D., Venkateswaran, J., Hemachandra, N.: Service levels, system cost and stability of
production-inventory control systems. Int. J. Prod. Res. 49(23), 7085–7105 (2011)
Birge, J.R., Louveaux, F.: Introduction to Stochastic Programming. Springer, New York (2011)
Blackhurst, J.V., Scheibe, K.P., Johnson, D.J.: Supplier risk assessment and monitoring for the
automotive industry. Int. J. Phys. Distrib. Logist. Manag. 38(2), 143–165 (2008)
Bojanc, R., Jerman-Blazic, B.: An economic modelling approach to information security risk man-
agement. Int. J. Inf. Manage. 28, 413–422 (2008)
Cakici, E., Mason, S.J., Kurz, M.E.: Multi-objective analysis of an integrated supply chain schedul-
ing problem. Int. J. Prod. Res. 50(10), 2624–2638 (2012)
Chahara, K., Taaffe, K.: Risk averse demand selection with all-or-nothing orders. Omega Int. J.
Manag. Sci. 37(5), 996–1006 (2009)
Che, Z.H., Wang, H.S.: Supplier selection and supply quantity allocation of common and non-
common parts with multiple criteria under multiple products. Comput. Ind. Eng. 55, 110–133
(2008)

© Springer International Publishing AG 2018 337

T. Sawik, Supply Chain Disruption Management Using Stochastic Mixed
Integer Programming, International Series in Operations Research
& Management Science 256, DOI 10.1007/978-3-319-58823-0
338 References

Chen, D.-S., Batson, R.G., Dang, Y.: Applied Integer Programming: Modeling and Solution. Wiley,
Hoboken (2010)
Cheng, B.-Y., Leung, J.Y.-T., Li, K.: Integrated scheduling of production and distribution to mini-
mize total cost using an improved ant colony optimization method. Comput. Ind. Eng. 83, 217–225
(2015)
Chen, Z.L.: Integrated production and outbound distribution scheduling: review and extensions.
Oper. Res. 58(1), 130–148 (2010)
Chen, H., Chau, M., Li, S.: Enterprise risk and security management: data, text and web mining.
Decis. Support Syst. 50, 649–650 (2011)
Chen, Z.L., Pundoor, G.: Order assignment and scheduling in a supply chain. Oper. Res. 54(3),
555–572 (2006)
Chen, Z.L., Vairaktarakis, G.L.: Integrated scheduling of production and distribution operations.
Manage. Sci. 51(4), 614–628 (2005)
Chopra, S., Sodhi, M.S.: Supply-chain breakdown. MIT Sloan. Manag. Rev. 46(1), 53–62 (2004)
Church, R.L., Scaparra, M.P.: Analysis of facility systems-reliability when subject to attack or a
natural disaster. In: Murray, A.T., Grubesic, T.H. (eds.) Reliability and Vulnerability in Critical
Infrastructure: A Quantitative Geographic Perspective. Springer, Berlin (2006)
Cohen, M.A., Kunreuther, H.: Operations risk management: overview of Paul Kleindorfer’s contri-
butions. Prod. Oper. Manag. 16(5), 525–541 (2007)
Cui, L.X.: Joint optimization of production planning and supplier selection incorporating customer
flexibility: an improved genetic approach. J. Intell. Manuf. (2014). doi:10.1007/s10845-014-
0932-5.Articleinpress
Deane, J.K., Ragsdale, C.T., Rakes, T.R., Rees, L.P.: Managing supply chain risk and disruption
from IT security incidents. Oper. Manag. Res. 2(1), 4–12 (2009)
Delage, E., Arroyo, S., Ye, Y.: The value of stochastic modeling in two-stage stochastic programs
with cost uncertainty. Oper. Res. 62(6), 1377–1393 (2014)
Demirtas, E.A., Ustun, O.: An integrated multiobjective decision making process for supplier selec-
tion and order allocation. Omega Int. J. Manag. Sci. 36, 76–90 (2008)
Dolgui, A., Proth, J.M.: Supply Chain Engineering: Useful Methods and Techniques. Springer,
London (2010)
Durbach, I.N., Stewart, T.J.: Using expected values to simplify decision making under uncertainty.
Omega Int. J. Manag. Sci. 37(2), 312–330 (2009)
Egan, M.: The Executive Guide to Information Security. Symantec Press, Indianapolis (2005)
Erenguc, S.S., Simpson, N.C., Vakharia, A.J.: Integrated production/distribution planning in supply
chains: an invited review. Eur. J. Oper. Res. 115(2), 219–236 (1999)
Erkut, E., Karagiannidis, A., Perkouidis, G., Tjandra, S.A.: A multicriteria facility location model
for municipal solid waste management in North Greece. Eur. J. Oper. Res. 187, 1402–1421 (2008)
Falasca, M., Zobel, C.W., Cook, D.: A decision support framework to assess supply chain resilience.
In: The Proceedings of the 5th International ISCRAM Conference, Washington, DC, USA, 596–
605 (2008)
Fiksel, J.: Sustainability and resilience: toward a systems approach. Sustain. Sci. Pract. Policy 2(2),
14–21 (2006)
Fourer, R., Grayand, D.M., Kernigham, B.W.: AMPL A Modeling Language for Mathematical
Programming, Duxbury (2003)
Fujimoto, T., Park, Y.W.: Balancing supply chain competitiveness and robustness through "virtual
dual sourcing": lessons from the Great East Japan Earthquake. Int. J. Prod. Econ. 147, 429–436
(2013)
Fuller, T.: Floodwaters are gone, but supply chain issues linger. The New York Times, NY, January
21, (2012)
Gabrel, V., Murat, C., Thiele, A.: Recent advances in robust optimization: an overview. Eur. J. Oper.
Res. 235, 471–483 (2014)
Gao, S., Qi, L., Lei, L.: Integrated batch production and distribution scheduling with limited vehicle
capacity. Int. J. Prod. Econ. 160, 13–25 (2015)
References 339

Ghodsypour, S.H., O’Brien, C.: The total cost of logistics in supplier selection under conditions of
multiple sourcing, multiple criteria and capacity constraints. Int. J. Prod. Econ. 73, 15–27 (2001)
Gordon, L.A., Loeb, M.P.: The economics of information security investment. ACM Trans. Inf.
Syst. Secur. 5, 438–457 (2002)
Gurnani, H., Mehrotra, A., Ray, S. (eds.): Supply Chain Disruptions: Theory and Practice of Man-
aging Risk. Springer, London (2012)
Hamdi, F., Ghorbel, A., Masmoudi, F., Dupont, L.: Optimization of a supply portfolio in the context
of supply chain risk management: literature review. J. Intell. Manuf. (2015). doi:10.1007/s10845-
015-1128-3.Articleinpress
Hammami, R., Temponi, C., Frein, Y.: A scenario-based stochastic model for supplier selection in
global context with multiple buyers, currency fluctuation uncertainties, and price discounts. Eur.
J. Oper. Res. 233, 159–170 (2014)
Haraguchi, M., Lall, U.: Flood risks and impacts: a case study of Thailand’s floods in 2011 and
research questions for supply chain decision making. Int. J. Disaster Risk Reduct. 14(3), 256–272
(2015)
Hausken, K.: Returns to information security investment: the effect of alternative information secu-
rity breach functions on optimal investment and sensitivity to vulnerability. Inf. Syst. Front. 8,
338–349 (2006)
He, H., Hongyan, X.: Dual sourcing and backup production: coexistence versus exclusivity. Omega
Int. J. Manag. Sci. 57, 22–33 (2015)
Heckmann, I., Comes, T., Nickel, S.: A critical review on supply chain risk - definition, measure
and modeling. Omega Int. J. Manag. Sci. 52, 119–132 (2015)
Heese, H.S.: Single versus multiple sourcing and the evolution of bargaining positions. Omega Int.
J. Manag. Sci. 54, 125–133 (2015)
Hezarkhani, B., Kubiak, W.: Coordinating contracts in SCM: a review of methods and literature.
Decis. Mak. Manuf. Serv. 4(1–2), 5–28 (2010)
Ho, W., Xu, X., Dey, P.K.: Multi-criteria decision making approaches for supplier evaluation and
selection: a literature review. Eur. J. Oper. Res. 202, 16–24 (2010)
Hoffmann, P., Schiele, H., Krabbendam, K.: Uncertainty, supply risk management and their impact
on performance. J. Purch. Supply Manag. 19(3), 199–211 (2013)
Hou, J., Zeng, A.Z., Zhao, L.: Coordination with a backup supplier through buy-back contract under
supply disruption. Transp. Res. Part E 46, 881–895 (2010)
Ivanov, D., Sokolov, B., Dolgui, A.: The Ripple effect in supply chains: trade-off ‘efficiency-
flexibility-resilience’ in disruption management. Int. J. Prod. Res. 52(7), 2154–2172 (2014)
Ivanov, D., Sokolov, B., Pavlov, A.: Optimal distribution (re)planning in a centralized multi-stage
network under conditions of ripple effect and structure dynamics. Eur. J. Oper. Res. 237(2),
758–770 (2015)
Ivanov, D., Pavlov, A., Dolgui, A., Pavlov, D., Sokolov, B.: Disruption-driven supply chain (re)-
planning and performance impact assessment with consideration of pro-active and recovery poli-
cies. Transp. Res. Part E (2016). doi:10.1016/j.tre.2015.12.007
Izakian, H., Abraham, A.: Fuzzy C-means and fuzzy swarm for fuzzy clustering problem. Expert
Syst. Appl. 38(3), 1835–1838 (2011)
Jenkins, L.: Selecting scenarios for environmental disaster planning. Eur. J. Oper. Res. 121, 275–286
(2000)
Kall, P., Mayer, J.: Stochastic Linear Programming: Models, Theory, and Computation. Springer,
New York (2011)
Kasilingam, R.G., Lee, C.P.: Selection of vendors - a mixed-integer programming approach. Com-
put. Ind. Eng. 31, 347–350 (1996)
Kleindorfer, P.R., Saad, G.H.: Managing disruption risks in supply chains. Prod. Oper. Manag.
14(1), 53–68 (2005)
Klibi, W., Martel, A.: Modeling approaches for the design of resilient supply networks under
disruptions. Int. J. Prod. Econ. 135, 882–898 (2012)
340 References

Klibi, W., Martel, A., Guitouni, A.: The design of robust value-creating supply chain networks: a
critical review. Eur. J. Oper. Res. 203, 283–293 (2010)
Knemeyer, A.M., Zinn, W., Eroglu, C.: Proactive planning for catastrophic events in supply chains.
J. Oper. Manag. 27(2), 141–153 (2009)
Kostreva, M.M., Ogryczak, W., Wierzbicki, A.: Equitable aggregation and multiple criteria analysis.
Eur. J. Oper. Res. 158, 362–377 (2004)
Kouvelis, P., Dong, L., Boyabatli, O., Li, R. (eds.): Handbook of Integrated Risk Management in
Global Supply Chains. Wiley, Hoboken (2011)
Larsen, C., Thorstenson, A.: A comparison between the order and the volume fill rate for a base
stock inventory control system under a compound renewal demand process. J. Oper. Res. Soc.
59(6), 798–804 (2008)
Larsen, C., Thorstenson, A.: The order and volume fill rates in inventory control systems. Int. J.
Prod. Econ. 147, 13–19 (2014)
Lee, H., Kim, M.S., Kim, K.K.: Interorganizational information systems visibility and supply chain
performance. Int. J. Inf. Manage. 34, 285–295 (2014)
Lei, L., Liu, S., Ruszczynski, A., Park, S.: On the integrated production, inventory and distribution
routing problem. IIE Trans. 38(11), 955–970 (2006)
Li, K.P., Ganesan, V.K., Sivakumar, A.I.: Synchronized scheduling of assembly and multi-
destination air-transportation in a consumer electronics supply chain. Int. J. Prod. Res. 43(13),
2671–2685 (2005)
Li, L., Zabinsky, Z.B.: Incorporating uncertainty into a supplier selection problem. Int. J. Prod.
Econ. 134, 344–356 (2011)
Li, X., Wang, Q.: Coordination mechanisms of supply chain systems. Eur. J. Oper. Res. 179, 1–16
(2007)
Li, X., Ouyang, Y., Peng, F.: A supporting station model for reliable infrastructure location design
under interdependent disruptions. Transp. Res. Part E 60, 80–93 (2013)
Liu, S., Papageorgiou, L.G.: Multiobjective optimisation of production, distribution and capacity
planning of global supply chains in the process industry. Omega Int. J. Manag. Sci. 41(2), 369–382
(2013)
Liu, S., Sawik, T., Papageorgiou, L.: Corrigendum to - multiobjective optimisation of production,
distribution and capacity planning of global supply chains in the process industry - [Omega 41
(2013) 369–382]. Omega Int. J. Manag. Sci. 44(1), 149 (2014)
MacKenzie, C.A., Barker, K., Santos, J.R.: Modeling a severe supply chain disruption and post-
disaster decision making with application to the Japanese earthquake and tsunami. IIE Trans.
46(12), 1243–1260 (2014)
Madadi, A.R., Kurz, M.E., Taaffe, K.M., Sharp, J.L., Mason, S.J.: Supply network design: risk-
averse or risk-neutral? Comput. Ind. Eng. 2014(78), 55–65 (2014)
Maggioni, F., Wallace, S.W.: Analyzing the quality of the expected value solution in stochastic
programming. Ann. Oper. Res. 200(1), 37–54 (2012)
Marszewska, J.R.: Implications of seismic hazard in Japan on Toyota supply chain disruption risks.
In: Proceedings of 13th International Conference on Industrial Logistics, September 28–October
1, 2016, Zakopane, Poland, AGH University of Science & Technology, 178–185 (2016)
Matsuo, H.: Implications of the Tohoku earthquake for Toyota’s coordination mechanism: supply
chain disruption of automotive semiconductors. Int. J. Prod. Econ. 161, 217–227 (2015)
Meena, P.L., Sarmah, S.P., Sarkar, A.: Sourcing decisions under risks of catastrophic event disrup-
tions. Transp. Res. Part E 47, 1058–1074 (2014)
Merzifonluoglu, Y.: Risk averse supply portfolio selection with supply, demand and spot market
volatility. Omega Int. J. Manag. Sci. 57, 40–53 (2015)
Nemhauser, G.L., Wolsey, L.A.: Integer and Combinatorial Optimization. Wiley, New York (1999)
Norrman, A., Jansson, U.: Ericsson’s proactive risk management approach after a serious sub-
supplier accident. Int. J. Phys. Distrib. Logist. Manag. 34(5), 434–456 (2004)
Noyan, N.: Risk-averse two-stage stochastic programming with an application to disaster manage-
ment. Comput. Oper. Res. 39, 541–559 (2012)
References 341

Ogryczak, W., Ruszczynski, A.: Dual stochastic dominance and related mean-risk models. SIAM
J. Optim. 13, 60–78 (2002)
Ogryczak, W., Tamir, A.: Minimizing the sum of the k largest functions in linear time. Inf. Process.
Lett. 85, 117–122 (2003)
Ogryczak, W., Wierzbicki, A., Milewski, M.: A multi-criteria approach to fair and efficient band-
width allocation. Omega Int. J. Manag. Sci. 36, 451–463 (2008)
Oke, A., Gopalakrishnan, M.: Managing disruptions in supply chains: a case study of a retail supply
chain. Int. J. Prod. Econ. 118, 168–174 (2009)
Parlar, M., Perry, D.: Inventory models of future supply uncertainty with single and multiple sup-
pliers. Nav. Res. Logist. 43, 191–210 (1996)
Park, Y., Hong, P., Roh, J.J.: Supply chain lessons from the catastrophic natural disaster in Japan.
Bus. Horiz. 56(1), 75–85 (2013)
Paul, S.K., Sarker, R., Essam, D.: Managing real-time demand fluctuation under a supplier-retailer
coordinated system. Int. J. Prod. Econ. 158, 231–243 (2014a)
Paul, S.K., Sarker, R., Essam, D.: Real time disruption management for a two-stage batch
production-inventory system with reliability considerations. Eur. J. Oper. Res. 237, 113–128
(2014b)
Paul, S.K., Sarker, R., Essam, D.: Managing disruption in an imperfect production-inventory system.
Comput. Ind. Eng. 84, 101–112 (2015a)
Paul, S.K., Sarker, R., Essam, D.: A disruption recovery plan in a three-stage production-inventory
system. Comput. Oper. Res. 57, 60–72 (2015b)
Paul, S.K., Sarker, R., Essam, D.: Managing risk and disruption in production-inventory and supply
chain systems: a review. J. Ind. Manag. Optim. 12(3), 1009–1029 (2016a)
Paul, S.K., Sarker, R., Essam, D.: A reactive mitigation approach for managing supply disruption in a
three-tier supply chain. J. Intell. Manuf. (2016b). doi:10.1007/s10845-016-1200-7.Articleinpress
Paul, S.K., Sarker, R., Essam, D.: A quantitative model for disruption mitigation in a supply chain.
Eur. J. Oper. Res. (2016). doi:10.1016/j.ejor.2016.08.035.Articleinpress
Pereira, J.V.: The new supply chain’s frontier: information management. Int. J. Inf. Manage. 29,
372–379 (2009)
Ponomarov, S.Y., Holcomb, M.C.: Understanding the concept of supply chain resilience. Int. J.
Logist. Manag. 20(1), 124–143 (2009)
PrasannaVenkatesan, S., Goh, M.: Multi-objective supplier selection and order allocation under
disruption risk. Transp. Res. Part E 95, 124–142 (2016)
Pundoor, G., Chen, Z.L.: Joint cyclic production and delivery scheduling in a two-stage supply
chain. Int. J. Prod. Econ. 119(1), 55–74 (2009)
Rakes, T.R., Deane, J.K., Rees, L.P.: IT security planning under uncertainty for high-impact events.
Omega Int. J. Manag. Sci. 40(1), 79–88 (2012)
Ravindran, A.R., Bilsel, R.U., Wadhwa, V., Yang, T.: Risk adjusted multicriteria supplier selection
models with applications. Int. J. Prod. Res. 48(2), 405–424 (2012)
Rawls, C.G., Turnquist, M.A.: Pre-positioning of emergency supplies for disaster response. Transp.
Res. Part B 44(4), 521–34 (2010)
Rees, L.P., Deane, J.K., Rakes, T.R., Baker, W.H.: Decision support for cybersecurity risk planning.
Decis. Support Syst. 51, 493–505 (2011)
Rockafellar, R.T., Uryasev, S.: Optimization of conditional value-at-risk. J. Risk 2(3), 21–41 (2000)
Rockafellar, R.T., Uryasev, S.: Conditional value-at-risk for general loss distributions. J. Bank.
Finance 26(7), 1443–1471 (2002)
Ruiz-Torres, A.J., Mahmoodi, F.: The optimal number of suppliers considering the costs of indi-
vidual supplier failures. Omega Int. J. Manag. Sci. 35(1), 104–115 (2007)
Ruiz-Torres, A.J., Mahmoodi, F., Zeng, A.Z.: Supplier selection model with contingency planning
for supplier failures. Comput. Ind. Eng. 66, 374–382 (2013)
Ryan, J.J.C.H., Mazzuchi, T.A., Ryan, D.J., de la Cruz, J.L., Cooke, R.: Quantifying information
security risks using expert judgment elicitation. Comput. Oper. Res. 39, 774–784 (2012)
342 References

Sajadieh, S., Fallahnezhad, M.S., Khosravi, M.: A joint optimal policy for a multiple-suppliers
multiple-manufacturers multiple-retailers system. Int. J. Prod. Econ. 146(2), 738–744 (2013)
Sarrafha, K., Rahmati, S.H.A., Niaki, S.T.A., Zaretalab, A.: A bi-objective integrated procurement,
production, and distribution problem of a multi-echelon supply chain network design: a new
tuned MOEA. Comput. Oper. Res. 54, 35–51 (2015)
Sarykalin, S., Serraino, G., Uryasev, S.: Value-at-risk vs. conditional value-at-risk in risk manage-
ment and optimization. Tutorials in Operations Research, INFORMS 270–294 (2008)
Sawik, B.: Conditional value-at-risk and value-at-risk for portfolio optimization model with weight-
ing approach. Automatyka 15(2), 429–434 (2011)
Sawik, B.: Bi-criteria portfolio optimization models with percentile and symmetric risk measures
by mathematical programming. Przegld Elektrotechniczny 88(10b), 176–180 (2012a)
Sawik, B.: Downside risk approach for multi-objective portfolio optimization. In: Klatte, D., Lthi,
H.-J., Schmedders, K. (eds.) Operations Research Proceedings 2011, pp. 191–196. Springer,
Berlin (2012b)
Sawik, B.: Conditional value-at-risk vs. value-at-risk to multi-objective portfolio optimization. In:
Lawrence, K.D., Kleinman, G. (eds.) Applications of Management Science 15, pp. 277–305.
Bingley, UK, Emerald (2012c)
Sawik, B.: Triple-objective models for portfolio optimisation with symmetric and percentile risk
measures. Int. J. Logist. Syst. Manag. 25(1), 96–107 (2016)
Sawik, T.: Stochastic optimal control of a multi-facility, multi-product production scheduling with
random times of supplies. Control Cybern. 6(3–4), 21–35 (1977)
Sawik, T.: A cyclic versus flexible approach to materials ordering in make-to-order assembly. Math.
Comput. Modell. 42(3–4), 279–290 (2005)
Sawik, T.: Hierarchical approach to production scheduling in make-to-order assembly. Int. J. Prod.
Res. 44(4), 801–830 (2006)
Sawik, T.: A lexicographic approach to bi-objective scheduling of single-period orders in make-to-
order manufacturing. Eur. J. Oper. Res. 180(3), 1060–1075 (2007)
Sawik, T.: Coordinated supply chain scheduling. Int. J. Prod. Econ. 120(2), 437–451 (2009a)
Sawik, T.: Monolithic versus hierarchical approach to integrated scheduling in a supply chain. Int.
J. Prod. Res. 47(21), 5881–5910 (2009b)
Sawik, T.: Multi-objective due-date setting in a make-to-order environment. Int. J. Prod. Res. 47(22),
6205–6231 (2009c)
Sawik, T.: Single vs. multiple objective supplier selection in a make to order environment. Omega.
Int. J. Manag. Sci. 38(3–4), 203–212 (2010)
Sawik, T.: Scheduling in Supply Chains Using Mixed Integer Programming. Wiley, Hoboken
(2011a)
Sawik, T.: Selection of supply portfolio under disruption risks. Omega Int. J. Manag. Sci. 39,
194–208 (2011b)
Sawik, T.: Supplier selection in make-to-order environment with risks. Math. Comput. Modell.
53(9–10), 1670–1679 (2011c)
Sawik, T.: Selection of a dynamic supply portfolio in make-to-order environment with risks. Comput.
Oper. Res. 38(4), 782–796 (2011d)
Sawik, T.: Selection of resilient supply portfolio under disruption risks. Omega Int. J. Manag. Sci.
41, 259–269 (2013a)
Sawik, T.: Selection and protection of suppliers in a supply chain with disruption risks. Int. J. Logist.
Syst. Manag. 15(2–3), 143–159 (2013b)
Sawik, T.: Integrated selection of suppliers and scheduling of customer orders in the presence of
supply chain disruption risks. Int. J. Prod. Res. 51(23–24), 7006–7022 (2013c)
Sawik, T.: Selection of optimal countermeasure portfolio in IT security planning. Decis. Support
Syst. 55, 156–164 (2013d)
Sawik, T.: Joint supplier selection and scheduling of customer orders under disruption risks: single
vs. dual sourcing. Omega. Int. J. Manag. Sci. 43(2), 83–95 (2014a)
References 343

Sawik, T.: Optimization of cost and service level in the presence of supply chain disruption risks:
single vs. multiple sourcing. Comput. Oper. Res. 51(1), 11–20 (2014b)
Sawik, T.: On the robust decision-making in a supply chain under disruption risks. Int. J. Prod. Res.
52(22), 6760–6781 (2014c)
Sawik, T.: Cost vs. customer service level in supply chains under major disruptions. In: Proceedings
of LISS 2014, 4th International Conference on Logistics, Informatics and Service Science, July
23-26, 2014, UC Berkeley, CA, Springer, NY, 1075–1079 (2014d)
Sawik, T.: On the fair optimization of cost and customer service level in a supply chain under
disruption risks. Omega Int. J. Manag. Sci. 53, 58–66 (2015a)
Sawik, T.: Integrated supply chain scheduling under multi-level disruptions. IFAC-Papers On Line
48(3), 1515–1520 (2015b)
Sawik, T.: Integrated supply, production and distribution scheduling under disruption risks. Omega
Int. J. Manag. Sci. 62, 131–144 (2016a)
Sawik, T.: On the risk-averse optimization of service level in a supply chain under disruption risks.
Int. J. Prod. Res. 54(1), 97–112 (2016b)
Sawik, T.: Resilient vs. robust supply portfolio under disruption risks. In: Proceedings of 13th
International Conference on Industrial Logistics, September 28 – October 1, 2016, Zakopane,
Poland, AGH University of Science & Technology, 255–263 (2016c)
Sawik, T.: Stochastic vs. deterministic approach to supply chain risks management. In: Proceedings
of MOTSP 2016: International Conference on Management of Technology, 11–3 June 2016,
Porec, Croatia, University of Zagreb (2016d)
Sawik, T.: A portfolio approach to supply chain disruption management. Int. J. Prod. Res. 55(7),
1970–1991 (2017)
Scaparra, M.P., Church, R.L.: A bilevel mixed-integer program for critical infrastructure protection
planning. Comput. Oper. Res. 35, 1905–1923 (2008)
Schilling, A., Werners, B.: Optimal selection of IT security safeguards from an existing knowledge
base. Eur. J. Oper. Res. 248(1), 318–327 (2016)
Schmitt, A.J.: Strategies for customer service level protection under multi-echelon supply chain
disruption risk. Transp. Res. Part B 45(8), 1266–1283 (2011)
Schmitt, A.J., Singh, M.: A quantitative analysis of disruption risk in a multi-echelon supply chain.
Int. J. Prod. Econ. 139, 22–32 (2012)
Schmitt, A.J., Snyder, L.V.: Infinite-horizon models for inventory control under yield uncertainty
and disruptions. Comput. Oper. Res. 39, 850–862 (2012)
Schmitt, A.J., Sun, S.A., Snyder, L.V., Shen, Z.J.M.: Centralization versus decentralization: risk
pooling, risk diversification, and supply chain disruptions. Omega Int. J. Manag. Sci. 52, 201–212
(2015)
Shao, X.F., Dong, M.: Comparison of order-fulfilment performance in MTO and MTS systems with
an inventory cost budget constraint. Int. J. Prod. Res. 50(7), 1917–1931 (2012)
Sheffi, Y.: The Resilient Enterprise. MIT Press, Cambridge (2005)
Sheffi, Y., Rice, J.B.: A supply chain view of the resilient enterprise. MIT Sloan Manag. Rev. 47(1),
41–48 (2005)
Shin, H., Collier, D., Wilson, D.: Supply management orientation and supplier/buyer performance.
J. Oper. Manag. 18, 317–333 (2000)
Simangunsong, E., Hendry, L.C., Stevenson, M.: Supply-chain uncertainty: a review and theoretical
foundation for future research. Int. J. Prod. Res. 50(16), 4493–4523 (2012)
Simchi-Levi, D., Schmidt, W., Wei, Y., Zhang, P.J., Combs, K., Ge, Y., Gusikhin, O., Sanders, M.,
Zhang, D.: Identifying risks and mitigating disruptions in the automotive supply chain. Interfaces
45(5), 375–390 (2015)
Snyder, L.V., Scaparra, P.M, Daskin, M.S., Church, R.L.: Planning for disruptions in supply chain
networks. Tutorials in Operations Research INFORMS (2005)
Snyder, L.V., Shen, Z.-J.M.: Fundamentals of Supply Chain Theory. Wiley, Hoboken (2011)
Snyder, L.V., Atan, Z., Peng, P., Rong, Y., Schmitt, A., Sinsoysal, B.: OR/MS models for supply
chain disruptions: a review. IIE Trans. 48(2), 89–109 (2016)
344 References

Sounderpandian, J., Prasad, S., Madan, M.: Supplies from developing countries: optimal order
quantities under loss risks. Omega Int. J. Manag. Sci. 36(1), 122–130 (2008)
Stecke, K.E., Kumar, S.: Sources of supply chain disruptions, factors that breed vulnerability, and
mitigating strategies. J. Mark. Channels 16(3), 193–226 (2009)
Steuer, R.E.: Multiple Criteria Optimization: Theory. Computation and Application. Wiley, New
York (1996)
Su, P., Liu, S.G.: Dual sourcing in managing operational and disruption risks in contract manufac-
turing. Int. J. Prod. Res. 53(1), 291–306 (2015)
Tang, C.S.: Perspectives in supply chain risk management. Int. J. Prod. Econ. 103, 451–488 (2006)
Tomlin, B.: On the value of mitigation and contingency strategies for managing supply chain
disruption risks. Manage. Sci. 52(5), 639–657 (2006)
Torabi, S.A., Baghersad, M., Mansouri, S.A.: Resilient supplier selection and order allocation under
operational and disruption risks. Transp. Res. Part E 79, 22–48 (2015)
Uryasev, S.: Conditional value-at-risk: optimization algorithms and applications. Financ. Eng. News
14(2) (2000)
Ustun, O., Demirtas, E.A.: An integrated multi-objective decision making process for multi-period
lot sizing with supplier selection. Omega Int. J. Manag. Sci. 36, 509–521 (2008)
Viduto, V., Maple, C., Huang, W., Lopez-Perez, D.: A novel risk assessment and optimisation model
for a multi-objective network security countermeasure selection problem. Decis. Support Syst.
53, 599–610 (2012)
Viergutz, C., Knust, S.: Integrated production and distribution scheduling with lifespan constraints.
Ann. Oper. Res. 213, 293–318 (2014)
Wang, G., Lei, L.: Polynomial-time solvable cases of the capacitated multi-echelon shipping network
scheduling problem with delivery deadlines. Int. J. Prod. Econ. 137, 263–271 (2012)
Wang, J., Chaudhury, A., Rao, H.R.: A value-at-risk approach to information security investment.
Inf. Syst. Res. 19, 106–120 (2008)
Weber, C.A., Current, J.R.: A multiobjective approach to vendor selection. Eur. J. Oper. Res. 68,
173–184 (1993)
Whitney, D.E., Luo, J., Heller, D.A.: The benefits and constraints of temporary sourcing diversifi-
cation in supply chain disruption and recovery. J. Purch. Supply Manag. 20, 238–250 (2014)
Wu, D., Olson, D.L.: Supply chain risk, simulation, and vendor selection. Int. J. Prod. Econ. 114,
646–655 (2008)
Xanthopoulos, A., Vlachos, D., Iakovou, E.: Optimal newsvendor policies for dual-sourcing supply
chains: a disruption risk management framework. Comput. Oper. Res. 39, 350–357 (2012)
Xia, W., Wu, Z.: Supplier selection with multiple criteria in volume discount environments. Omega
Int. J. Manag. Sci. 35, 494–504 (2007)
Yager, R.R.: On ordered weighted averaging aggregation operators in multi criteria decision making.
IEEE Trans. Syst. Man Cybern. 18, 183–190 (1988)
Yu, H., Zeng, A.Z., Zhao, L.: Single or dual sourcing: decision-making in the presence of supply
chain disruption risks. Omega Int. J. Manag. Sci. 37, 788–800 (2009)
Yue, J., Xia, Y., Tran, T.: Selecting sourcing partners for a make-to-order supply chain. Omega Int.
J. Manag. Sci. 38, 136–144 (2010)
Zeng, A.Z., Xia, Y.: Building a mutually beneficial partnership to ensure backup supply. Omega
Int. J. Manag. Sci. 52, 77–91 (2015)
Zobel, C.W., Khansa, L.: Characterizing multi-event disaster resilience. Comput. Oper. Res. 42,
83–94 (2014)
Index

A Constraints
AMPL, 34, 65, 93, 129, 137, 144, 176, 178, budget, 146, 318, 324
182, 208, 234, 265, 268, 303, 331 capacity and parts availability, 108
countermeasure selection, 319, 322, 323
customer order scheduling, 156
B demand allocation, 21, 50, 109, 197, 216
Bill-of-material constraints, 249, 284 demand fulfillment, 284
dual sourcing strategy, 111
due-date meeting, 157
C dynamic supply portfolio selection, 49–
Capacity 52
of plant, 273, 274, 280, 287 equality, 21, 50, 76, 111
of producer, 106, 121, 151, 193, 201, 213, flow conservation, 280
246 inequality, 76
of supplier, 17, 26, 44, 45, 70, 71, 84 inventory balance, 200, 209, 270
per period of plant, 273, 307 multiple batch shipping, 159
per period of producer, 241, 256, 266 multiple sourcing strategy, 110
remaining, 69, 81, 94, 98 non-delayed delivery, 156
time-varying, 107, 194 order quantity and emergency inventory
total of producer, 241, 256 allocation, 75
transportation, 152, 159, 167 order-to-period assignment, 110, 112,
Capacity-to-demand ratio, 256, 264, 265 197, 216
Conditional Cost-at-Risk, 5, 235 parts availability, 146
Conditional Service-at-Risk, 5, 235 primary supply portfolio selection, 245,
Conditional Value-at-Risk (CVaR), 4, 16, 23, 251, 278
24, 52, 53, 66, 67, 214, 220 producer capacity, 110, 197, 216
Confidence level, 5, 11, 17, 22, 26–28, 31, production capacity, 246, 249, 279, 284
35, 36, 38, 44, 51, 56, 58, 62, 64, 67, quantity and emergency inventory allo-
71, 77, 85, 87–89, 91, 93, 96, 100, cation, 77
106, 114, 117–119, 121–124, 128, recovery supply and demand portfolio
131–135, 137, 138, 141–144, 154, selection, 278, 289
161, 178, 212, 214, 215, 217, 218, recovery supply portfolio selection, 245
220–222, 224, 226, 228, 231, 234, risk, 23, 24, 52, 53, 77, 115–117, 161,
244, 252, 253, 268, 276, 304, 305, 162, 217, 218, 253, 290, 291, 323
308, 323–326, 328, 330, 332, 333 single batch shipping, 156

© Springer International Publishing AG 2018 345

T. Sawik, Supply Chain Disruption Management Using Stochastic Mixed
Integer Programming, International Series in Operations Research
& Management Science 256, DOI 10.1007/978-3-319-58823-0
346 Index

supplier capacity, 200 Countermeasure portfolio, 318, 323–325,

supplier selection, 112 329, 330, 332–334
supplier selection and protection, 75, 77 CPLEX, 34, 93, 129, 137, 176, 208, 234
supply and demand portfolio balance, Cybersecurity portfolio, 315, 318, 319, 323
284 mean-risk, 325
supply and demand portfolio selection, risk-averse, 323, 324
280, 284 risk-neutral, 319, 322, 324
supply portfolio selection, 20, 22–24,
155
surviving threats balance, 322, 323 D
Coordinating constraints Decision variables
production-distribution, 150, 156, 157, first stage, 9, 10, 19, 46, 73, 81, 109, 154,
160, 184 195, 214, 244, 276, 319
supply-distribution, 150, 160, 184 second stage, 10, 73, 109, 154, 195, 214,
supply-production, 150, 156, 157, 184, 244, 276
246, 249 Decision-making
supply-transshipment-production, 279, expected value-based, 150, 178
284 fair mean-risk, 211, 235
Cost mean-risk, 9, 11, 24
expected, 12, 15, 16, 18–20, 24, 32, 40, mixed mean-risk, 236
43, 104, 119, 155, 164, 189, 195, 207, risk-averse, 9, 11, 22, 39, 93, 146, 150,
212, 215, 219, 295, 324 177, 225, 240, 252, 266, 290, 304
risk-neutral, 9, 10, 18, 39, 46, 109, 114,
expected delay penalty, 48
115, 146, 150, 166, 167, 222, 225, 240,
expected transportation, 159, 167
256, 276, 294
expected worst-case, 22, 24, 32, 123,
robust, 222, 225, 235
212, 215, 220, 222, 291
Delays
expected worst-case of unfulfilled cus-
independent
tomer order, 137
local, 45
of countermeasure, 317, 324
Demand portfolio
of ordering, 17, 18, 20, 26, 34, 44, 45,
recovery, 271, 272, 276–278, 281, 284–
106, 121, 129, 151, 189, 193, 195, 212,
286, 288–292, 294, 295, 298, 303, 304
213
Disruption
penalty, 17, 18, 21, 229 correlated
penalty for unfulfilled demand, 229, 234, global, 7, 71
273, 282, 284, 295, 301, 304 regional, 7, 18, 45, 106, 152, 189,
penalty of delayed customer order, 44, 194, 240, 273
45, 106, 107, 121, 151, 193, 195, 201, independent
212, 213, 219 local, 7, 8, 18, 45, 71, 105, 152, 189,
penalty of unfulfilled customer order, 44, 194, 240, 273
45, 106, 107, 121, 151, 193, 195, 201, multi-level, 8
212, 213, 219 probability, 7, 17, 26, 32, 35, 55, 56, 71,
plant recovery, 275 72, 80, 85–88, 91, 93, 94, 96, 105, 106,
production, 273, 282, 284, 292 120–123, 128, 130, 131, 137, 151, 152,
production setup, 273 166, 193, 194, 201, 202, 213, 222, 256,
purchasing, 20, 36 273
shortage, 169, 175 total, 86, 163, 164
tail, 10, 11, 19, 23, 109, 115, 116, 154, two-level, 8
161, 162, 214, 215, 217, 276, 290 Disruption management strategy, 3, 268, 310
transportation, 151, 152, 154, 159, 200
transshipment, 273, 292
Cost-at-risk, 5 E
Cost-to-recover, 241–243, 256, 268, 273– Emergency inventory, 69, 71–75, 77, 80–82,
275, 293 85, 86, 93, 97, 98, 100
Index 347

Expected value approach, 149, 184 CP_CVB, 324

DSP_CV(c), 52
DSP_CV(sl), 52
F DSupport_CV(c), 290
Fortification, 69 DSupport_CV(sl), 291
cost, 81, 83, 94, 99 RDSupport_CV(c), 292
of supplier, 81, 97, 98 RDSupport_CV(sl), 292
RSP_CV, 77
RSP(mlp)_CV, 83
G SCS1_CV, 161
Gurobi, 65, 93, 137, 144, 176, 178, 182, 265, SCS2_CV, 162
268, 303, 331 SCS3_CV, 162
SP_CV(c), 23
SP_CV(sl), 23
H SPS1_CV(c), 116
Hierarchical approach, 146, 183, 243, 249, SPS1_CV(sl), 117
256, 258–261, 263–265, 271, 272, SPS1_CV(sl)+, 118
291, 294–296, 299, 301–303, 307, SPS2_CV(c), 116
308 SPS2_CV(sl), 117
SPS2_CV(sl)+, 118
SPS_CV(c), 220
I
SPS_CV(sl), 220
Integrated approach, 145, 243, 258–260,
SPSm_CV(c), 115
263–265, 271, 290, 294, 295, 297,
SPSm_CV(sl), 116
301–304, 308, 311
SPSm_CV(sl)+, 118
Support_CV(c), 253
Support_CV(sl), 253
L
Lexicographic minimax, 190–192, 204, 208, risk-neutral
209, 224, 231, 234 WSPS_E, 206
LP relaxation, 62, 265, 308 CP_E, 322
CP_EB, 324
DSP_E(c), 49
M DSP_E(sl), 50
Mixed integer programming, 129 DSupport_E, 278
Model DSupportMP_E, 282
deterministic ESPS_E, 196
ESCS1, 164 NCP_E, 319
PSupport, 250, 285 RDSupport_E, 288
RDSupport(s), 286 RSP_E, 74
RSupport(s), 251 RSP(mlp)_E, 82
mean-risk RSupport_E, 252
CP_EBCV, 325 SCS1_E, 155
DSP_ECV(c), 53 SCS2_E, 158
DSP_ECV(sl), 53 SCS3_E, 159
RSP_ECV, 78 SP_E(c), 20
RSP(mlp)_ECV, 84 SP_E(sl), 21
SP_ECV(c), 25 SPS1_E(c), 112
SP_ECV(sl), 25 SPS1_E(sl), 114
SPS_ECV(c), 231 SPS2_E(c), 111
SPS_ECV(sl), 231 SPS2_E(sl), 113
SPSm_E(c)CV(sl), 119 SPS_E(c), 198, 219
risk-averse SPS_E(c,α), 220
CP_CV, 323 SPS_E(sl), 198, 219
348 Index

SPS_E(sl,α), 221 Service-at-Risk, 5

SPSm_E(c), 109 Sourcing
SPSm_E(sl), 113 multi-region, 16, 34, 36
Support_E, 245 single-region, 16, 25, 27, 35
SupportMP_E, 248 Stochastic mixed integer programming, 9
robust Supplier fulfillment rate, 240, 244, 254, 274
RSPS_ECV(c), 216 all-or-nothing, 163
RSPS_ECV(sl), 218 expected, 163, 180, 245, 256
Supplier protection
multi-level, 80, 82–84, 93
N single-level, 70, 73, 85
Network flow Supply chain
model, 322, 334 customer-driven, 15, 149, 183, 193, 212,
problem 235
embedded, 271, 280, 309 electronics, 165, 175, 182, 183, 268
Non-disruption probability, 7, 22, 51, 163 global, 1, 145, 189, 268, 315, 334
Nondominated solution, 25, 32, 34, 53, 78, multi-echelon, 2, 145, 146, 151
90, 91, 141, 155, 180, 189, 205–207, network, 1, 6, 39, 98, 235
209, 210, 231, 232, 325, 330 three-echelon, 104, 193
two-echelon, 39
visibility, 315
P Supply chain risk
Production scheduling, 239, 247, 248, 250– delay, 43
252, 254, 271, 272, 276, 278, 281, disruption, 43
282, 285, 286, 288, 290–292 Supply portfolio
Protection index, 69, 80, 90, 100 diversification, 27, 31, 36, 57, 134, 137,
141, 144, 178, 267
dynamic, 49, 50, 52
R equitably efficient, 189, 190, 193, 195,
Recovery 196, 198, 201, 204
cost, 243, 268, 270, 293, 304, 305, 307, mean-risk, 25, 53, 69
310, 311 nondominated, 92, 169, 182, 207
mode, 242, 270, 310 primary, 239, 240, 243, 245, 247, 249–
time, 3, 268, 270, 293, 307, 310, 311 254, 256, 259, 260, 262, 265–269, 272,
276–278, 284, 290–292, 294, 295, 298,
304
S recovery, 239, 240, 243–245, 247, 249,
Scheduling of customer orders, 104, 108, 251–254, 256, 260–262, 265, 266, 268,
109, 111–117, 119, 145, 147, 149, 269, 272, 276–278, 284–286, 288–292,
190, 192, 195, 198, 200, 201, 205, 294, 295, 298, 303, 304
209, 210, 212, 214, 216, 218–220 resilient, 70, 73, 74, 76–78, 82–84, 88,
Service level 93, 98
demand fulfillment rate, 20, 49, 104, 113, risk-averse, 23, 52, 69, 85–87, 91, 92, 96,
114, 117–119, 137, 138, 142, 144, 146, 115, 116, 125–127, 134, 141, 144, 224
149, 154, 189, 195, 199, 202, 245, 277 risk-neutral, 20, 21, 49, 50, 69, 86, 92, 96
expected, 15, 16, 18–21, 24, 40, 43, 157, robust, 213, 214, 218, 224, 229–231
164, 189, 195, 207, 212, 219, 295 static, 16–19, 25
expected worst-case, 22, 24, 119, 137,
212, 220, 222, 291
order fulfillment rate, 104, 113, 119, 137, T
138, 142, 144, 146, 189, 195, 199, 202, Time
213 delivery, 106, 121, 151, 193, 213, 241,
tail, 10, 19, 23, 109, 154, 161, 162, 214, 246, 254, 263, 273, 286, 292, 294, 300,
276, 291 303
Index 349

disruption start, 242, 243, 256, 263–265, V

267, 268, 288, 294, 303 Value-at-Risk (VaR), 4, 16, 66, 214
plant recovery, 275
start of disruptive event, 241, 263, 273,
298
W
transition, 242, 270, 275
Wait-and-see approach, 11, 149, 182, 184,
transportation, 105, 151–153, 158, 194,
247, 281
199, 240, 275
transshipment, 273, 275, 292
Time-to-recover, 241, 242, 247, 256, 273–
275, 293 X
Two-stage approach, 11 XPRESS, 176