This article was downloaded by: [111.68.97.

37] On: 04 June 2024, At: 03:43

Publisher: Institute for Operations Research and the Management Sciences (INFORMS)
INFORMS is located in Maryland, USA

INFORMS Journal on Optimization

Publication details, including instructions for authors and subscription information:
http://pubsonline.informs.org

The Nutritious Supply Chain: Optimizing

Humanitarian Food Assistance
Koen Peters, Sérgio Silva,Rui Gonçalves,Mirjana Kavelj,Hein Fleuren, Dick den
Hertog, Ozlem Ergun, Mallory Freeman

To cite this article:

Koen Peters, Sérgio Silva,Rui Gonçalves,Mirjana Kavelj,Hein Fleuren, Dick den Hertog, Ozlem Ergun, Mallory
Freeman (2021) The Nutritious Supply Chain: Optimizing Humanitarian Food Assistance. INFORMS Journal on
Optimization 3(2):200-226. https://doi.org/10.1287/ijoo.2019.0047

Full terms and conditions of use: https://pubsonline.informs.org/Publications/Librarians-Portal/PubsOnLine-

Terms-and-Conditions

This article may be used only for the purposes of research, teaching, and/or private study. Commercial use or
systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher approval,
unless otherwise noted. For more information, contact [email protected].

The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitness
for a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, or
inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or support
of claims made of that product, publication, or service.

Copyright © 2021, INFORMS

Please scroll down for article—it is on subsequent pages

With 12,500 members from nearly 90 countries, INFORMS is the largest international association of operations
research (O.R.) and analytics professionals and students. INFORMS provides unique networking and learning
opportunities for individual professionals, and organizations of all types and sizes, to better understand and use
O.R. and analytics tools and methods to transform strategic visions and achieve better outcomes.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
 INFORMS JOURNAL ON OPTIMIZATION
Vol. 3, No. 2, Spring 2021, pp. 200–226
http://pubsonline.informs.org/journal/ijoo ISSN 2572-1484 (print), ISSN 2575-1492 (online)

The Nutritious Supply Chain: Optimizing Humanitarian

Food Assistance
Koen Peters,a,b Sérgio Silva,a Rui Gonçalves,a Mirjana Kavelj,a Hein Fleuren,b Dick den Hertog,c Ozlem Ergun,d
Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

Mallory Freemane
a
Supply Chain Planning and Optimization Unit, World Food Programme, 00148 Rome, Italy; b Department of Econometrics and Operations
Research, Tilburg University, 5000 LE Tilburg, Netherlands; c Amsterdam Business School, University of Amsterdam, 1018 TV Amsterdam,
Netherlands; d Department of Mechanical and Industrial Engineering, Northeastern University, Boston, Massachusetts 02115;
e
Advanced Technology Group, United Parcel Service, Atlanta, Georgia 30328
Contact: [email protected], https://orcid.org/0000-0002-9685-1808 (KP); [email protected] (SS); [email protected] (RG);
[email protected] (MK); h.fl[email protected], https://orcid.org/0000-0002-8902-0555 (HF);
[email protected], https://orcid.org/0000-0002-1829-855X (DDH); [email protected],
https://orcid.org/0000-0002-3420-3338 (OE); [email protected] (MF)

Received: May 21, 2020 Abstract. The World Food Programme (WFP) is the largest humanitarian agency fighting
Revised: September 24, 2020 hunger worldwide, reaching approximately 90 million people with food assistance across
Accepted: October 30, 2020 80 countries each year. To deal with the operational complexities inherent in its mandate,
Published Online in Articles in Advance: WFP has been developing tools to assist its decision makers with integrating supply chain
April 1, 2021 decisions across departments and functional areas. This paper describes a mixed integer
https://doi.org/10.1287/ijoo.2019.0047 linear programming model that simultaneously optimizes the food basket to be delivered,
the sourcing plan, the delivery plan, and the transfer modality of a long-term recovery
Copyright: © 2021 INFORMS operation for each month in a predefined time horizon. By connecting traditional supply
chain elements to nutritional objectives, we are able to make significant breakthroughs in
the operational excellence of WFP’s most complex operations. We show three examples of
how the optimization model is used to support operations: (1) to reduce the operational
costs in Iraq by 12% without compromising the nutritional value supplied, (2) to manage
the scaling-up of the Yemen operation from three to six million beneficiaries, and (3) to
identify sourcing strategies during the El Niño drought of 2016.

Keywords: supply chain • nutrition • MILP • humanitarian logistics • WFP

1. Introduction
Humanitarian organizations are currently facing an incredible amount of complex crises. Conflicts in countries
such as Syria, South Sudan, and Yemen have been unprecedentedly long and large in scale, and many African
countries are suffering from droughts and poor harvests. These crises result in rapidly deteriorating living
conditions for everyone in the vicinity, threatening millions of innocent people with hunger, malnutrition, and
worse. For decades, humanitarian organizations, such as the United Nations (UN), Médecins sans Frontières,
and the International Committee of the Red Cross, have been doing everything in their power to provide
assistance to those in need.
There are nearly 690 million undernourished people in the world today (Food and Agriculture Organization
of the United Nations et al. 2020). This means that 1 in 11 people in the world do not have access to enough
food to be healthy and lead an active life. Currently a global pandemic (COVID-19) is threatening the health
and livelihoods of these people even further, so these numbers are expected to increase. However, the UN
considers hunger one of the greatest solvable problems (Food and Agriculture Organization of the United
Nations et al. 2015). The second sustainable development goal is zero hunger, and it calls on all people to do
their part to eliminate hunger in our lifetimes.
One of the key players in responding to emergencies and eliminating hunger is the United Nations World
Food Programme (WFP). WFP is the world’s largest humanitarian organization fighting hunger worldwide
with more than 18,000 employees across the globe. In emergencies, it distributes food where it is needed,
saving the lives of victims of war, civil conflict, and natural disasters. After the cause of an emergency has
passed, WFP uses food assistance to help communities rebuild their lives and return to a semblance of
normality. In 2019, a total of 97 million people spread across 88 countries were reached directly by WFP
assistance, covering 70% of those facing critical hunger levels (World Food Programme 2020).
Delivering food to this number of people in these complex environments requires a supply chain that is
agile, adaptable, and aligned (Lee 2004). Every year, WFP procures, transports, and distributes around four

200
 Peters et al.: The Nutritious Supply Chain
INFORMS Journal on Optimization, 2021, vol. 3, no. 2, pp. 200–226, © 2021 INFORMS 201

million metric tons (four billion kilograms) of food to people in need. The poor infrastructure conditions and
high levels of insecurity in conflict areas necessitate creativity and flexibility in delivering this food, and WFP
has been known to employ every variety of transportation method—from elephants and camels to airplanes
and barges. In 2017, on any given day WFP had 5,000 chartered trucks on the road, 92 chartered planes in the
air, and 20 chartered ships at sea (World Food Programme 2018). By virtue of its excellent logistical per-
formance, WFP is even mandated to lead logistics operations whenever a humanitarian agency requires a joint
response from UN agencies and the humanitarian community.
Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

1.1. Literature
There are several phases in responding to a disaster. In the literature, we generally see that disaster timelines
are split into four stages: mitigation, preparedness, response, and recovery (Altay and Green 2006, Van
Wassenhove 2006, Ergun et al. 2011). Mitigation focuses on the prevention of a disaster and the reduction of its
intensity, for instance, by setting up alert systems that warn against floods or by defining building guidelines
for areas that are vulnerable to earthquakes or hurricanes. Preparedness is more concerned with setting up the
appropriate resources (both physical and human), such as the building and stocking of strategic warehouses or
the training of personnel. Response starts once the disaster has occurred; it includes activities such as the
delivery of food and services to those in need, evacuation of the affected region, and the collection of debris.
Finally, the recovery phase aims to return a semblance of normality to the area affected.
Holguı́n-Veras et al. (2012) split up the recovery phase into short- and long-term recovery, in which short-
term recovery is a transitional stage covering damage assessments, repairs, housing, etc. Long-term recovery
may span multiple years and includes the rebuilding of infrastructure and distribution of medical and food
supplies to prevent disease and malnutrition. This long-term recovery is one of the focus areas of WFP. Most
literature on humanitarian logistics focuses on the preparedness and response side of a disaster, whereas long-
term recovery is a topic that does not receive much attention.
Many researchers have characterized and discussed the challenges and opportunities of humanitarian
logistics. Van Wassenhove (2006) discusses how the private sector can learn from the agility and adaptability
inherent in humanitarian supply chains and how the humanitarian sector can learn from the established
supply chain management (SCM) best practices in the private sector. In a follow-up paper, Van Wassenhove
and Pedraza Martinez (2012) illustrate the potential of operations research in particular for adapting such SCM
best practices to humanitarian logistics. Apte (2010) discusses research issues and potential actions sur-
rounding the field of humanitarian logistics and reviews analytical models from the literature to understand
the state of the art in humanitarian logistics. Apte (2010) mentions the sustaining of long-term developmental
aid when discussing research gaps. Çelik et al. (2012) also mention that there is a research gap in relation to
long-term recovery, using the term “long-term humanitarian development” instead. Additionally, they
highlight that there is a lack of good implementation of decision support tools in humanitarian operations.
Holguı́n-Veras et al. (2012) pinpoint research gaps that need to be filled to enhance both the efficiency of
humanitarian logistics and the realism of the mathematical models designed to support it. They argue that
humanitarian logistics is too broad a field to fit neatly into a single definition of operational conditions and
urge researchers to treat these different operational conditions separately.
Despite the qualitative attention to long-term recovery, there are few mathematical formulations available
that cover the entire scope of a humanitarian supply chain. Most existing research is focused on (a com-
bination of) three subproblems, namely that of facility location (Balcik and Beamon 2008), distribution
(Haghani and Oh 1996, Özdamar and Demir 2012, Rottkemper et al. 2012, Rancourt et al. 2015), and inventory
control (Beamon and Kotleba 2006, Pérez-Rodrı́guez and Holguı́n-Veras 2016). Humanitarian researchers have
extended the traditional models for these three subproblems with constraints, objectives, and solution
methods to facilitate the special requirements of a humanitarian supply chain. These extensions include (but
are not limited to) research on the appropriate objective function (Holguı́n-Veras et al. 2013, Gralla et al. 2014,
Gutjahr and Nolz 2016) and modeling uncertain demand, prices, and capacities (Ben-Tal et al. 2011, Rawls and
Turnquist 2012, Bozorgi-Amiri et al. 2013). For an in-depth discussion of what is necessary to make a tra-
ditional (i.e., commercial) supply chain model work in a humanitarian context, we refer to Holguı́n-
Veras et al. (2012).
For food assistance in particular, the design of food baskets is an important topic. Many papers illustrate
how mathematical models can be used to generate food baskets or nutritious products that satisfy all nu-
tritional requirements (Briend et al. 2003, Fleige et al. 2010, Ryan et al. 2014, Deptford et al. 2017, Bose et al.
2019). For example, Carlson et al. (2003) develop a quadratic optimization model that, for each age–gender
group, selects the optimal food plan that meets the dietary standards, adheres to the budget constraints, and
 Peters et al.: The Nutritious Supply Chain
202 INFORMS Journal on Optimization, 2021, vol. 3, no. 2, pp. 200–226, © 2021 INFORMS

resembles the reported food consumption for that specific age–gender group (therefore, making it more likely
that the food plan “sticks”). Similarly, Chastre et al. (2007) develop linear programming routines to generate
hypothetical diets using a combination of foods that enable a family to meet their energy and nutrient re-
quirements as recommended by the World Health Organization and the Food and Agriculture Organization of
the UN at the lowest possible cost. As the software (Cost of the Diet) can account for the frequency at which
foods are eaten, for example, by specifying that a particular food is eaten three times a day, the food baskets
can be adjusted to reflect typical dietary patterns.
One aspect of humanitarian aid that is not yet covered in the operations research literature is that hu-
Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

manitarian organizations have several methods at their disposal for delivering assistance. Whereas, in the
past, transfers were done exclusively in-kind (i.e., the organization buys, transports, and distributes com-
modities), these days, it has become common to provide beneficiaries with cash or vouchers instead, allowing
them to purchase their own commodities at local markets or selected retailers. Lentz et al. (2013) discuss the
rise of these new food-assistance instruments. There are multiple interacting effects to be considered, and it is
important to weigh the benefits (reduced transportation costs, increase in the beneficiary’s dignity) against the
dangers (the influx of cash or vouchers may disrupt the local economy). They state that no single tool is always
and everywhere preferable and seek to educate the reader on their appropriate use. In particular, the economic
repercussions of choosing one transfer modality over another are notoriously hard to measure, making it
difficult to select the appropriate instrument. Ryckembusch et al. (2013) discuss an analytical tool that is able
to compare the cost-effectiveness of transfer modalities. Their tool, the “omega value,” considers the trade-off
between total costs (procurement, transportation, services, etc.) and the nutrient value score (NVS). The NVS
is a weighted score function that shows to what extent all nutritional requirements are met (for nutrients such
as energy, protein, vitamins, etc.). The omega value shows the nutritional value per dollar spent and can assist
policy makers in making the right choice of food basket and transfer modality.
For the successful application of the types of mathematical models described, availability of good data are
crucial. Paradoxically, humanitarians face the double burden of insufficient data and information overload. On
the one hand, it is often difficult to find the data that is required to support analyses; Altay and Labonte (2014)
describe eight data challenges in particular, such as the inconsistency between data sources (e.g., different
definitions, master data, geographical scopes, etc.) and the inaccessibility of vital information (e.g., because of
physical access constraints or sudden disruptions). On the other hand, and in particular for sudden-onset
emergencies, information is being collected by dozens of agencies and thousands of volunteers, all of whom
are trying to support the relief activities. It is becoming increasingly complex to analyze all this data to distill
the critical information needed for decision making (United Nations Office for the Coordination of Hu-
manitarian Affairs 2002, Harvard Humanitarian Initiative 2010).

1.2. Research Questions

The literature survey highlights several challenges and research gaps. In particular, we find
a. Long-term humanitarian assistance has received little attention.
b. Most papers focus on subproblems (e.g., facility location).
c. Alternative transfer modalities (e.g., cash, vouchers) have not been addressed yet.
d. There is a lack of models that are actually being implemented (e.g., because of data challenges).
In this paper, we try to close these four gaps by addressing our main research question: how can opti-
mization help WFP to improve the cost-effectiveness of its operations? In answering this question, we pay
particular attention to the integration of WFP’s key supply chain decisions (i.e., food basket design, transfer
modality selection, and sourcing and delivery plan) into a single tractable mathematical model and how
such a model can be implemented in practice. Although the focus of the paper is on WFP, many of the models
and lessons learned can be applied to other humanitarian supply chains.
The rest of the paper is structured as follows: in Section 2, we provide a brief description of the key
components of WFP’s supply chain. In Section 3, we develop a mixed integer linear programming (MILP)
model that optimizes the food basket design, transfer modality selection, sourcing plan, and delivery plan of a
long-term recovery operation for each month in a predefined time horizon. In Section 4, we discuss various
insights that can be gained from this integrated model and how it influences decision making. In Section 5, we
apply this model to three WFP use cases: Iraq, Yemen, and El Niño, demonstrating how we can use the
optimization model to reduce the operational costs and improve the effectiveness of complex operations. In
Section 6, we extend the core model with additional features and reflect on future research. In Section 7, we
provide our concluding remarks. The appendix contains the detailed model specifications (constraints,
variables, etc.) and a short history of the project (Appendix C).
 Peters et al.: The Nutritious Supply Chain
INFORMS Journal on Optimization, 2021, vol. 3, no. 2, pp. 200–226, © 2021 INFORMS 203

2. WFP’s Supply Chain

Before defining the mathematics, it would be prudent to offer a brief discussion of WFP’s supply chain and the
components of our model. We consider the food basket, transfer modalities, sourcing, and logistics network.

2.1. Food Basket

WFP’s food assistance depends on the context of the crisis and the beneficiaries’ health, demographics, and
access to non-WFP food. The bulk of food assistance is provided through General Food Distribution (GFD), a
monthly parcel that provides enough food for a standard-sized family. Additional food assistance is supplied
Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

to more vulnerable beneficiaries, such as those suffering from malnutrition, young children, and pregnant or
nursing women. Nutritionists track the nutrients that WFP’s food baskets provide and make sure that they
align with the needs of the different beneficiary types (e.g., a young child may receive a fortified cereal on top
of the generic ration). Traditionally, demand was defined in terms of predetermined food baskets per
beneficiary type, but in this paper, we take one step back and define demand as nutritional requirements,
allowing us to optimize the food baskets rather than using the predefined ones. For the applications in this
paper, we consider 11 nutrients (three macronutrients and eight micronutrients), but the developed meth-
odology works for any number of nutrients.

2.2. Transfer Modalities

There are multiple ways of satisfying the demand—be it a predefined food basket or a nutritional requirement.
Whereas WFP used to supply the necessary commodities itself, in the last decade, there has been a transition
toward cash-based transfers (also known as cash and vouchers). Traditional food transfers require WFP to
source, store, transport, and distribute the commodities, whereas a cash-based transfer allows beneficiaries to
obtain the commodities themselves from local markets and shops. Food transfers allow for more control over
the food assistance and benefit from economies of scale (e.g., making use of bulk purchases, price seasonality,
tax waivers, etc.). Cash-based transfers (CBT), on the other hand, are more direct (so more of the donation is
spent on procurement), but they are dependent on local market availability, and their impact is harder to
measure up front. We consider three types of CBT assistance:
1. Commodity vouchers: These vouchers entitle a beneficiary to a specific ration of one or more com-
modities (e.g., 15 kg of maize) to be redeemed at WFP-contracted stores.
2. Value vouchers: These vouchers entitle a beneficiary to purchase a range of commodities from WFP-
contracted stores up to a certain value (e.g., US$20). Each beneficiary can choose what combination of
commodities best fits their needs.
3. Cash: This is an unrestricted form of assistance, and beneficiaries have full control over where they spend
their money and what they spend it on.
Note that each of these transfer modalities can be delivered to the beneficiary in different ways (e.g.,
distribution of physical vouchers, e-money, SMS vouchers, etc.), and the most appropriate transfer mechanism
is heavily dependent on the location and what (financial) tools are already being used by the assisted
population. In this paper, we develop and apply a model that considers commodity vouchers as an alternative
transfer modality to in-kind assistance, allowing us to replace or expand traditional rations with locally
available commodities that would be difficult to supply in-kind (e.g., fresh vegetables, dairy, meat, etc.). In
Section 6, we develop a model that also incorporates the value vouchers and cash and include some pre-
liminary results.

2.3. Sourcing
Depending on the transfer modality, multiple sources are available. Cash-based transfers (such as e-vouchers
and direct cash transfers to debit cards) allow beneficiaries to purchase commodities at local markets. For food
transfers, WFP has more purchasing options. We distinguish three supplier types: international suppliers,
regional suppliers, and local suppliers. Local suppliers (wholesale) can be found in the recipient country;
regional suppliers can be found in neighboring countries, and transport from the supplier to the recipient
country is usually done by land. Procuring from international suppliers involves shipping the commodities
to a discharge port (DP).

2.4. Logistics Network

The handover between suppliers and WFP is flexible and dependent on the Incoterms (Ramberg 2011).
Usually, WFP takes charge of the commodities at a loading port (for international purchases) or at one of its
 Peters et al.: The Nutritious Supply Chain
204 INFORMS Journal on Optimization, 2021, vol. 3, no. 2, pp. 200–226, © 2021 INFORMS

hubs (for local and regional purchases). From discharge ports, WFP moves commodities into the country to so-
called extended delivery points (EDPs), transshipment points at which commodities can be stored, packaged,
consolidated, etc. From the EDPs, the commodities are transported (usually by truck) to the final delivery
points (FDPs), at which they are handed over to WFP’s cooperating partners: local NGOs that take care of
what is called the last mile distribution (Balcik et al. 2008) in order to reach the final distribution points (e.g.,
schools, villages, hospitals, etc.). Note that the logistics network setup may vary between countries and is very
context dependent; WFP may take charge of the entire delivery network (from pickup at the supplier to last
mile distribution), or it may outsource the network partially or entirely to local logistics providers or
Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

sometimes even the government. For example, WFP has about 5,000 trucks on the road every day, but its own
fleet consists of “only” 1,000 trucks, illustrating that much of the transportation is outsourced.

3. Model and Implementation

In this paper, we extend a capacitated, multicommodity, multiperiod network flow model with nutritional
components. For an introduction to network flow models, we refer to Ball et al. (1995); an example tailored to a
humanitarian supply chain can be found in Haghani and Oh (1996). This type of model is commonly used to
optimize sourcing and delivery strategies, two of the four components that we aim to integrate. The supply
chain network is sketched as per Figure 1. The suppliers are modeled as source nodes with the discharge ports
and WFP warehouses acting as transshipment nodes. The demand (or sink) nodes are the FDPs, at which
demand is dependent on the (variable) food basket and the number of beneficiaries.
To integrate the food basket design into the network flow model, we define a ration variable that governs
the commodities flowing into an FDP, ensuring that the food WFP sends addresses the nutrient gap, is
distributable, and is palatable. This link between nutrients, commodities, and rations resembles the flexible bill
of materials sometimes found in the manufacturing industry (Ram et al. 2006), for which the end product
is a (monthly) food basket and the number of beneficiaries in the FDP is the demand for this end product. An
interesting distinction is that, in the manufacturing industry, the end product is fixed and the fulfilled demand
is variable, whereas here we usually consider the fulfilled demand fixed (100% of beneficiaries should
receive a food basket) and make the quality of the end product variable instead. So, when there is a funding
shortfall, WFP typically prefers to supply a less nutritious food basket to all beneficiaries rather than sup-
plying the full food basket to fewer beneficiaries. Product design and sourcing are traditionally done sep-
arately, and there is much potential to improve end-to-end performance through joint decision making (Novak
and Eppinger 2001).
Integrating the transfer modality selection is possible but not straightforward. The most pragmatic approach
is to model local markets as source nodes that are linked directly to FDPs. Beneficiaries can then receive some
or all of their commodities from these local markets through commodity vouchers (one of the transfer
modalities). We discuss a more advanced approach that also covers value vouchers and cash in Section 6.
Note that the model we are using at WFP is vast and needs to be able to handle a plethora of (mathematically
trivial) constraints, such as sourcing restrictions, capacity utilization, beneficiary preferences, funding allocation,
etc. Although this comes with an increased risk of conflicting constraints, in practice, it is rarely an issue. Ad-
ditionally, the core of the model is based on two traditional mathematical problems: a network flow model (Ball
et al. 1995) combined with a diet problem (Dantzig 1990). Because these two formulations are well known, in
the main body of the paper, we discuss only our unique contribution: how to connect the two. A more
comprehensive description of the model that was used for the case studies can be found in Appendix A.

Figure 1. High-Level Overview of the Modeled Supply Chain Network

 Peters et al.: The Nutritious Supply Chain
INFORMS Journal on Optimization, 2021, vol. 3, no. 2, pp. 200–226, © 2021 INFORMS 205

3.1. Defining Demand

Typically in multicommodity network flows, the demand is defined as a parameter demikt that specifies the
quantity that should be delivered for every location i, commodity k, and period t. In order to connect such a
network flow to a diet problem, here we break up the demand using several parameters:

demit = number of beneficiaries at location i in period t

dayst = feeding days in period t
Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

nutreql = daily requirements for nutrient l

nutvalkl = contribution to nutrient l per gram of commodity k.

We define variables Rk (continuous, nonnegative) to optimize the food basket composition, where Rk
represents the gram per person per day of commodity k that WFP supplies. These ration variables ensure that
beneficiaries receive the same food basket and allow us to impose restrictions on the composition of the ration.
For example, this formulation makes it easy to specify that beneficiaries should receive at least 300 grams of
rice per day (Rrice ≥ 300) or that split peas are not accepted (Rsplit peas  0). Generally speaking, Rk should be
such that the nutrient gap is closed:
∑
nutvalkl × Rk ≥ nutreql ∀l. (1)
k

Rather than a fixed demand parameter demikt , we end up with a variable demand:
Dikt  demit × dayst × Rk ∀i, k, t. (2)
Note that the ration variable Rk can easily be extended to allow more flexibility; for example, defining rations
as Rkt allows for a food basket that is dynamic over time. This helps WFP to make the most of seasonal price
windows (e.g., the harvest season basket could be different from the lean season basket) or temporarily reduce
rations (e.g., in case of a pipeline break). Similarly, we can define the rations as Rbk to cover the needs of
different beneficiary types (e.g., different nutritional requirements). Differentiation by location (Rik ) is less
common for in-kind food baskets as humanitarian organizations strive to harmonize distribution across
operations (based on principles of fairness and equality). When consumption patterns differ within assisted
populations (e.g., refugee camps versus local population), alternative rations are typically modeled using the
Rbk approach; that is, they are considered a different beneficiary type. Within each beneficiary type, the aim is
typically to harmonize distribution across locations, but differentiation could open up opportunities to di-
versify the basket using commodities that are only available in a handful of locations (e.g., procurement from
local markets). An example of location differentiation for CBT (with which such an approach is more
commonly accepted) is discussed in Section 6.

3.2. Objective and Goals

Conventional network flow models usually revolve around profit maximization or cost minimization, but for
humanitarian operations, there are often additional considerations. We refer to Holguı́n-Veras et al. (2013)
for a comprehensive treatise on this subject.
In this paper, we consider four main classes of objectives:
1. Efficiency relates to resource utilization and covers objectives such as the cost of the operation and the
utilization of port capacity.
2. Effectiveness relates to the nutritional impact of WFP’s assistance and covers objectives such as the
remaining nutrient gap and the dietary diversity score.
3. Development relates to the impact on the local economy and covers objectives such as the dollars spent on
local procurement and CBT assistance.
4. Agility relates to the responsiveness of the solution and covers objectives such as the maximum and
average lead time of the supply chain.
Typically, users of the model specify a range of programmatic goals (e.g., they request a solution that covers
100% of the nutrient gap, covers 30% of the needs through local procurement with a lead time of at most three
months). The model then identifies the solution that meets these targets at the lowest possible cost (i.e., Pareto
optimal). Some examples of common goals and their mathematical formulation are provided in Appendix A.4.
Depending on the goals and objective chosen, additional variables and constraints are added to the base
model. Many goals can be formulated as linear constraints even if the measure itself is nonlinear (e.g.,
percentage local procurement). Note that not all combinations of objectives and goals are possible in a
 Peters et al.: The Nutritious Supply Chain
206 INFORMS Journal on Optimization, 2021, vol. 3, no. 2, pp. 200–226, © 2021 INFORMS

linear model. For example, maximizing the number of beneficiaries would conflict with a variable food basket
because, in the demand constraint of the network flow model, we multiply the number of beneficiaries with
their demand; if both are variable, the resulting model is nonlinear.

3.3. Data
One of the biggest challenges of applying prescriptive analytics to humanitarian operations is finding and
processing the necessary data, which is typically of poor quality or even nonexistent. Often, we see that
humanitarian applications fail postpilot because their data requirements are too difficult to collect or sustain
Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

over a longer period of time. To avoid this common pitfall, from the start, the optimization model was rooted
in data that was already being collected across the organization for other purposes (e.g., reporting, budgeting,
payments, etc.). Although every country was using a different combination of systems and tools to manage
their operations, there were enough commonalities to ensure that core parameters, such as procurement costs,
beneficiary figures, and transportation options, were always available. The underlying model was designed in
such a way that many parameters are optional, so even if a country does not have part of the data (e.g., price
forecasts) and there is no time to collect it, they can still apply the model.
Although data availability can be very limiting, the existence of data does not mean it is easily usable. In
practice, data are spread over many different sources, systems, and technologies. In the worst case, data has to
be extracted from scanned copies of reports (pdf files), but more likely it comes from Excel files or databases.
To ensure the sustainability of an analytics-heavy tool, a lot of emphasis should be on the integration of data
from disparate sources. For any parameter, there are multiple possible data sources, each following a different
taxonomy (i.e., master data). Properly consolidating data from multiple sources, therefore, requires significant
investments in data integration.
For the first three years, the data pipeline required for this optimization model was managed in Excel,
involving labor-intensive manual collection and integration of data sets prior to any analysis. During 2018 and
2019, the entire data logic was automated with hundreds of data sets across more than a dozen systems
automatically being refreshed and integrated every night. By minimizing the dreaded “data request” that
typically precedes optimization exercises, users can now quickly start interacting with the model to identify
improvements to their operation.

3.4. Solution Approach

The mathematical model is implemented in Python and solved using the COIN-OR solver (Saltzman 2002)
with the PuLP module (Mitchell et al. 2011) acting as the interface between the two. PuLP does not support
multithreading, so all calculations are performed using a single processor. Parallel processing is supported,
however, so it is possible to run multiple scenarios simultaneously (asynchronous). The coding language,
solver, and module are all freeware, making the tool easy to implement in a humanitarian context.
WFP staff can interact with the mathematical model through a web application (Optimus) coded in Python
and hosted using Django. The web app allows them to add operational constraints and specify goals for the
key outputs and gives them access to a range of automated analyses. Results from their analyses are displayed
in interactive graphics (e.g., maps, pivot tables, etc.) that provide quick insights into all of the important
decision variables (e.g., food basket composition, sourcing plan, delivery plan, etc.) and the resulting per-
formance (e.g., cost, nutritional contents, etc.).
Optimized plans are typically generated in an iterative way. Starting from a baseline scenario that mirrors
the current plans for the operation (i.e., the food basket, sourcing strategy, and delivery plan), alternative
scenarios are generated (e.g., by relaxing some of the constraints) to identify potential areas of improvement.
Local experts validate the resulting plans and key performance indicators to assess their feasibility and
relevance, which often leads to additional constraints being included in subsequent runs of the model. Some
practical examples and outputs of this approach are described in the “Applications” section.

3.5. Verification and Validation

When using analytics to (re)design or manage an operation, it is of paramount importance that the model and
data are thoroughly verified and validated. In the humanitarian field in particular, many people are ana-
lytically averse—there used to be a general consensus that humanitarian operations are too difficult to model
and optimize. This made extensive and continual verification and validation crucial. The optimization model
and tool were developed over the course of many years (starting in 2013) in close collaboration with the end
users and consumers of the outputs: WFP decision makers in the field. Iterative development helped identify
the required capabilities, model them, and ensure that they were implemented correctly.
 Peters et al.: The Nutritious Supply Chain
INFORMS Journal on Optimization, 2021, vol. 3, no. 2, pp. 200–226, © 2021 INFORMS 207

For example, when collaborating with WFP’s country offices, a recurring validation step is to ensure that the
model outputs mirror the budgeted costs when imposing the current plan on the optimization model. This
allows users to quickly compare the model’s calculations with theirs, which immediately highlights any
components that are estimated differently or for which some unmodeled costs are incurred. The latter
triggers a series of in-depth discussions during which we identify the unmodeled costs and find a way to
introduce them into the mathematical model.
Another form of validation employed regularly is comparing the current operational plan to unrestricted
solutions from the optimization tool, allowing analysts running the optimization tool to rapidly identify
Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

operational constraints and preferences. This approach also generates strong buy-in from the local experts
because they start appreciating the scope, flexibility, and speed of the optimization tool.
The frequent rounds of verification and validation contribute significantly to the adoption rate of the tool
and ensure that the model is built on data that is actually available and that its results are intuitive and usable
for WFP decision makers.

3.6. Assumptions and Limitations

Every optimization tool of this size comes with a long list of assumptions and limitations; we highlight the
most important ones.
We assume that every beneficiary (of a certain type) receives the same food basket and that no beneficiary
goes unfed. This assumption is based on the humanitarian principles of humanity and impartiality; it would
be unethical to feed only those beneficiaries that are conveniently located, for instance.
We do not model the transportation network between suppliers and loading ports. WFP floats tenders in
specific countries to ensure competitive prices, and this means that price forecasts are done at the country
level. As a result of the tendering process, the suppliers may differ every time WFP orders, making the cost
from a supplier to the loading port difficult to estimate. Generally, WFP requests free-on-board offers, which
means that the supplier takes care of the transportation up to the loading of the ship at a major port in
their country.
Costs are captured using dollars per metric ton rates. Given WFP’s economies of scale and the type of
contracts that it usually has with suppliers and contractors, these linear rates are representative of the ac-
tual costs.
The modeled nutrition measure, the nutrient value score, is a simple average across all nutrients. We are
working with WFP’s nutrition department to come to better measures, reflecting the importance of some
nutrients over others, and covering a larger range of nutrients overall. For general food distribution, for
example, a shortfall in macronutrients (energy, protein, fat) is often considered more severe than a shortfall in
minerals or vitamins. In programs that prevent or treat malnutrition, on the other hand, we want to favor
these micronutrients more heavily.
We consider the supply chain network to be fixed. WFP has a large and well-established presence in most
countries, and although questions surrounding the setup of these networks arise occasionally (e.g., opening or
closing of corridors and hubs), they are typically not addressed using this particular model (which is more
focused on the design of the food basket and other programmatic decisions).
The current model is not robust against uncertain parameters (such as costs, lead times, and capacities).
Although we recognize the potential of robustness in optimizing humanitarian supply chains (see, for ex-
ample, the El Niño application in the next section), the introduction of robust optimization has been put on
hold for the moment. Explorations of robustness extensions by masters students indicate that many types of
uncertainty can be introduced without significantly affecting the tractability of the model (which becomes
conical but remains convex). However, these extensions come with significant interpretation challenges and
make it difficult for decision makers to understand (and, therefore, validate and implement) the resulting
plans. In the meantime, users of the model verify with WFP’s experts that the parameter values used are the
current “best guess” and base the analyses on that.

4. Analyses and Insights

The model described in Section 3 is very flexible and supports a wide range of analyses that can help evaluate
and improve the efficiency and/or effectiveness of WFP operations. In this section, we explore some of the
insights that can be obtained from this integrated optimization approach and how this is supporting decision
making at WFP.
 Peters et al.: The Nutritious Supply Chain
208 INFORMS Journal on Optimization, 2021, vol. 3, no. 2, pp. 200–226, © 2021 INFORMS

Figure 2. The Pareto-Efficient Curve for the NVS

Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

Notes. For different levels of NVS, the optimization tool finds the cheapest food basket, sourcing plan, and delivery plan. On the y-axes, we have
the cost per beneficiary per month (left) and the cost of the entire operation per month (right).

4.1. Trade-offs
To optimize humanitarian operations, it is important to evaluate and balance a wide range of (often con-
flicting) objectives. Although the mathematical model has a single objective (usually costs), solutions are
evaluated using multiple metrics (efficiency, effectiveness, agility, and development). A goal programming
approach can be used to find Pareto-efficient curves for the various metrics, providing insight into what kind
of outcomes are achievable and how this would affect the cost of the operation. We highlight two of these
trade-off analyses in particular: one for the nutritional value score (effectiveness) and one for the cash-based
transfer ratio (development).
The nutritional value score trade-off curve (Figure 2) shows the lowest cost at which WFP can attain
different levels of nutrition (effectiveness). Each plot represents a solution to the mathematical model; that is, it
corresponds to an optimized food basket, sourcing plan, and delivery plan that takes into account all user-
added constraints. Ideally, beneficiaries receive 100% of the recommended daily intake (RDI), but depending
on the context (available funding, number of beneficiaries in need, access to nutritious commodities), this may
not be realistic. With this trade-off graph, decision makers can see the price tag of supplying different levels of
nutrition. In Figure 2, for instance, we can see that supplying 95% of the required nutrients reduces the cost of
the operation from US$10.8M to US$8.8M (18.7% reduction). This means that WFP could still supply all
beneficiaries with 95% of their requirements if its funding is 18.7% short or, alternatively, that it could reach
15%–20% more beneficiaries with 95% of their requirements if there is no funding shortfall. Because of this
significant difference in cost and potential outreach, it is important to understand what level of nutritional
requirements is most appropriate. If supplying 95% is still OK for the majority of beneficiaries, it may have a
big impact on how humanitarian organizations design their programs.
The cash-based transfer ratio trade-off (Figure 3) shows the lowest cost at which WFP can attain different
levels of cash and voucher transfers. CBT are very popular among many donors at the moment, but depending
on the operation, CBT may be more expensive than supplying the food in-kind, so it is important to assess
whether the premium price weighs up against its benefits. This graph displays the most cost-efficient level of
CBT and how the cost increases as one deviates further from the optimal ratio. We can observe (Figure 3) that,
 Peters et al.: The Nutritious Supply Chain
INFORMS Journal on Optimization, 2021, vol. 3, no. 2, pp. 200–226, © 2021 INFORMS 209

Figure 3. The Pareto-Efficient Curve for the Cash-Based Transfers Metric

Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

Notes. For different levels of CBT (as a percentage of the total metric tons distributed), the optimization tool finds the cheapest food basket,
sourcing plan, and delivery plan. On the y-axes, we have the cost per beneficiary per month (left) and the cost of the entire operation per
month (right).

initially, the cost decreases when increasing CBT as the model starts supplying the most cost-effective local
commodities through commodity vouchers. Increasing CBT further means that the model has to start using
less cost-effective local commodities in order to supply the required nutrients. The example shows that the
optimal ratio for this operation is 33% but that it is possible to scale up to 70% CBT with only a small increase
in cost (US$0.9M; 4.4%). Decision makers have to weigh the added value of increasing CBT for local markets
and beneficiaries against this increase in cost. Insight into the relative cost of CBT compared with in-kind and
to what extent the local markets can absorb this increased demand is very important when deciding on the
optimal transfer modality.

4.2. Cost-Effectiveness of Commodities

Optimal food baskets are highly dependent on the available commodities, their nutritional profile, and the
nutrient gap that we aim to close. For example, let us consider four cereals that are viable for the Nigeria
operation: rice, sorghum, wheat, and (white) maize.
Table 1 displays the nutritional contents of these four cereals for 11 nutrients. By optimizing the sourcing
strategy for these four cereals, we can obtain their minimum cost (end to end). Combining the optimal costs
with the nutritional contents allows us to calculate the relative cost-effectiveness of each of these commodities.
Table 2 shows how much (in U.S. dollars) WFP would have to spend on each cereal to supply 100% of the RDI.
For example, we can observe that it requires $0.66 worth of rice to supply a days’ worth of kilocalories (2,100)
to a beneficiary as opposed to $0.54 in the case of sorghum. Similarly, we can learn that wheat is the most cost-
effective when it comes to supplying proteins, white maize for riboflavin, et cetera. Across all nutrients,
sorghum is the most cost-effective cereal for Nigeria, and rice is the least cost-effective. Of course, a food
basket is composed of more than just cereals, so the availability of commodities, such as oils and pulses, and
their respective cost-effectiveness decides what the optimal cereal is. All commodities have a unique nu-
tritional profile, and the optimization model described in this paper can help balance these nutritional values
when designing food baskets while taking into account the resulting costs.
 Peters et al.: The Nutritious Supply Chain
210 INFORMS Journal on Optimization, 2021, vol. 3, no. 2, pp. 200–226, © 2021 INFORMS

Table 1. The Nutritional Value of Four Cereals (per 100 g) with the RDI for Reference

Energy, Protein, Fat, Calcium, Iron, Iodine, Vitamin A, Thiamine, Riboflavin, Niacin, Vitamin C,
Cereal kcal g g mg mg μg μg mg mg mg mg

Rice 360.00 7.00 0.50 9.00 1.70 0.00 0.00 0.10 0.03 5.58 0.00
Sorghum 335.00 11.00 3.00 26.00 4.50 0.00 0.00 0.34 0.15 5.00 0.00
Wheat 330.00 12.30 1.50 36.00 4.00 0.00 0.00 0.30 0.07 8.92 0.00
White maize 350.00 10.00 4.00 7.00 2.71 0.00 0.00 0.39 0.20 2.20 0.00
Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

RDI 2100.00 52.50 40.00 450.00 22.00 150.00 500.00 0.90 1.40 13.86 28.00

Note. The units differ by nutrient (e.g., energy is measured in kilocalories, whereas iron is measured in milligrams).

One of the benefits of the integrated model is that these cost-effectiveness calculations can be done at the
location level; for any location hosting beneficiaries, the end-to-end costs are calculated and incorporated into
the final recommended ration. Traditionally, the design of the ration was decoupled from the supply chain
execution, which means “optimal” rations were designed based on the nutritional values of commodities and a
simpler estimation of the costs for each commodity—typically, just the procurement cost. In this Nigeria
example, sorghum and wheat cost approximately the same ($0.01 per kilogram difference) to provide to
beneficiaries (considering end-to-end costs). However, the average procurement cost for sorghum during the
last four years has been $0.36 per kilogram, whereas wheat was purchased on average for $0.24—a third
cheaper. If the design were based on such simpler cost estimations, the optimized rations would be heavily
skewed toward wheat instead of the sorghum that comes out on top if we consider the operational costs more
holistically.
Rather than redesigning food baskets from scratch using these insights, in practice, it is often more
pragmatic to stay close to the current design of an operation (making any recommendations much easier to
implement). We highlight one particular analysis that allows WFP to identify slight changes to the current
food basket that can significantly improve the performance of an operation: a commodity swap analysis.
The commodity swap analysis (Figure 4) shows the impact of swapping one commodity in the current food
basket on the performance of the operation—in this case, on the supplied nutrients (effectiveness) and op-
eration costs (efficiency). The graph shows an analysis of cereals and grains for one of WFP’s activities. In each
iteration of the model, one of the cereals/grains in the current food basket is replaced with a cereal/grain that
is not currently included (using the same ration size). We can observe, for example, that replacing the wheat in
the current basket with white maize would increase the NVS by 0.3%, and reduces the cost of the operation by
US$4.5M (26.5%). Decision makers can combine these quantitative insights with context-specific constraints
(such as donor and beneficiary preferences) to find the best food basket design. For instance, changing to
sorghum/millet makes no sense if the beneficiaries have no idea how to prepare this grain no matter how cost-
effective the commodity is. In this analysis, we stay close to what the beneficiaries are already receiving. One
of the major benefits of this approach is that we can arrive at solutions that, although not mathematically
optimal, are easy to validate and have a significantly higher likelihood of actually being implemented.

5. Applications
The mathematical model presented in this paper is the result of iterative development over many years (since
2013). Through regular pilots we were able to identify what data are reliably available, what kind of analyses
are the most impactful, and how the optimization tool fits into WFP’s business processes. Additionally, these
pilots were used to continually verify and validate the mathematical model underlying the tool by scheduling

Table 2. The Relative Cost-Effectiveness of Four Cereals with Respect to Supplying Different Nutrients

Cost, Energy, Protein, Calcium, Iron, Vitamin Thiamine, Riboflavin, Niacin, Vitamin
Cereal $/kg $ $ Fat, $ $ $ Iodine A $ $ $ C

Rice 1.13 0.66 0.84 9.01 5.63 1.46 1.01 5.25 0.28
Sorghum 0.86 0.54 0.41 1.14 1.48 0.42 0.23 0.80 0.24
Wheat 0.87 0.55 0.37 2.32 1.09 0.48 0.26 1.74 0.14
White maize 0.93 0.56 0.49 0.93 5.96 0.75 0.22 0.65 0.58

Notes. Each value indicates how much U.S. dollars we have to spend on the cereal every day to supply 100% of the nutritional requirement, so the
lower the value, the more cost-effective the commodity is for that particular nutrient. The cereals do not provide iodine, vitamin A, or vitamin C
so no cost-effectiveness measure can be calculated for these nutrients..
 Peters et al.: The Nutritious Supply Chain
INFORMS Journal on Optimization, 2021, vol. 3, no. 2, pp. 200–226, © 2021 INFORMS 211

Figure 4. This Commodity Swap Analysis Shows the Impact on Nutrition and Cost of Swapping One Commodity (from the
Cereals and Grains Food Group)
Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

Notes. The y-axes show the efficiency with the cost per beneficiary per month on the left and the monthly operation costs on the right. On the x-
axis, we show the effectiveness as a percentage of the minimum daily nutrient requirements. Each plot corresponds to a unique food basket
(including its optimal sourcing and delivery plan) with the circle representing the current food basket. All plots that are below and to the right of
the circle are, therefore, strict improvements (from a cost and nutrition perspective, respectively).

in-depth sessions with experts throughout the organization and by comparing the tool’s outputs to historical
performance.
Since we started applying the tool, optimization gained significant traction within WFP. At present, the
software is used on a regular basis to provide WFP’s biggest and most complex operations with optimization
support, such as Syria, Ethiopia, South Sudan, and Yemen. In this section, we highlight some of the results we
achieved in Iraq, Yemen, and the South African region and show how different types of analyses are being
used to support strategic decision making for WFP’s most complex operations. For each of the analyses, we
provide some context, such as the cause of the crisis and the number of people affected, but please note that
ongoing crises evolve rapidly and, therefore, the provided statistics may no longer be up to date at the time
of reading.

5.1. Iraq
Years of conflict have hindered Iraq’s economic development. Since 2014, the occupation of the Islamic State of
Iraq and the Levant in Iraq has resulted in the displacement of more than three million people. When, two
years later, the Iraqi Security Forces launched a military offensive to regain control, fighting deepened in-
security, rolled back development, and exacerbated vulnerabilities. Many Iraqis sought refuge in neighboring
countries and in Europe. Beset by violence, social disruption, and economic hardship, thousands of Iraqi
families were left in desperate need of food assistance.
Although many are now returning home every month, around 700,000 Iraqis are still living in camps with
few possibilities to earn an income that enables them to put food on the table. In addition, an estimated
quarter of a million Syrian refugees have sought refuge in northern Iraq, placing additional pressure on
limited resources. At the end of 2017, around 800,000 people still require some sort of food assistance in Iraq
every month; some 10 million need humanitarian assistance in general. The security landscape remains
volatile, sometimes posing access challenges for humanitarian actors (World Food Programme 2015a).
 Peters et al.: The Nutritious Supply Chain
212 INFORMS Journal on Optimization, 2021, vol. 3, no. 2, pp. 200–226, © 2021 INFORMS

WFP has been operating in Iraq since 1991 and has provided food assistance in the country since April 2014
through emergency operations to assist those hundreds of thousands of Iraqis forced from their homes by
recent violence. In the face of further mass displacement from major Iraqi cities, such as Mosul and Ramadi,
WFP scaled up activities to reach an average of 1.5 million people per month in all 18 governorates, including
hard-to-reach areas. WFP is assisting people through monthly family food parcels for those with access to
cooking facilities, food vouchers that can be redeemed at local shops, and ready-to-eat food known as im-
mediate response rations that provide a family of five with food for three days.
WFP’s operations in Iraq often face funding shortfalls, so it is vital that the operation’s design is as cost-
Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

effective as possible. In October 2015, Iraq’s country office requested optimization support to redesign the food
basket for different levels of funding, allowing it to supply as many kilocalories as possible from the donations
that it receives.
We worked intensively with the field staff to gather all the necessary data and to identify operational
constraints (with respect to procurement, logistics, transfer modalities, beneficiary preferences, etc.). Once we
had a good grasp of the situation, we ran hundreds of analyses using the optimization tool to identify al-
ternative designs that could improve the performance of the operation. Many of the solutions were pushing
the envelope, so there was much going back and forth to verify the feasibility of these solutions. This intense
collaboration and the iterative process resulted in the field staff feeling a strong sense of ownership of the final
outcomes, resulting in a rapid implementation of the final recommendation.
One of the main deliverables to Iraq’s management was Figure 5, containing our recommendations for Iraq’s
family food parcel (FFP)—supplied to half a million Iraqis every month. The first column represents the latest
official food basket (BR4), supplying 1,844 kilocalories (daily) at a cost of US$13.13 per beneficiary per month.
Each subsequent column is an optimized basket that uses different interventions (commodity swaps, ration
adjustments, etc.) to improve the cost-effectiveness of the operation. Iraq officially adopted the food basket
represented by the green column, supplying 1,841 kilocalories (only three kilocalories fewer) and 69% of the

Figure 5. An Overview of Alternative Food Baskets for Iraq’s FFP

Notes. We compare the kilocalories and cost per beneficiary per month of each option (as given by the acronyms on the x-axis). The acronyms
show how the food basket changed compared with the original food basket (BR4). RIC = rice, BUL = bulgur wheat, WHF = wheat flour, SRL =
split red lentils, WRL = whole red lentils, WIB = white beans, CHK = chickpeas, HOB = horse beans, OIL = sunflower oil, SUG = white sugar. A
“+” or “−” denotes an increase or decrease in the current ration size. A “0” denotes a removal of the commodity, whereas the “2” denotes that we
change commodity a to commodity b using the same ration size. Ration size increments are predefined and based on commercial packaging
types for that specific commodity. The green option has been implemented in practice, representing a cost reduction of 12% while supplying only
three kilocalories fewer.
 Peters et al.: The Nutritious Supply Chain
INFORMS Journal on Optimization, 2021, vol. 3, no. 2, pp. 200–226, © 2021 INFORMS 213

nutrient gap (same as BR4) at a cost of US$11.54 per beneficiary per month (12% reduction). At the time of
implementation, the FFP was being supplied to 800,000 beneficiaries every month, so this adoption corre-
sponded to US$1.3M monthly savings. This meant that WFP could supply the same amount of beneficiaries
with a nutritious food basket in the case of up to a US$1.3M funding shortfall per month, or these 12% savings
could be used to supply an additional 109,000 (14%) beneficiaries with an FFP every month if funding
remained at the same level. Note that, in this analysis, all savings result from changing the food basket; we did
not identify any significant savings on the existing supply chain network. The optimized food basket was
distributed throughout 2016 and 2017, representing total savings of more than US$25M, with positive
Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

feedback from the beneficiaries regarding the changes in commodities and ration sizes compared with the
original food basket.

5.2. Yemen
Even before fighting broke out in early 2015, Yemen was one of the poorest countries in the Arab world. With
an average life expectancy below 64, the nation is ranked 178th out of 188 on the human development index, a
composite statistic of life expectancy, education, and per capita income indicators that is used to rank
countries into four tiers of human development.
Nearly four years of conflict have left thousands of civilians dead and three million internally displaced. Its
impact on the country’s infrastructure has been devastating with major overland routes and airports severely
damaged. Despite ongoing humanitarian assistance, 15.9 million people wake up hungry every day, which is
more than half of Yemen’s population of 28 million. It is estimated that, in the absence of food assistance, this
number would go up to 20 million (World Food Programme 2015b).
WFP has been active in Yemen since 1967. In 2019, WFP is scaling up to provide 12 million people with
monthly food assistance through direct food distributions or vouchers that people can use at retailers in areas
where the markets are functioning. Each family of six gets a monthly ration of wheat flour, pulses, vegetable
oil, sugar, and salt. Beneficiaries include internally displaced persons and returnees, vulnerable populations in
the most food-insecure areas, people affected by transient crises, infants, pregnant and nursing women af-
fected by acute and chronic malnutrition, and school-age children.
Since the start of the main operation, WFP has gradually been scaling up toward the 12 million beneficiaries
and this is proving to be a Herculean task in light of the limited resources available and the escalation of
conflict within Yemen. In December 2015, WFP was still reaching three million beneficiaries with in-kind food
assistance every month with the aim of increasing this to five million in 2016. Cash-based transfers were being
scaled up to one million beneficiaries in parallel, allowing WFP to reach six million people with life-saving
assistance. In November 2015, Yemen’s management team requested optimization support for their pending
scale-up from three to six million beneficiaries.
Similar to our approach in Iraq (and all other applications), we worked intensively with local experts and
management to gather the necessary data and identify all operational constraints from a supply chain and
donor/beneficiary preference perspective. We were requested to keep the commodities in the food basket the
same and focus mainly on the optimization of ration sizes. Given that distribution is carried out in active

Figure 6. This Beneficiary Matrix (Yemen) Shows the Interdependence of Funding Levels, Nutritional Targets,
and Beneficiary Numbers

Notes. Columns (2)–(4) show the performance of a food basket (column (1)). We display the cost per beneficiary per month and two effectiveness
measures: the kilocalories and the percentage of all nutrients supplied (as percentages of the minimum daily requirements). The remaining
columns show how many beneficiaries we can supply (in million beneficiaries per month) with these food baskets under different funding
scenarios (ranging from US$20 to 100M per month). For example, the current food basket can supply 4M beneficiaries if WFP receives US$55M
per month and 5M beneficiaries in the case WFP receives US$70M per month.
 Peters et al.: The Nutritious Supply Chain
214 INFORMS Journal on Optimization, 2021, vol. 3, no. 2, pp. 200–226, © 2021 INFORMS

conflict zones, we designed the food baskets in such a way that the monthly ration for each family (of six)
corresponds to an industry-standard packaging size, making distribution as quick and seamless as possible.
We presented a breakdown of our recommendations in the form of a column chart similar to the one used in
Iraq, but the most powerful deliverable is a “beneficiary matrix.”
The beneficiary matrix (Figure 6) shows the interdependence of funding levels, nutritional targets, and
beneficiary numbers. It shows decision makers how much it costs to supply different numbers of beneficiaries
with the current food basket and how alternative food baskets allow WFP to either cope with funding
shortfalls (without cutting down on beneficiary numbers) or to increase the number of beneficiaries reached
Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

with the available funding (by cutting down on nutrition). For example, suppose WFP is supplying four
million beneficiaries with the current food basket and would like to scale up to five million. The matrix shows
us that supplying five million beneficiaries with the current basket would cost US$70M per month (US$15M or
27% more than the current cost). Alternatively, WFP could reach the five million beneficiaries by slightly
reducing the nutrition levels of the food basket. In this case, we can observe that there is a food basket (WHE50
SYP10 OIL7 WSB10) that supplies five million beneficiaries with 80% kilocalories and 90% NVS for the same
level of funding (US$55M). Insight into this trichotomy allows WFP to manage its scale-up properly and
enables it to communicate clearly to donors what is required from them if WFP is to reach all people at risk
within Yemen with nutritious food. Every few months, the matrix is updated to reflect the latest data changes
to the point at which WFP was able to reach 13 million beneficiaries in 2019.

5.3. El Niño
In 2016–2017, the El Niño climate pattern, which is strongly linked to weather fluctuations around the globe,
was fueling an international food-security crisis for millions of people. By disturbing rainfall and temperature
patterns, El Niño affected agriculture, water supplies, and the spread of disease and was threatening the food
security and livelihoods for some 60 million people worldwide. Particular areas of concern included nearly all
of Southern Africa, which was the hardest hit region; Ethiopia and its neighbors Somalia and Sudan in East
Africa; central America’s “dry corridor”, nearby Haiti, and the northern region of South America; and many of
Asia’s island nations, including Indonesia, Papua New Guinea, and Philippines.
Between October 2015 and January 2016, El Niño conditions caused the lowest recorded rainfall in at least
35 years across many regions of Southern Africa. The same period also recorded the hottest temperatures in
the past 10 years. El Niño’s impact on rain-fed agriculture is severe. Poor rainfall combined with excessive
temperatures created conditions that were unfavorable for crop growth in many areas. In Lesotho, South
Africa, Swaziland, Zambia, and Zimbabwe, planting was delayed by up to two months or more, which
severely impacted maize yields. It became evident that the 2015–2016 maize harvest would be insufficient to
cover full cereal needs for the region without significant importation.
As more information about the impact became available, we were asked to provide support by drafting
various contingency analyses. Given that the local cereal availability would be too limited to meet all needs,
we used the model to explore global sourcing strategies under different capacity and export assumptions. For
example, if some of the countries barely have enough to meet their own needs, they may impose export bans
to ensure availability for their own population. We found that many countries appeared to be self-sufficient,
but WFP would need to source about 20,000 metric tons of white maize from international markets—more
than a quarter of the total needs. Given the market prices and the strong GMO restrictions in the South African
region, Argentina and Mexico were identified as the optimal sourcing locations by our model. In Figure 7, we
show the global sourcing plan.
With slow-onset disasters, such as the El Niño crisis, the assumptions and forecasts are very prone to change
as the crisis develops. Because of this, it is important to perform sensitivity and robustness analyses so that
WFP can ensure its plans are still valid if the forecasts change. At the time of the analysis, South Africa was
procuring significant quantities locally, in particular for cereals and pulses. If the availability of commodities
reduces because of poor harvests, we often see significant price spikes in local markets, so it was important to
evaluate whether a local procurement strategy still made sense if the local prices increased drastically. In
Figure 8, we show an optimized sourcing strategy under different price scenarios with Figure 9 showing the
detailed plan (aggregated by origin country) for white maize. We find that, for the current prices, the optimal
breakdown is to purchase 29.7% international, 31.9% local, and 38.4% regional (across all commodities). Price
increases as little as 10% already increase the optimal international procurement ratio to 40% and up to 78%
when local/regional prices double. If we consider the cost of the baseline plan (i.e., optimal plan under the
assumption that the prices remain stable) under different price scenarios, we see that it is up to US$700,000
more expensive per month (7.8%) than the optimized plan for scenarios in which prices increase. We see that,
 Peters et al.: The Nutritious Supply Chain
INFORMS Journal on Optimization, 2021, vol. 3, no. 2, pp. 200–226, © 2021 INFORMS 215

Figure 7. (Color online) An Optimized Global Sourcing Plan in Case Local Cereal Production Is Not Sufficient to Meet All
Needs in the Most Affected South African Countries
Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

depending on El Niño’s effect on the local markets, which is likely to increase local prices, we may have to
consider transitioning to a more international-heavy sourcing plan. The large differences between the sce-
narios mean that the current sourcing strategy for white maize is not robust against price uncertainty.
Considering that a shift toward international (and more price-stable) procurement would increase lead times
by up to two months for many commodities, it is important to start shifting straight away.

6. Extensions and Future Research

The underlying model described in this paper is very adaptable, so there are many opportunities to extend it
to cover additional levels of complexity or to apply it to non-WFP supply chains. For example, the model was
initially developed to manage two transfer modalities: in-kind transfers and commodity vouchers. Since then,
the humanitarian community and WFP, in particular, have seen a steady shift toward less restricted forms of
assistance, such as value vouchers and cash. In this section, we describe how the model can be extended to
deal with these transfer modalities as well. Additionally, we describe some other use cases that we are
exploring as future research.

6.1. Cash-Based Transfers

Even though cash-based transfers were introduced more than a decade ago (2007), it is still difficult to
properly evaluate and compare the various transfer modalities from a cost-effectiveness perspective. The main
reason for this is that, for two of the three CBT approaches (value vouchers and cash), the effectiveness of the
transfer is dependent on the beneficiary’s purchasing behavior and local market conditions (cost, availability).
For commodity vouchers, we know exactly what the beneficiary will receive and base our effectiveness on
that. In practice, we consider two main approaches for evaluating the effectiveness of CBT assistance: ex ante
and ex post.
In an ex ante analysis, we make a best guess about the purchasing behavior of the beneficiary (e.g., 15 kg
maize, 5 kg pulses, 2 L oil, etc.) and calculate the effectiveness based on the assumption that all beneficiaries
purchase and consume this basket. This “best guess” basket is usually based on local expert knowledge and
 Peters et al.: The Nutritious Supply Chain
216 INFORMS Journal on Optimization, 2021, vol. 3, no. 2, pp. 200–226, © 2021 INFORMS

Figure 8. Optimized Sourcing Strategies Under Different Price Scenarios

Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

Notes. On the left axis, we show the percentage of the total need supplied through the three sourcing types (local, regional, and international). On
the x-axis, we show the different price scenarios (from very favorable at −50% to very unfavorable at +100%). The red line is plotted against the
right axis and shows how the optimal monthly costs differ for these scenarios (US$7M currently and US$8.7M if local/regional prices increase by
100%). The dotted line shows the costs under these scenarios if we implement the optimized plan as per the current price estimates (i.e., the 0% plan).

household surveys. This projection of the effectiveness requires little data and can provide some initial insights
into the potential cost-effectiveness of CBT assistance.
In practice, however, the expected purchasing behavior can be very different from the actual purchasing
behavior. In an ex post analysis, we analyze the effectiveness of CBT assistance based on actual expenditure
data that is increasingly being collected by several branches within WFP (e.g., through point-of-sale–level
stock-keeping unit (SKU) data). We can link each purchase to a nutritional contribution and cost, and this
allows us to define the actual expenditure pattern of beneficiaries and the effectiveness of WFP assistance for
any location and period in which actual expenditure data are available (again assuming that what beneficiaries
buy is also what they consume). This is a very data-heavy approach (SKUs have to be converted to weights
and commodities, and all entries have to be translated to English) and can only be applied after CBT assistance
has been implemented.

Figure 9. Optimized Sourcing Strategies for White Maize Under Different Price Scenarios

Notes. The table shows how the 73,000 metric tons of maize are purchased predominantly locally/regionally for scenarios in which prices are
stable or become cheaper. Then, as the local/regional prices increase, we can observe a steady shift toward an international source: Argentina.
The large differences between the scenarios mean that the sourcing strategy for white maize is not robust against price uncertainty.
 Peters et al.: The Nutritious Supply Chain
INFORMS Journal on Optimization, 2021, vol. 3, no. 2, pp. 200–226, © 2021 INFORMS 217

Outputs from these ex ante and ex post analyses are generally compared against the cost-effectiveness of in-
kind food assistance to decide on the most appropriate transfer modality. In practice, the transfer modality
decision is made for each FDP (i.e., either the FDP receives an in-kind basket or CBT assistance), and ideally,
decision makers would like to switch flexibly between the modalities during the planning horizon (e.g., CBT
during harvest season, in-kind during the lean season). Note that commodity vouchers are sometimes
considered as a form of in-kind assistance because beneficiaries do not have the freedom to choose what to
spend their vouchers on. To illustrate the flexibility of the underlying model, we show how to incorporate the
transfer modality selection more explicitly in the model, seeing as the initial model covered only the com-
Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

modity voucher assistance.

6.2. Location Differentiation Approach

The idea behind this extension (location differentiation) is to decide for each FDP individually whether it
should use an in-kind food basket (may include commodity vouchers) or a CBT contribution (value voucher
or cash); the in-kind basket is optimized as before, and the CBT contribution would result in performance
(cost, effectiveness) as per our best guess (i.e., data input). The CBT basket (which may differ by location and
by month) would ideally be based on an ex post evaluation of prior CBT distributions so that we can compare
the actual performance of CBT in each location against the actual performance of an optimized in-kind basket.
The mathematical formulation for this extension can be found in Appendix B.
The main outputs of the location differentiation approach are an optimized food basket for the selected
activity and an indication for each FDP for each month whether it should opt for CBT assistance (for which the
effectiveness may differ by location and by month) or the in-kind assistance (for which the effectiveness may
differ by month but not by location). The model chooses CBT in those FDPs in which it makes the most sense
(e.g., well-established markets) and/or during the periods that make the most sense (e.g., postharvest season).
This allows a very flexible approach and a very elegant way to trade off between CBT and in-kind while
utilizing all of the functionality of the core model (e.g., specifying programmatic goals, funding restric-
tions, etc.).
We implemented the location differentiation approach in our web app and tested it using data from WFP’s
Mali operation. We found that the optimal cost was US$20.22 per beneficiary per month if we allow for in-kind
only. When we allowed for unrestricted CBT (location differentiation approach) as per an ex post evaluation,
we found that the cost decreased to US$17.65 (13% reduction). In the optimized scenario, about 53% of the
beneficiaries received unrestricted CBT with an average NVS of 73% for FDPs that receive CBT and 100% for
those that receive in-kind. We found that the optimal transfer modality is heavily dependent on the location as
prices (and expenditure patterns) may vary significantly in different settlements.
Note that this extension is very focused on evaluating the nutritional cost-effectiveness of different mo-
dalities, but recently, we see that many humanitarian agencies are shifting away from using such analyses as
drivers for the transfer modality selection. Cash-based transfers bring unique benefits to beneficiaries—from
financial inclusion to increased dignity and autonomy over the assistance they receive—often difficult to
capture in a quantitative framework. As the humanitarian community continues exploring and researching the
benefits that each transfer modality brings to beneficiaries, it is important to integrate the latest findings and
approaches into mathematical models such as this one to support decision makers in identifying the best
modality choices.

6.3. Other Possible Applications

WFP supplies life-saving assistance to some 100 million beneficiaries every year, but it is not the only actor
when it comes to providing food assistance. For example, national governments often have their own school
feeding programs in place, supplying children in primary and secondary education with nutritious meals.
Sometimes they also provide takeaway rations to further incentivize parents to continue sending their children
to school. Consider that India alone is already providing some 120 million children with school lunches
through its mid-day meal scheme. Applying the methods developed in this paper to such school meal
programs has a huge potential, and we have been engaging with several countries to help them improve the
cost-effectiveness of their school feeding initiatives (e.g., Benin, Colombia, Egypt, India, etc.).
Another humanitarian area in which the methodology developed has merit is in the supply chain for
nonfoods (e.g., UN Humanitarian Response Depot; water, sanitation, and hygiene; health kits; etc.). Currently,
beneficiary needs are defined in terms of nutrients, but needs for nonfood items, such as mosquito nets and
vaccines, could be modeled in very similar ways. This becomes particularly interesting when some nonfood
items supply multiple needs; for example, you could provide a kit instead of individual items if it makes sense
 Peters et al.: The Nutritious Supply Chain
218 INFORMS Journal on Optimization, 2021, vol. 3, no. 2, pp. 200–226, © 2021 INFORMS

from a cost perspective. Additionally, it would allow for coordination across multiple humanitarian actors and
avoids duplication of efforts (e.g., there is need for one mosquito net regardless of how many agencies aim to
provide one).

7. Conclusion
In this paper, we explore the possibilities of integrating key supply chain decisions, such as the food basket
design, the transfer modality selection, and the sourcing and delivery plan, into a single model to support
Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

decision making at WFP. We define a MILP model based on a multicommodity, multiperiod network flow that
covers the full spectrum of operational decisions at WFP and is rooted in data that is reliably available.
Applications in some of WFP’s most complex emergencies (Iraq, Yemen, and El Niño) demonstrate the
added value of optimization in managing humanitarian operations. In Yemen, optimization enabled the
supply chain management team to scale up its operation from three to six million beneficiaries through rapid
evaluation of the trichotomy between beneficiary numbers, food basket quality, and available funding. This
insight allowed the supply chain management team to scale up the operation responsibly while providing
donors with evidence-based information about the resources necessary to reach all people in need and the
impact of lack and/or untimeliness of resources on beneficiaries and operational costs. In Iraq, optimization
enabled the supply chain management team to reduce the cost of general food distribution (supplying 800,000
beneficiaries with their daily nutritional needs) by 12% (representing more than US$25M savings over the
course of two years) while only providing three kilocalories fewer than the original food basket. In South
Africa, we used optimization to generate sourcing and delivery plans for different contingency scenarios, such
as limited local availability and significant price spikes because of El Niño.
Through the development of the model presented in this paper, optimization has become an enabler for
cross-functional collaboration at WFP. By integrating key decisions, performance measures, and operational
constraints for all major components of WFP’s supply chain into one model, WFP is now able to rapidly assess
the impact of operational decisions on its overall performance. By connecting the composition of the food
basket to the supply chain that is necessary to deliver it, we are able to increase the efficiency and effectiveness
of WFP operations, ensuring that, with the funds available, WFP can continue to supply life-saving assistance
to as many people in need as possible.

Acknowledgments
The work described in this paper is the result of more than seven years of iterative development in collaboration with uni-
versities, private sector partners, and WFP colleagues around the world. Without their enthusiasm, pragmatism, and con-
structive criticism, this initiative would have fallen flat a long time ago. Instead, the authors are able to see it grow every month
with more and more operations looking to optimization when faced with complex operational challenges, and many staff
having started using the web application autonomously. The authors thank everyone who helped them better understand WFP
operations, those who invited them to support their operations, those who have championed their work, and those who have
contributed funding so they could continue this work.

Appendix A. Full Model Specification

In Section 3, we focus on describing the differences with standard network flow models. For the sake of completeness, here
we provide a full breakdown of parameters and constraints that are used for the applications in Iraq, Yemen, and South
Africa. This means that the formulation is based on commodity vouchers but not yet on value vouchers and cash
transfer modalities.

A.1. Sets
We define multiple subsets for the nodes in particular to ease the definition of constraints and statistics.

N = set of nodes (i, j ∈ N )

N S = set of source nodes
N IS = set of international suppliers
N RS = set of regional suppliers
N LS = set of local suppliers
N LM = set of local markets
N P = set of procurement nodes (N IS ∪ N RS ∪ N LS ∪ N LM )
N DP = set of discharge ports
 Peters et al.: The Nutritious Supply Chain
INFORMS Journal on Optimization, 2021, vol. 3, no. 2, pp. 200–226, © 2021 INFORMS 219

(Continued)

N EDP = set of extended delivery points

N T = set of transshipment nodes (N P ∪ N DP ∪ N EDP )
N FDP = set of final delivery points
K = set of commodities (k ∈ K)
G = set of food groups ( g ∈ G)
L = set of nutrients (l ∈ L)
B = set of beneficiary types (b ∈ B)
Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

T = set of months (t ∈ T )

Note that the suppliers (N IS , N RS , N LS ) provide access to commodities that WFP ships to final delivery points (N FDP )
using its network of discharge ports (N DP ) and extended delivery points (N EDP ). Local markets (N LM ) provide bene-
ficiaries with direct access to commodities when they receive a cash-based transfer. Each procurement node (N P ) may be
linked to multiple source nodes (N S ), which we use to model procurement restrictions. Procurement nodes represent
physical handover points between WFP and a supplier, whereas in our implementation, the source node represents the
country of origin for the commodity (which may differ from the country where WFP takes ownership).

A.2. Parameters
Because of the unique nature of long-term food distribution, we require some nonconventional parameters, such as
nutritional values and feeding days.

α = conversion rate from metric tons (mt) to grams (g) (106 )

ben = the beneficiary type to be optimized (ben ∈ B)
capHit = handling capacity (in mt) of node i ∈ N EDP ∪ N DP in month t
capTijt = transportation capacity (in mt) from node i to node j in month t
capPikt = procurement capacity (in mt) of commodity k from source i ∈ N S in month t
costijkt = cost (in $/mt) of moving commodity k from node i to node j in month t
daysb = current number of feeding days per month for beneficiaries of type b
dembit = number of beneficiaries of type b at node i ∈ N FDP in month t
durij = duration (in days) of moving from node i to node j
groupk = the food group to which commodity k belongs (groupk ∈ G)
hci = handling costs (in $/mt) at node i ∈ N DP ∪ N EDP ∪ N FDP
invikt = incoming arrivals (in mt) of commodity k at node i ∈ N EDP ∪ N DP in month t
ltpij = maximum duration (in days) to supply all FDPs if we procure from source i ∈ N S at node j ∈ N P
nutreqbl = nutritional requirement of beneficiaries of type b for nutrient l
nutvalkl = nutritional value per gram of commodity k for nutrient l
odocCV = other direct operational costs rate (in %) for C&V transfers
odocF = other direct operational costs rate (in $/mt) for food transfers
ratbk = current ration size (gram/person/day) of commodity k for beneficiaries of type b
sci = storage costs (in $/mt/month) at node i ∈ N DP ∪ N EDP

Handling/storage capacities (capH it ) are only tracked for DPs and EDPs although we can easily extend it to other nodes.
Basically, the suppliers take care of any handling/storage taking place at procurement nodes, and WFP’s cooperating
partners do the same at the FDP level. Costs (costijkt ) are captured for all movements through the supply chain network;
movements between source and procurement nodes incur procurement costs, and other movements are transportation
costs. Costs may change over time because of seasonality and stock market movements for procurement and because of
rainy seasons for transportation. We split demand (dembit ) per location and beneficiary type, and allow it to change over
time so we can phase out or scale up operations. It is important to distinguish between beneficiary types because their
nutritional requirements are vastly different. The needs of a child are very different from those of a newly arrived refugee
or a nursing woman for example. The food baskets currently used (ratbk ) and their (monthly) feeding days (daysb ) are
predefined for each beneficiary type. The model optimizes one of these baskets at a time (ben) and supplies the other
beneficiary types with their predefined basket. Optimizing all food baskets simultaneously is mathematically possible (and
tractable) but rarely done in practice because prioritization across beneficiary types is difficult to quantify. We use invikt to
model initial inventories and incoming shipments, allowing us to easily integrate the model’s decisions with the current
status of an operation. Other direct operational costs (ODOC) (odocF and odocCV ) are surcharges per transfer modality that
take care of additional costs that are incurred beyond the procurement, transport, storage, and handling of the com-
modities (such as milling, packaging, monitoring, reporting, etc.), and odocF is usually defined in dollars per metric ton and
odocCV as a percentage surcharge for every dollar transferred to beneficiaries.
 Peters et al.: The Nutritious Supply Chain
220 INFORMS Journal on Optimization, 2021, vol. 3, no. 2, pp. 200–226, © 2021 INFORMS

A.3. Variables
The two major variables are the ones governing commodity flows (Fijkt ) and ration sizes (Rkt ); the others are mainly used to
calculate statistics.

Fijkt = flow of commodity k from node i to node j in month t (metric ton)

Rkt = ration of commodity k in the food basket of month t (gram/person/day)
Slt = shortfall (slack variable) of nutrient l in month t (fraction)
Olt = overshoot (slack variable) of nutrient l in month t (fraction)
Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

SFIlt = binary indicator: one if there is a shortfall for nutrient l in month t

Pijt = binary indicator: one if food is procured from source i at procurement node j in month t
LTt = maximum lead time of commodities purchased in month t (days)

A.4. Objectives
In order to adequately track and constrain the model’s outputs, we define dozens of statistics (most of them linear
combinations of the variables). Here, we focus on the two most important statistics for each of the four goal classes
(efficiency, effectiveness, development, and agility).

TOC = total operation cost (efficiency 1)

FCR = full cost recovery rate (efficiency 2)
CAL = kilocalories (effectiveness 1)
NVS = nutritional value score (effectiveness 2)
LOC = local procurement (development 1)
CBT = cash-based transfers (development 2)
ALT = average lead time (agility 1)
MLT = maximum lead time (agility 2)

A multiobjective approach that allowed users to optimize any of the statistics was included in earlier versions of the
model, but we discarded it for the sake of solvability. We found that the solution times are significantly higher (3–20×)
when incorporating performance measures other than costs into the objective function. For cost objectives, little branch
and bound is necessary to find the optimal solution; this is not the case for other objectives. Because of this big difference in
solution speed, it is more practical to use the other statistics as constraints (akin to a goal programming approach).
We define the eight main measures as follows:
∑∑
TOC  costijkt × Fijkt , (A.1)
i∈N S j,k,t
∑ ∑∑
+ costijkt × Fijkt , (A.2)
∈N S ji k,t
i/
∑ ∑
+ sci × Fiikt , (A.3)
i∈N DP ∪N EDP k,t
∑ ∑ ∑ ∑
+ hci × Fijkt + hci × Fijkt , (A.4)
i∈N DP ∪N EDP ji,k,t j∈N FDP i,k,t
∑ ∑ ∑
+ odocF × Fijkt , (A.5)
∈N LM k,t
i∈N S j/
∑ ∑
+ costijkt × odocCV × Fijkt , (A.6)
j∈N LM i,k,t
∑
costijkt × Fijkt
i,j,k,t
FCR  ∑ , (A.7)
i,j,k,t Fijkt
∑
k nutvalk,energy × Rkt
CAL  , (A.8)
|T |
∑
l,t (1 − Slt )
NVS  × 100%, (A.9)
|L||T |
∑ ∑
j∈N LS i,k,t Fijkt
LOC  ∑ ∑ × 100%, (A.10)
j∈N P i,k,t Fijkt
∑ ∑
j∈N LM i,k,t Fijkt
CBT  ∑ ∑ × 100%, (A.11)
j∈N P i,k,t Fijkt
∑
i,j,k,t durijk × Fijkt
ALT  ∑ , (A.12)
i,j,k,t Fijkt
MLT  max LTt . (A.13)
t
 Peters et al.: The Nutritious Supply Chain
INFORMS Journal on Optimization, 2021, vol. 3, no. 2, pp. 200–226, © 2021 INFORMS 221

The usual optimization objective, the total operation cost, consists of the following components: the procurement costs
(A.1), the transportation costs (A.2), the storage costs (A.3), the handling costs (A.4), the ODOC costs for food (A.5), and
the ODOC costs for C&V (A.6). For WFP, this efficiency measure makes the most sense because a lower cost per
beneficiary means that it is able to supply more beneficiaries from the available funding. The other measures usually are
used as constraints (goals) in this paper to ensure that the resulting solutions are in line with the objectives of the operation
that is being optimized.
A second efficiency measure is the full cost recovery (FCR) rate, the average dollars per metric ton that WFP spends on
food. The FCR rate varies wildly between countries and is particularly high when countries are landlocked (e.g., more
Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

transport necessary to get the commodities to the beneficiaries) or when WFP is forced to adopt suboptimal strategies (e.g.,
sourcing, transfer modality) by donors or the local government. The global average FCR is about a dollar per kilogram
of food.
We measure effectiveness using kilocalories and the nutritional value score (as developed by Ryckembusch et al. 2013).
The number of kilocalories (A.8) that a food basket provides is very easy to capture and is still one of the most important
and easy-to-understand performance measures when it comes to measuring effectiveness. For a more holistic view of all
macronutrients and micronutrients we use the NVS (A.9). We know the nutritional profile of the targeted beneficia-
ries and the nutritional contents of our food basket; the NVS measures how close the current food basket is to the
required nutrients. It is defined as the sum of the delivered percentages for each nutrient l, and we truncate the delivered
nutrients at 100% (i.e., the score does not improve if you supply more than necessary). In practice, this means that we can
just sum the (reversed) shortfalls. The maximum value for this statistic is then the amount of nutrients that are being
tracked (11 for the applications in Iraq and Yemen), which is not a very intuitive measure. We adjust the measure slightly
by dividing the original NVS by the amount of nutrients. To measure the NVS performance adequately across the time
horizon, we also average it over the months. The NVS defined here, therefore, measures the average requirements supplied
across all nutrients with a maximum value of 100% when we supply at least 100% of the requirements for each of the
tracked nutrients in each of the months in the time horizon. Note that the NVS does not favor some nutrients over others,
and the NVS score may hide the fact that one nutrient is not supplied at all. Because of these limitations, we are working
with WFP’s nutrition department to come up with better measures.
Humanitarian organizations prefer using local businesses to source and supply food as this contributes to the de-
velopment of the country. As a performance measure, we consider the percentage of commodities purchased locally, either
through wholesale procurement (A.10) or commodity vouchers (A.11). Note that even though the “percentage purchased”
measure is nonlinear, upper and lower bounds (goals) for this measure can be included as linear constraints.
For the lead times (which are notoriously hard to capture in linear programming models) we use two proxy measures.
One is a maximum lead-time proxy variable (A.13), which is used when WFP needs to respond to a disaster quickly
(i.e., they need to get a food basket inside the country within x days). In practice, however, we generally apply this model
to long-term recovery operations. This means that there is always food on its way to the recipient country, there are
inventories in the warehouses, commodities with a long lead time are being prepositioned, etc. Maximum lead times are of
less concern in such a scenario. Additionally, if one commodity has a long lead time, but the others can be supplied
quickly, the maximum lead time does not reflect the agility of the supply chain well. The second proxy measure is,
therefore, the average number of days that it takes for a metric ton of food to arrive at its destination after being ordered
(A.12). Note that this is again a nonlinear measure that can be modeled as a linear constraint.

A.5. Constraints
In the implementation of the optimization model, we make a distinction between constraints that always hold and optional
constraints. For the minimum working example, we focus on the former. Through a graphical user interface, WFP officers
may impose additional context-specific constraints, such as beneficiary preferences, long-term supplier agreements,
funding restrictions, etc. Most of those (optional) constraints are mathematically trivial though.
∑ ∑
Fijkt  invikt + Fjikt∗ ∀i ∈ N T ∀k, t. (A.14)
j ( j,i,k,t∗ )∈A(i,k,t)

Constraint (A.14) is the traditional “flow in = flow out” constraint for the transshipment nodes. The parameter invikt
contains initial inventories and incoming shipments that are already underway when running the tool, and A(i, k, t) are
pregenerated sets that contain the indices of all flow variables Fijkt that arrive in transshipment node i with commodity k at
time t, generated based on their lead times (durij ). They are currently defined as follows:
{( ) ( ) }
A(i, k, t)  j, i, k, t∗ : t∗ + f durj,i  t ∀i, k, t
⌊ ⌋
t
f ( t)  + 1{t mod 30>20} ,
30

so we look at the duration of each arc and transform the lead time in days to a lead time in months, using a cutoff point of
20 days. This cutoff point is very context-specific and may need to be revised when applying this algorithm to non-WFP
 Peters et al.: The Nutritious Supply Chain
222 INFORMS Journal on Optimization, 2021, vol. 3, no. 2, pp. 200–226, © 2021 INFORMS

supply chains. In practice, we use this formula to generate the parameter initially, but users can override the “rounded”
lead time for any arc themselves if the default cutoff results in poor behavior.
∑ ∑
Fjikt × α  demben, it × daysben × Rkt + dembit × daysb × ratbk ∀i ∈ N FDP ∀k, t. (A.15)
j bben

Constraint (A.15) states that the flow into an FDP must equal its demand (number of beneficiaries times feeding days times
daily ration for commodity k plus any additional requirements for that commodity from other beneficiary types). Note that
we multiply incoming flows by α  106 , converting the metric ton values to grams (which is how food baskets
are specified).
Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

∑
Fijkt ≤ capPikt ∀i ∈ N S ∀k, t,
j

Fijkt ≤ capPikt × Pijt ∀i ∈ N S ∀j, k, t. (A.16)

Constraint (A.16) specifies that the procurement flow must remain lower than its capacity. Multiplying the capacity by the
binary variable Pijt allows us to track maximum lead times as follows:
LTt ≥ Pijt × ltpij ∀i ∈ N S ∀j, t. (A.17)
Recall that ltpij is a proxy for the lead time when purchasing commodities in node j from source i. We currently generate
this value by applying Dijkstra’s shortest path algorithm to get the fastest route to each of the FDPs and taking the
maximum of those values. The lead-time proxy then reflects the shortest time it takes to reach all FDPs when a commodity is
procured this way. If this is the only source for this commodity, the proxy value is accurate. If multiple sources are used to
supply the same commodity in some month, this proxy is an upper bound. In practice, however, it is uncommon to have
multiple sources for the same commodity in the same month (the sources do change over time to benefit from price seasonality).
∑
Fijkt ≤ capTijt ∀i, j, t. (A.18)
k

Constraint (A.18) bounds the flow between nodes so that the flow cannot exceed the (transportation) capacity of an arc.
Note that arc capacities may change over time because of hazardous conditions (security, rainy seasons, etc.).
∑
Fjikt ≤ capH
it ∀i ∈ N DP ∪ N EDP ∀t. (A.19)
jk

Constraint (A.19) bounds the flow through a node (based on handling capacity rather than transportation capacity).
∑
nutvalkl × Rkt  nutreqben, l × (1 − Slt + Olt ) ∀l, t. (A.20)
k

Constraint (A.20) states that the supplied nutrients must cover the requirements. To ensure that the slack variables for
shortfalls (Slt ) and overshoots (Olt ) are working as intended, we need to make sure that they are never positive at the same
time. For this, we use the shortfall indicator:
Slt ≤ SFIlt ∀l, t,
Olt ≤ (1 − SFIlt ) × 10 ∀l, t. (A.21)
Note that Slt is a continuous variable between zero and one, whereas the overshoot is only bounded from below
(maximum nutrient intakes are not as well defined as minimum nutrient intakes).
When leaving all food basket choices up to the model, it chooses the most cost-effective way to supply all nutrients
regardless of the palatability of the resulting food basket. This may mean that a daily ration consists of half a kilogram of
peas, that it includes large amounts of fortified food (not very palatable and with limited supply), or even 200 milliliters of
oil. All these instances and more were found during test scenarios. Intuitively, these solutions make no sense and are,
therefore, not credible (despite being the cost-optimal way of delivering the nutritional requirements). Through interviews
with nutrition experts and food basket designers, the “unwritten rules” for food baskets were brought to light and
quantified. The following constraints result in sensible food basket outputs:
∑
Rkt ≥ minratg ∀g, t
k:groupk g
∑
Rkt ≤ maxratg ∀g, t
k:groupk g

RIodised Salt,t ≥ 5 ∀t (A.22)

SEnergy,t ≤ 0.1 ∀t
SProtein,t ≤ 0.1 ∀t
SFat,t ≤ 0.1 ∀t,
 Peters et al.: The Nutritious Supply Chain
INFORMS Journal on Optimization, 2021, vol. 3, no. 2, pp. 200–226, © 2021 INFORMS 223

Table A.1. The Sensible Food Basket Constraints (in Prams per Person per Day) for Each Food Group

Food group minratg maxratg

Cereals and grains 250 500

Pulses and vegetables 30 130
Oils and fats 15 40
Mixed and blended foods 0 60
Meat and fish and dairy 0 40
Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

Note. These ration sizes are tailored for GFD, which is the bulk of WFP distribution.

with minratg and maxratg values as given in Table A.1. Following these constraints, the rations consist of plenty of staple foods
(grains and pulses) to make two or three dishes per day, oil to satisfy fat requirements and for cooking, some fortified foods to
prevent malnutrition, and iodized salt to satisfy iodine and sodium requirements. Additionally, we make sure that the three
most important nutrients have at most a very small shortfall. Naturally, extra constraints should be added to capture the
preferences of the targeted beneficiaries (that may differ a lot between different tribes or countries of origin). Note that these
constraints are geared toward so-called GFD, which is the ration that we supply to most (if not all) targeted beneficiaries in a
country. Beneficiaries with additional needs (children, pregnant and nursing women, newly arrived refugees, etc.) receive the
required supplements on top of this GFD basket. These constraints are disabled when we optimize non-GFD baskets.
Finally, we add some nonnegativity constraints:

Fijkt ≥ 0 ∀i, j, k, t
Rkt ≥ 0 ∀k, t
Slt ≥ 0 ∀l, t (23)
Olt ≥ 0 ∀l, t
LTt ≥ 0 ∀t.

Appendix B. Extension: Location Differentiation Approach

To track the choice between CBT and in-kind for each FDP, we create new binary variables CBTit , which are one if FDP i
receives CBT assistance in month t and zero if it receives in-kind assistance. We need to make a small adjustment to the
demand constraint in the network flow model to account for this choice. For this, we need an auxiliary variable (con-
tinuous) that keeps track of the in-kind demand for the activity to be optimized:

SADikt ≥ demit × Rkt − M × CBTit ∀i, k, t, (B.1)

∑
OADikt  dembit × ratbk ∀i, k, t, (B.2)
b

where M is a sufficiently large number. The selected activity demand (SADikt ) is equal to the in-kind demand as per the
optimized food basket (Rkt ) if the FDP chooses in-kind and zero otherwise. The other activity demand (OADikt ) keeps track
of the in-kind demand for the other activities (b) that still need to be supplied. Note that we do not need a variable for this;
the explicit specification is just for notational convenience. Similarly, we are ignoring the conversion factors in these
formulas (metric ton to gram conversion and daily ration to monthly ration conversion).
With these new variables defined, the new demand constraint becomes
∑
Fijkt ≥ SADjkt + OADjkt ∀j, k, t. (B.3)
i

Note that, in practice, we may need to limit the number of switches between IK and CBT assistance (to make the
solution more implementable); for example, if we decide to switch, it should be at least for n months. A potential solution
for this is to track the decision to switch as a(n) (additional) binary variable (by location, by month), for which the original
CBTit variables are now defined implicitly by these switch variables. The approach can be modeled using the following
constraints.
We need new binary variables STit (“switch to”) that are equal to one if location i swaps transfer modality (to CBT or IK)
in month t:

STitCBT , STitIK ∈ {0, 1} ∀i, t. (B.4)

For any location, in any month, we can swap at most once:

STitCBT + STitIK ≤ 1 ∀i, t. (B.5)

 Peters et al.: The Nutritious Supply Chain
224 INFORMS Journal on Optimization, 2021, vol. 3, no. 2, pp. 200–226, © 2021 INFORMS

We have to start with one of the two modalities:

STitCBT
0
+ STitIK0  1 ∀i. (B.6)

The original CBT variables are now determined implicitly by the new swap variables:

CBTit  CBTi,t−1 + STitCBT − STitIK ∀i, t, (B.7)

CBTit0  STitCBT
0
∀i. (B.8)
Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

Ensure that swaps hold for at least n months (n would probably be something like four to six months):

∑
t
STitIK ≤ 1 − STitCBT
∗
∀i, t, (B.9)
t∗ t−n
∑
t
STitCBT ≤ 1 − STitIK∗ ∀i, t. (B.10)
t∗ t−n

During implementation of the location differentiation approach, we found that the solution speed increased drastically.
The core model requires little branch and bound with the binary variables mostly impacting capacities but not the costs
(e.g., if during branch and bound a binary variable is 0.7, increasing it to one would not affect the cost of the node). With
the new binary variables governing the CBT decisions, the problem becomes significantly more difficult to solve. With our
default solver, we were unable to find solutions for more than three months of demand. We were able to define a custom
solver strategy using Gurobi that makes the problem tractable again even for large planning horizons, but there is still
some research and experimentation to be done to reduce the optimality gap to reasonable levels. We will be refining our
approach moving forward.

Appendix C. Author Biographies and Project History

Historically, WFP has had a siloed approach to managing and optimizing its supply chains. With humanitarian operations
becoming increasingly complex, the need arose for cross-functional innovations that consider WFP’s operations more
holistically. Initial ad hoc analyses by WFP’s logistics development unit (now called the supply chain planning & op-
timization unit) showed that there was a huge potential for savings in Syria by integrating the decisions of the pro-
curement and logistics teams; a thorough cross-functional analysis of the sourcing strategy, corridor allocation, and supply
chain network design led to savings of two million U.S. dollars. The analysis was performed in 2013 by modeling Syria’s
operation in Excel and using the built-in Excel Solver to find improvements.
From this initial analysis, two things became clear. First, although it was possible to optimize sourcing and delivery
for a WFP operation using the Excel solver, it was a very labor-intensive and cumbersome process. Second, it was noted
that the sourcing and delivery solution was heavily dependent on the project design, that is, the food basket that
beneficiaries were supposed to receive and the way in which they were supposed to receive it (food, cash, or voucher).
Changing even one commodity would have effects that rippled throughout the entire sourcing and delivery strategy.
Integrating project design decisions into the Excel solver proved too difficult, however, so WFP reached out to its partner
universities to develop a model and tool that could cope with the complexity of connecting project design decisions to
traditional sourcing and delivery decisions.
In 2014, students from Georgia Tech (United States) and Tilburg University (Netherlands) developed a tractable
prototype that connected the major decisions, and this was then piloted in Syria during a redesign of WFP’s operation
there. Through optimization, we were able to identify concrete savings opportunities, and we recommended a range of
food basket adjustments that led to significant savings throughout 2015 (US$23 million). Our collaborative approach to
optimization and our ability to rapidly analyze, quantify, and visualize key decisions resulted in traction for optimization
initiatives. Since then, we have been focusing on improving the accuracy, scope, and user-friendliness of the prototype
while institutionalizing optimization as a management tool and supporting WFP’s biggest and most complex operations
with ad hoc analyses using the model.
The project was initiated in 2013 by WFP by Rui Gonçalves and Sérgio Silva under the guidance of Mirjana Kavelj. Rui
Gonçcalves is a supply chain officer working in the humanitarian sector for more than 10 years. He is an engineer
interested in lean principles and in the optimization of resources that benefit the most vulnerable and food-insecure
populations in complex humanitarian emergencies. Sérgio is a graduate of mechanical engineering and management and
has been working for the World Food Programme for more than ten years. He is currently the chief of WFP’s supply chain
planning and optimization unit and is the founder and current council president of G.A.S. Porto, an NGO with around 400
volunteers. Mirjana was the chief of WFP’s logistics development unit and has been working for the World Food
Programme for more than 20 years. She is currently the chief of the HR technology & analytics unit.
In 2014, WFP requested support from two universities, first Georgia Tech and then Tilburg University. A team of
undergraduate students in the Stewart School of Industrial & Systems Engineering (Maria Ayers, Lakshmi Sangeeta
Gadepalli, Ashfaque Kachwala, Tahsin Munir, Cane Punma, Gabriel Rodriguez, and Yuvraj Singh) from Georgia Tech
worked on the first prototype, supported by Ozlem Ergun and Mallory Soldner, with the aim of integrating the various
 Peters et al.: The Nutritious Supply Chain
INFORMS Journal on Optimization, 2021, vol. 3, no. 2, pp. 200–226, © 2021 INFORMS 225

decision parameters into a single model. Ozlem Ergun is a professor in mechanical and industrial engineering at
Northeastern University. Her research focuses on design and management of large-scale and decentralized networks,
focusing on organizations that respond to emergencies and humanitarian crises around the world including the U.S.
Agency for International Development, WFP, UN High Commissioner for Refugees, Centers for Disease Control, and
many more. She currently serves as the area editor at the Operations Research journal for policy modeling and the public
sector area. Mallory Freeman completed her graduate work with the Center for Health & Humanitarian Systems. She is
now lead data scientist at the United Parcel Service (UPS) advanced technology group, where she works on research and
development projects and consults within the company.
Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

The next prototype iteration (also in 2014) was with Tilburg University with the aim of operationalizing the model by
integrating it better with operational data and fine-tuning its decision granularity. This activity was conducted by masters
student Koen Peters, supported by Hein Fleuren and Dick den Hertog. Koen Peters is a graduate of operations research &
management science, pursuing a Ph.D. at Tilburg University’s Zero Hunger Laboratory. He is currently the head of
development for WFP’s supply chain planning and optimization unit, where he has been working for more than five years.
Hein Fleuren is a professor of application of operations research at Tilburg University. Over the past 35 years, he has been
involved in many practical applications in production and transportation and, in the last years, also the humanitarian
world. Hein shares the passion of applying data science and optimization to obtain a better world with Dick den Hertog
and Koen Peters. Together with Perry Heijne, he is the founder of the Zero Hunger Lab. Dick den Hertog is a professor of
operations research at University of Amsterdam. His research interests cover various fields in prescriptive analytics, in
particular, linear and nonlinear optimization. He is specifically interested in applications that contribute to a better society.
Since 2015, WFP has taken ownership of the maintenance and development of the model, working together with
internal experts and private sector partners to support its implementation. The main contributors (in terms of expertise and
funding) from the private sector are Palantir, who supported us in the automation of our data feed through their Foundry
platform, and UPS, who supported us with change management expertise related to the rollout of optimization tools.

References
Altay N, Green W (2006) OR/MS research in disaster operations management. Eur. J. Oper. Res. 175(1):475–493.
Altay N, Labonte M (2014) Challenges in humanitarian information management and exchange: Evidence from Haiti. Disasters 38(S1):S50–S72.
Apte A (2010) Humanitarian Logistics: A New Field of Research and Action (Now Publishers Inc., Monterey, CA).
Balcik B, Beamon B (2008) Facility location in humanitarian relief. Internat. J. Logist. 11(2):101–121.
Balcik B, Beamon B, Smilowitz K (2008) Last mile distribution in humanitarian relief. J. Intelligent Transportation Systems 12(2):51–63.
Ball M, Magnanti T, Monma C, Nemhauser G (1995) Handbooks in Operations Research and Management Science, vol. 7, Network Models
(Elsevier Science, Amsterdam).
Beamon B, Kotleba S (2006) Inventory modelling for complex emergencies in humanitarian relief operations. Internat. J. Logist.: Res. Appl.
9(1):1–18.
Ben-Tal A, Do Chung B, Mandala S, Yao T (2011) Robust optimization for emergency logistics planning: Risk mitigation in humanitarian relief
supply chains. Transportation Res. Part B: Methodological 45(8):1177–1189.
Bose I, Baldi G, Kiess L, de Pee S (2019) The “Fill the Nutrient Gap” analysis: An approach to strengthen nutrition situation analysis and decision
making towards multisectoral policies and systems change. Maternal Child Nutrition 15(3):e12793.
Bozorgi-Amiri A, Jabalameli M, Al-e Hashem S (2013) A multi-objective robust stochastic programming model for disaster relief logistics under
uncertainty. OR Spectrum 35(4):905–933.
Briend A, Darmon N, Ferguson E, Erhardt J (2003) Linear programming: A mathematical tool for analyzing and optimizing children’s diets
during the complementary feeding period. J. Pediatric Gastroenterology Nutrition 36(1):12–22.
Carlson A, Lino M, Gerrior S, Basiotis P (2003) Revision of USDA’s low-cost, moderate-cost, and liberal food plans. Family Econom. Nutrition Rev.
15(2):43–51.
Çelik M, Ergun Ö, Johnson B, Keskinocak P, Lorca Á, Pekgün P, Swann J (2012) Humanitarian logistics. New Directions in Informatics, Opti-
mization, Logistics, and Production (INFORMS, Catonsville, MD), 18–49.
Chastre C, Duffield A, Kindness H, LeJeune S, Taylor A (2007) The minimum cost of a healthy diet: Findings from piloting a new methodology in
four study locations. Technical report, Save the Children UK, London.
Dantzig GB (1990) The diet problem. Interfaces 20(4):43–47.
Deptford A, Allieri T, Childs R, Damu C, Ferguson E, Hilton J, Parham P, et al. (2017) Cost of the diet: A method and software to calculate the
lowest cost of meeting recommended intakes of energy and nutrients from local foods. BMC Nutrition 3(1):26.
Ergun Ö, Keskinocak P, Swann J (2011) Introduction to the special issue on humanitarian applications: Doing good with good OR. Interfaces
41(3):215–222.
Fleige L, Moore W, Garlick P, Murphy S, Turner E, Dunn M, Van Lengerich B, Orthoefer F, Schaefer S (2010) Recommendations for optimization
of fortified and blended food aid products from the United States. Nutrition Rev. 68(5):290–315.
Food and Agriculture Organization of the United Nations, International Fund for Agricultural Development, World Food Programme (2015)
Achieving zero hunger: The critical role of investments in social protection and agriculture. Accessed January 21, 2021, http://
www.fao.org/policy-support/tools-and-publications/resources-details/en/c/422255/.
Food and Agriculture Organization of the United Nations, International Fund for Agricultural Development, UNICEF, World Food Programme,
World Health Organization (2020) The state of food security and nutrition in the world 2020. Transforming food systems for affordable
healthy diets. Accessed January 20, 2021, http://www.fao.org/documents/card/en/c/ca9692en/.
Gralla E, Goentzel J, Fine C (2014) Assessing trade-offs among multiple objectives for humanitarian aid delivery using expert preferences.
Production Oper. Management 23(6):978–989.
Gutjahr W, Nolz P (2016) Multicriteria optimization in humanitarian aid. Eur. J. Oper. Res. 252(2):351–366.
 Peters et al.: The Nutritious Supply Chain
226 INFORMS Journal on Optimization, 2021, vol. 3, no. 2, pp. 200–226, © 2021 INFORMS

Haghani A, Oh S (1996) Formulation and solution of a multi-commodity, multi-modal network flow model for disaster relief operations.
Transportation Res. Part A Policy Practice 30(3):231–250.
Harvard Humanitarian Initiative (2010) Disaster relief 2.0: The future of information sharing in humanitarian emergencies (Harvard Hu-
manitarian Initiative; United Nations Foundation; Office for the Coordination of Humanitarian Affairs; The Vodafone Foundation). Accessed
January 21, 2021, https://reliefweb.int/report/world/disaster-relief-20-future-information-sharing-humanitarian-emergencies.
Holguı́n-Veras J, Jaller M, Van Wassenhove L, Pérez N, Wachtendorf T (2012) On the unique features of post-disaster humanitarian logistics.
J. Oper. Management 30(7):494–506.
Holguı́n-Veras J, Pérez N, Jaller M, Van Wassenhove L, Aros-Vera F (2013) On the appropriate objective function for post-disaster humanitarian
logistics models. J. Oper. Management 31(5):262–280.
Downloaded from informs.org by [111.68.97.37] on 04 June 2024, at 03:43 . For personal use only, all rights reserved.

Lee H (2004) The triple-a supply chain. Harvard Bus. Rev. 82(10):102–113.
Lentz E, Barrett C, Gómez M, Maxwell D (2013) On the choice and impacts of innovative international food assistance instruments. World
Development 49:1–8.
Mitchell S, O’Sullivan M, Dunning I (2011) PuLP: A linear programming toolkit for Python. Accessed January 20, 2021, http://
www.optimization-online.org/DB_FILE/2011/09/3178.pdf.
Novak S, Eppinger S (2001) Sourcing by design: Product complexity and the supply chain. Management Sci. 47(1):189–204.
Office for the Coordination of Humanitarian Affairs (2002) Symposium on best practices in humanitarian information exchange. Accessed
January 21, 2021, https://reliefweb.int/report/world/symposium-best-practices-humanitarian-information-exchange.
Özdamar L, Demir O (2012) A hierarchical clustering and routing procedure for large scale disaster relief logistics planning. Transportation Res.
Part E Logist. Transportation Rev. 48(3):591–602.
Pérez-Rodrı́guez N, Holguı́n-Veras J (2016) Inventory-allocation distribution models for postdisaster humanitarian logistics with explicit
consideration of deprivation costs. Transportation Sci. 50(4):1261–1285.
Ram B, Naghshineh-Pour M, Yu X (2006) Material requirements planning with flexible bills-of-material. Internat. J. Production Res. 44(2):399–415.
Ramberg J (2011) ICC Guide to Incoterms 2010 (ICC Publishing S.A., Paris).
Rancourt M, Cordeau J, Laporte G, Watkins B (2015) Tactical network planning for food aid distribution in Kenya. Comput. Oper. Res. 56:68–83.
Rawls C, Turnquist M (2012) Pre-positioning and dynamic delivery planning for short-term response following a natural disaster. Socio-Econom.
Planning Sci. 46(1):46–54.
Rottkemper B, Fischer K, Blecken A (2012) A transshipment model for distribution and inventory relocation under uncertainty in humanitarian
operations. Socio-Econom. Planning Sci. 46(1):98–109.
Ryan K, Adams K, Vosti S, Ordiz M, Cimo E, Manary M (2014) A comprehensive linear programming tool to optimize formulations of ready-to-
use therapeutic foods: An application to Ethiopia. Amer. J. Clinical Nutrition 100(6):1551–1558.
Ryckembusch D, Frega R, Silva M, Gentilini U, Sanogo I, Grede N, Brown L (2013) Enhancing nutrition: A new tool for ex-ante comparison of
commodity-based vouchers and food transfers. World Development 49:58–67.
Saltzman M (2002) COIN-OR: An open-source library for optimization. Programming Languages and Systems in Computational Economics and
Finance (Springer, Dordrecht, Netherlands), 3–32.
Van Wassenhove L (2006) Humanitarian aid logistics: Supply chain management in high gear. J. Oper. Res. Soc. 57(5):475–489.
Van Wassenhove L, Pedraza Martinez A (2012) Using OR to adapt supply chain management best practices to humanitarian logistics. Internat.
Trans. Oper. Res. 19(1–2):307–322.
World Food Programme (2015a) The World Food Programme’s operations in Iraq. Accessed January 20, 2021, http://www.wfp.org/
countries/iraq.
World Food Programme (2015b) The World Food Programme’s operations in Yemen. Accessed January 20, 2021, http://www.wfp.org/
countries/yemen.
World Food Programme (2018) WFP year in review 2018. Accessed January 20, 2021, https://publications.wfp.org/2018/en/annual-
report/home.html.
World Food Programme (2020) Annual performance report for 2019. Accessed January 20, 2021, https://www.wfp.org/publications/annual-
performance-report-2019.