You are on page 1of 30

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/324135924

Assortment Optimisation Problem: A Distribution-Free Approach

Preprint · April 2018

CITATIONS READS

0 622

3 authors:

Dmytro Matsypura Erick Li


The University of Sydney The University of Sydney
32 PUBLICATIONS   763 CITATIONS    30 PUBLICATIONS   539 CITATIONS   

SEE PROFILE SEE PROFILE

Rebecca Chan
The University of Sydney
2 PUBLICATIONS   9 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Assortment Optimisation View project

Incremental Network Design View project

All content following this page was uploaded by Dmytro Matsypura on 15 May 2019.

The user has requested enhancement of the downloaded file.


Assortment Optimisation Problem: A Distribution-Free Approach

Rebecca Chan, Zhaolin (Erick) Li, Dmytro Matsypura∗


The University of Sydney

Abstract
Assortment optimisation is a critical decision that is regularly made by retailers. The decision involves a
trade-off between offering a larger assortment of products but smaller inventories of each product and offering
a smaller number of varieties with more inventory of each product. We propose a robust, distribution-free
formulation of the assortment optimisation problem such that the assortment and inventory levels can be
jointly optimised without making specific assumptions on the demand distributions of each product. We
take a max-min approach to the problem that provides a guaranteed lower bound to the expected profit
when only the mean and variance of the demand distribution are known. We propose and test three heuristic
algorithms that provide solutions in O(n log(n)) time and identify two cases where one of the heuristics is
guaranteed to return optimal policies. Through numerical studies, we demonstrate that one of the heuristics
performs extremely well, with an average optimality gap of 0.07% when simulated under varying conditions.
We perform a sensitivity analysis of product and store demand attributes on the performance of the heuristic.
Finally, we extend the problem by including maximum cardinality constraints on the assortment size and
perform numerical studies to test the performance of the heuristics.
Keywords: max-min approach; static substitution; heuristic; cardinality constraints

1 Introduction
Increased attention to the assortment optimisation problem reflects a general industry trend of growing
product variety, which in turn is due to several reasons, including technology and company drive to fulfil
changing consumer needs (El Maraghy et al., 2013) and low margins due to increased competition (Kurtuluş
and Nakkas, 2011). From a managerial perspective, the incentive is clear: suboptimal assortment policies
lead to missed sales or increased inventory costs. There is significant empirical evidence to support the
importance of the problem. Studies by Chong et al. (2001) across 8 food categories and 5 stores found that
reconfiguring 32 products with suboptimal policies led to an increase in profits of 25.1%, while industry
trends have seen retailers reduce assortment size. For example, large U.S. chains, such as Walgreen Co. and
Walmart Stores Inc., have chosen to reduce the number of varieties of measuring tape from 24 to 4 and
the number of varieties of superglue from 25 to 11 (Brat et al., 2009). Indeed, the problem continues to be
relevant due to the continual increase in the number of products on the market, for example, Chong et al.
(2001) show that the risk of suboptimal policies increases with the size of the problem.
Assortment optimisation, also referred to as assortment planning, has its roots in the revenue management
of the airline industry (c.f. McGill and Van Ryzin, 1999). The literature on assortment optimisation can
be split into two main substitution mechanisms (static and dynamic) and three key consumer behaviour
models (multinomial logit, exogenous demand and locational choice). Static or out of assortment substitution

? Authorsare in alphabetical order.


∗ Corresponding author.
Email address: dmytro.matsypura@sydney.edu.au (Dmytro Matsypura)

Preprint submitted to Elsevier May 15, 2019


assumes that consumers make choices with knowledge of the assortment offered but without knowledge of
the inventory levels, making no attempt to substitute if their first choice is out of stock (Smith and Agrawal,
2000). This can be contrasted with dynamic or out of stock substitution, wherein a consumer makes an
attempt to find a substitute if their chosen product is out of stock.
Three models are commonly used to model consumer behaviour processes: exogenous demand, locational
choice and multinomial logit. The exogenous demand model specifies known demand behaviour and substitu-
tion behaviour for each product, with the probability of substitution modelled explicitly through a matrix
of substitution probabilities. Smith and Agrawal (2000) study static substitution effects in an exogenous
demand context by allowing for general substitution rather than deterministic or one-way patterns. Kök and
Fisher (2007) provide an empirical approach to estimating the substitution and demand parameters under an
exogenous demand model while allowing for stockouts. Hübner et al. (2016) provide an analytical solution to
the problem studied by Kök and Fisher (2007) and propose a more efficient heuristic. A key characteristic of
the exogenous demand model is that no assumptions are made regarding consumer behaviour. However,
while the exogenous demand model is able to accurately handle consumer behaviour, this results in the need
for greater data collection to estimate the substitution matrix.
The locational choice model, originally developed by Lancaster (1966) by extending the work of Hotelling
(1929), represents variants and consumers using vectors in multi-dimensional (‘attribute’) space, where
each dimension corresponds to a product characteristic. The substitution probability is determined by the
proximity of the consumer’s ideal vector to those of the surrounding products. Lancaster (1966) models
uniformly distributed consumer preferences such that the entire attribute space is covered, and Gaur and
Honhon (2006) extends this model by studying the problem under both static and dynamic substitution
in the context of horizontally differentiated products. While the locational choice model allows for greater
control over the specification of substitution rates and heterogeneity levels for consumers, it considers only
continuously valued attributes, does not account for randomness in consumer preferences and does not allow
demand to be a function of the assortment size (c.f. Gaur and Honhon, 2006).
A third commonly used consumer choice model is the multinomial logit (MNL) model, one of a class of
random utility models that represent the utility of a product as a random variable. Talluri and van Ryzin
(2004) study the problem when consumer behaviour is modelled either explicitly or using the MNL and show
that the optimal policy is one of an ordered sequence of “efficient” sets, also referred to as ‘revenue-ordered
sets’. Their result is extended by Liu and Van Ryzin (2008), who prove that under the same setup as that of
Talluri and van Ryzin (2004), subsets nested by the marginal profit guarantee optimality. A robust model is
studied by Rusmevichientong and Topaloglu (2012), where the parameter of the MNL is not deterministic
and instead is taken from a compact set of likely values. Rusmevichientong et al. (2014) extend this model
by modelling the parameter of the MNL as a random variable. The problem where variants are identically
priced and demand is normally distributed is studied by van Ryzin and Mahajan (1999), who make static
choice assumptions and show that the optimal solution in this problem is a subset formed by adding the
most popular items. Their model is extended by Li (2007), who allows products to be differently priced
under a normally distributed demand.
One important extension to the assortment optimisation problem is the inclusion of cardinality constraints.
These constraints play an important role for retailers with limited shelf capacity and for e-retailers who, for
example, face the problem of choosing a limited number of ads to show and wish to maximise the number of
clicks on ads placed (Rusmevichientong et al., 2010). Rusmevichientong et al. (2010) jointly approximate
parameter values and by solving the assortment optimisation problem, show that optimality of subsets
nested by marginal profit, as shown by Talluri and van Ryzin (2004), does not hold for constrained problems.
Instead, their problem can be solved using a polynomial time algorithm, which they call ‘StaticMNL’. Davis
et al. (2013) show that assortment optimisation problems with five general classes of constraints, including
cardinality bounds on the offered assortment, can be reformulated as linear programs.
A common approach to the assortment optimisation problem is to model the inventory-level decision
using the single-period newsvendor model. Two common streams of classical newsvendor modelling exist.
The first assumes an underlying demand distribution. However, this assumption may be restrictive and
may not provide sufficient protection against possible changes to future demand. The second stream of
newsvendor modelling attempts to overcome such uncertainty by studying distribution-free models.
2
The max-min approach is commonly used in a distribution-free newsvendor setting. Early literature
comes from the seminal work by Scarf (1958), who shows that for a class of distributions, described by only
their first and second moments, the worst-case demand is satisfied by a distribution with positive mass at
only two points. His result is extended by Gallego and Moon (1993), who study the same problem and
provide a closed form for the upper bound on the profit function. They also consider important extensions
to Scarf’s result, including allowing for second purchasing opportunities, fixed ordering costs and multi-item
cases. Several notable extensions of the distribution-free newsvendor problem include Alfares and Elmorra
(2005), who further study the extensions proposed by Gallego and Moon (1993) by taking into account
shortage costs beyond lost profits. Yue et al. (2006) provide a method for calculating the maximum expected
value of distribution information (EVDI) over the class of distributions with a given mean and variance.
Perakis and Roels (2008) investigate the newsvendor model under the minimax regret approach, where only
partial information about the demand distribution is given.
The MNL faces two key criticisms. The first is its independence from irrelevant alternatives (IIA)
assumption, under which different degrees of substitutability are not taken into account. The model instead
assumes that the new variant cannibalises from all existing variants in proportion to market penetration rates.
The second key limitation is that the attractiveness of the product determines both its market penetration
and substitution rates (c.f. Kök et al., 2008). One method proposed to address the IIA assumption is the
nested multinomial logit model (NMNL), which was first introduced by Williams (1977). The NMNL is
a special case of nested attraction models and is an important extension of the MNL that overcomes the
IIA assumption by nesting equally dissimilar products together such that the consumer first chooses a nest
followed by a product from this nest. However, the assortment optimisation problem under the NMNL is
NP-hard except for some special cases (Davis et al., 2013, 2014).
Both the MNL and the NMNL are particular instances of a more general model, known as the mixed
multinomial logit (MMNL) model, which was first introduced by Boyd and Mellman (1980). The key
advantage of MMNL is its ability to closely approximate all other discrete-choice random-utility maximisation
models, including the MNL and NMNL. Furthermore, its ability to overcome the aforementioned limitations
of the MNL makes it an attractive model, even though it is NP-hard (Bront et al., 2009; Rusmevichientong
et al., 2014).
Despite its shortcomings, the MNL model is among the most commonly used models in the assortment
optimisation literature. In this paper, we use MNL and static substitution assumptions to develop a max-min
approach to the assortment optimisation problem. We propose three heuristic algorithms and identify two
cases where two of these heuristics are guaranteed to return optimal solutions in O(n log(n)) time. Numerical
studies demonstrate that one of the heuristics performs extremely well in all test instances and returns
solutions with an average optimality gap of 0.07%. Then, we consider an extension of the original problem
by introducing maximum cardinality constraints on the subset to be offered and test our proposed heuristics
using simulated data. Our simulations show that the two cases that guarantee optimality in the unconstrained
problem no longer guarantee optimality when the problem is constrained. These findings are consistent with
those of Rusmevichientong et al. (2010), who show that the optimal set is not necessarily composed of the
few products with the highest marginal profit when the problem includes a capacity constraint. Nevertheless,
our heuristic provides solutions with an average optimality gap of < 0.3%. The remainder of the paper is
organised as follows. Section 2 presents the model formulation used, as well as several key results that we
apply to the problem. Section 3 outlines the proposed heuristics and identifies two cases where our heuristics
return optimal results. In Section 4, we conduct comprehensive numerical studies of both the constrained
and unconstrained problems and demonstrate the efficiency of the proposed heuristics. Finally, Section 6
discusses possible future directions of research and concludes the paper.

2 Model Formulation
This section presents a formulation of the distribution-free problem using the max-min approach. We
first model consumer behaviour using the MNL with static substitution assumptions. Then, we define and
apply Scarf’s max-min approach and ordering decision to the assortment problem.
3
2.1. Multinomial Logit Model
We model the underlying consumer behaviour using the MNL first proposed by Luce (1959). This
model is a special case of a broader group of random utility models. MNL models the utility of a product
i to a consumer as a random variable Ui . The utility is decomposed into a sum of two parts given by
Ui = ui + i , where ui is the expected utility and i represents a consumer’s deviation from the expected
2 2
utility. i is modelled using a Gumbel random variable with mean 0 and variance µ 6π , where π ≈ 3.14159 is
a mathematical constant and µ > 0 is a constant that corresponds to the level of heterogeneity amongst
consumers.
Let N = {1, ..., n} be the set of all possible variants and S ⊆ N be the subset stocked by the retailer.
The no-purchase option is product i = 0, such that consumers choose one of the options from the set S ∪ {0}.
Since utility is ordinal, without loss of generality, we let u0 = 0. Under the MNL, the consumer preferences
for each product i = 0, ..., n are given by
 
ui
νi = exp ,
µ

with the preference for the no-purchase option being ν0 = 1. The probability that a given product i is chosen
from set S is therefore defined as
νi
pi (S) = P .
1+ νk
k∈S

It is relatively simple to show that expanding the set S by adding a new product always increases the
overall probability that a consumer will purchase a product. A consumer purchases a product from the
assortment with probability
P
νk
k∈S
p= P .
νk + 1
k∈S

The marginal change in the probability of purchase after adding a new product j ∈
/ S to the existing
assortment is therefore given by
P P
νj + νk νk
k∈S k∈S
P − P (1)
1 + νj + νk 1+ νk
k∈S k∈S
 
1 1
= 1− P − 1 − P  (2)
1 + νj + νk 1+ νk
k∈S k∈S
1 1
= P − P . (3)
1+ νk 1 + νj + νk
k∈S k∈S

Since νj > 0, we find that Eq. (3) > 0; therefore, the addition of a new product j to an existing assortment
k ∈ S always increases the probability that a purchase is made. Indeed, we find that increasing the assortment
size usually increases the chance that a consumer will find a suitable product. However, due to substitution
effects, increasing the assortment size also decreases the probability that a specific product is selected. When
products with different profit margins are offered, the addition of all products is rarely favourable. To
illustrate this point, consider adding a product with a low profit margin to an existing set of products. The
new product will increase the probability of a consumer purchasing an item from the set. However, due to
substitution and cannibalisation, if the added demand is not sufficient to cover the profits, then it is not in
the retailer’s interest to add the product.
4
2.2. Static Substitution
Following the work of Smith and Agrawal (2000), we make two static or out-of-assortment substitution
assumptions.
Assumption 1. A consumer has knowledge of only the offered set S and not of the inventory levels available
at each point in time. If a consumer chooses a product that is not offered in the set S, the consumer makes
an attempt to substitute within the offered set.
Assumption 2. If a consumer’s chosen product is out of stock, the consumer makes no attempt to substitute
and the sale opportunity is lost.
While not realistic in all retail settings, there is considerable literature surrounding the static substitution
assumptions and the benefits of their use in assortment optimisation.
Specifically, the assumptions are valid for scenarios such as catalogue retailing, in which the consumer has
prior knowledge of only the set of offered products and not of each product’s inventory at the time of ordering
(c.f. van Ryzin and Mahajan, 1999). Moreover, van Ryzin and Mahajan argue that Assumption 2 is likely to
hold under circumstances where consumers make choices based on floor models. In this scenario, consumers
make a choice after reviewing the set of models on offer. If their choice is out of stock, the consumer may
then opt to leave the store and look elsewhere for their chosen product.

2.3. Scarf ’s Ordering Decision


We now consider the ordering rule from Scarf (1958) using the proof from Gallego and Moon (1993).
First, we define for each product i = 1, ..., n the following:

ci > 0 Unit cost


ti = ci (1 + mi ) Unit selling price
si = ci (1 − di ) Unit salvage price
qi Order quantity

where mi ≥ 0 and di ≥ 0 are the unit mark-up and discount rates, respectively, and the salvage cost is
assumed to be inclusive of any additional marginal costs such as disposal and holding that may be incurred
as a result of over-stocking. We define the observed store demand D, and the expected store demand and
coefficient of variation are denoted by λ and σ, respectively. The individual product demand expectation
and standard deviation are defined, respectively, as:

Di = Dpi =⇒ E[Di ] = λpi = λi


σi = (λpi )β σ i ∈ S ∪ {0},

where β ∈ [0, 1] is a shape parameter. A key benefit of defining the product standard deviation in this way is
that we are able to model different underlying distributions by varying the values of σ and β. For example,
when β = 1, σ can be interpreted as the coefficient of variation for each variant i since σ = λσii . Moreover,
when β = 0.5 and σ = 1, we have σi = λi and can therefore approximate the decision when the underlying
demand is represented by a Poisson distribution. No other assumptions about the underlying distribution of
the store demand D are made, other than that it comes from a class of distributions G with mean λ and
coefficient of variation σ.
The profit function of each product is given by:

ΠG (qi ) = ti E[min (Di , qi )] + si E [qi − Di ]+ − ci qi , (4)

where x+ = max {x, 0}. That is, if qi > Di , then there is remaining stock at the end of the selling period
and the salvage value is positive for each unit of remaining stock. However, if q ≤ D, then there are no
remaining stock units and hence no salvage value for the period.
We then solve for the worst-case profit scenario for all distributions F ∈ G for each possible subset S ⊆ N .
The ordering quantity that maximises the worst-case profit is given by the following theorem:
5
Theorem 1. (Scarf, 1958) The optimal ordering policy that maximises the worst-case profit for all distribu-
tions F ∈ G is given by:
r r 
? 1 t i − ci ci − si
qi = E[Di ] + σi − , (5)
2 ci − si t i − ci

with the associated worst-case profit under this quantity given by


n p o
max 0, (ti − ci )E(Di ) − σi (ti − ci )(ci − si ) , (6)

Proof. See appendix A.

The expected profit of assortment S is defined as the sum of the expected profits of each product under the
optimal ordering decision given by Scarf’s ordering rule in Eq. (5). Therefore, when making the assortment
decision, we seek to find the subset S ? ⊆ N that maximises the total expected profit for the assortment; that
is, we seek to solve the following optimisation problem:
Xh p i+
maximize (ti − ci )E(Di ) − σi (ti − ci )(ci − si ) (7)
S⊆N
i∈S

Using the definitions of the individual product demand expectation and standard deviation, we can
rewrite Problem (7) in the following form:
Xh p i+
maximize (ti − ci )λpi (S) − (λpi (S))β σ (ti − ci )(ci − si ) (8)
S⊆N
i∈S

where the decision variable S more explicitly enters the objective function. Furthermore, Problem (7) can be
written as a multiple-ratio fractional 0-1 programming problem
  β 
X  λνi zi (ti − ci ) λνi zi p
maximize P − P  σ (ti − ci )(ci − si )
. (9)
1+ νk zk 1+ νk zk
n

z∈{0,1}
i∈N
k∈N k∈N

where the decision variable is z ∈ {0, 1}n , and zi = 1 if and only if product i is included in the optimal
assortment S ? . This form is very common in the assortment optimisation literature; see Borrero et al. (2017)
for a recent survey. Problem (9) is known to be NP-hard in the general case for β = 0 and β = 1; see Hansen
et al. (1990). Prokopyev et al. (2005) show that the problem remains NP-hard for the case in which the
denominator is positive and enumerator is non-negative, which is the case here. However, the complexity of
Problem (9) for β ∈ (0, 1) is unknown.
To provide the intuition behind the problem complexity, we recall that adding a product j ∈ / S to the
existing assortment S changes the probabilities of purchase pi for each product i ∈ S; see Section 2.1. The
change in probabilities of purchase pi affects the profit function (4) and, ultimately, the worst-case profit (6),
which we are attempting to maximise.

3 Heuristics
Due to the combinatorial nature of the problem defined in Eq. (7), the number of subsets grows
exponentially with the input size. For small n, the optimal solution can be obtained by an exhaustive
search of all 2n − 1 subsets. However, this approach quickly becomes intractable for large problems. For
example, while a search of 20 variants requires approximately 4 seconds, the increase to 24 variants increases
the computational time to 100+ seconds when performed in MATLAB 2016b using an Intel(R) Core(TM)

6
i5-6300U CPU processor. Therefore, the key motivation for the following heuristics is to approximate the
problem efficiently and effectively when the problem size is large. In particular, we propose three heuristics
and examine their performance. The proposed heuristics reduce the exhaustive search to n subsets, nested
by indicators of the product’s potential success.

3.1. Heuristic 1
The first heuristic, inspired by the results obtained in Talluri and van Ryzin (2004) and Li (2007),
computes the theoretical profit for each product under the assumption that the variant was purchased by all
consumers and creates n subsets nested by this theoretical profit. We first define the ‘profit rate’ for each
product as:
q
(1)
Rj = λ(tj − cj ) − σλ (tj − cj )(cj − sj ) j = 1, . . . , n.

Then, we sort the products and, if necessary, re-index the variants such that the profit rate is monotonically
(1) (1) (1)
decreasing in the index: R1 ≥ R2 ≥ . . . ≥ Rn . The first subset is formed by taking the product with the
highest profit rate and calculating the optimal quantity under Scarf’s ordering rule and the overall profit
under this policy. Then, in each iteration, the product with the next highest profit rate is added and the
optimal ordering quantity and corresponding expected profit are calculated. Using this policy, n nested
subsets are created as follows:

Sj = {1, ..., j}, j = 1, ..., n.

Finally, we compare the expected profits for each of the n resulting policies and take the assortment policy
with the highest expected profit. The algorithm is summarised below:

Algorithm 1 Heuristic based on profit rate


1: procedure
2: for j = 1 to |N | do
(1) p
3: Rj = λ(tj − cj ) − σλ (tj − cj )(cj − sj )
4: j =j+1
5: end for
(1) (1) (1) (1)
6: Sort and re-index Rj such that R1 ≥ R2 ≥ · · · ≥ Rn
0
7: S = ∅, S = ∅
8: Π(S) = 0, Π(S 0 ) = 0
9: for i = 1 to |N | do
10: S = S ∪ {i} n o
P p
11: Π(S) = i∈S max 0, (ti − ci )E(Di ) − σi (ti − ci )(ci − si )
qi ≥0
12: if Π(S) > Π(S 0 ) then
13: S0 = S
14: Π(S 0 ) = Π(S)
15: end if
16: i=i+1
17: end for
18: return S 0 , Π(S 0 )
19: end procedure

Talluri and van Ryzin (2004) show that the optimal policy for the assortment optimisation problem is
one of a set of “efficient” subsets, where efficiency is given by the following definition.

7
Definition 1. (Talluri and van Ryzin, 2004): A set is inefficient if the same or greater expected revenue
can be achieved using a combination of other sets without achieving a greater probability of purchase. A set is
efficient if it is not inefficient.
Using this definition of efficiency, we identify two circumstances, β = 1 and σ = 0, under which the
heuristic creates efficient sets and is therefore guaranteed to produce the optimal policy. Our result is similar
to, and draws upon results from, those of Gallego and Phillips (2004) and Liu and Van Ryzin (2008), who
use this definition of efficiency to show that nesting by the marginal profit rate in their static optimisation
problem produces efficient sets. However, the optimality does not hold for all realisations of σ and β in our
objective function. Instead, we achieve optimality using Algorithm 1 only when σ = 0 and/or β = 1.
We formalise these observations in the following theorems.
Theorem 2. If β = 1, the profit function can be written as:
X p
Π(S) = (ti − ci )λi − σi (ti − ci )(ci − si )
i∈S
(1)
X
= pi Ri ,
i∈S

and the optimal policy is one of the subsets S1 , S2 , ..., Sn corresponding to the subsets nested by the profit rate
(1)
Ri .
Proof. See Appendix B.
Theorem 3. If σ = 0, the optimal policy is one of the subsets S1 , S2 , ..., Sn corresponding to the subsets
(1)
nested by the profit rate Ri .
Proof. See Appendix C.

3.2. Heuristic 2
The second heuristic is formulated using a procedure similar to the first, except that we use marginal
revenue instead of the profit to generate nested subsets. That is, we create nested subsets ordered by the
marginal revenue rate, which is equal to the expected profit when the assortment offered contains only that
given product. First, the rate can be defined as:
(2)
Ri = Π(S(i) )
 β
qi qi p
= λ(ti − ci ) − λ σ (ti − ci )(ci − si ), i = 1, ..., n,
1 + qi 1 + qi

where S(i) implies that set S contains only product i. Again, the rates are re-ordered and re-indexed, if
(2) (2) (2)
necessary, such that R1 ≥ R2 ≥ ... ≥ Rn . We begin by computing the profit of the assortment for the
single product with the highest rate. Then, in each iteration we add the variant with the next highest rate,
taking the assortment only if the assortment policy produces a higher profit. The heuristic again allows for
comparison of n potential subsets, taking the subset with the highest expected profit as our solution. We
present the pseudocode in Algorithm 2.

3.3. Heuristic 3
We define Heuristic 3 as the strategy where we pick the subset obtained by either Heuristic 1 or Heuristic
2 that yields the highest profit; that is, Heuristic 3 returns

S ? = arg max Π(S 1? ), Π(S 2? ) .




Note that since Heuristic 3 is the maximum of Heuristics 1 and 2, it is also guaranteed to be optimal when
σ = 0 and β = 1.
8
Algorithm 2 Heuristic based on marginal revenue
1: procedure
2: for j in 1:|N | do
(2) qj qj p
3: Rj = 1+q j
λ(tj − cj ) − ( 1+q j
λ)β σ (tj − cj )(cj − sj )
4: j =j+1
5: end for
(2) (2) (2) (2)
6: Sort Rj and re-index such that R1 ≥ R2 ≥ · · · ≥ Rn
0
7: S = ∅, S = ∅,
8: Π(S) = 0, Π(S 0 ) = 0
9: for i in 1:|N | do
10: S = S ∪ {i} n o
P p
11: Π(S) = i∈S max 0, (ti − ci )E(Di ) − σi (ti − ci )(ci − si )
qi ≥0
12: if Π(S) > Π(S 0 ) then
13: S0 = S
14: Π(S 0 ) = Π(S)
15: end if
16: i=i+1
17: end for
18: return S 0 , Π(S 0 )
19: end procedure

4 Numerical Study
4.1. Setup and Performance Measures
This section details a numerical study of the performance of the heuristics using simulated data. For
the purposes of brevity, the variables are divided into two sets referred to as the item parameters and
the store parameters. The item parameters include the item price, cost, salvage value, attractiveness, and
any other variables directly linked to the individual item, whereas the store parameters refer to the store
demand λ, shape parameter β and coefficient of variation σ. In the following numerical studies, we perform
simulations over equally spaced values of the store parameters. The results are aggregated by measuring
average performance using two measures: the hit rate and optimality gap.

4.1.1. Hit Rate


The hit rate of each heuristic is the fraction of instances where the heuristic returns exactly the optimal
result. Holding the store parameters constant, M problem instances of the item parameters are simulated,
and the hit rate is denoted as:
PM
hi
HR = i=1 · 100%,
M
where
(
(j)
1 if ΠE − ΠH = 0
hi =
0 otherwise,

(j)
and ΠE and ΠH , j = {1, 2, 3} are the profits of the assortments obtained using the Exhaustive Search
method and the Heuristics, respectively.

9
4.1.2. Optimality Gap
The optimality gap (∆E%) is the percent difference between the profits obtained under the Heuristics
and the Exhaustive Search methods. The measure is defined as:
(j)
ΠE − ΠH
∆E% = · 100%.
ΠE
In the following simulations, unless otherwise specified, the reported optimality gap is the average optimality
gap over M instances where the store parameters are constant and the item parameters are generated
randomly.

4.1.3. Critical Fractile


The data used to simulate the item parameters are given in Table 1. These values are selected such
that the critical fractile (CF) for all products lies within the interval [0.6, 0.9] for the studies performed in
Sections 4.2 and 5, where the fractile is calculated as
ti − ci
CF = .
(ci − si ) + (ti − ci )

The numerator ti − ci is the understock cost, or the lost profits, incurred when a retailer is unable to meet
demand. The denominator is the sum of the understock cost and the overstock cost, i.e., the cost associated
with excess stock levels (including any holding or disposal costs and salvage value).

Variable Distribution
Selling Price (ti ) ∼ U(50, 105)
Cost (ci ) ∼ U(15, 20)
Salvage (si ) ∼ U(0, 5)
Attractiveness (νi ) ∼ U(1, 4)

Table 1: Data distributions for the numerical studies

We select the critical fractile to be in the interval [0.6, 0.9], a range consistent with both existing research
(e.g., Song and Zipkin (1996) use the range [0.5, 0.8]) and practical applications. When the critical fractile is
sufficiently small (< 0.5), the costs of overstocking far outweigh the costs of understocking, often resulting in
a conservative policy of not stocking at all. In general, for products with overstock costs significantly higher
than understock costs, the retailer may opt for an alternative strategy, such as adopting a make-to-order
system, instead of a make-to-stock system as we consider here.

4.2. Results
In this section, we perform a simulation study with varying parameter values to investigate the effect of
certain parameters on the heuristic performance. More specifically, we vary values of the store parameters
λ, σ and β across equally spaced intervals in each iteration. That is, for 0 ≤ β ≤ 1, 50 ≤ λ ≤ 250 and
0 ≤ σ ≤ 1.5, the parameters β, λ and σ take on appropriately spaced values, and M instances are simulated
for each combination of these parameters with n = 20. Beginning with β = 0, λ = 50, and σ = 0, we
simulated M = 100 problem instances of product data ti , ci , si and νi from the uniform distributions given
in Table 1 and measured the performance of the heuristics.

10
1.8 100

1.6 90

1.4 80

70
1.2
60
1
50
0.8
40
0.6
30
0.4
20

0.2 10

0 0
50 90 130 170 210 250 50 90 130 170 210 250

(a) Optimality gap relative to expected store demand (b) Hit rate relative to expected store demand

1.8 100

1.6 90

1.4 80

70
1.2
60
1
50
0.8
40
0.6
30
0.4
20

0.2 10

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

(c) Optimality gap relative to shape parameter (d) Hit rate relative to shape parameter

1.8 100

1.6 90

1.4 80

70
1.2
60
1
50
0.8
40
0.6
30
0.4
20

0.2 10

0 0
0 0.3 0.6 0.9 1.2 1.5 0 0.3 0.6 0.9 1.2 1.5

(e) Optimality gap relative to coefficient of variation (f) Hit rate relative to coefficient of variation

Figure 1: Heuristic performance on simulated instances of the unconstrained problem with the total number of
variants n = 20. The plots show the relative optimality gap (left) and the mean hit rate (right) for different values of
store parameters. All values are averages across 100 instances simulated for each combination of store parameters λ,
β and σ.
11
Figure 1 shows the observed hit rates and optimality gaps, averaged over values of the parameters λ, β
and σ.
For Heuristic 1, we observe that increased expected store demand has a minimal but positive effect on
performance, improving the average optimality gap from 0.23% at λ = 50 to 0.14% at λ = 250, with a
similar decreasing trend for the hit rate (see Figs. 1e and 1f). For a constant store variance, higher store
demand decreases the coefficient of variation σ, resulting in more accurate performance of the heuristic.
We observe the opposite effect for Heuristic 2, that is, marginally decreasing performance with increasing
expected demand. The average optimality gap is 0.88% and 0.98% for λ = 50 and λ = 250, respectively,
with an overall average gap of 0.94%. Heuristic 3 appears to be robust to changes in λ, with little change in
either the hit rate or optimality gap.
When β = 1, the hit rates of Heuristics 1 and 3 reach 100% (and therefore a 0% optimality gap) regardless
of all other parameter values, as stated in Theorem 2. Similarly, as per Theorem 3, these simulations confirm
that Heuristic 1 is optimal when σ = 0 (Fig. 1f); therefore, Heuristic 3 is also optimal at σ = 0. While it
is not necessarily realistic to assume that the coefficient of variation of demand is zero in practice, we can
examine the behaviour of the heuristics surrounding this point. From Fig. 1e, we can see that the optimality
gap for Heuristic 3 increases only marginally between 0.0 and 0.9, reaching an average of 0.01% at σ = 0.3
and 0.06% at σ = 0.9 (see Table 6 for tabulated values of the results shown in Figs. 1a to 1f). A large value of
σ is extremely risky for the retailer; for example, σ = 1.5 suggests that the variance is 1.5 times larger than
the expected demand. However, in the context of retailing, the variance of demand is usually finite because
consumer demand is bounded. Therefore, we still find that Heuristic 3 performs well under reasonably large
demand variance.
We observe that the accuracy of Heuristics 1 and 3 is relatively robust to changes in λ, the store demand.
However, for Heuristic 2, the probability of selecting the correct subset decreases marginally as λ increases.
For β, Heuristic 3 improves the hit rate compared to that of Heuristic 1. In general, when β is small (< 0.7),
the optimality gap is relatively robust to changes in β. However, as β approaches 1, the probability of
choosing the correct subset approaches 1 as expected.
Heuristic 3 clearly outperforms the other heuristics under the scenarios simulated above. Indeed, Heuristic
3 performs exceptionally well. Even when the average hit rate drops to 69.92% at σ = 1.5, the average
optimality gap increases only to 0.20%. Similarly, under suitable conditions for λ and β, the average
optimality gap remains less than 0.1% (see Table 6 in Appendix D for tabulated results). Therefore, we
consider only Heuristic 3 in the following simulations of the constrained problem.

5 Cardinality Constraints
With ever-increasing product variety, it may be the case that the optimal assortment to offer from a large
set of products is still large in size. However, retailers, particularly those who operate in brick-and-mortar
retailing, often face shelf-space constraints. Even if the optimal assortment of a large potential set of variants
is significantly smaller than the set of all potential variants, it may not be feasible to stock the entire optimal
set when accounting for other units that the retailer must stock simultaneously. This problem extends to
scenarios such as e-retailing, where a company must decide which advertisements to place on a website with
limited advertising space to maximise the click-through rate. In such cases, it is often sensible to consider a
cardinality constraint on the assortment size, limiting the assortment size to a maximum cardinality k such
that |S| ≤ k ≤ n. We therefore consider an extension of the original problem, where the optimal assortment
must contain no more than k products. The constrained problem is given by:
Xh p i+
maximize (ti − ci )E(Di ) − σi (ti − ci )(ci − si )
i∈S
(10)
subject to S ⊆ N
|S| ≤ k

12
Pk
The constrained problem is reduced in size from a search of 2n − 1 subsets to a search of i=1 Cni
subsets, where k is the maximum cardinality of the assortment to be offered, and the heuristic algorithm
reduces to a search of k possible subsets.
We present the results of a numerical study conducted to test the efficiency and efficacy of Heuristic 3 in
solving the constrained problem. Again, following procedures similar to those in earlier sections, we simulate
data from the distributions outlined in Table 1. We let the possible number of products be n = 20 for all
iterations. While setting the cardinality constraint to some constant value k, we vary the values of the store
parameters in the same fashion as above. For k = {4, 8, 12}, M = 100 instances are simulated under each set
of conditions, and the optimality gap and hit rate are used to compare the results.

5.1. Results
Figure 2 shows the plots of the mean hit rate and optimality gap for M = 100 instances simulated for
each combination of store parameters β, λ and σ for each value of k = {4, 8, 12}, as well as the unconstrained
case (k = 20).
As k grows towards |N |, the heuristic performance converges to the results obtained in the unconstrained
case (see Figures 2e, 2a and 2c). However, for small k, the heuristic performance decreases. Specifically,
when the maximum cardinality k is smaller than the size of the optimal set S ? in the unconstrained problem
(k < |S ? |), the heuristic does not necessarily produce the optimal solution even under optimal conditions in
the unconstrained problem (i.e., when β = 1 or σ = 0). When β = 1 and the heuristic is optimal under the
unconstrained setting, for k = 4, the average optimality gap is 0.05% with a mean hit rate of 90.67%. As
k → n, these values tend to 0 and 100 percent, respectively. Indeed, for k = 8 and n = 20, the simulated
values return a 0% optimality gap for β = 1 (see Table 7 in Appendix D for the tabulated values).
Our results are consistent with the findings of Rusmevichientong et al. (2010), who showed that under a
profit maximising policy with a capacity constraint modelled with a dynamic substitution and the MNL,
the optimal solution does not follow the intuition from Liu and Van Ryzin (2008). That is, the optimal
assortment does not necessarily contain the few products with the highest marginal profit. Similarly, under the
distribution-free profit-maximising problem studied here, the optimal solution under cardinality constraints
does not follow the intuition of Talluri and van Ryzin (2004), and the subsets nested by profit rate as created
by Heuristic 1 do not necessarily guarantee optimality. Nonetheless, the heuristic performs well, with average
optimality gaps of 0.16%, 0.07% and 0.07% for k = 4, 8, and 12, respectively.
As observed in the optimality gap graphs, the similarity index approaches the values obtained in the
unconstrained problem as k → n. Again, we note that although Heuristic 3 guarantees optimality when
σ = 0 and β = 1 in the unconstrained case, in the constrained case, optimality is not guaranteed. Instead,
for k = 4, the hit rate is 83.42% at σ = 0, increasing to 99.89% for k = 8 and 100% for k = 12.
We also present here a summary table of the optimality gap for the constrained problem, averaged over
values of the maximum cardinality constraint, k, for equally spaced values of the parameters β, σ, and λ.
The results shown in Table 2 are consistent with those in Figures 2c-2f. These figures show that for small
values of k, (i.e., for k = 4), the average optimality gap is larger. However, for values of k close to n, the
results tend towards the unconstrained problem. For more detailed results, we refer the reader to Tables 6
and 7 in Appendix D.

Optimality Gap (%)


β σ λ
k 0 0.2 0.4 0.6 0.8 1.0 0 0.3 0.6 0.9 1.2 1.5 50 90 130 170 210 250 Average
4 0.23 0.23 0.20 0.15 0.11 0.05 0.08 0.10 0.12 0.17 0.20 0.29 0.14 0.15 0.17 0.17 0.18 0.16 0.16
8 0.09 0.10 0.09 0.08 0.05 0.00 0.00 0.00 0.02 0.06 0.12 0.21 0.07 0.07 0.07 0.08 0.08 0.06 0.07
12 0.10 0.10 0.09 0.08 0.05 0.00 0.00 0.00 0.02 0.06 0.13 0.21 0.07 0.06 0.07 0.07 0.08 0.07 0.07
20 0.09 0.10 0.09 0.08 0.05 0.00 0.00 0.00 0.02 0.06 0.14 0.20 0.07 0.07 0.06 0.07 0.07 0.07 0.07

Table 2: Summary table of the average optimality gap for the constrained problem. This table shows the average
optimality gap for various values of parameters. Problems were simulated with n = 20, and hence, k = 20 is equivalent
to the unconstrained problem and included for comparison.

13
0.3 100

90
0.25
80

70
0.2
60

0.15 50

40
0.1
30

20
0.05
10

0 0
50 90 130 170 210 250 50 90 130 170 210 250

(a) Optimality gap relative to expected store demand (b) Hit rate relative to expected store demand

0.3 100

90
0.25
80

70
0.2
60

0.15 50

40
0.1
30

20
0.05
10

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

(c) Optimality gap relative to shape parameter (d) Hit rate relative to shape parameter

0.3 100

90
0.25
80

70
0.2
60

0.15 50

40
0.1
30

20
0.05
10

0 0
0 0.3 0.6 0.9 1.2 1.5 0 0.3 0.6 0.9 1.2 1.5

(e) Optimality gap relative to coefficient of variation (f) Hit rate relative to coefficient of variation

Figure 2: Performance of Heuristic 3 on simulated instances of the constrained problem with the total number of
variants n = 20. The plots show the relative optimality gap (left) and the mean hit rate (right) for different values of
store parameters. All values are averages across 100 instances simulated for each combination of store parameters λ,
β and σ for each value of maximum cardinality k = {4, 8, 12, 20}.
14
5.2. Run Times
Table 3 presents the average run times for the Exhaustive Search, Heuristic 1 and Heuristic 2 methods
for the unconstrained and constrained problems. For each set of parameters, we simulated 100 instances
and recorded the average run time. The results of our tests for large n confirm that the run time of both
heuristics grows approximately linearly with n. For these instances, we did not run the Exhaustive Search.

Run times of Proposed Methods (seconds)


n k Exhaustive Search Heuristic 1 Heuristic 2
20 4 0.70 < 0.01 < 0.01
20 8 9.95 < 0.01 < 0.01
20 12 31.91 < 0.01 < 0.01
20 20 36.36 < 0.01 < 0.01
1,000 200 - 0.06 0.06
10,000 200 - 0.62 0.52
100,000 200 - 7.97 5.66

Table 3: Table of average run times (given in seconds) for the Exhaustive Search and Heuristic methods. All values
are averages over 100 instances.

5.3. Expected Value of Additional Information


We consider what happens when retailers stock the optimal quantity under the distribution-free approach
in comparison to when the true distribution is known and can be used to find the optimal quantity. This
measure of robustness is commonly referred to as the Expected Value of Additional Information (EVAI) (e.g.,
Gallego and Moon, 1993) or Expected Value of Distributional Information (EVDI) (e.g., Yue et al., 2006)
and measures the expected value of the added knowledge of the true demand distribution. The EVAI can be
calculated as
Πf (S † ) − Πf (S ? )
EVAI = ,
Πf (S † )

where Πf (S † ) is the profit under the distribution f ∈ G where the knowledge of the distribution has been
used to obtain the optimal stocking policy and Πf (S ? ) denotes the optimal profit obtained using the proposed
approach where demand follows the true distribution f .
Here, we perform simulations to calculate the EVAI when the true demand is both normally distributed
and exponentially distributed. We chose these distributions because they have two distinctly different shapes.
Our proposed method, i.e., Heuristic 3, is based on Scarf’s rule and it has been documented in the literature
that Scarf’s rule performs very well under normal distribution (see, for example, Gallego and Moon, 1993).
Hence, we chose a normal distribution to confirm this finding. Exponential distribution, on the other hand,
is skewed with a peak at 0 and is an excellent candidate to reduce the performance of Scarf’s rule and, hence,
our method.
Normal distribution. If we assume that the demand for each product follows a normal distribution
with a mean λi = λpi and a standard deviation σi = (λpi )β σ, we use this information to calculate the optimal
quantity; then, for each product,
 
† −1 t i − ci
qi = F ,
(ci − si ) + (ti − ci )

15
where F −1 (·) refers to the inverse of normal CDF with mean λi and standard deviation σi . Then,

qi† − λi
Zi =
σi
and the expected sales for a given product i is 0 if the product is not offered or

salesi = λi + σi (−f (Zi ) + Zi (1 − F (Zi )))

if the product is offered, where f (·) refers to the normal pdf with mean λi and standard deviation σi as
above. Then, the profit for each product is

Πi = salesi (ti − ci ) − (qi† − salesi )(ci − si )

with the total profit X


Π(S † ) = Πi .
i∈N

Exponential distribution. If we assume that the demand for each product follows an exponential
distribution with mean 1/λi , then the optimal quantity qi† is 0 if product i is not offered, or if the product is
offered, it is given by  
ti − ci
qi† = −λi log 1 − .
(ci − si ) + (ti − ci )
Note that the exponential distribution has a constant coefficient of variation. Hence, we must set β = σ = 1.
Then, the expected sales quantity is given by
1 †
salesi = (1 − e−λi qi ).
λi
Similarly, the profit for each product is

Πi = salesi (ti − ci ) − (qi† − salesi )(ci − si )

with the total profit X


Π(S † ) = Πi .
i∈N

Table 4 shows the values of the average percentage EVAI using 100 simulated instances for each value
of β ∈ (0, 1], σ ∈ (0, 1.5] and λ ∈ [50, 250] for normally distributed demand. Under these conditions,
the proposed method produces policies that are on average within 0.42% of the optimal decision when
the policy has been optimised using the true known distribution. Similarly, Table 5 shows values of the
average percentage EVAI using 100 simulated instances for each value of β = 1, σ = 1 and λ ∈ [50, 250] for
exponentially distributed demand. Under these conditions, the proposed method produces policies that are
on average within 2.31% of the optimal decision when the policy has been optimised using the true known
distribution.
These results indicate that the proposed method, i.e., Heuristic 3, performs very well under the normal
distribution of demand. However, its performance decreases slightly when the true distribution of demand is
exponential.

5.4. Effect of First and Second Moments on Chosen Assortment


We also investigate the effect of the first and second moments of the distribution on the chosen assortment.
The following plots detail the average chosen assortment size for the Exhaustive Search and the three proposed
Heuristic methods and their relationship with the mean demand (λ) and the demand variance, measured here

16
EVAI for Normal Distribution (%)
β σ λ
k 0.2 0.4 0.6 0.8 1.0 0.3 0.6 0.9 1.2 1.5 50 90 130 170 210 250 Average
4 0.06 0.12 0.24 0.48 1.13 0.14 0.26 0.40 0.54 0.68 0.47 0.42 0.41 0.39 0.37 0.37 0.41
8 0.07 0.12 0.24 0.45 1.21 0.14 0.29 0.45 0.54 0.67 0.48 0.42 0.43 0.40 0.38 0.39 0.42
12 0.07 0.12 0.25 0.47 1.13 0.14 0.28 0.43 0.54 0.66 0.47 0.43 0.41 0.40 0.36 0.37 0.41
20 0.07 0.12 0.23 0.49 1.13 0.14 0.28 0.42 0.54 0.66 0.46 0.44 0.38 0.40 0.40 0.38 0.41

Table 4: Summary table of the average EVAI under the normal distribution of demand for the constrained problem.
Problems were simulated with n = 20; hence, k = 20 is equivalent to the unconstrained problem and is included for
comparison.

EVAI for Exponential Distribution (%)


λ
k 50 90 130 170 210 250 Average
4 2.10 2.31 2.25 2.27 2.54 2.26 2.29
8 1.97 2.01 2.28 2.13 2.34 2.23 2.16
12 2.44 2.43 1.84 1.95 2.17 2.21 2.17
20 2.57 2.14 2.59 2.13 2.39 2.06 2.31

Table 5: Summary table of the average EVAI under the exponential distribution of demand. Problems were simulated
with n = 20; hence, k = 20 is equivalent to the unconstrained problem and is included for comparison.

using the coefficient of variation (σ). Problems were simulated randomly, using λ ∈ [50, 250], σ ∈ [0, 1.5], and
β ∈ [0, 1] (as per the remainder of the paper). We simulated 30 problems of size n = 15 for each combination
of σ, λ and β and calculated the average problem size for each value of λ and σ. The findings shown in
Figures 3a and 3b reflect the commonly noted behaviour in the assortment optimisation literature. That is,
for increases in the variability of store demand (and therefore product demand), a store tends to offer smaller
assortments. Moreover, increases in the mean store demand lead to increases in the size of the assortment
offered. This behaviour is apparent for all Heuristic methods, as well as the Exhaustive Search.

5.5. Shelf Space


One interesting extension of the assortment optimization problem we study is the inclusion of shelf-space
constraints. In retail practice, assortment and shelf-space decisions are typically sequential steps in the
category planning process. Assortment planning is usually executed by the marketing department, whereas
shelf planning is a subordinate planning problem generally owned by the sales department (Hübner and
Schaal, 2017). However, these two problems are mutually dependent if the shelf space is limited. Here, we
briefly describe the inclusion of a shelf-space constraint in our model and leave the detailed analysis for
future work.
From the proof of Theorem 1 we know that the retailer’s worst-case profit function is concave and
piecewise continuous, see Eq. (18). To simplify the modelling of the piecewise nature of the objective function,
we split the inventory level into two parts as qi (S) = qi1 (S) + qi2 (S), where qi1 (S) ≥ 0 and qi2 (S) ≥ 0. Then,
the shelf-space constraint is X
(qi1 (S) + qi2 (S)) ≤ K, (11)
i∈S

where K ≥ 0 represents the available shelf space. We introduce an indicator variable yi such that yi = 1 if

17
4 4

3.5 3.5

3 3

2.5 2.5

2 2

1.5 1.5

1 1

0.5 0.5

0 0
50 90 130 170 210 250 0 0.3 0.6 0.9 1.2 1.5

(a) Assortment size over expected store demand (b) Assortment size over coefficient of variation

Figure 3: The effect of demand parameter λ and coefficient of variation σ on the size of the chosen assortment.

i ∈ S and yi = 0 if i ∈
/ S. The constraints on inventory level of product i can be formulated as:

λ2i (S) + σi2 (S)


qi1 (S) ≤ , (12)
2λi (S)
qi1 (S) + qi2 (S) ≤ M yi . (13)

These two constraints properly account for the piecewise nature of the objective function. Notably, the
objective function in Eq. (18) is continuous, concave, and first-order differentiable. Therefore, if constraint (12)
is not binding, qi2 (S) = 0. The worst-case expected profit of variant i then becomes

λ2 (S)
Πi (qi1 (S), qi2 (S)) = −(ci − si )(qi1 (S) + qi2 (S)) + (ti − si ) 2 i qi1 (S)
σi (S) + λ2i (S)
(ti − si ) λi (S) + σi2 (S)
 2 
+ + qi2 (S)
2 2λi (S)
s
2
λ2 (S) + σi2 (S)

(ti − si )
− qi2 (S) + i − λi (S) + σi2 (S).
2 2λi (S)

The shelf-space model is then transformed to


X
maximize Π(qi1 (S), qi2 (S))
i∈S
X
subject to (qi1 (S) + qi2 (S)) ≤ K
i∈S
λ2i (S) + σi2 (S)
qi1 (S) ≤ for all i ∈ N
2λi (S)
qi1 (S) + qi2 (S) ≤ M yi for all i ∈ N
S ⊆ N, |S| ≤ k
yi ∈ {0, 1} for all i ∈ N

This model is more challenging than the models we study in the previous sections. In a special case, where

18
K is small, such that qi2 (S) = 0 for all i, the model reduces to a special case of Problem 9 or Problem 10,
depending on whether the cardinality constraint is present or not. In particular, the worst-case expected
profit of variant i takes the following form:

λ2 (S)
Πi (qi1 (S), qi2 (S)) = −(ci − si )(qi1 (S)) + (ti − si ) 2 i qi1 (S)
σi (S) + λ2i (S)
(ti − si ) λi (S) + σi2 (S)
 2 
+
2 2λi (S)
s
 2 2
(ti − si ) λi (S) + σi2 (S)
− − λi (S) + σi2 (S)
2 2λi (S)
λ2i (S)
= −(ci − si )qi1 (S) + (ti − si ) qi1 (S)
σi (S) + λ2i (S)
2

(ti − si ) λ2i (S) + σi2 (S)


+ ·
2 2λi (S)
(ti − si ) λ (S) + σi2 (S)
2
− · i
2 2λi (S)
λ2i (S)
= −(ci − si )qi1 (S) + (ti − si ) qi1 (S)
σi (S) + λ2i (S)
2

and hence, we recover the first part of Eq. (18).


In another special case, when K is very large, such that the shelf-space constraint (11) is not binding, the
problem reduces to what we study in previous sections. In this case, we do not have to split the inventory
level into two parts, the worst-case profit of variant i takes the form of Eq. (18), and the shelf-space model
takes form of Problem 9 or Problem 10, depending on whether the cardinality constraint is present or not.
However, in a general case, the piecewise nonlinear nature of the objective function makes the shelf-space
model more difficult than the problems we study in previous sections. In future research, we shall investigate
how the heuristics perform in the shelf-space model.

6 Concluding Remarks
The assortment optimisation problem is an important area of study with wide applications to retail. When
retailers lack sufficient information or ability to accurately fit the product demand distribution, optimising
over an assumed distribution may result in suboptimal policies and the potential for lost sales and excess
overstock costs. We propose a new approach to the assortment optimisation problem using max-min decision
rules and by making multinomial logit choice and static substitution assumptions. This approach maximises
the worst-case profit over all possible realisations of the true demand distribution when only the first and
second moments of the distribution are known.
We propose three heuristic algorithms for situations in which the problem size is large and exhaustive
search methods become intractable. Numerical studies demonstrate that the proposed heuristics perform
well under the simulated conditions. When products have critical fractiles between 0.6 and 0.9, Heuristic 3
performs extremely well, with a mean optimality gap of 0.07% when simulated over variable values of the
coefficient of variation shape index (β), the coefficient of variation of store demand (σ) and the expected
store demand (λ). In fact, the heuristic performs well under most simulated conditions, with the exception
of extremely large coefficients of variation. When the coefficient of variation is between 1 and 1.5, the mean
optimality gap increases rapidly, with the largest optimality gap of any simulated instance (0.36%) occurring
when σ = 1.5. Nonetheless, when σ = 1.5, the average optimality gap is 0.20%. In a retailing context, the
variance of demand is bounded because consumer demand is bounded, and a value of σ larger than 1 implies

19
that demand variation is at least as large as the demand itself. Although we have considered values of σ on
the interval [0, 1.5], it is rare for retailers to encounter sufficiently large variance to result in a coefficient
of variation of 1.5; hence, we expect that our heuristics will perform well for most realistic values of the
coefficient of variation. When the problem size is small, exhaustive search methods can be used to find the
optimal solution. However, as the problem size increases, Heuristic 3 can be implemented in O(n log(n))
time, allowing for an efficient estimate of the optimal solution when the problem becomes intractable.
Based on the formulation used in this paper, we are able to identify several opportunities for further
research. First, we studied a single-period model in this paper. However, a single-period model is often
implemented when a product’s life cycle is short. The model requires that the initial inventory is zero
and assumes that any remaining units are ‘disposed’ of at the end of the selling period. Furthermore, a
key assumption of the single-period model is that demand for products is stationary or increasing. When
applied to multiple periods without dynamically adjusting for changes in product and store demand, the
single-period model may not produce optimal policies. Indeed, Song and Zipkin (1993) find that although
the single-period model is optimal for non-decreasing demand, when expected demand is decreasing over
time, for example, under future product obsolescence, the single-period model becomes the upper bound and
is therefore suboptimal. Therefore, an extension of the single-period model to a multi-period model may be
beneficial when these assumptions are violated.
In this paper, we used the max-min approach under the assumption of known mean and variance. However,
when access to information is further limited and accurate estimates of the moments cannot be obtained,
it may be useful to consider approaches that relax these moment assumptions. In particular, it may be
interesting to consider the minimax regret approach, which minimises the maximum opportunity cost (or
regret) associated with making suboptimal policies. This approach will be useful to retailers who might only
have a rough estimate of the true store demand with confidence intervals.
Finally, obvious extensions to the model would be to consider additional constraints such as inventory
or shelf space constraints that limit the total combined inventory of all products. Moreover, in this paper,
we considered the joint optimisation of the assortment composition and inventory levels. In doing so, we
assumed that the prices are exogenously determined. Therefore, a useful extension of the model may be to
consider the joint optimisation of product pricing under the max-min approach when the pricing structure is
not exogenously determined.

References
Alfares, H.K., Elmorra, H.H., 2005. The distribution-free newsboy problem: Extensions to the shortage penalty case. International
Journal of Production Economics 93, 465–477. doi:102.
Borrero, J.S., Gillen, C.P., Prokopyev, O.A., 2017. Fractional 0-1 programming: Applications and algorithms. Journal of Global
Optimization 69, 255–282.
Boyd, J.H., Mellman, R.E., 1980. The effect of fuel economy standards on the us automotive market: an hedonic demand
analysis. Transportation Research Part A: General 14, 367–378.
Brat, I., Byron, E., Zimmerman, A., 2009. Retailers cut back on variety, once the spice of marketing. Wall Street Journal .
Bront, J.J.M., Méndez-Dı́az, I., Vulcano, G., 2009. A column generation algorithm for choice-based network revenue management.
Operations Research 57, 769–784.
Chong, J.K., Ho, T.H., Tang, C.S., 2001. A modeling framework for category assortment planning. Manufacturing & Service
Operations Management 3, 191–210.
Davis, J., Gallego, G., Topaloglu, H., 2013. Assortment planning under the multinomial logit model with totally unimodular
constraint structures. Department of IEOR, Columbia University. Available at http://www. columbia. edu/gmg2/logit const.
pdf , 335–357.
Davis, J., Gallego, G., Topaloglu, H., 2014. Assortment optimization under variants of the nested logit model. Operations
Research 62, 250–273.
El Maraghy, H., Schuh, G., El Maraghy, W., Piller, F., Schönsleben, P., Tseng, M., Bernard, A., 2013. Product variety
management. CIRP Annals-Manufacturing Technology 62, 629–652.
Gallego, G., Moon, I., 1993. The distribution free newsboy problem: review and extensions. Journal of the Operational Research
Society 44, 825–834.
Gallego, G., Phillips, R., 2004. Revenue management of flexible products. Manufacturing & Service Operations Management 6,
321–337.
Gaur, V., Honhon, D., 2006. Assortment planning and inventory decisions under a locational choice model. Management Science
52, 1528–1543.

20
Hansen, P., Poggi de Aragão, M.V., Ribeiro, C.C., 1990. Boolean query optimization and the 0-1 hyperbolic sum problem. Annals
of Mathematics and Artificial Intelligence 1, 97–109. URL: https://doi.org/10.1007/BF01531072, doi:10.1007/BF01531072.
Hotelling, H., 1929. Stability in competition. The Economic Journal 39, 41–57.
Hübner, A., Kuhn, H., Kühn, S., 2016. An efficient algorithm for capacitated assortment planning with stochastic demand and
substitution. European Journal of Operational Research 250, 505–520.
Hübner, A., Schaal, K., 2017. Effect of replenishment and backroom on retail shelf-space planning. Business Research 10,
123–156. URL: https://doi.org/10.1007/s40685-016-0043-6, doi:10.1007/s40685-016-0043-6.
Kök, A.G., Fisher, M.L., 2007. Demand estimation and assortment optimization under substitution: Methodology and
application. Operations Research 55, 1001–1021.
Kök, A.G., Fisher, M.L., Vaidyanathan, R., 2008. Assortment planning: Review of literature and industry practice, in: Retail
Supply Chain Management. Springer, pp. 99–153.
Kurtuluş, M., Nakkas, A., 2011. Retail assortment planning under category captainship. Manufacturing & Service Operations
Management 13, 124–142.
Lancaster, K.J., 1966. A new approach to consumer theory. Journal of Political Economy 74, 132–157.
Li, Z., 2007. A single-period assortment optimization model. Production and Operations Management 16, 369.
Liu, Q., Van Ryzin, G., 2008. On the choice-based linear programming model for network revenue management. Manufacturing
& Service Operations Management 10, 288–310.
Luce, R.D., 1959. Individual choice behavior: A theoretical analysis. Wiley.
McGill, J.I., Van Ryzin, G.J., 1999. Revenue management: Research overview and prospects. Transportation science 33,
233–256.
Perakis, G., Roels, G., 2008. Regret in the newsvendor model with partial information. Operations Research 56, 188–203.
Prokopyev, O.A., Huang, H.X., Pardalos, P.M., 2005. On complexity of unconstrained hyperbolic 0–1 programming problems.
Operations Research Letters 33, 312 – 318. URL: http://www.sciencedirect.com/science/article/pii/S016763770400094X,
doi:https://doi.org/10.1016/j.orl.2004.05.011.
Rusmevichientong, P., Shen, Z.J.M., Shmoys, D.B., 2010. Dynamic assortment optimization with a multinomial logit choice
model and capacity constraint. Operations Research 58, 1666–1680.
Rusmevichientong, P., Shmoys, D., Tong, C., Topaloglu, H., 2014. Assortment optimization under the multinomial logit model
with random choice parameters. Production and Operations Management 23, 2023–2039.
Rusmevichientong, P., Topaloglu, H., 2012. Robust assortment optimization in revenue management under the multinomial
logit choice model. Operations Research 60, 865–882.
van Ryzin, G., Mahajan, S., 1999. On the relationship between inventory costs and variety benefits in retail assortments.
Management Science 45, 1496–1509.
Scarf, H., 1958. A min max solution of an inventory problem. Studies in the Mathematical Theory of Inventory and Production .
Smith, S.A., Agrawal, N., 2000. Management of multi-item retail inventory systems with demand substitution. Operations
Research 48, 50–64.
Song, J.S., Zipkin, P., 1993. Inventory control in a fluctuating demand environment. Operations Research 41, 351–370.
Song, J.S., Zipkin, P.H., 1996. Inventory control with information about supply conditions. Management science 42, 1409–1419.
Talluri, K., van Ryzin, G., 2004. Revenue management under a general discrete choice model of consumer behavior. Management
Science 50, 15–33.
Williams, H.C., 1977. On the formation of travel demand models and economic evaluation measures of user benefit. Environment
and Planning A 9, 285–344.
Yue, J., Chen, B., Wang, M.C., 2006. Expected value of distribution information for the newsvendor problem. Operations
Research 54, 1128–1136.

A Proof of Theorem 1
To make this paper self-contained, in this section, we present a proof of Scarf’s result for the worst-case
maximising ordering policy. Moreover, we find that the approach we provide here using the Karush-Kuhn-
Tucker (KKT) and tangent conditions is a more comprehensive proof than that presented in Scarf (1958).

First, the expected profit function for a product i is given by:

ΠG (qi ) = (ti − ci )E[min(Di , qi )] + si E[qi − Di ]+ − ci qi

where [x]+ = max{0, x}. We note that the expected profit function can be rewritten as:

E min(Di , qi )] = E[Di − [Di − qi ]+


 

= λi − E[Di − qi ]+

21
and
(qi − Di )+ = (qi − Di ) + [Di − qi ]+
Hence, the profit function becomes

ΠG (qi ) = (ti − si )λi − (ci − si )qi − (ti − si )E[Di − qi ]+

To minimise the expected profit function, we must maximise the lost sales, given by

maximize E[Di − qi ]+ ,

subject to the following constraints


Z ∞
f (x) δx = 1
Z0 ∞
xf (x) δx = λi
Z0 ∞
x2 f (x) δx = λ2i + σi2
0

where Pr(Di = x) = f (x). That is, the demand must satisfy the given first and second moment conditions,
with the sum of probabilities equal to 1. We solve the dual problem, given by the following:

minimize γ1 + γ2 λi + γ3 (λ2i + σ 2 )
subject to γ1 + γ2 Di + γ3 Di2 − [Di − qi ]+ ≥ 0

Scarf (1958) showed that for every quantity qi , there exists a two-point distribution that satisfies the
worst-case profit. Indeed, the constraint verifies that the function (Di − qi )+ is a tangent to the curve
γ1 + γ2 Di + γ3 Di2 at two distinct points. In order to solve for this distribution, we denote the two points as
Di,1 and Di,2 and reformulate the optimisation problem as follows:

minimize γ1 + γ2 λi + γ3 (λ2i + σ 2 )
2
subject to γ1 + γ2 Di,1 + γ3 Di,1 − (Di,1 − qi ) ≥ 0
2
γ1 + γ2 Di,2 + γ3 Di,2 ≥0

Solving for Di,1 and Di,2 , we have 2 possible cases;

A.0.1. Case 1: Di,1 6= 0


The Lagrangian of the reformulated dual is given by

L = γ1 + γ2 λ + γ3 (λ2i + σi2 ) + ν1 (γ1 + γ2 Di,1 + γ3 Di,1


2 2
− (Di,1 − qi )) + ν2 (γ1 + γ2 Di,2 + γ3 Di,2 ).

22
Under the KKT conditions, the gradient of the Lagrangian vanishes w.r.t. γ1 , γ2 , γ3 , ν1 , ν2 at the optimal
point. That is,
δL
= 1 + ν1 + ν2 = 0
δγ1
δL
= λi + ν1 Di,1 + ν2 Di,2 = 0
δγ2
δL
= λ2i + σ 2 + ν1 Di,1
2 2
+ ν2 Di,2 =0
δγ3
δL
= ν1 γ2 + 2γ3 Di,1 − 1 = 0
δDi,1
δL
= ν2 γ2 + 2γ3 Di,2 = 0
δDi,2

By the tangent conditions in the constraint function,


2
γ1 + γ2 Di,1 + γ3 Di,1 − (Di,1 − qi ) = 0 =⇒ (γ2 − 1)2 − 4γ3 (γ1 + qi ) = 0,
2
γ1 + γ2 Di,2 + γ3 Di,2 = 0 =⇒ γ22 − 4γ1 γ3 = 0

which yields
q

Di,1 = qi − (λi − qi )2 + σi2
q

Di,2 = qi + (λi − qi )2 + σi2

with associated probabilities


 
1 qi − λi
q
Pr Di,1 = qi − (λi − qi ) + σi = ν1∗ = + p
2 2
2 4(λi − qi )2 + 4σi2
 
1 qi − λi
q
Pr Di,2 = qi − (λi − qi )2 + σi2 = ν2∗ = − p
2 4(λi − qi )2 + 4σi2

λ2i +σi2
Hence, for qi ≥ 2λi , we have
  +
1 qi
q
E[Di − qi ]+ = +p qi − (λi − qi )2 + λ2i − qi
2 4(λi − qi )2 + 4σi2
  +
1 qi − λi
q
+ −p qi + (λi − qi )2 + σi2 − qi
2 4(λi − qi )2 + 4σi2
  +
1 qi − λi
q
= −p q + (λ − q )2 + σ2 − q
i i i i i
2 4(λi − qi )2 + 4σi2
  q
1 qi − λi
= 1− p · (λi − qi )2 + σi2
2 4(λi − qi )2 + 4σi2
q 
1
= (λi − qi )2 + σi2 + λi − qi
2

23
A.0.2. Case 2: Di,2 = 0
With Di,2 = 0, the Lagrangian becomes

L = γ1 + γ2 λ + γ3 (λ2i + σi2 ) + ν1 (γ1 + γ2 Di,1 + γ3 Di,1


2
− (Di,1 − qi )) + ν2 γ1

Again by the KKT conditions, the gradient of the Lagrangian vanishes at the optimal point. Therefore, we
have:
δL
= 1 + ν1 + ν2 = 0
δγ1
δL
= λi + ν1 Di,1 = 0
δγ2
δL
= λ2i + σ 2 + ν1 Di,1
2
=0
δγ3
δL
= ν1 γ2 + 2γ3 Di,1 − 1 = 0
δDi,1

and by the tangent conditions,

γ1 = 0
2
γ1 + γ2 Di,1 + γ3 Di,1 − (Di,1 − qi ) = 0 =⇒ (γ2 − 1)2 − 4γ3 (γ1 + qi ) = 0

Hence, solving for Di,2 , ν1 , ν2 yields

λ2i + σi2 λ2
 
Pr Di,1 = = 2 i 2
λi λi + σi
σ2
Pr (Di,2 = 0) = 2
λi + σi2

λ2i +σi2
Therefore, for qi ≤ 2λi ,

+ 
λ2i + σi2 λ2i σ2
   
+ +
E[Di − qi ] = − qi · + (−qi ) (14)
λi λ2i + σi2 λ2i + σi2
λ2
= λi − 2 i 2 · qi (15)
σi + λ i

λ2i +σi2
and for qi ≥ 2λi ,

  q
1 qi − λi
E[Di − qi ]+ = −p · (λi − qi )2 + σi2 (16)
2 4(λi − qi )2 + 4σi2
q 
1 2 2
= (λi − qi ) + σi + λi − qi (17)
2

Combining both cases, we find that the retailer’s worst-case profit equals
 2
 −(ci − si )qi + (ti − si ) 2λi 2 qi λ2i +σi2
G σi +λi
if qi ≤ 2λi ,
Π (qi ) = (t −s )
 p 
λ2i +σi2
(18)
 −(ci − si )qi + i i 2 2
λi + qi − (λi − qi ) + σi if qi ≥
2 2λi ,

We note that the worst-case profit function is concave and that it is continuous and first-order differentiable

24
everywhere. Finally, the optimal ordering policy (Eq. (5)) and the maximum worst-case profit in Eq. (6) are
obtained by optimising the worst-case profit from Eq. (18).

B Proof of Theorem 2
The proof for this result is derived from the findings of Li (2007).

Proof. The profit for a set S is given by:


X n p o
Π(S) = max 0, (ti − ci )λi − σi (ti − ci )(ci − si ) .
i∈S

In the non-trivial case, it is reasonable to assume that the expected profit for each individual product is
non-negative, and hence, the profit function can be rewritten as:
X p
Π(S) = (ti − ci )λi − σi (ti − ci )(ci − si ).
i∈S

When β = 1, the profit of set S is:


X p
Π(S) = (ti − ci )λi − σi (ti − ci )(ci − si )
i∈S
X p
= (ti − ci )pi λ − pi λσ (ti − ci )(ci − si )
i∈S
ν
Pi (1)
X
= Ri .
1+ i∈S νi
i∈S

(1)
For the purposes of brevity in this proof, the profit rate from Heuristic 1 Ri is denoted simply by Ri . We
now show that when β = 1, the optimal assortment is one of the nested subsets created using Heuristic 1.
First, consider an arbitrary set S. The difference in expected profit under the policy of including (versus not
including) a product j is given by:
" #
X νi X νi νj
Π(S) − Π(S ∪ j) = P Ri − P Ri + P Rj (19)
1 + i∈S νi 1 + νj i∈S νi 1 + νj i∈S νi
i∈S i∈S
" P #
X νi (1 + νj +
νj i∈S νi ) νi
= P P Ri − Ri − Rj (20)
1 + νj + i∈S νi νj (1 + i∈S νi ) νj
i∈S
" #
νj X νi Ri
= P P − Rj . (21)
1 + νj + i∈S νi 1 + i∈S νi
i∈S

Since the first term in Eq. (21) is nonnegative by construction, the expected profit grows by adding product
j only if Π(S) − Π(S ∪ j) < 0, i.e., if the second term is negative:
P
i∈S νi Ri
P − Rj < 0. (22)
1 + i∈S νi

25
For j = 1, the LHS is:
P P P
νi Ri
i∈S νi Ri − Rj (1 + i∈S νi )
i∈S
P − R1 = P (23)
1 + i∈S νi 1 + i∈S νi
P P
i∈S νi Ri − R1 i∈S νi − Rj
= P (24)
1 + i∈S νi
P

i∈S i (R i − R1 )] − R1
= P . (25)
1 + i∈S νi

Recall that R1 ≥ R2 ≥ . . . ≥ Rn , and therefore, the numerator of Eq. (25) is negative for all i 6= 1. Hence,
the optimal assortment always includes the product with the highest profit rate.
We prove by contradiction that if a set is optimal and contains product j + 1, then it must also contain
product j. First, consider an optimal set S ∗ that contains product j + 1 but not j. Then, from Eq. (22), the
following must hold:
P
i∈S ∗ νi Ri
P − Rj ≥ 0 (26)
1 + i∈S ∗
P P
i∈S ∗ νi Ri − Rj (1 + i∈S ∗ (νi ))
=⇒ P ≥ 0. (27)
1 + i∈S ∗ (νi )

Since the denominator of the equation above is positive by construction, we must prove that the numerator
is positive. The numerator is given by:
X X X X
νi Ri − Rj (1 + νi ) = νi Ri − Rj − Rj νi (28)
i∈S ∗ i∈S ∗ i∈S ∗ i∈S ∗
X X
= (νi Ri ) − Rj − Rj (νi ) − Rj νj+1 + Rj+1 νj+1 (29)

i∈S{−(j+1)} ∗
i∈S{−(j+1)}
 
X X
= (νi Ri ) − Rj 1 + (νi ) + νj+1 (Rj+1 − Rj ), (30)
i∈S ∗
{−(j+1)} i∈S ∗ {−(j+1)}

where S ∗ denotes the optimal set S ∗ minus the element j + 1. In order to show this result, we must first
{−(j+1)}

consider that if S ∗ is optimal, then adding j + 1 to the set S ∗ must indeed be profitable. That is, we
{−(j+1)}

must have
P

i∈S{−(j+1)} νi Ri
P − Rj+1 < 0
1 + i∈S{−(j+1)}
∗ νi
 
X X
=⇒ νi Ri − Rj+1 1 + νi  < 0.

i∈S{−(j+1)} ∗
i∈S{−(j+1)}

Since Rj ≥ Rj+1 , then


   
X X X X
νi Ri − Rj 1 + νi  ≤ νi Ri − Rj+1 1 + νi  < 0.

i∈S{−(j+1)} ∗
i∈S{−(j+1)} ∗
i∈S{−(j+1)} ∗
i∈S{−(j+1)}

Hence, in Eq. (30), the sum of the first two terms is negative. Furthermore, since Rj ≥ Rj+1 , the final term
is also nonpositive and Eq. (27) does not hold. Therefore, if a product j + 1 is an element of the optimal
assortment, product j must also be offered and the optimal assortment is one of the nested subsets created

26
by Heuristic 1 when β = 1.

C Proof of Theorem 3
Proof. The profit for a set S is given by:
X n p o
Π(S) = max 0, (ti − ci )λi − σi (ti − ci )(ci − si ) .
i∈S

Again, each product is assumed to have a non-negative expected profit, giving the following assortment profit
function:
X p 
Π(S) = (ti − ci )λi − σi (ti − ci )(ci − si ) .
i∈S

When σ = 0, the profit function can be written as


X
Π(S) = (ti − ci )λi
i∈S
ν
Pi
X
= λ(ti − ci )
1+ i∈S νi
i∈S
(1)
X νi Ri
= P .
1 + i∈S νi
i∈S

Again, we consider the profit gained (or lost) under the policy of including product j
" #
X νi (1)
X νi (1) νj (1)
Π(S) − Π(S ∪ j) = R − R + R .
1 + i∈S νi i 1 + νj i∈S νi i 1 + νj i∈S νi j
P P P
i∈S i∈S

We omit the remainder of the details, except to say that the proof follows the same procedure as in the proof
of Theorem 2 above, and when σ = 0, the heuristic again guarantees that the optimal assortment is one of
(1)
the subsets nested by the profit rate Ri .

27
D
Tables

Unconstrained Problem Simulation Summary


Heuristic 1 Heuristic 2 Heuristic 3
Hit Rate Optimality Gap (%) Hit Rate Optimality Gap (%) Hit Rate Optimality Gap (%)
β Min Mean Max Min Mean Max Min Mean Max Min Mean Max Min Mean Max Min Mean Max
0.0 0.35 0.73 1 0 0.17 0.94 0.14 0.23 0.39 0.77 1.02 1.29 0.50 0.80 1 0 0.09 0.36
0.2 0.31 0.73 1 0 0.20 1.09 0.11 0.25 0.44 0.62 1.00 1.20 0.50 0.79 1 0 0.10 0.35
0.4 0.36 0.73 1 0 0.21 0.96 0.18 0.27 0.52 0.52 0.95 1.28 0.53 0.81 1 0 0.09 0.31
0.6 0.29 0.74 1 0 0.22 1.24 0.18 0.31 0.56 0.41 0.90 1.29 0.62 0.83 1 0 0.08 0.27
0.8 0.51 0.80 1 0 0.15 0.61 0.15 0.33 0.52 0.57 0.88 1.31 0.73 0.88 1 0 0.05 0.18
1.0 1.00 1.00 1 0 0.00 0.00 0.15 0.33 0.59 0.50 0.90 1.18 1.00 1.00 1 0 0.00 0.00

σ
0.0 1.00 1.00 1 0 0.00 0.00 0.11 0.20 0.27 0.85 1.07 1.29 1.00 1.00 1 0 0.00 0.00
0.3 0.91 0.97 1 0 0.00 0.01 0.15 0.21 0.30 0.91 1.07 1.31 0.92 0.97 1 0 0.00 0.01

28
0.6 0.78 0.87 1 0 0.03 0.06 0.15 0.26 0.35 0.79 0.96 1.20 0.81 0.90 1 0 0.02 0.04
0.9 0.64 0.75 1 0 0.10 0.19 0.18 0.30 0.46 0.60 0.91 1.25 0.72 0.82 1 0 0.06 0.11
1.2 0.36 0.62 1 0 0.27 0.61 0.19 0.34 0.51 0.49 0.86 1.12 0.55 0.73 1 0 0.14 0.26
1.5 0.29 0.53 1 0 0.55 1.24 0.16 0.41 0.59 0.41 0.78 1.16 0.50 0.70 1 0 0.20 0.36

λ
50 0.32 0.77 1 0 0.23 1.24 0.17 0.33 0.56 0.41 0.88 1.29 0.59 0.86 1 0 0.07 0.36
90 0.39 0.80 1 0 0.15 0.77 0.19 0.30 0.56 0.60 0.89 1.20 0.56 0.86 1 0 0.07 0.30
130 0.40 0.81 1 0 0.15 0.81 0.11 0.28 0.58 0.50 0.98 1.24 0.56 0.86 1 0 0.06 0.29
170 0.31 0.79 1 0 0.15 0.77 0.14 0.27 0.52 0.57 0.95 1.29 0.50 0.85 1 0 0.07 0.32
210 0.29 0.79 1 0 0.14 0.73 0.15 0.27 0.59 0.55 0.97 1.31 0.50 0.85 1 0 0.07 0.35
250 0.36 0.78 1 0 0.14 0.69 0.15 0.27 0.51 0.60 0.98 1.27 0.53 0.84 1 0 0.07 0.35
Overall 0.79 0.16 0.29 0.94 0.85 0.07

Table 6: Summary of Simulated Values from Figs. 1a and 1b. Minimum, average and maximum optimality gap and hit rate values are given versus the expected store demand (λ),
store coefficient of variation (σ) and shape parameter (β)
View publication stats
Constrained Simulation Summary Values for Heuristic 3 using n = 20 over k = {4, 8, 12}
k=4 k=8 k = 12
Hit Rate Optimality Gap (%) Hit Rate Optimality Gap (%) Hit Rate Optimality Gap (%)
β Min Mean Max Min Mean Max Min Mean Max Min Mean Max Min Mean Max Min Mean Max
0.0 0.49 0.67 0.87 0.10 0.23 0.51 0.50 0.79 1 0 0.09 0.34 0.44 0.79 1 0 0.10 0.38
0.2 0.42 0.67 0.87 0.10 0.23 0.64 0.45 0.79 1 0 0.10 0.45 0.50 0.81 1 0 0.10 0.34
0.4 0.50 0.72 0.88 0.00 0.20 0.50 0.52 0.81 1 0 0.09 0.36 0.51 0.81 1 0 0.09 0.37
0.6 0.54 0.76 0.91 0.00 0.15 0.41 0.60 0.83 1 0 0.08 0.28 0.57 0.83 1 0 0.08 0.27
0.8 0.69 0.80 0.90 0.00 0.11 0.25 0.74 0.88 1 0 0.05 0.24 0.69 0.89 1 0 0.05 0.26
1.0 0.75 0.91 1.00 0.00 0.05 0.14 0.99 1.00 1 0 0.00 0.00 1.00 1.00 1 0 0.00 0.00

σ
0.0 0.75 0.84 0.91 0.00 0.08 0.15 0.99 1.00 1 0 0.00 0.00 1.00 1.00 1 0 0.00 0.00
0.3 0.71 0.82 0.90 0.00 0.10 0.17 0.92 1.00 1 0 0.00 0.00 0.93 0.98 1 0 0.00 0.00

29
0.6 0.65 0.78 0.92 0.00 0.12 0.23 0.79 0.90 1 0 0.01 0.00 0.77 0.89 1 0 0.02 0.05
0.9 0.59 0.74 0.94 0.00 0.17 0.34 0.67 0.81 1 0 0.06 0.13 0.68 0.81 1 0 0.06 0.12
1.2 0.51 0.72 0.98 0.00 0.20 0.39 0.56 0.74 1 0 0.12 0.23 0.56 0.75 1 0 0.13 0.24
1.5 0.42 0.67 1.00 0.00 0.29 0.64 0.45 0.70 1 0 0.21 0.45 0.44 0.69 1 0 0.21 0.38

λ
50 0.60 0.78 0.98 0.01 0.14 0.32 0.59 0.87 1 0 0.07 0.29 0.56 0.87 1 0 0.07 0.31
90 0.57 0.77 1.00 0.00 0.15 0.42 0.54 0.86 1 0 0.07 0.33 0.57 0.86 1 0 0.06 0.32
130 0.42 0.76 0.98 0.00 0.17 0.64 0.49 0.85 1 0 0.07 0.33 0.51 0.86 1 0 0.07 0.35
170 0.44 0.75 1.00 0.00 0.17 0.59 0.45 0.84 1 0 0.08 0.45 0.44 0.84 1 0 0.07 0.35
210 0.48 0.75 1.00 0.00 0.18 0.54 0.52 0.84 1 0 0.08 0.36 0.47 0.85 1 0 0.08 0.38
250 0.49 0.76 0.96 0.02 0.16 0.51 0.50 0.85 1 0 0.06 0.29 0.51 0.84 1 0 0.08 0.37
Overall 0.76 0.16 0.85 0.07 0.85 0.07

Table 7: Summary of values for simulations of the constrained problem using Heuristic 3. For each combination of β, λ and σ, 100 instances were simulated with n = 20, and
k = 4, 6, 12. Results are averaged over values of β, λ and σ.

You might also like