Professional Documents
Culture Documents
Interfaces
Publication details, including instructions for authors and subscription information:
http://pubsonline.informs.org
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 permissions@informs.org.
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.
INFORMS is the largest professional society in the world for professionals in the fields of operations research, management
science, and analytics.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
Vol. 45, No. 5, September–October 2015, pp. 444–461
ISSN 0092-2102 (print) ISSN 1526-551X (online) http://dx.doi.org/10.1287/inte.2015.0802
© 2015 INFORMS
Ming Zhao
Interfaces 2015.45:444-461.
Department of Decision and Information Sciences, Bauer College of Business, University of Houston,
Houston, Texas 77204, mzhao@bauer.uh.edu
The Procter & Gamble (P&G) fabric-care business is a multibillion dollar organization that oversees a global
portfolio of products, including household brands such as Tide, Dash, and Gain. Production is impacted by
a steady stream of reformulation modifications, imposed by new-product innovation and constantly changing
material supply conditions. In this paper, we describe the creation and application of a novel analytical frame-
work that has helped P&G determine the ingredient levels and product and process architectures that enable the
company to create some of the world’s best laundry products. Modeling cleaning performance and other key
properties such as density required P&G to develop innovative quantitative techniques based on visual statisti-
cal tools. It used advanced mathematical programming methods to address challenges that the manufacturing
process imposed, product performance requirements, and physical constraints, which collectively result in a
hard mixed-integer nonlinear (nonconvex) optimization problem. We describe how P&G applied our framework
in its North American market to identify a strategy that improves the performance of its laundry products,
provides targeted consumer benefits, and enables cost savings in the order of millions of dollars.
Keywords: pooling; blending; optimization; response surface; design of experiments.
History: This paper was refereed.
and semi-empirical models that predict chemical reac- products as a group of products with a common set
tions during manufacturing, in-use physical prop- of characteristics. In this paper, we define portfolio as
erties of the product, technical performance of the set of formula-unique powder laundry detergents
the product, and even consumer acceptance rates. manufactured in our North American site.
These tools enable researchers to instantly predict a Figure 1 illustrates the mixing architecture and
product’s physical properties and performance, inte- problem structure of the laundry detergent blending
grate models, and balance production trade-offs using process. A large portfolio of products is created from
a variety of predictive and prescriptive capabilities. a relatively small number of intermediate batches (i.e.,
Until recently, the complexity of laundry-formulation 1 to 8 in Figure 1); an intermediate batch, also called
and manufacturing processes limited us to consider an intermediate, is a mixture that is shared by var-
reformulating only a single product at a time; how- ious finished products. Each product is created by
ever, breakthroughs in mathematical optimization blending a portion of its mixture from exactly one
technology have made possible system-wide portfolio intermediate batch with as many finishing additives
reformulation. This is critically important because it as required. Intermediates and finished products are
permits us to model and optimize product differentia- chemical mixtures of one or more ingredients or fin-
tion within a portfolio and consider sharing common ishing additives. Ingredients or finishing additives are
materials within the manufacturing process. In this sourced in the form of chemical mixtures, which we
paper, we present the scope of laundry-portfolio mod- refer to as premixes. The ingredient composition of
Interfaces 2015.45:444-461.
eling and optimization at P&G, the creation of capa- each premix is given, whereas the proportion of pre-
bilities we developed to address this scope, and its mixes to be combined to produce a desirable mixture
application to innovate P&G’s North American pow- of ingredients must be specified (as a decision vari-
der laundry portfolio. able). Costs are specified at the premix level, whereas
product properties are determined by the ingredient
composition.
Problem Definition and Challenges The goal of production is to minimize portfo-
The P&G North American laundry detergent busi- lio annual material spend across a network, which
ness comprises three product forms: powders, liquids, currently includes about 40 products and up to
and pods. Powder detergents, which generate annual 40 ingredients; material costs typically account for
sales of several hundred million U.S. dollars, are a approximately 60 percent of the total cost of produc-
critical part of P&G’s North American business. Even tion. P&G imposes many constraints to ensure that its
as we focus on powders as the primary application, targeted levels of quality and manufacturing feasibil-
the framework for these tools can (and must) be eas- ity are achieved. These include requirements for stain
ily extendable to other forms. Therefore, although we removal and whiteness performance, material bal-
focus on the powder problem in this paper, we note ance and density of intermediate batches and product
that the liquid form is a simplified version of this mixtures, manufacturing site (i.e., plant) throughput,
problem. water content, and raw material usage. Decisions to
Laundry-product formulation can occur in one or be made include: assignments of products to interme-
more manufacturing sites to supply multiple markets. diates, intermediate-proportion contributions to each
Identical product formulations are commonly made product, mixture compositions of intermediates, and
in three or four different sites to fulfill the demand additive proportions in final products. In addition,
of an entire region, such as Western Europe or North for laundry detergent powders, intermediate batches
America. Because each manufacturing site defines its must conform to unique evaporation rules that make
own set of products, the possibility exists that 80 per- the problem more complex. Making the intermediate
cent of the products produced in two different man- batch requires mixing ingredient premixes in a slurry,
ufacturing sites may coincide, whereas the remaining and then evaporating the excess water to form a free-
20 percent are small-volume formulations that only flowing powder, which is mixed with finishing addi-
one site supplies. We typically refer to a portfolio of tives to create the final product.
Esquejo et al.: Laundry Portfolio Optimization at P&G
446 Interfaces 45(5), pp. 444–461, © 2015 INFORMS
Interfaces 2015.45:444-461.
Figure 1: (Color online) The production of laundry detergent mixtures creates a blending process.
To formalize a solution to this complex problem, we polynomial functions that capture the two perfor-
separate the analysis into two categories: predictive mance qualities of a mixture: stain removal and
models and optimization. Predictive models are used whiteness.
to quantify the various relationships within the system, Empirical models for stain removal and whiteness
and optimization incorporates these predictive mod- were created using experimental design procedures,
els into the mathematical formulation to determine the an efficient means of model creation for controlled
ideal values of decision variables. Next, we provide experiments (Box et al. 2005, Kutner et al. 2004), using
details about each component of the problem. JMP software for the design and analysis. Figure 2
shows an example of an experimental design for a
Predictive Models three-variable model for a coffee stain. The table in
Predictive models are either empirical or semi- the figure lists the set of 16 test treatments we ran
empirical in nature. Empirical models are third-order and the associated coffee-stain response. The image to
60 60
7 100 100 0 83.46
8 50 100 50 85.21
40 40
9 100 0 0 83.48
10 100 0 100 91.88 20 20
11 50 50 50 84.00
12 50 50 100 88.99 0 0
100
13 50 0 50 84.58 100 80
80 60
14 0 50 50 84.52 60 40 1
Var 40 ble
15 0 0 100 88.84 iab 20 20 ria
50 50 50
le 2 0 Va
16 81.62 0
Figure 2: (Color online) This experimental design for a stain-removal index (SRI) for a coffee stain is
characterized by SRI coefficients (left) and can be visualized as response-surface models (right).
Esquejo et al.: Laundry Portfolio Optimization at P&G
Interfaces 45(5), pp. 444–461, © 2015 INFORMS 447
the right of the table is a graphical representation of Stain before wash Stain after wash
the treatments, with the shading corresponding to the
value of the stain-removal index (SRI) for coffee. In
this example, higher SRI values (darker shading) are
more desirable.
Our experimental designs were based on i-optimal
criteria; such designs minimize average variance of
prediction over the region of experimentation (Goos
and Jones 2011, Johnson et al. 2011). They also in-
cluded 16 variables with all two-way and selected
three-way interactions, producing third-order designs
Figure 4: (Color online) Stains are scanned before and after a wash exper-
with approximately 300 model terms. These variables iment; in this example, we use a coffee stain.
included all the key cleaning ingredients and wash
conditions of interest (e.g., surfactants and wash tem-
models. We accomplished model selection for each
perature). We used this design procedure for all stain
and whiteness models (approximately 60 responses). response using three stepwise regression techniques:
We generated empirical data for the design by P -value threshold, minimum-corrected Akaike infor-
making laboratory-scale formula prototypes, which mation criterion, and minimum Bayesian information
we physically tested in standardized wash proto- criterion; see Burnham and Anderson (2002, 2004),
Interfaces 2015.45:444-461.
cols. The stain removal and whiteness procedures we and Miller (1990) for a description of using Akaike
used were similar to ASTM method D4265-14 (ASTM and Bayesian information criteria for model selec-
International 2014), which involves creating standard tion. We used multivariate regression to quantify the
stain sets and characterizing their color before and selected models and validation metrics to determine
after wash (E) using image analysis. Figure 3 illus- the best model for each response, and we conducted
trates this procedure for stains for which measures are several levels of validation for each model to char-
assigned to each of several standard technical stains acterize prediction quality. One of the common tech-
that are washed together in a single experiment. Fig- niques involved quantifying standard fitting diagnos-
ure 4 shows an example of a standard coffee stain that tics for the data set used for model creation. These
has been processed with a given product test mixture metrics include R square, R square adjusted, root
(before and after wash). mean square error, lack-of-fit p-values, and other sim-
With the experiments in the design completed, ilar metrics. R square adjusted for the models ranged
our next step was to evaluate and select response from 0.55 to 0.95 with an average of 0.84. We used
models with R square adjusted below 0.70 only if
Before wash color – After wash color
Stain removal = –
Before wash color
our technology experts agreed that trends that the
model displayed were acceptable for business pur-
poses. All models in this design were deemed accept-
able. Figure 5 shows an example of these diagnostics
for coffee SRI.
Before After
wash Semi-empirical models are based, to some extent,
wash
on physical relationships. In this paper, we define
Temperature = t semi-empirical models as functional forms derived
Hardness = h
Soil level = s
from known equations (typically based on physical
laws and theorems), whose coefficients were deter-
Washer
mined by fitting the equation to a set of experimental
data. Finished-product density and intermediate den-
Figure 3: (Color online) A standard wash protocol is used when testing the sity were the primary semi-empirical models used;
stain-removal effectiveness of a mixture. both are nonconvex functions of mixture-ingredient
Esquejo et al.: Laundry Portfolio Optimization at P&G
448 Interfaces 45(5), pp. 444–461, © 2015 INFORMS
Actual by predicted plot is based on the material balance of water in the drying
process.
Optimization
INSTANT ACTUAL
2OOT MEAN SQUARE ERROR
formulation in the appendix.
-EAN OF RESPONSE
Intermediate batches and final products consist of
/BSERVATIONS OR SUM WGTS
mixtures of ingredients. Although ingredient propor-
tions determine the properties of these mixtures, the
Analysis of variance
manufacturing process does not simply blend pure
Sum of ingredients. Rather, premixes (mixtures of a small
Source DF squares Mean square F ratio
subset of ingredients) are sourced and mixed to
-ODEL achieve a final-ingredient mixture. The eligibility of
%RROR Prob > F premixes to be added to an intermediate batch versus
# TOTAL * a final product is determined primarily by the nature
of the premix. Premixes with high water content are
Lack of fit typically added to intermediate batches so that the
Sum of F ratio water can evaporate. Some premixes (e.g., perfumes)
Source DF squares Mean square can be unique to one product or available to all prod-
,ACK OF FIT Prob > F ucts, either in the intermediate batch, postevaporation
0URE ERROR stage, or both.
4OTAL ERROR Max r sq Figure 6 illustrates the analytical representation
of material flowing through various stages of the
manufacturing process for a simple two-intermediate,
Figure 5: (Color online) Fit diagnostics, shown for the SRI response func- three-product example. Individual premixes may be
tion for a coffee stain, show strong predictive model performance and are eligible to go into intermediate mixtures, directly into
typical of stain results.
the final-product mixtures (as additives), or either. For
example, the figure shows that premix Pre1 can be
proportions. The appendix includes detailed func- assigned only to intermediates (B1 and B2), whereas
tional forms of these models. premix Pre3 can be assigned to intermediates and
Finally, we used a first-principles model of evapo- product Prod2. The additive Pre5 can be assigned
rative load for any given formulation to estimate the only to Prod3 directly.
impact on evaporative rate, an important manufactur- Percentages by weight of premixes used in total
ing consideration for detergent powders. This model production of intermediate batches (B1 and B2) and
Esquejo et al.: Laundry Portfolio Optimization at P&G
Interfaces 45(5), pp. 444–461, © 2015 INFORMS 449
the set of products into subsets, such that an inter- implement a new work process. The use of optimiza-
mediate is assigned to each subset of products that tion tools requires a set of skills best defined by an
must satisfy various nonlinear constraints, such as optimization triad (Figure 7) that includes functional,
performance requirements. This interpretation of the data, and optimization experts.
problem is closely related to the pooling problem Functional experts are typically a small number of
(Bodington and Baker 1990), and can be shown to individuals from different R&D functions, including
be an NP-hard mixed-integer nonlinear (nonconvex) consumer, formulation, and process. Data experts are
program. one or two individuals who have access to all the
Previous approaches to solving pooling problems necessary information, such as material pricing and
include relaxation and discretization strategies (Gupte material balances, and are typically skilled in visual
et al. 2013), Benders decomposition (Floudas and analytics tools, such as JMP statistical tools. Optimiza-
Aggarwal 1990), Lagrangian relaxation (Visweswaran tion experts are staff members who have the neces-
and Floudas 1990), branch and cut (Audet et al. 2004), sary programming skills to interact with the optimiza-
and mixed integer linear programming (MILP) (Dey tion models at the SAS code level. Figure 7 defines
and Gupte 2013). These approaches do not directly the responsibilities of expert in each category.
apply to the variant we consider in this paper. Most of We defined a new work process (Figure 8), which
the effort to date has focused on addressing the bilin- has enabled this multifunctional, multiskilled team
ear terms in the problem; however, the evaporation to efficiently use the new capability; we describe the
Interfaces 2015.45:444-461.
7ORK PROCESS
!SSUMPTION
0ROBLEM DEFINITION +NOWLEDGE -ODEL DEVELOPMENT
FUNCTIONAL EXPERT DEVELOPMENT FUNCTIONAL EXPERT
FUNCTIONAL EXPERT
/PTIMIZATION 0ORTFOLIO DATA
7RITE OPTIMIZATION
PROBLEM FORMULATION GATHERING EG VOL
CODE
OPTIMIZATION EXPERT COST FORMULATION
OPTIMIZATION EXPERT
DATA EXPERT
)TERATIVE TESTING /PTIMIZED
-ATHEMATICAL #HURN!NALYSIS
RECOMMENDATION
VALIDATION !LL
FUNCTIONAL EXPERT
OPTIMIZATION EXPERT
Interfaces 2015.45:444-461.
Figure 8: (Color online) The optimization work process at P&G follows a sequence of nine interrelated steps.
density model to use. Additional data or analysis may new research question, constantly verifying assump-
be required if the team cannot come to an agreement tions and validating preliminary results with other
because of missing or conflicting information. experts and stakeholders.
3. Model development: Functional experts develop 7. Iterative testing and mathematical validation:
models for new-product parameters or adapt (e.g., Optimization experts participate with other members
linearize) current models to make them more suitable of the team in a cycle of iterative testing to mathemat-
for optimization. ically validate models, ensuring that constraints are
4. Portfolio data gathering: Data experts collect and correctly interpreted and observed and that optimum
format all the information needed for the framework. solutions are robust. In this step, parameters (e.g.,
This proved to be a challenging step because we the number of multistart points and stopping criteria)
encountered multiple databases with missing rela- are tuned, and model infeasibilities, often caused by
tional keys. The process can be quite manual in some products with too-stringent performance constraints,
cases. are commonly found. At this stage, steps must be
5. Optimization problem formulation: Optimiza- taken to ensure feasibility; revisiting the constraint
tion experts interpret the research question(s) and, bounds is an example.
using the knowledge collected, models, and portfolio 8. Churn and analysis: In this step, the entire team
information as context, create or adapt the mathemat- exercises the optimization engine to optimize multiple
ical framework for optimization by defining or mod- scenarios to answer the research question(s). Scenarios
ifying variables, objective functions, and constraints. can include changes in performance requirements, the
Optimization experts start the modeling work as soon number of intermediates allowed, constraint bounds,
as the research question has been defined, and further materials allowed to be used (and where they can be
refine the optimization models as additional knowl- used), or a combination of these scenarios. For highly
edge is generated and the portfolio data are compiled. complex research questions, P&G implements a churn
6. Write optimization code: Optimization experts event; in such an event, all members of the team are
write or adapt SAS/OR code as needed to address the colocated for two to four days. They focus all their
Esquejo et al.: Laundry Portfolio Optimization at P&G
452 Interfaces 45(5), pp. 444–461, © 2015 INFORMS
time on a common set of problems, and their objec- stopping criterion at three intermediates at an esti-
tive is to produce data to inform decision making. mated gap of three percent for all optimization runs.
During churn, the team occupies a common room Figure 9 shows solution objective values (total cost
equipped with physical and digital visualization tools difference for annual production) of a typical optimi-
that aid the work process. Poster-sized printouts dis- zation, which we represent as the difference between
play a master list of all scenarios, and key parame- our solution output objective and the annual cost of
ters are captured and color coded to differentiate the a historical production run. An obvious feature of the
scenarios. They analyze completed scenarios on a dis- problem we consider here is that as the number of
play that consists of eight 42" high-definition televi- intermediate batches increases, the globally optimal
sion screens configured to behave as a single monitor, objective must not increase, because the model allows
allowing high-resolution visualizations to be spread more flexibility in tuning each product’s assigned
across a large area. The team typically reviews the intermediate. The figure reflects this, and we note that
results using JMP software, which permits interactive the algorithm’s flow was motivated to ensure that this
visualization and analysis. check to verify rationality would never be violated
9. Optimized recommendation: Functional experts despite the presence of nonconvexity.
take results from the analysis and formulate a recom- Reporting such a multiple intermediate solution
mendation. Recommendations can be as simple as a proved valuable to management because an immedi-
new set of formulations to meet a new requirement, ate comparison of neighboring solutions could show
or as complex as a multistage strategy for a portfolio
Interfaces 2015.45:444-461.
in the order of magnitude shown (between $0.5 and Table 2: The table shows results for instance “3bii.” The bolded values
$6 million). represent a best comparison to the benchmark solution, which has 12
Few instances currently exist at P&G because sta- intermediate batches.
tistical and semi-empirical models have only recently
been developed for entire product portfolios. Table 1
Interfaces 2015.45:444-461.
lists two important powder instances we used in bolded row represent a direct comparison with the
this work and also shows a number of ingredi- benchmark intermediate configuration.
ents, premixes, products, response-surface models, The most direct application of our work has been in
and benchmark intermediates that represent typical P&G’s North American dry laundry portfolio, which
problem sizes. consists of more than 20 unique formulations. Our
The optimization tool produces significantly bet- objective was to study different strategies for sim-
ter results than the benchmark for both instances in plification and savings, while delivering the same or
the table; its run times are very reasonable, given the better consumer-relevant cleaning performance. The
needs of the P&G work process. Tables 2 and 3 sum- churn team spent two days running and analyzing
marize the results for these instances, listing most
intermediate configurations from three to the num-
ber of products in the instance. The tables show cost Cost diff. vs. bench
differences from the benchmark (in millions) and run Intermediates ($ in millions) Bounds (%) Run time (sec)
times at which each intermediate configuration satis-
fied its stopping criterion. The longest run time is the 3 −7604 2046 25606
4 −7908 1042 25607
total run time of the process. Note the cost improve-
5 −8109 0079 25608
ments when compared to the benchmark even when 6 −8206 0058 25700
running at the minimum number of intermediate 7 −8300 0043 25702
batches (three) for both instances. The values in the 8 −8305 0028 25703
9 −8307 0021 25704
10 −8309 0015 25705
11 −8400 0012 25707
12 −8401 0009 25708
Instance name Ingredients Premixes Products RS Intermediates
13 −8402 0007 25800
14 −8403 0005 25801
3bii 54 38 21 50 12 15 −8403 0002 25803
Wenlock 42 34 25 38 13 16 −8404 0001 25804
25 −8404 0000 6805
Table 1: This table summarizes the size of two representative instances
in P&G’s portfolio. The term RS refers to the number of response-surface Table 3: In these results for the Wenlock instance, the bolded values rep-
models, and the Intermediates column represents the number of interme- resent a best comparison to the benchmark solution, which has 13 inter-
diate batches in the benchmark solution. mediate batches.
Esquejo et al.: Laundry Portfolio Optimization at P&G
454 Interfaces 45(5), pp. 444–461, © 2015 INFORMS
more than 30 scenarios. Here, we discuss two exam- • Redefining the cleaning vision: Differentiating
ples from the study results. performance between brands and maintaining brand
• Increasing the number of intermediates allowed equity.
from five to 10 would result in additional formula cost • Exploring out-of-the-box concepts: Exploring cur-
savings. Compared to a five-intermediate benchmark, rent production constraints and raw material ingredi-
using the optimization procedure would result in ents, but with the flexibility to integrate deviations
annual cost improvements of $2 million for five that may lead to better results in performance and
intermediates and $5 million for 10 intermediates. process.
Although a consequence of this strategy is to increase • Managing what-if scenarios: Saving time and
complexity at the manufacturing site for handling resources.
a higher number of intermediate batches, this is a • Multifunctional integration: Performing a more
justifiable decision based on this demonstrable cost robust proposition versus isolated optimization ef-
reduction. forts based on function.
• The introduction of a new active ingredient (cur-
rently not in the North American powder formula- Summary of Benefits
tion, but available in empirical and semi-empirical Portfolio optimization is changing the way we do
models) across the whole portfolio would generate product development at P&G. Previously, we were
annual savings in excess of $20 million, while deliv- often limited to one of two strategies; we developed
Interfaces 2015.45:444-461.
ering target performance profile (i.e., cleaning of dif- each product individually, resulting in highly complex
ferent stains) for each product. P&G has incorporated portfolios that required high numbers of intermediate
this knowledge into its short- to mid-term strategies batches, or we imposed simplification strategies that
for this part of our business. resulted in higher formulation costs.
We have estimated that without the portfolio opti- Portfolio optimization allows us to test formulation
mization tool, twice as many staff members would and simplification strategies against the whole port-
have to work for far longer periods of time to run folio, giving us a realistic estimate of the potential
30 scenarios on 20 formulations using the former impact of these strategies. Fast iteration cycles allow
us to evaluate multiple strategies in a short time, dis-
single-product-at-a-time approach and would pro-
carding elements that will bring little value and com-
duce inferior results.
bining elements that provide significant advantages.
User sponsorship and adoption have been instru-
The ability to analyze an entire portfolio simultane-
mental to the success of this project. In an email to
ously is also changing the way product designers
the team, Christian Becerra, P&G senior researcher
think about performance, because we can now more
and lead formulator for the North American powder
easily differentiate performance among the products.
business, provided a set of additional benefits of our
Thus, we can make smarter decisions about formu-
portfolio optimization framework that go beyond the lation and simplification strategies and respond with
savings described earlier (Becerra 2014): agility when needed.
• Smart optimization: Identifying a formula and As lead strategies are identified and the formula-
process strategy that meets our criteria, removing tion for the full portfolio is generated, some phys-
chemistry that will not deliver the desired perfor- ical testing is required to confirm that the required
mance profile to the consumer. This leads to smart performance and physical properties of the products
savings. are indeed met. This is especially true for solutions
• Next level of optimization: Going beyond the tra- that are near the minimum or maximum values of
ditional single-formula to full-portfolio optimization the input ranges, because the confidence intervals of
(i.e., the ability to see the big picture). the predictions are typically at their widest in these
• Flexibility: Integrating technical features and an ranges. Future phases of this project will address this
understanding of consumer needs to make more need, and continuously expand and improve the qual-
robust portfolio propositions. ity of predictive and optimization models.
Esquejo et al.: Laundry Portfolio Optimization at P&G
Interfaces 45(5), pp. 444–461, © 2015 INFORMS 455
Because mixing decisions are made at a premix level, Empirical models impose constraints on product perfor-
variables v, z, and y take a p ∈ P index, thus specifying pre- mance. The term Fk 4k 5 represents a vector for each product
mix composition. The evaporation process is expressed in of all SRI and whiteness functions, each row characterized
terms of ingredient compositions, requiring us to calculate by third-order polynomial expressions of the product’s com-
ingredient mass percentage quantities pre- and postevapo- position ik and parameters temperature, hardness, and soil
ration for each ingredient i in each intermediate b: level. Products created by the optimization must achieve a
X minimum level of performance, which is defined by vector
ib = ip vpb 1 ∀ i ∈ I1 b ∈ B1 (1) fk for each product:
p∈P
0
ib = eb ib 1 ∀ i ∈ I\8w91 b ∈ B1 (2) Fk 4k 5 ≥ fk 1 ∀ k ∈ K0 (10)
0
eb 41 − wb 5 = 1 − wb 1 ∀ b ∈ B1 (3) Semi-empirical models impose constraints on product
density,
where the subscript w is used to denote water. Similarly,
empirical and semi-empirical functions are expressed in F Dk ≤ F Dk 4zk 1 xk 1 ybk 5 ≤ F Dk 1 ∀ k ∈ K1 (11)
terms of ingredient mass percentage at the product level, where we recall that F Dk 4zk 1 xk 1 ybk 5 is a nonlinear function
requiring us to calculate these values for each ingredient i representing the density of product k, which is dependent
in each product k: on ybk through its corresponding intermediate batch den-
X X 0 sity CDb .
ik = ip zpk + xk ybk ib 1 ∀ i ∈ I1 k ∈ K0 (4)
p∈P b∈B
Finally, we recall that a requirement of the process is that
each product is assigned to exactly one intermediate batch:
Constraints. Fundamental physical requirements that
Interfaces 2015.45:444-461.
X
characterize the mixing process are modeled as constraints. ybk = 11 ∀ k ∈ K0 (12)
b∈B
Intermediate and final-product mixtures have mass percent-
age ingredient contributions that must add to 100 percent: Objective. The optimization objective is to minimize the
X X total cost of premixes used in the mixing process, weighted
vpb = 11 ∀ b ∈ B and zpk + xk = 11 ∀ k ∈ K0 (5) by product dosage per stat unit (qks , a unit of demand) and
p∈P p∈P
product production-volume targets (qkv ), using at most n
The rate of evaporation Rk is limited by intermediates:
X s vX X
Rk = qkv xk 0 min qk qk cp zpk + xk ybk eb vpb 0 (13)
X
4w1 b − w1 b 5ybk ≤ R̄k 1 ∀ k ∈ K0 (6)
b∈B k∈K p∈P b∈B
Additionally, there are physical limitations on the amount A useful quantity in the presentation to follow is the cost
of water that can be evaporated in the intermediate batch: of production of a subset of products A ⊆ K for a given set
of values x, y, e, v, and z:
0
w1 b − w1 b ≤ r¯b 1 ∀ b ∈ B0 (7) X s vX
X
cA = qk qk cp zpk + xk ybk eb vpb 0
Empirical and semi-empirical models are based on exper- k∈A p∈P b∈B
imental designs that are valid only within specific values,
and accuracy can degrade severely if extrapolated beyond Optimization Solution Methodology
these bounds. Furthermore, compositions cannot differ
Our algorithm is based on a column-generation heuris-
drastically from benchmark mixtures, and the water content
tic, where a set-covering master problem interacts with
of powder products must be strictly controlled. Therefore,
independent subproblems that prescribe intermediate-to-
ingredient mass percentage values must lie within prede-
product groupings and mixture compositions. The algo-
fined lower and upper bounds:
rithm is based on the following sequence of steps:
ik ≤ ik ≤ ¯ ik 1 ∀ i ∈ I1 k ∈ K0 (8) 1. Singleton
2. Grouping
Similarly, bounds must be imposed on ingredient composi- 3. Configuration
tions in the intermediate batches: 4. Selection
5. Return to step 2 until convergence.
0 0 0
ib ≤ ib ≤ ¯ ib 1 ib ≤ ib ≤ ¯ ib 1 ∀ i ∈ I1 b ∈ B1 (9)
Singleton Step
which include, most importantly for the role they play in To start the process, we solve for the artificial case in
the manufacturing of intermediate batches, constraints on which each product is allowed to have its own dedi-
water content. cated intermediate batch. This is equivalent to specifying K
Esquejo et al.: Laundry Portfolio Optimization at P&G
Interfaces 45(5), pp. 444–461, © 2015 INFORMS 457
Premixes SRI requirements for the same stains often share an inter-
mediate to reduce costs; in isolation, they would have very
Pre1 similar intermediate batches.
Common intermediates To exploit this observation, we define a metric of simi-
larity between optimal (or best-known) intermediate com-
Pre2 e1 positions of singleton solutions. For any two singletons k
B1 (pre) B1 (post)
Products and l,
kl = cp pk − pl (14)
Pre3 Prod1 p∈
and P3 . Figure A.1 shows the corresponding Prod1 sin- preted to specify product groups that are appended to
gleton of the problem illustrated in Figure 6. (each composed of a subset of products).
We use the SAS/OR interior point nonlinear program-
ming solver for the singleton subproblems, which are inde- Sets
pendent and can therefore be solved in parallel by enabling • : Current groupings pool.
the SAS cofor multithreaded processing capability. Compli- • = k l ∈ × k ≥ l.
cations exist, however, primarily because of bilinear terms
Parameters
and nonconvex empirical and semi-empirical functions.
• kl : Similarity measure; see Equation (14).
We address these complications by employing the multi-
start mechanism of the SAS nonlinear programming (NLP) Variables
solver (also threaded), which aims to improve the likelihood • vl = 1 if the intermediate of product l is used in calcu-
of finding globally optimal solutions. Note that the single- lating kl for all k to be grouped with l, 0 otherwise.
ton step subproblem is a specialization of the configuration • ukl = 1 if product k is assigned to intermediate of sin-
step problem, where = and ykk = 1. gleton l, 0 otherwise.
Although global optimality is neither provable nor guar- •
A = 1 if group A ∈ is selected into the solution, 0
anteed, we will describe a method for improving the ulti- otherwise.
mate quantity that is to be derived from these solutions: the
Grouping Problem Formulation
globally optimal cost ck of each singleton. For now, we begin
to build a groupings pool as the union of all singletons. Minimize kl ukl (15)
In the previous example, = P1 P2 P3 . k l∈
subject to ukl = 1 ∀ k ∈ (16)
Grouping Step l∈
We exploit a physical observation to generate a relatively
small number of promising product groups to approximate vl = n (17)
l∈
this idea. The observation relies on the premise that expect-
ing products with similar performance requirements to be ukl ≤ vl ∀ k l ∈ (18)
able to benefit from extracting a portion of their mixture vk ≤ 1 − ukl ∀ k l ∈ k = l (19)
from the same intermediate is reasonable. Singletons are
ideal chemical compositions for products because they ben- ukl ∈ 0 1 ∀ k l ∈ (20)
efit from dedicated intermediates. Grouping finished prod- vl ∈ 0 1 ∀ l ∈ (21)
ucts based on the similarity of their singleton intermediate
batch compositions is known to be beneficial by observation Inequality (19) ensures that if a product has been assigned
in practice. For example, products that must meet stringent to an intermediate batch of another product’s singleton,
Esquejo et al.: Laundry Portfolio Optimization at P&G
458 Interfaces 45(5), pp. 444–461, © 2015 INFORMS
= 1 1
1
This variant of the problem has no integer variables,
Because the purpose of this step is to enrich to provide allowing us to use standard NLP solvers. Because the sin-
a better approximation of the optimal members of , gleton problem can be interpreted as a special case of the
each iteration (i.e., each time the grouping problem is called) configuration problem where only one product exists, we
must produce n groups of which at least one is not currently implement the singleton computation by using the configu-
in . We accomplish this by adding to the above formula- ration model code.
tion variable
A and constraints In the configuration step, we solve the related NLPs to
identify (locally) minimum-cost values of producing these
1 − ulk + ulk ≥ 1 −
A ∀ A ∈ k ∈ (22) independent groups of mixtures. These problems are solved
l∈A l∈\A in parallel using SAS NLP solver-threaded capability.
Initially, when we start with singletons and the first round ated NLPs. We illustrate this using the following example.
of groupings, A contains at most K + n subsets. We note Consider a scenario in which we are solving a five-
that the similarity-based grouping step attempts to heuristi- product, two-intermediate problem and have built A de-
cally exploit an observation of optimal intermediate compo- fined in Equation (33). Such a pool would be constructed
sitions but is not guaranteed to lead to optimal intermediate by the solution of the singleton step with one augmentation
batch groups. Furthermore, because of the nonconvexity of from the grouping step that produces 88P2 1 P3 91 8P1 1 P4 1 P5 99.
the problem, the configuration step likely produces solu- We solve an NLP for each member of A to calculate its cor-
tions that are not globally optimal. responding cost. Although we have not solved the NLP for
To address these issues, we iterate through the group- the group 8P1 1 P2 1 P3 9, we can estimate its lower bound as a
consequence of Inequality (32),
ing, configuration, and selection steps to accomplish two
goals: (1) improve our approximation of the optimal region
c8P1 1 P2 1 P3 9 2= max 4c8P1 9 + c8P2 9 + c8P3 9 51 4c8P1 9 + c8P2 1 P3 9 5 0
of P4K5 by augmenting A; and (2) improve the global opti-
mality of cost values by evaluating subgroup relationships. We define A = A ∪ 88P1 1 P2 1 P3 99 and the corresponding opti-
Item (1) occurs automatically within the formulation by mal cost (derived by calculating T ∗A in the selection step):
the inclusion of constraints that require at least one new
c ∗A1 n =
X
group not already in A to be generated, ensuring that at c Ai 0
Ai ∈T ∗A1 n
least one next-best product group is appended to A for sub-
sequent steps. Item (2) requires more explanation, which we We could augment A with all members of P4A5 because
illustrate in the following example. the lower-bound estimate can always be calculated from
Splitting a group A into a partition TA allows each sub- singletons. For large problems, P4A5 is prohibitively large;
group of the partition to have its own intermediate batch therefore, we instead augment based on building supersets
(and therefore more flexibility in the choice of mixture com- of existing groups.
position). Global optimality thus requires that We have been careful to call this an estimate of the lower
X bound because nonconvexity could prevent us from accu-
cA ≥ cAi 0 (32) rately calculating the costs of each Ai , because cAi is only an
Ai ∈TA
upper bound. Therefore, a true lower bound cannot be guar-
When we restrict solution sA to the products in Ai ⊆ A (by anteed (although it continues to improve as the algorithm
eliminating any k ∈ A\Ai ), it becomes feasible for any such iterates). Given some desired optimality gap n , we use
subgroup. It is therefore advantageous to replace sAi by sA ˜ n 2= 4cA1
∗ ∗ ∗
n − cA1 n 5/cA1 n ≤ n (34)
whenever the condition cA < cAi is detected for any subset
of A. This check is performed for all groups in the solution as a heuristic stopping criterion, with the expectation that
pool A and is motivated by our knowledge that a known the algorithm might terminate prior to achieving true global
locally optimal solution of Ai might be inferior (more costly) optimality within n .
Esquejo et al.: Laundry Portfolio Optimization at P&G
460 Interfaces 45(5), pp. 444–461, © 2015 INFORMS
Algorithm Summary in 1996 and is a section head in R&D at Procter & Gamble.
Here, we summarize the steps in the algorithm. She has extensive experience in both process and formu-
1. Singleton step. Let iteration count m = 0. Initialize A lation design in the fabric and home care business and is
with all singletons, solve singleton NLP configuration prob- working in the modelling and simulation group, focused on
lems, and calculate similarities ãkl for all pairs of singletons. integration and optimization of models, and application of
2. Grouping step. Let m = m + 1. Solve the grouping big data techniques. She also conducts training and consults
problem and append candidate groups to A, including par- on design of experiments and process control techniques
titions of each group. Also, append to A additional groups internally in P&G.
derived from combinations of members of A. Kevin Miller received his bachelor’s degree in chemistry
3. Configuration step. In parallel, solve independent con- from Xavier University in 1999. He is a principal researcher
figuration problem NLPs for new groups of A. at Procter & Gamble working in Fabric and home care mod-
4. Selection step. Solve set-covering problems for upper elling and simulation. He started in laundry product design
∗ ∗
bounds cA1n and lower-bound estimates cA1n , respectively. for North American Granules products and continued prod-
5. Terminate the algorithm if Inequality (34) holds; else uct design in Central Eastern European and Latin American
go to step 2. Granules. His current focus in modelling and simulation is
on model integration and product optimization.
Kevin Norwood received his PhD in physical chemistry
References from Iowa State University in 1990 and is a research fellow
ASTM International (2014) Standard Guide for Evaluating Stain in R&D at Procter & Gamble. He leads technical work to
Removal Performance in Home Laundering (ASTM International, create and apply modeling approaches to formulate prod-
West Conshohocken, PA). ucts within the fabric and home care businesses. His current
Audet C, Brimberg J, Hansen P, Le Digabel S, Mladenovic N (2004) work is focused on integration of models across disciplines.
Interfaces 2015.45:444-461.
(AAOS) group at SAS. The projects he has worked on earned his PhD from University at Buffalo in 2008 and
include chemical mixture portfolio optimization for P&G, was a postdoctoral researcher at IBM T.J. Watson Research
operating rooms scheduling, renewable energy integration Center, where he worked primarily on unit commitment
and power system operations, retail inventory replenish- problem and supply chain management in the mining
ment and pricing, and optimization for data mining. He industry.
Interfaces 2015.45:444-461.