International Journal of Production Research, 2017

Vol. 55, No. 22, 6801–6823, https://doi.org/10.1080/00207543.2017.1351635

Adjustable Robust Optimisation approach to optimise discounts for multi-period supply chain
coordination under demand uncertainty
Viktoryia Buhayenkoa∗ and Dick den Hertogb
a Department of Economics and Business Economics, Cluster for Operations Research And Logistics (CORAL), Aarhus University,
Aarhus, Denmark; b Department of Econometrics and Operations Research, Tilburg University, Tilburg, Netherlands
(Received 15 November 2016; accepted 27 June 2017)

In this research, a problem of supply chain coordination with discounts under demand uncertainty is studied. To solve the
problem, an Affinely Adjustable Robust Optimisation model is developed. At the time when decisions about order periods,
ordering quantities and discounts to offer are made, only a forecasted value of demand is available to a decision-maker. The
proposed model produces a discount schedule, which is robust against the demand uncertainty. The model is also able to
utilise the information about the realised demand from the previous periods in order to make decisions for future stages in an
Downloaded by [EP- IPSWICH] at 00:55 21 November 2017

adjustable way. We consider both box and budget uncertainty sets. Computational results show the necessity of accounting for
uncertainty, as the total costs of the nominal solution increase significantly even when only a small percentage of uncertainty
is in place. It is testified that the affinely adjustable model produces solutions, which perform significantly better than the
nominal solutions, not only on average, but also in the worst case. The trade-off between reduction of the conservatism of
the model and the uncertainty protection is investigated as well.
Keywords: Robust optimisation; supply chain coordination; demand uncertainty; Adjustable Robust Optimisation; discount

1. Introduction
Traditionally, the supplier tries to meet the actual demand of his customers. But can he gain more profit if he is able to
influence the ordering pattern of his customers? In the course of the years, the study of numerous means of supply chain
coordination has showed that not only individual, but also mutual benefits can be achieved by means of synchronisation. The
variety of research in this area can be seen in a review paper by Arshinder, Kanda, and Deshmukh (2008).
In this research, we consider the problem of multi-period supply chain coordination by means of temporary time-
based discounts with multiple heterogeneous customers and dynamic uncertain demand. This article extends the research of
Buhayenko and den Hertog (2016).
To reduce the complexity while modelling the supply chain, it is often assumed that the prices, which the supplier charges
his customers, are constant. In reality, however, prices can vary within the time horizon and from customer to customer (see
Phillips 2005, 18–20). Changing prices with the aim of influencing the customers’ total demand and boosting the supplier’s
profit has been widely studied by dynamic pricing (Elmaghraby and Keskinocak 2003; Bitran and Caldeney 2003).
Significant gains can also be achieved if discounts are viewed as a way to coordinate the supply chain. Here, the main type
is quantity discounts, which are offered, if the customer orders more products. Among the earliest papers to study channel
coordination by means of quantity discounts are Jeuland and Shugan (1983), Lal and Staelin (1984), and Lee and Rosenblatt
(1986). Quantity discounts for supply chain coordination are also researched by Weng (1995), Corbett and de Groote (2000),
Tsay and Agrawal (2000) and Chen, Federguen, and Zheng (2001). Further, Qi et al. (2004) apply a quantity discount policy
for one-supplier-one-retailer supply chain with demand disruption. Li and Liu (2006) study a single-buyer single-supplier
system with one type of product with multi-period probabilistic customer demand and develop a quantity discount policy to
achieve supply chain coordination. Qin et al. (2007) consider quantity discounts and franchise fees as coordination mechanism
in one-buyer-one-supplier supply chain with price sensitive demand. Tsai (2007) presents linearisation techniques for single
breakpoint, step and multiple breakpoint functions of quantity discounts. Xie et al. (2010) develop a wholesale price-discount
scheme for two-level supply chain consisting of a single manufacturer and multiple independent retailers and introduce early
order commitment strategy. Huang et al. (2011) study a reverse one-buyer-one-supplier supply chain and develop a quantity
discount contract, where the payment to the customer is exponentially decreasing in the number of false failure returns.
Du, Banerjee, and Kim (2013) compare two means of coordination, i.e. a quantity discount and a delay in payments. They
consider two-echelon one-buyer-one-supplier supply chains with price sensitive market demand.
∗ Corresponding author. Email: [email protected]

© 2017 Informa UK Limited, trading as Taylor & Francis Group

 6802 V. Buhayenko and D. den Hertog

Although many researchers (including the papers listed above) focus on studying deterministic problems, in real life,
demand parameters are often not known exactly and are difficult to estimate. Future demand is often forecasted based on
historical data with the help of expert opinions, and is therefore, prone to prediction or assessment errors. This is considered
to be one of the main sources of uncertainty in any supply chain system. The problem is also known as system nervousness,
which appears, when downstream players continuously change the timing of their orders (Tunc et al. 2013). As a result,
significant costs associated with re-planning can occur, if the realised demand differs considerably from predicted. The
uncertain nature of demand is discussed by Higle and Kempf (2011). Bernstein and DeCroix (2015) quantify the value of
advanced demand information. The problem of incorporating demand uncertainty into the production-planning processes of
companies is discussed by Fisher et al. (1994) and Ben-Tal et al. (2004). Arshinder, Kanda, and Deshmukh (2011) stress that
uncertainty is growing along with supply chain complexity and name coordination to be one of the means to improve the
performance of the individual players in supply chains. They provide a comprehensive review of supply chain coordination
mechanisms under both supply and demand uncertainty, as well as production disruptions.
Our research aims to develop a discount scheme that is robust against the uncertainty in demand.
Aclassical way to address a problem of this kind is stochastic programming (SP) (Birge and Louveaux 2011). Nevertheless,
this approach is notorious for its severe computational difficulties. It also requires insight into the distribution of the uncertain
parameter, which is almost always unknown in practice, and speculations about it are not precise. At the same time, wrong
assumptions about the distribution can result in a solution significantly deviating from the optimal one (Shapiro, 1994).
Moreover, the resulting feasible regions of SP problems are often non-convex. For example, Shapiro and Philpott (2007) state
Downloaded by [EP- IPSWICH] at 00:55 21 November 2017

that optimisation models with chance constraints have in general non-convex feasible regions. To gain a better understanding
of the complexity of stochastic programming problems, the reader may refer to Shapiro and Nemirovski (2005).
On the contrary, while applying Robust Optimisation (RO), no speculation on the uncertain parameter distribution is
needed. Instead, the unknown parameter is assumed to reside within the bounds of a user-specified uncertainty set. It is
ensured that the constraints hold for all possible values of the parameter in the uncertainty set. As opposed to Stochastic
Optimisation, in case of RO computational tractability is preserved for many optimisation problems and choices of the
uncertainty set. Due to these advantages, it has been decided to use RO in this research.
Although some concepts of RO date back to the seventies (Soyster 1973; Soyster 1974), the major contributions to the
theory appear in the late nineties (Ben-Tal and Nemirovski 1997; Ben-Tal and Nemirovski 1998; El-Ghaoui and Lebret 1997;
El-Ghaoui, Oustry, and Lebret 1998). The general theory and applications of robust optimisation can be found in Ben-Tal,
El-Ghaoui, and Nemirovski (2009). For practical considerations regarding application of the robust optimisation approach,
the reader may refer to the paper by Gorissen, Yanıkoğlu, and den Hertog (2015). Here, only a short explanation of the basic
concepts is given.
An uncertain linear programming problem can be mathematically written as follows:

min{c T x|Ax ≤ b}(c,A,b)∈U ,

x

where c ∈ Rn , b ∈ Rm and A ∈ Rm × Rn are the uncertain parameters residing in a compact uncertainty set U . The
inequality constraints Ax ≤ b must be satisfied for all the realisations of the uncertain parameters inside the uncertainty set
U . Therefore, there is not a single deterministic problem but a number of problems. If there is uncertainty in the objective or
the right-hand-side of the constraints, the problem can be reformulated. Thus, without loss of generality, it can be assumed
that only coefficients A are uncertain. In the so-called robust counterpart problem it is ensured that all constraints are satisfied
for all the uncertain parameters:
min{c T x|Ax ≤ b}, ∀A ∈ U.
x
The difficulty of a robust program depends on computational tractability of its robust counterpart.
Moreover, robustness with respect to U can be reformulated constraint-wise:

a T x ≤ b, ∀a ∈ U,

and the uncertain parameter a can be written in terms of a primitive factor ζ ∈ R L :

a = ā + Pζ, ζ ∈ Z,

where ā ∈ R n , P ∈ R n×L and Z is the uncertainty set for primitive factors. The vector ā is fixed and is called the nominal
value.
The easiest type of uncertainty regions is interval (or box) uncertainty. This uncertainty set leads to the following robust
counterpart:
(ā + Pζ )T x ≤ b, ∀ζ : ||ζ ||∞ ≤ ρ.
 International Journal of Production Research 6803

It can easily be proven that this semi-infinite inequality can be rewritten as:

āx + ρ||P T x||1 ≤ b,

where ρ||P T x||1 is an extra ‘safety’ term that accounts for uncertainty.
The advantage of the robust counterpart is that it does not have a semi-infinite structure anymore and can be easily
modelled as a set of linear constraints. However, the number of variables and constraints grows.
Despite being simple and tractable, the box uncertainty very often results in a too conservative solution. In order to
reduce the conservatism, other uncertainty sets were proposed, e.g. ellipsoidal uncertainty (Ben-Tal and Nemirovski 1999).
Unfortunately, the resulting robust counterpart is a Conic Quadratic Programming constraint. Another uncertainty set that
gives a possibility to control the conservatism, while having a linear robust counterpart, is the budget uncertainty set introduced
by Bertsimas and Sim (2004) and also used in this paper:

U = {ζ : ||ζ ||∞ ≤ 1, ||ζ ||1 ≤ }.

In this uncertainty set, only  of uncertain parameters are allowed to take their worst-case values. The resulting robust
counterpart is:
 T
Downloaded by [EP- IPSWICH] at 00:55 21 November 2017

ā x + v 1 1 + v 2 ∞ ≤ b
v 1 + v 2 = P T x,

which can easily be rewritten as a set of linear constraints.

If some knowledge is available about the distribution of the parameter, it can be utilised by means of Distributionally
Robust Optimisation, where the user does not commit to a specific distribution but instead deals with a set of possible
distributions (Wiesemann, Kuhn, and Sim 2014).
Adjustable Robust Optimisation, which is an important extension of Robust Optimisation for multi-stage problems, was
proposed by Ben-Tal et al. (2004). According to this approach, some of the decision variables are modelled as functions
(or decision rules) of the information on the values of the uncertain parameters revealed in the previous periods. If general
decision rules are substituted with affine decision rules, the approach remains computationally tractable for large linear
problems with fixed recourse and appear to be near optimal or, in some special cases, optimal (Bertsimas, Iancu, and Parrilo
2010; Iancu, Sharma, and Sviridenko 2013). The main advantage of this approach is that not all the optimisation variables
have to be determined immediately; instead of, the decision on some of them can be postponed until more information on
the uncertain parameters is available. This extra information can improve the quality of the solution to a great extent. This
attribute is especially important for multi-period problems, which can often become infeasible if classic RO is applied.
Robust Optimisation has been successfully applied to various optimisation problems. An extensive overview of research
in the field of robust optimisation can be found in the survey papers by Bertsimas, Brown, and Caramanis (2011) and
Gabrel, Murat, and Thiele (2014). Among applications that are relatively close to the current research can be distinguished
applications to supply chain management (e.g. Ben-Tal et al. 2005; Bertsimas and Thiele 2006), a dynamic pricing and
inventory control problem for a make-to-stock manufacturing system (e.g. Adida and Perakis 2006), an inventory control
problem (e.g. Ben-Tal, Golany, and Shtern 2009), emergency supply chain planning (e.g. Ben-Tal et al. 2011), an inventory
routing problem (e.g. Solyalı et al. 2012), supply chain network design (e.g. Pishvaee et al. 2011; Zokaee et al. 2014), a
supply chain downsising problem (e.g. Ashayeri, Ma, and Sotirov 2014), a production planning and pricing problem (e.g.
Ardjmand et al. 2016). To the best of our knowledge, the robust optimisation approach has never been applied to the problem
under consideration.
This research extends an earlier paper by Buhayenko and van Eikenhorst (2016), who developed a deterministic mixed-
integer model for supply chain coordination by discounts. The algorithm for discount calculation is based on the classic
Wagner–Whitin model, which is an important methodological contribution that provides an optimal solution for problems
with deterministic dynamic demand (Wagner and Whitin 1958). Incorporation of demand uncertainty into the Wagner–Whitin
problem has been considered as an important practical extension (Vargas 2009). In the current paper, we offer a discount
schedule that ensures supply chain coordination when uncertainty in future demand values is present. The distribution of
this parameter is not known, and the parameter resides in a user-specified uncertainty set. As the classic RO model does
not produce a feasible solution for our problem, we derive an Affinely Adjustable Robust Optimisation model with integer
‘here-and-now’ decision variables and continuous ‘wait-and-see’ variables that are approximated by affine decision rules.
The following research questions are answered in this paper:
 6804 V. Buhayenko and D. den Hertog

• What is the effect of different uncertainty levels on the nominal solution? Is the feasibility of the nominal solution
preserved, when demand uncertainty is in place?
• Is it possible to apply ARO to the problem under consideration in order to get feasible solutions with good
performance?
The computational experiments conducted during this research testify that the affinely adjustable model produces solutions
which perform better than the nominal solutions, not only on average, but also in the worst case. Already with a simple box
uncertainty set (Ben-Tal, El-Ghaoui, and Nemirovski 2009), which ensures maximal uncertainty protection, the robust
solutions outperform the nominal solutions. The objective function can be further improved assuming budget uncertainty
(Bertsimas and Sim 2004) which helps to regulate the conservativism of the solution. Box and budget uncertainty sets
are chosen because they result in tractable linear problems. The box uncertainty set has an advantage of simplicity of
representation but is too conservative. The budget uncertainty set allows to control the level of conservatism.
The contributions of this research can be summarised as follows:
• To the best of our knowledge, this study is the first to consider a robust supply chain problem, where discounts are
means of coordination. We develop an Affinely Adjustable Robust Optimisation model for this problem. Utilising
the proposed model, the supplier is able to coordinate his customers’ orders with the help of discounts even if the
demand is uncertain.
• We show that, even in the face of significant uncertainty, supply chain coordination by offering discounts is possible
Downloaded by [EP- IPSWICH] at 00:55 21 November 2017

and beneficial if the Adjustable Robust Optimisation methodology is applied. On the contrary, if uncertainty is not
taken into account at the modelling stage, the resulting solutions do not perform well.
The rest of the paper is organised in the following way. In Section 2, the pricing problem under consideration is described,
and the nominal model is presented. Section 3 contains the Affinely Adjustable Robust model. In Section 4, extensive
computational experiments, comparing the quality of the nominal and robust solutions, are presented. The conclusions are
given in Section 5.

2. Problem formulation and nominal mathematical model

Here, we discuss the main assumptions of the problem, define the parameters used, and introduce the nominal model.

2.1 Main assumptions and definitions of parameters and variables

We consider a single-product multi-period two-echelon supply chain problem with a supplier, who can be a manufacturer
or a wholesaler, and a number of heterogeneous customers of different sizes, who might have their own customers. In this
research, we are aiming to develop the discount schedule that is beneficial for the supplier. At the same time, the customers
are also better off if they accept the discounts offered by the supplier. The supplier possesses full information about the costs
of every customer, whereas information about their demand is available in the form of forecasted values. As a result, the
supplier is able to determine the discount he needs to offer to change the order pattern of his customers in a way that is
beneficial for him. Although the inventory costs of the customers may grow, the decrease in their purchasing costs makes the
new order pattern attractive for them.
At the beginning of the planning horizon, the supplier has to decide on the discount schedule, as well as in which periods
he is going to replenish. Also the customers have to decide on the periods in which they will place an order. At the beginning
of each period, the decisions about production quantities, if the supplier chose to replenish in this period, and order quantities,
if the customers decided to order in this period, are made.
The temporary time-based discount represents a single price reduction and is given temporarily at a certain order period
for a certain customer. The purpose of the discount is to change the timing of customers’ orders, not the order quantity.
Therefore, the discount per item does not depend on the customer’s order quantity. However, the total discount, of course,
depends on the amount of items purchased in a period with a discount. The customers are informed about their discount
schedules in advance. No backorders are possible, and all the customers’ demand has to be satisfied in the period it occurs.
Customers and the supplier have no capital, capacity or warehouse restrictions.
The total demand is not affected with prices and stays the same. Practical situations with the total demand, insensitive to
changes in prices, are discussed by Buhayenko and van Eikenhorst (2016). However, discounts influence the distribution of
the demand between periods. The customers move their orders to periods with a discount and order less in other periods. As
a result, supplier’s inventory costs decrease.
The problem under consideration is divided into three parts or sub-problems. Each part is an optimisation model in itself.
One of the parameters in the main model from Section 2.2 is the optimal value of a different optimisation model, which
 International Journal of Production Research 6805

Figure 1. Procedure scheme.

is explained in Section 2.3. We start our explanation from the most central model that determines replenishment schedules
for the supplier and customers (Section 2.2). The final discount for every period is calculated in Section 2.4. The procedure
scheme is shown in Figure 1.
Downloaded by [EP- IPSWICH] at 00:55 21 November 2017

The following sets are used in this paper:

(I) the set of all customers i (i = 1, . . . , I or i ∈ I ) and supplier i = 0;
(T) the set of all periods t (t = 1, . . . , T or t ∈ T ).
The models are described with the help of the following parameters:
dit demand for every customer i in every period t. The demand for supplier i = 0 is a summation of all his
customers’ demands;
pi initial unit price without a discount for every customer i and supplier i = 0. With the help of the unit price for
the supplier, his inventory holding costs are calculated;
si order processing costs for every customer i and supplier i = 0;
h i inventory holding costs for each customer i and for supplier i = 0 to carry a unit of inventory from period t to
period t + 1;
ci total costs for customer i before discounts are offered.
Parameters si and h i are assumed to be constant over time.
The variables in the models are as follows:
Hit inventory at the end of period t for every customer i and supplier i = 0;
Q it order quantity for customer i and supplier i = 0 in period t;
Bi total amount of compensation from supplier i = 0 to customer i;
Sit binary variable; 1, when the order for customer i or supplier i = 0 is placed in period t; 0, otherwise;
Ritk binary variable; 1, if the total demand of customer i from period t to period k is placed in period t; 0, otherwise.
H0 , S0 , Q 0 , B, R stand for vectors or arrays of corresponding optimisation variables.

2.2 Determination of the replenishment patterns

One of the ways to formulate the central part of the nominal model is proposed by Buhayenko and van Eikenhorst (2016):

T 
T 
I
minimise h0 H0t + s0 S0t + Bi (1)
H0 ,S0 ,Q 0 ,B,R
t=1 t=1 i=1


I 
T
subject to S0t diu − Q 0t ≥ 0 ∀t (2)
i=1 u=t


I 
T 
k
Q 0t + H0t−1 − H0t = Ritk diu ∀t (3)
i=1 k=t u=t
 6806 V. Buhayenko and D. den Hertog


T
Ri1k = 1 ∀i > 0 (4)
k=1


t−1 
T
Rikt−1 = Ritk ∀i > 0 , ∀t > 2 (5)
k=1 k=t


T 
T 
T 
T 
k 
T
si Ritk + h i Ritk diu (u − t) + pi diu − Bi ≤ ci ∀i > 0 (6)
t=1 k=t t=1 k=t u=t u=1

H0t ≥ 0, Q 0t ≥ 0, S0t , Ritk ∈ {0, 1} ∀i, t, k.

In the objective (1), the costs for the supplier are minimised. These costs consist of inventory holding and order costs as
well as the compensation, which models discounts here. Constraint (2) ensures that if an order is not placed by the supplier in
period t, then his replenishment quantity in period t is 0. Inventory balance constraint (3) guarantees that the demand of the
Downloaded by [EP- IPSWICH] at 00:55 21 November 2017

customers in period t is satisfied either from the inventory in period t − 1, or from the replenishment in period t; unsold items
from the inventory for period t. For the customers, the problem is formulated using the shortest path reformulation developed
by Eppen and Martin (1987). Constraint (4) ensures that, since it is assumed that there is no initial inventory, the demand for
period 1 has to be ordered in period 1 and can be ordered up to any future period k. Constraint (5) guarantees that when the
demand is ordered up to the previous period, then a new order must be placed in this period up to some future period k, thus
the demand for all periods will be satisfied. Constraint (6) ensures that the total costs of customer i after the introduction of
the compensations are less or equal to the total costs of the customers before the introduction of compensations. As a result,
the compensation Bi is equal to the difference between them at the optimum. Variables H0t and Q 0t are non-negative, and
variables S0t and Ritk are binary.

2.3 Calculation of parameter ci

With the help of the model explained in Section 2.2, the new beneficial order pattern is determined. However, parameter
ci has to be calculated in advance. This can be done using the algorithm described by Buhayenko and van Eikenhorst
(2016). Unfortunately, the algorithm cannot be applied together with Robust Optimisation techniques. Therefore, a mixed-
integer programming formulation for the dynamic lot sizing problem is used in this paper Wagner and Whitin (1958). The
mixed-integer programming model to calculate parameter ci for every customer i is as follows:

minimize ci
ci ,Hi ,Si ,Q i


T 
T 
T
subject to hi Hit + si Sit + pi dit ≤ ci (7)
t=1 t=1 t=1

M Sit − Q it ≥ 0 ∀t (8)

Q it + Hit−1 = Hit + dit ∀t (9)

Hit ≥ 0, Q it ≥ 0, Sit ∈ {0, 1} ∀t.

Parameter ci represents the total costs of the customer i (inventory, order, and purchasing costs) which are minimised
(constraint (7)). Constraint (8) is a “big M” constraint linking variables Sit and Q it . Constraint (9) is an inventory balance
constraint. Nonnegativity constraints for variables Hit and Q it are in place.
 International Journal of Production Research 6807

2.4 Calculation of the actual discount for every period to achieve the desired replenishment pattern
The goal here is to calculate the discounts offered by the supplier in such a way that his profit loss due to introduction of
discounts is minimised. On the other hand, the discounts should be big enough to make the desired order pattern advantageous
for the customer, which means that his total costs while ordering according to this new desired pattern should be lower than
his total costs while utilising any other order schedule. To achieve this, the model ensures that the cost of ordering up to
a certain period is minimal when the order is placed in the desired order period. Here, we propose a reformulation of the
Wagner–Whitin problem (Wagner and Whitin 1958) into a linear model that calculates the actual discounts.
After application of the model described in Section 2.2, the new desired order pattern for the supplier and customers is
received. The periods in which the order is placed, t ∗ , and the last periods for which the demand is included in this order, k ∗ ,
according to this new desired order pattern form a two-dimensional set W consisting of pairs (t ∗ , k ∗ ). Prices in the periods
we want the customer to place an order, denoted as pit ∗ , are variables, the rest of the prices are fixed equal to the initial price
pi . Thus, the desired order pattern is fixed and the prices in order periods are flexible.
Here, parameter h i is the inventory costs percentage. Additionally, the following variables are used in the model:
Citk the total costs of placing an order at period t to cover the demand from period t to period k together with the
lowest costs up to period t − 1 for customer i;
L ik the lowest total costs of placing an order from any period to period k for customer i.
Downloaded by [EP- IPSWICH] at 00:55 21 November 2017

The linear programming model that the supplier has to solve for customer i is as follows:

T
maximise pit (10)
pi ,Ci ,L i
t=1

subject to Cit ∗ k ∗ ≤ L ik ∗ ∀(t ∗ , k ∗ ) ∈ W (11)


k 
k
Citk = L it−1 + si + diu pit + diu (u − t)h i pit ∀t, k (12)
u=t u=t

L ik ≤ Citk ∀t, k (13)

0 ≤ pit ≤ pi ∀t (14)

pit = pi ∀t = t ∗ . (15)
In objective (10), prices in periods where the orders are placed according to the new desired order pattern are maximised.
Constraint (11) ensures that the costs of ordering in the new desired period t ∗ up to period k ∗ are not greater than the costs of
ordering the demand of period k ∗ in any other period. In constraint (12), the costs of placing an order from period t to period
k, Citk , are calculated as a summation of the following elements:
• the minimal costs to order up to period k − 1,
• order costs, since customer i places an order in period t,
• the total purchasing costs to order all the demand from period t to period k, and the total inventory costs to hold the
demand quantities in storage u − t periods.
Constraint (13) makes sure that L ik is lower than or equal to any of costs Citk . Constraint (14) guarantees that the new
prices are not higher than the initial prices. With the help of constraint (15), the rest of the prices are fixed equal to the initial
price pi .
Due to the fact that the supplier has to decide on the discount schedule in advance and announce it to the customers at
the beginning of the planning horizon, future adjustments of discounts are not possible. Therefore, the application of the
nominal model in the rolling horizon way will generally yield poor solutions.

3. Affinely Adjustable Robust model

In this section, we develop the Affinely Adjustable Robust model for the problem under consideration.
 6808 V. Buhayenko and D. den Hertog

3.1 Demand uncertainty

In reality, demand dit is often not deterministic, therefore, in this paper, we consider demand uncertainty. The demand belongs
to an uncertainty set which is created based on historical data and experts’ opinions. In alignment with RO principles, we
ensure that the constraints hold for all the demand realisations from this uncertainty set.
Here, we consider two uncertainty set structures: interval or box uncertainty and budget uncertainty. The interval
uncertainty set can be written as Ui = {di | dit ∈ [d¯it − εit d¯it , d¯it + εit d¯it ]}, with uncertain parameter dit , given positive εit
and positive nominal demand d¯it . The nominal demand d¯it can be determined as a point estimate of the expected value of
dit , or a forecast value.
In case of the interval uncertainty, all the parameters are taking their worst case values, which is unlikely in the reality.
This is the reason why this approach often results in a huge increase in total costs, which is undesirable. In order to decrease
the conservatism of the developed robust model, an alternative uncertainty set is used. It was decided to apply a budget
polyhedral uncertainty set introduced by Bertsimas and Sim (2004), which results in a linear robust counterpart. Another
possibility would be application of an ellipsoid uncertainty set, introduced by Ben-Tal and Nemirovski (1998). In this case,
the resulting robust counterpart is a conic quadratic program that might not work well together with integer variables.
For the case of budget uncertainty, we assume that the demand in each period dit can take a value from interval [d̄it −
εit d̄it ; d̄it + εit d̄it ], where d̄it is a nominal value and εit is a maximum deviation of the demand. To control the degree
of robustness, the budget of uncertainty i is defined for every customer i, which allows only the uncertain parameters
Downloaded by [EP- IPSWICH] at 00:55 21 November 2017

for some of the time periods simultaneously to deviate from their nominal value. Providing that the maximum possible
number of the uncertain parameters dit for every customer i is bounded by T , the number of time periods t in our planning
horizon, i , can take values in the interval [0, T ]. In case of i = 0, the model is deterministic. If i = T , the maximal
protection against uncertainty is in place, which is equal to interval uncertainty. The final uncertainty set is Ui = {di | dit ∈
 |dit −d̄it |
[d̄it − εit d̄it ; d̄it + εit d̄it ], ∀t; Let dit
= dεit −d̄d̄it , then the uncertainty set can be normalised in the
t∈T εit d̄it ≤ i }.  it it

following way: Ui = {di | dit ∈ [−1 ≤ dit ≤ 1], ∀t;

t∈T |dit | ≤ i }.

Alternatively, the budget uncertainty set can be defined as Ut = {dit ∈ [d̄it − εit d̄it ; d̄it + εit d̄it ]∀t, i∈I |dεit −d̄d̄it | ≤ t }.
it it
Here, the budget t is defined for every period t and determines for how many customers i the uncertain coefficients will
take their worst case values. This could only be relevant for the model described in Section 2.2, where all the customers are
considered simultaneously.

3.2 Determination of the replenishment patterns in a robust way

The following steps are performed in order to obtain robust counterparts and make them computationally tractable:
Step 1 Equality constraint elimination. This is performed only for those constraints that contain uncertain parameters,
because equality constraints with uncertainty are very restrictive and difficult to satisfy (Gorissen, Yanıkoğlu, and
den Hertog 2015).
Step 2 Introduction of linear decision rules. Since decisions about some of the variables can be postponed until the demand
in past periods will be known, the principles of Adjustable Robust Optimisation are applied here.
Step 3 Derivation of computationally tractable robust counterparts. As known, for fixed recourse situations, the robust
counterpart can be reformulated as a linear programming model (Ben-Tal, El-Ghaoui, and Nemirovski 2009). At this
step, we get rid of the universal quantifier, ensuring that every constraint holds for any scenario belonging to the
uncertainty set.
The steps listed above are realised for all the three parts of the model. For the sake of space, here, we restrict ourselves
to demonstrating the reformulations on the model from Section 2.2 and further on one of the constraints of this model.

Step 1
Initially,
n theuncertain parameter dit is present in constraints (2), (3) and (6). Constraint (2) is a so-called ‘big M’ constraint,
m
and i=1 u=t diu is an upper bound for the value of Q 0t . In a robust model, the value of the big M should be increased, but
this does not have to be done in a robust way, as long as it remains a valid upper bound to Q 0t . Constraint (3) is an equality
constraint which can be reformulated into the following inventory balance constraint (the initial inventory is assumed to be
equal to 0):
t I  t  T  k
H0t = Q 0v − Rivk diu ∀t.
v=1 i=1 v=1 k=v u=v
 International Journal of Production Research 6809

Therefore, H0t can be eliminated in the robust model.

The fact that, at the optimum, the compensation Bi is equal to the difference between the total costs of customer i before
and after the introduction of discounts allows us to rewrite the inequality (6) as follows:

T 
T 
T 
T 
k 
T
Bi = si Ritk + h i Ritk diu (u − t) + pi diu − ci ∀i > 0.
t=1 k=t t=1 k=t u=t u=1

Therefore, Bi can be eliminated as well.

Now, we concentrate only on objective function (1), for which the epigraphic form is used. After elimination is performed,
this constraint becomes as follows:

T  t 
t 
I 
T 
k 
T
h0 [ Q 0v − Rivk diu ] + s0 S0t
t=1 v=1 v=1 i=1 k=v u=v t=1

I 
T 
T T T 
k 
T
+ [si Ritk + h i Ritk diu (u − t) + pi diu − ci ] ≤ τ. (16)
i=1 t=1 k=t t=1 k=t u=t u=1
Downloaded by [EP- IPSWICH] at 00:55 21 November 2017

Step 2
In order to decrease the conservatism, the Adjustable Robust Optimisation approach is adopted in this research. Binary
variables S0t and Ritk are ‘here and now’ decisions, and continuous variables Bi , Q 0t and H0t are ‘wait and see’ or adjustable
variables. This division is made due to practical considerations. The periods in which the supplier replenishes and the
customers place an order up to a certain period are determined in advance and the decisions are fixed. The decisions about
how many items to order exactly for period t can be postponed until the beginning of this period, when the demand for periods
1 to t is realised. H0t and Bi are analysis variables and are already eliminated which is equivalent to the introduction of linear
decision rules, if the coefficients of the eliminated variables do not contain uncertain parameters, and the equality constraint
is linear in the uncertain parameters (Gorissen, Yanıkoğlu, and den Hertog 2015), which is the case here. Furthermore,
elimination does not lead to an increase in the number of the constraints and variables and is, therefore, a better choice than
the introduction of linear decision rules.
The following linear decision rule is used for the adjustable variables Q 0t :

I 
t
Q 0t = α0t + βitu diu .
i=1 u=1

It was decided to confine to linear decision rules due to computational intractability of adjustable robust optimisation
problems involving general decision rules. For many practical cases, linear decision rules were proven to produce optimal
or near optimal solutions (Bertsimas, Iancu, and Parrilo 2010). Decision variable Q 0t is now dependent on the realisations
of all the customers’ demand up to period t inclusively. Coefficients α0t and βitu are our new non-adjustable variables.
Constraint (16) transforms into:

T t  v
I  
t 
I 
T 
k 
T
h0 [ [α0v + βivu diu ] − Rivk diu ] + s0 S0t
t=1 v=1 i=1 u=1 v=1 i=1 k=v u=v t=1

I T  T T 
T 
k 
T
+ [si Ritk + h i Ritk diu (u − t) + pi diu − ci ] ≤ τ.
i=1 t=1 k=t t=1 k=t u=t u=1

To obtain a computationally tractable robust counterpart, all the coefficients for uncertain parameter dit are collected:

T 
t 
I 
T 
k 
I 
T 
u 
T
h0 Rivk diu = diu h 0 (T + 1 − v)Rivk
t=1 v=1 i=1 k=v u=v i=1 u=1 v=1 k=u


I 
T 
T 
k 
I 
T 
u 
T
hi Ritk diu (u − t) = diu h i (u − t)Ritk
i=1 t=1 k=t u=t i=1 u=1 t=1 k=u

T  t  v
I  
I 
T  T
βivu diu = diu (T + 1 − v)βivu .
t=1 v=1 i=1 u=1 i=1 u=1 v=u
 6810 V. Buhayenko and D. den Hertog

The adjustable robust model for the model in Section 2.2 is:

minimise τ
subject to


T 
t 
T 
I 
T 
T 
I
h0 α0v + s0 S0t + si Ritk − ci
t=1 v=1 t=1 i=1 t=1 k=t i=1


I 
T  T 
u T
+ diu h 0 (T + 1 − v)βivu − h 0 (T + 1 − v)Rivk
i=1 u=1 v=u v=1 k=u


u 
T
+ hi (u − t)Ritk + pi ≤ τ, ∀di ∈ Ui
t=1 k=u


I 
t
M S0t − α0t − βitu diu ≥ 0 ∀t, di ∈ Ui
i=1 u=1
Downloaded by [EP- IPSWICH] at 00:55 21 November 2017


T
Ri1k = 1 ∀i > 0
k=1

t−1 
T
Rikt−1 = Ritk ∀i > 0 , ∀t > 2
k=1 k=t
⎛ ⎞

t 
I 
T 
min (u,t) 
T
α0v + diu ⎝βiu

− Rivk ⎠ ≥ 0 ∀t, di ∈ Ui
v=1 i=1 u=1 v=1 k=u

I 
t
α0t + βitu diu ≥ 0 ∀t, di ∈ Ui (17)
i=1 u=1
S0t ∈ {0, 1}, Ritk ∈ {0, 1} ∀i, t, k
⎧
⎨ t

β u≤t
where βiu = v=u ivu
⎩
0 otherwise.

Step 3
The universal quantifier, which is present in the adjustable robust model above, is not computationally tractable. To get rid
of this, we perform the following reformulations leading to a linear programming problem.
Constraint (17) is equivalent to:
I  t
α0t + min βitu diu ≥ 0 ∀t.
di ∈Ui
i=1 u=1
Since diu =

εiu d̄iu diu + d̄iu , we have:


I 
t

α0t + min

βitu (εiu d̄iu diu + d̄iu ) ≥ 0 ∀t

di ∈Ui
i=1 u=1


I 
t 
I 
t

α0t + βitu d̄iu + min

(βitu εiu d̄iu )diu ≥0 ∀t

di ∈Ui
i=1 u=1 i=1 u=1


I 
t 
I 
t

α0t + βitu d̄iu − max −(βitu εiu d̄iu )diu ≥0 ∀t.

di
∈Ui

i=1 u=1 i=1 u=1

 International Journal of Production Research 6811

It can be noticed that the budget uncertainty set is an intersection of two simple polyhedral uncertainty sets. Moreover,
the resulting maximisation problem is linear in di
. Therefore, the dual can be taken as described in Ben-Tal et al. (2005):
⎧  I t I
⎨ α0t + i=1 u=1 βitu d̄iu − i=1 [v 1 + i v ∞ ] ≥ 0
it1 it2 ∀t
vu + vu = −βitu εiu d̄iu
it1 it2 i > 0, ∀t, u = 1, . . . , t
⎩ it1
vu + vuit2 = 0 i > 0, ∀t, u = t + 1, . . . T.
 ⎧
 I t  I t I
⎪
⎪ α0t + i=1 u=1 βitu d̄iu − u=1 witu − i=1 i z it ≥ 0 ∀t
⎨ it1 i=1
vu + vuit2 = −βitu εiu d̄iu i > 0, ∀t, u = 1, . . . , t
⎪
⎪ vu + v u = 0
it1 it2 i > 0, ∀t, u = t + 1, . . . T
⎩
witu ≥ vuit1 , witu ≥ −vuit1 , z it ≥ vuit2 i > 0, ∀t, u = 1, . . . , t.
The final robust counterpart for the whole model is presented in Appendix A.1.
In this paper, we restrict ourselves to presenting reformulations only for the budget uncertainty set. Derivations for the
interval uncertainty set are straightforward. Moreover, the budget uncertainty set with  = T guarantees full protection
against uncertainty and is equal to the interval uncertainty set.
A similar procedure is performed for the two other parts of the model presented in Sections 2.3 and 2.4. This is done in
Sections 3.3 and 3.4, respectively.
Downloaded by [EP- IPSWICH] at 00:55 21 November 2017

3.3 Calculation of parameter ci in a robust way

For the model in Section 2.3, equality constraint (9) is eliminated and variables Q it are made adjustable with the help of the
following linear decision rule:
t
Q it = αit + βitu diu . (18)
u=1
In the end, the following adjustable robust model is obtained:
minimise ci
subject to

T 
t 
T 
T 
T
hi αiv + si Sit + [ pi − h i (T + 1 − u) + h i (T + 1 − v)βivu ]diu ≤ ci , ∀di ∈ Ui
t=1 v=1 t=1 u=1 v=u


t
M Sit − αit − βitu diu ≥ 0 ∀t, di ∈ Ui
u=1

t 
t t
αiv + diu ( βivu − 1) ≥ 0 ∀t, di ∈ Ui
v=1 u=1 v=u


t
αit + βitu diu ≥ 0 ∀t, di ∈ Ui
u=1
Sit ∈ {0, 1} ∀t.
The pre-calculation of parameter ci can be done in two ways. The parameter is completely pre-calculated and fixed, which
corresponds to a situation in practice, when the supplier does not possess the necessary information about the customers
and cannot determine their total costs before introduction of discounts. Therefore, he asks the customers to determine this
parameter and report it to him. According to the second approach, which is used in this research, the supplier is in possession
of the full information about the customers and instead of calculating the parameter in advance, he determines a linear
decision rule which depends on the uncertain parameter dit of the form:
 

T  t 
T T 
T
ci = h i αiv + si Sit + pi − h i (T + 1 − u) + h i (T + 1 − v)βivu diu , diu ∈ Uit ,
t=1 v=1 t=1 u=1 v=u

where h i , αiv , si , Sit , pi , and βivu are fixed. This decision rule is used instead of parameter ci in the model of Section 2.2.
The final robust counterpart for this model with budget uncertainty can be seen in Appendix A.2.
 6812 V. Buhayenko and D. den Hertog

3.4 Calculation of the actual discount in a robust way

For the model in Section 2.4, equality constraint (12) is eliminated and variable L ik is made adjustable. As L ik is an analysis
variable, the linear decision rule for this variable depends on all the realisations of the uncertain demand:


T
L ik = αik + βiku diu .
u=1

The following adjustable robust model is received in the end:


maximise pit ∗
t ∗ ∈W

subject to


T

αit ∗ −1 + si + diu (βit ∗ −1u + n iu [ pit ∗ + (u − t ∗ )h i pit ∗ ] − βik ∗ u ) ≤ αik ∗ ∀(t ∗ , k ∗ ) ∈ W, dit ∈ Uit
u=1
Downloaded by [EP- IPSWICH] at 00:55 21 November 2017


T

αik ≤ αit−1 + si + diu (βit−1u + n iu [ pit + (u − t)h i pit ] − βiku ) ∀t, k, dit ∈ Uit
u=1

pit ∗ ≤ pi ∀t ∗

pit = pi ∀t = t ∗



1 t ≤u≤k
where n iu =
0 otherwise.
The final robust counterpart for this model, when budget uncertainty is assumed, is given in A.3.

4. Computational results
In this section, the numerical results for the approach are presented. All the computations are performed on Intel i7 2.4 GHz
dual-core with 8 GB of RAM. The model is optimised using solver IBM ILOG CPLEX version 12.51 within the modelling
language AMPL. The parameters for the test problems are generated in the same way as in Buhayenko and van Eikenhorst
(2016). Here, we provide a short summary of the formulas used.
Demand for customer i for period t is generated using the following formula to differentiate between the customers with
high and low average demands:

U (5, 150) t =1
dit = ∀i > 0.
max(N (di1 , 0.25di1 ), 0) t >1
Replenishment costs for the supplier and customers are generated using the following formula:

5000n i =0
si =
U (2500, 5000) i > 0.
The inventory holding costs for the supplier and customers are generated using the following formula:

h i = U (0.05, 0.1) ∀i ≥ 0.

The initial inventory is assumed to be zero for all the parties. The initial price, before discounts are introduced, is 100 for
every customer. The price of the items for the supplier, which is used to calculate supplier’s inventory holding costs, is
supposed to be equal to 70.
 International Journal of Production Research 6813

4.1 Simulation Results on a set of 100 problems with 5 customers and 20 periods
First, we test the AAR model on a set of 100 experiment problems with 5 customers and 20 periods. Several uncertainty
levels are considered – 1, 2, 4, 8, 16 and 32%, which means that ε = 0.01, 0.02, 0.04, etc. The uncertainty rate of 32% is
considered, as demand could be prone to significant fluctuations. The quality of the solutions received applying the developed
Affinely Adjustable Robust model is assessed using simulation.
Two sets of 100 demand replications are generated for every customer according to a Uniform distribution in the interval
[d̄it − εit d̄it ; d̄it + εit d̄it ] with mean equal to the nominal value. They represent the realised demand. Each demand replication
has 20 entries to account for a 20 periods time horizon. One set corresponds to a simple box uncertainty, and the other
set represents budget uncertainty. The box uncertainty set is chosen for its simplicity of representation. The downside
of this uncertainty structure is conservatism. To receive less conservative solutions, the budget uncertainty set is used.
Both uncertainty sets result in tractable linear programs. The generation of sets corresponding to the box uncertainty is
straightforward, as the demand in all the periods can take its worst case values at the same time. However, it is more
challenging to generate the budget uncertainty sets, especially when i is small, since only a certain number of parameters
can simultaneously deviate from their nominal values. Therefore, it is necessary to check the deviation from the nominal
 |dit −d̄it |
value in a demand replication set: t∈T εit d̄it , and if this value exceeds i , another demand set should be generated.
For both the box uncertainty and the budget uncertainty simulated demand replications, three models are solved: a perfect
hindsight model, the nominal model given in Section 2, and the Affinely Adjustable Robust model presented in Section 3.
Downloaded by [EP- IPSWICH] at 00:55 21 November 2017

The perfect hindsight costs correspond to the costs which would have been obtained if the realised demand was known in
advance. The perfect hindsight solution is obtained by solving the nominal model given in Section 2 with the simulated
demand to get the lowest possible costs. This is by all means a value that cannot be received in reality and represents a
lower bound on the costs. The perfect hindsight costs are used as a benchmark to assess the increase or decrease in costs
for the nominal and robust solutions. The nominal solution is, most often, not feasible in the case of uncertain demand,
especially if the demand is higher. The plan has to change to stay feasible and satisfy all the demand. The assumption is that
the procurement quantity can always increase to fulfill the demand. So, if the demand is higher than forecasted in a period
where a production order was already planned, the procurement quantity in this period can always be increased without a
trouble. If, however, we face a demand in a period where no production was planned, and this demand cannot be fulfilled
out of stock, an extra procurement order has to be planned to acquire the needed quantity.
The price of uncertainty (PoU), the mean of difference between the perfect hindsight and robust total costs, which was
first defined by Ben-Tal et al. (2005), is calculated. The PoU can be viewed as the maximum amount of investments to reduce
the uncertainty, which is still acceptable. Here, the PoU is given in percentage increase in total costs compared to the perfect
hindsight solution. The difference between the total costs in the robust solution, compared to the nominal solution, gives an
indication on how much cost can be saved by taking uncertainty into account when planning procurements and developing
a discount schedule.
Table 1 compares robust and nominal results in case of box uncertainty and presents the percentage increase or decrease
in costs of nominal and robust solutions compared to perfect hindsight solutions. For the box uncertainty set, inventory
holding and order costs as well as discount costs are presented separately. The discount costs are obtained by multiplying
the discount given with the amount of items sold with this discount. This is done to account for the total revenue loss as a
result of introduction of the discounts.
It can be noticed that in the nominal solution, total costs grow immediately by more than 45%, when uncertainty of only
1% is in place. At the same time, further increase in uncertainty results only in 5% extra increase in costs. This happens
because the extra replenishment orders are needed and, as a result, the order costs increase dramatically. This increase is
independent of the actual quantities of an extra replenishment. Therefore, the scale of uncertainty does not play a role here,
since a higher uncertainty level mainly increases the quantities of an extra replenishment, but not the number of times it is
needed. Discount costs decrease slightly. This can be explained by the fact that, due to changes in their demand, the customers
do not perceive the discount offered by the supplier as an attractive offer any more and instead place their order in the period
with no discount. Although the revenue increases by about 24% on average, even in case of the smallest uncertainty, the
discounts are not an efficient method to coordinate the supply chain if they are determined using the nominal approach. As
a result, the orders are placed in periods without replenishment, requiring the need for inventory (if the order now comes
in later than was expected) or extra replenishments (if the order comes in earlier than expected). The nominal discounts are
very sensitive to uncertainty, because the method for calculating the discounts in the nominal solution sets them to be just
high enough to make it slightly cheaper to order in the period according to the nominal order plan.
On the contrary, the robust optimisation solution is created taking uncertainty into account. As a result, this solution
provides a possibility for supply chain coordination, even if there is a significant demand uncertainty. The robust solution
results in higher discount costs, as in uncertain environment, the supplier has to offer more discount to change the ordering
 6814 V. Buhayenko and D. den Hertog

Table 1. Average percentage differences (over 100 replications) in supplier’s costs between Robust and Nominal results for box uncertainty
(% to the perfect hindsight result).

Order costs Inventory costs Discount costs Total costs Total costs
Uncertainty Average Average Average Average Worst case

Nominal
0.01 40.33 267.40 −23.56 45.11 86.13
0.02 41.55 236.42 −21.09 44.12 89.91
0.04 42.38 258.61 −22.75 46.45 108.91
0.08 38.78 267.92 −23.04 44.54 96.84
0.16 42.19 319.25 −23.84 49.13 83.64
0.32 42.27 344.59 −28.60 50.19 113.94
Robust
0.01 0.00 2.30 1.50 0.58 2.58
0.02 0.28 1.98 3.01 1.11 8.89
0.04 0.28 5.53 6.41 2.27 10.32
0.08 0.28 14.65 11.66 4.41 9.48
0.16 −0.82 43.15 23.02 8.80 17.70
Downloaded by [EP- IPSWICH] at 00:55 21 November 2017

0.32 0.00 73.80 41.70 17.05 41.47

Table 2. Average percentage differences (over 100 replications) in total costs for customers between Robust and Nominal results for box
uncertainty (% to the perfect hindsight result).

Nominal Robust
Uncertainty Average Worst case Average Worst case

0.01 −1.67 0.29 −3.28 −1.14

0.02 −1.58 0.12 −3.34 −1.15
0.04 −1.62 1.21 −3.41 −1.19
0.08 −1.73 1.75 −3.53 −1.29
0.16 −1.23 3.38 −3.83 −1.50
0.32 −1.16 8.42 −4.07 1.11

pattern of his customers. For low uncertainty levels, this increase is insignificant, as a slightly higher discount can already
make it cheaper to order in the period according to the new replenishment plan for all possible demand realisations. Order
costs remain close to the values of the perfect hindsight solution for all uncertainty levels, since the new replenishment
plan remains feasible, and extra replenishments do not occur. Inventory holding costs do not increase much for moderate
uncertainty levels. In cases of high uncertainty level, the increase in inventory holding costs in the robust solution is less
than a quarter of the nominal solution. As a result, the increase in total costs is always significantly smaller for the robust
solution. While in the worst case, the increase in total costs for the nominal solution is close to 100% for many uncertainty
levels or in some cases even greater, for the robust solution, this measure is significantly smaller.
We conducted 200 additional experiments for 32% uncertainty for both the nominal and robust solutions to make sure that
the reported averages are reliable and we do not need to have more computations. The error for total costs, which accounts
here for the difference between the average total costs received when 100 experiments were conducted and those received
after 300 experiments were conducted, did not exceed 0.3% for the robust solution and 3.5% for the nominal solution.
Table 2 shows how the customers’ total costs are affected. The conclusion is that in the robust solution, the total costs
decrease more than in the nominal solution, both on average and in the worst case. Only in case of 32% uncertainty, the
customers’ total costs may increase in the robust solution in the worst case. However, the observed changes are much less
significant than for the supplier. It is logical that the robust solution is better for the customers than the perfect hindsight
solution or the nominal solution, since they almost always get a higher discount. The nominal solution normally results in
lower costs for the customer on average than the perfect hindsight because the discounts offered are never optimal for the
realised demand, they are either too high, and then the customers easily profit accepting them; or too low and the customers
ignore them choosing a different order pattern.
 International Journal of Production Research 6815

Table 3. Average percentage differences (over 100 replications) in total costs between Robust and Nominal results for budget uncertainty
(% to the perfect hindsight result).

Uncertainty =2 =4 =8  = 12  = 16  = 20

Nominal
0.01 38.74 44.96 46.49 45.86 45.28 45.11
0.02 42.49 47.61 45.86 44.23 44.10 44.12
0.04 45.09 48.00 45.23 44.26 46.03 46.45
0.08 46.98 46.08 48.87 45.34 45.86 44.54
0.16 46.63 47.37 47.12 49.68 50.69 49.13
0.32 47.74 49.67 48.99 51.95 52.40 50.19
Robust
0.01 0.33 0.39 0.47 0.48 0.58 0.58
0.02 0.67 0.85 0.96 1.11 1.11 1.11
0.04 1.28 1.75 2.07 2.18 2.25 2.27
0.08 2.65 3.59 4.15 4.24 4.29 4.41
0.16 5.37 7.27 8.00 8.31 8.31 8.80
0.32 10.52 14.06 15.76 16.70 16.90 17.05
Downloaded by [EP- IPSWICH] at 00:55 21 November 2017

Table 4. Worst case percentage differences (over 100 replications) in total costs between Robust and Nominal results for budget uncertainty
(% to the perfect hindsight result).

Uncertainty =2 =4 =8  = 12  = 16  = 20

Nominal
0.01 82.62 96.09 88.48 86.62 86.48 86.13
0.02 81.93 78.59 83.52 87.99 94.22 89.91
0.04 80.74 89.93 82.77 90.88 93.83 108.91
0.08 95.83 96.18 106.74 103.83 96.33 96.84
0.16 94.22 97.96 96.77 101.45 97.05 83.64
0.32 94.68 111.87 100.70 94.10 105.06 113.94
Robust
0.01 0.58 0.77 2.47 2.50 1.48 2.58
0.02 1.19 2.86 1.87 4.73 4.77 8.89
0.04 3.37 3.86 6.35 6.36 5.15 10.32
0.08 4.81 11.24 8.69 8.87 11.19 9.48
0.16 10.90 16.34 16.00 17.94 18.45 17.70
0.32 19.41 26.03 28.94 34.34 34.65 41.47

Tables 3 and 4 compare the average and worst case performance of nominal and robust solutions, if budget uncertainty
is assumed. These results indicate the trade-off between increase in total costs and decrease in the level of protection against
uncertainty.  = T means maximal uncertainty protection, which is equal to box uncertainty. While decreasing , the level
of protection decreases. At the same time, for small values of , the increase in the average total costs for robust solution is
less significant. With 32% of uncertainty and  = 4, only 14% increase in total costs on average is experienced compared
to 17% when  = T and full protection is in place. With  = 2, the increase in total costs is even smaller on average, but
at the same time, the level of protection against uncertainty is low. On the contrary, for smaller values of uncertainty, this
difference is less significant, and full protection against uncertainty can be preferable. It can be seen as well that the nominal
model again performs much worse than the robust model, both on average and in the worst case.
Table 5 shows the decrease in inventory holding and discount costs for different levels of budget uncertainty compared to
box uncertainty. As it was expected, the decrease is higher for lower values of . The order costs do not change substantially
and are therefore not included in the analysis.
De Ruiter, Brekelmans, and den Hertog (2016) warn that multiple solutions, that have an equivalent worst case objective
value, but differ on the mean objective value, may exist. We checked for such a possibility, but it turned out that this problem
has only one optimal solution.
 6816 V. Buhayenko and D. den Hertog

Table 5. Average percentage decrease (over 100 replications) in costs for the Robust solution with budget uncertainty (% to the robust
results with box uncertainty).

Uncertainty =2 =4 =8  = 12  = 16

Inventory costs
0.01 −0.97 −0.66 −0.30 −0.18 −0.08
0.02 −1.81 −1.22 −0.53 −0.24 −0.03
0.04 −3.46 −2.25 −0.98 −0.34 −0.09
0.08 −6.46 −4.33 −1.78 −0.68 0.11
0.16 −11.67 −7.93 −3.50 −2.12 −1.18
0.32 −18.94 −12.83 −6.04 −3.86 −1.30
Discount costs
0.01 −0.59 −0.26 −0.13 −0.03 −0.01
0.02 −1.15 −0.49 −0.26 −0.07 −0.02
0.04 −2.23 −0.96 −0.51 −0.13 −0.04
0.08 −4.19 −1.80 −0.95 −0.24 −0.08
0.16 −7.50 −3.23 −1.72 −0.43 −0.14
0.32 −12.79 −5.65 −3.03 −0.76 −0.24
Downloaded by [EP- IPSWICH] at 00:55 21 November 2017

Table 6. Percentage decrease in prices for the Robust solution with box uncertainty (% to the Nominal solution).

Period
Customer 1 3 5 8 10 11 13 15 16 17 19

1 −3.72 – – – 0 – – −0.66 – – –
2 −5.50 – −2.94 −4.53 −4.34 −2.96 −1.44 −2.14 – −3.46 −3.58
3 −2.07 −3.00 −2.43 – −2.39 – −1.89 – – – –
4 −2.95 – – – −3.67 – – – – – –
5 −8.96 – – – −8.80 – – – – – –

4.2 Comparison of solution structures

Here, a numerical example is given in order to visualise and compare solution structures of robust and nominal solutions, i.e.
replenishment patterns for the supplier and customers, as well as the discount schedule. In Figure 2, changes in the solution
structure are shown on an example which was randomly taken from the 100 scenarios considered in Section 4.1. The problem
has 5 customers and 20 periods. The replenishment pattern of the supplier can be seen in row 0; the rest of the rows represent
the order patterns of the customers. For the nominal solution, the demand is supposed to be deterministic. For the robust
solution, the box uncertainty set with 32% of uncertainty is supposed.
On average, the replenishment patterns of the robust solution do not differ substantially from those of the nominal solution.
In this example case, only the order pattern of customer 2 in the robust solution (Figure 2(c)) is slightly different from his
order pattern in the nominal solution (Figure 2(a)). At the same time, simulation shows that if 32% uncertainty is in place, the
nominal solution becomes infeasible. Due to changes in their demand, the customers do not perceive the supplier’s discounts
as an attractive offer any more and start placing orders in the periods with no discount. Although the actual discount costs for
the supplier decrease, this situation is not optimal for him, as he has higher inventory holding costs and has to introduce extra
replenishments to be able to satisfy his customers’ demand. The introduction of extra replenishments explains the significant
increase in order costs for the nominal solution. The final situation after simulation for the nominal solution is shown in
Figure 2(b). At the same time, the robust solution stays feasible, even when the demand is uncertain, due to higher discounts
which ensure that customers do always order in the anticipated periods. The decrease in prices due to higher discounts in
the robust solution compared to the nominal solution can be seen in Table 6. Only periods, where a discount was given to at
least one customer, are included. The average decrease in prices is 3.4%.
 International Journal of Production Research 6817
Downloaded by [EP- IPSWICH] at 00:55 21 November 2017

Figure 2. Difference between the Nominal and Robust replenishment patterns. a – Nominal solution. b – Nominal solution after a
simulation. c – Robust solution. The replenishment patterns do not change after a simulation.

4.3 Heuristic robust solution procedure for solving larger problems

The conducted computations indicate that the Affinely Adjustable Robust model developed for the problem produces good
solutions, but is not fast enough to be effectively applied for large instances. For comparison, the average computational time
for the nominal model for 5 customers and 20 periods is 1.68 s, while the robust model is almost 6 times slower.
Here, a heuristic robust procedure is offered for the application on larger instances. As it has been discovered that the
replenishment patterns of the robust solution do not differ on average significantly from those of the nominal solution, it
has been decided to determine them in the nominal way. As a result, only the discounts and order quantities are calculated
in the robust way. Thus, we use the nominal models from Sections 2.2 and 2.3 and the robust model from Section 3.4 here.
Moreover, to speed up the calculation process even more, we compress the information basis to avoid a large increase in
the number of variables because of the linear decision rule. Instead of having a linear decision rule that depends on all the
known demand values, we restrict the information basis to the known demand of the last periods from the period when the
previous supplier’s replenishment takes place until the current period. This helps to reduce the number of decision variables
significantly without negative consequences for the quality of the decision rule.
First, the heuristic procedure is tested on the same set of 100 instances with 5 customers and 20 periods to compare
its performance with the performance of the AAR model. The average increase in total costs for the supplier was 0.43%
on average for the highest uncertainty level of 32%. At the same time, the heuristic was almost five times faster than the
AAR model, and, thus, close to the computational time of the nominal model. Afterwards, the computational experiments
are conducted for a set of 100 instances with 20 customers and 30 replenishment periods and a set of 100 instances with
30 customers and 20 periods for all the uncertainty levels. The results for the nominal model and the heuristic procedure
are summarised in Table 7. It can be concluded that the developed heuristic results in robust solutions that outperform the
 6818 V. Buhayenko and D. den Hertog

Table 7. Average percentage differences (over 100 replications) in supplier’s total costs between Robust and Nominal results for box
uncertainty (% to the perfect hindsight result).

20 × 30 30 × 20
Uncertainty Average Worst case Average Worst case

Nominal
0.01 45.30 68.13 45.47 89.37
0.02 43.46 85.05 51.34 93.00
0.04 43.62 73.37 47.97 80.29
0.08 47.01 81.97 49.02 79.35
0.16 46.17 73.21 51.17 97.78
0.32 46.06 102.73 49.96 88.16
Robust
0.01 0.64 1.96 0.52 0.98
0.02 1.18 2.35 1.11 2.55
0.04 2.13 3.68 1.86 3.41
0.08 4.56 7.42 4.19 7.05
0.16 8.95 14.19 8.10 15.19
Downloaded by [EP- IPSWICH] at 00:55 21 November 2017

0.32 17.87 26.96 17.65 29.65

nominal solutions both on average and in the worst case in case of uncertain demands. Due to the low computational time
the heuristic can be successfully applied to large instances.

5. Conclusions and future research

In this paper, a problem of multi-period supply chain coordination with the help of a temporary time-based discount under
demand uncertainty is studied.
We develop an Affinely Adjustable Robust model to determine replenishment patterns and discount schedules to minimise
total costs of the supplier and customers when demand uncertainty is in place.
The model is tested on a set of 100 problems with 5 customers and 20 periods. As the computational time grows faster
with the increase in periods than in customers, it is possible to increase the number of customers significantly, shrinking the
time horizon to 15 periods. We have also developed a heuristic procedure that produces good solutions faster. The procedure
was tested on 2 sets of 100 problems with 20 customers and 30 periods, and 30 customers and 20 periods.
From the conducted computational experiments, it can be concluded that even for a relatively small uncertainty, the
nominal solution results in a total cost increase of more than 40% on average. Therefore, the development of an alternative
approach taking uncertainty into consideration at the modelling stage is well justified, also in cases when uncertainty does
not exceed 5%.
Computational experiments show that the Affinely Adjustable Robust model, developed in this research, produces a
solution with significantly lower total costs, both on average and in the worst case, as the increase in total costs is under
2.5% as compared to PH solution. The received solution is robust against demand uncertainty and remains feasible as the
uncertain data change.
The increase in total costs can be significantly reduced if the budget uncertainty set is considered. However, in this case,
the protection against uncertainty may decrease.
Our research can be beneficial in many situations, both when the supplier is a producer with high set-up costs (a producer
of specialised subcomponents for heavy industries) and a distribution centre with over-seas suppliers and high replenishment
costs. Moreover, as uncertainty is treated at the modelling stage, our approach is advantageous in situations when system
nervousness, i.e. constant change of the order timing by downstream players, is significant, and the costs of re-planning are
high.
Furthermore, our research provides managerial insights into the value of reliable demand information.
Thus, the contributions of this research are as follows:
• We show that the nominal approach results in a huge increase in total costs when in reality the demand has uncertain
nature.
 International Journal of Production Research 6819

• We are the first to develop an Affinely Adjustable Robust Optimisation model for a multi-period supply chain coor-
dination problem with multiple heterogeneous customers and uncertain dynamic demand by means of temporarily
time-based discounts.
• We conduct extensive computational experiments that show benefits of our approach, even in cases when the demand
uncertainty is low.
• In addition, we propose a robust heuristic procedure that can be applied for larger instances and is able to produce
a good solution faster.
For future research, some alternative representations of uncertainty sets can be considered. One possibility has been
mentioned in Section 3.1.
Instead of the nominal model, it can be interesting to take the nominal model with separate buffering techniques, i.e.
safety stock, safety time or inflating demand parameters, as a benchmark. In this case the benchmark can be more realistic
and probably harder to outperform.

Disclosure statement
No potential conflict of interest was reported by the authors.
Downloaded by [EP- IPSWICH] at 00:55 21 November 2017

References

Adida, E., and G. Perakis. 2006. “A Robust Optimization Approach to Dynamic Pricing and Inventory Control with no Backorders.”
Mathematical Programming 107: 97–129.
Ardjmand, E., G. R. Weckman, W. A. Young II, O. S. Bajgiran, and B. Aminipour. 2016. “A Robust Optimisation Model for Production
Planning and Pricing under Demand Uncertainty.” International Journal of Production Research 54 (13): 3885–3905.
Arshinder, K., A. Kanda, and S. G. Deshmukh. 2008. “Supply Chain Coordination: Perspectives, Emperical Studies and Research
Directions.” International Journal of Production Economics 115 (2): 316–335.
Arshinder, K., A. Kanda, and S. G. Deshmukh. 2011. A Review on Supply Chain Coordination: Coordination Mechanisms, Managing
Uncertainty and Research Directions, 39–82. Berlin, Heidelberg: Springer, Berlin Heidelberg.
Ashayeri, J., N. Ma, and R. Sotirov. 2014. “Supply Chain Downsizing under Bankruptcy: A Robust Optimization Approach.” International
Journal of Production Economics 154: 1–15.
Ben-Tal, A., B. Do Chung, S. R. Mandala, and T. Yao. 2011. “Robust Optimization for Emergency Logistics Planning: Risk Mitigation in
Humanitarian Relief Supply Chain.” Transportation Research Part B: Methodological 45: 1177–1189.
Ben-Tal, A., L. El-Ghaoui, and A. Nemirovski. 2009. Robust Optimization. Princeton, NY: Princeton Press.
Ben-Tal, A., B. Golany, A. Nemirovski, and J-Ph Vial. 2005. “Retailer-supplier Flexible Commitment Contracts: A Robust Optimization
Approach.” Manufacturing & Service Operations Management 7: 248–271.
Ben-Tal, A., B. Golany, and S. Shtern. 2009. “Robust Multi-echelon Multi-period Inventory Control.” European Journal of Operational
Research 199: 922–935.
Ben-Tal, A., A. Goryashko, E. Guslitzer, and A. Nemirovski. 2004. “Adjustable Robust Solutions of Uncertain Linear Programs.”
Mathematical Programming 99 (2): 351–376.
Ben-Tal, A., and A. Nemirovski. 1997. “Stable Truss Topology Design via Semidefinite Programming.” SIAM Journal on Optimization
7 (4): 991–1016.
Ben-Tal, A., and A. Nemirovski. 1998. “Robust Convex Optimization.” Mathematics of Operations Research 23 (4): 769–805.
Ben-Tal, A., and A. Nemirovski. 1999. “Robust Solutions of Uncertain Linear Programs.” OR Letters 25 (1): 1–13.
Bernstein, F., and G. A. DeCroix. 2015. “Advance Demand Information in a Multiproduct System.” Manufacturing & Service Operations
Management 17 (1): 52–65.
Bertsimas, D., D. Brown, and C. Caramanis. 2011. “Theory and Applications of Robust Optimization.” SIAM Review 53 (3): 464–501.
Bertsimas, D., D. A. Iancu, and P. A. Parrilo. 2010. “Optimality of Affine Policies in Multistage Robust Optimization.” Mathematics of
Operations Research 35: 363–394.
Bertsimas, D., and M. Sim. 2004. “The Price of Robustness.” Operations Research 52 (1): 35–53.
Bertsimas, D., and A. Thiele. 2006. “A Robust Optimization Approach to Supply Chain Management.” Operations Research 54: 150–168.
Birge, J. R., and F. Louveaux. 2011. Introduction to Stochastic Programming. Springer, NY: Springer Science & Business Media.
Bitran, G., and R. Caldeney. 2003. “Comissioned Paper: An Overview of Pricing Models for Revenue Management.” Manufacturing and
Service Operations Management 5: 203–229.
Buhayenko, V., and D. den Hertog, . 2016. “Adjustable Robust Optimization Approach to Optimize Discounts for a Multi-period Supply
Chain Problem under Demand Uncertainty.” 28th European Conference on Operational Research EURO 2016.
Buhayenko, V., and E. van Eikenhorst. 2016. “Synchronization of Customers’ Orders Using Discounts.” International Journal of Systems
Science: Operations & Logistics 3 (3): 148–162.
Chen, F., A. Federguen, and Y. Zheng. 2001. “Coordination Mechanisms for a Distribution System with One Supplier and Multiple
Customers.” Management Science 47: 693–708.
 6820 V. Buhayenko and D. den Hertog

Corbett, C. J., and X. de Groote. 2000. “A Supplier’s Optimal Quantity Discount Policy Under Asymmetric Information.” Management
Science 46 (3): 444–450.
De Ruiter, F. J. C. T., R. C. M. Brekelmans, and D. den Hertog. 2016. “The Impact of the Existence of Multiple Adjustable Robust
Solutions.” Mathematical Programming 160 (1): 531–545.
Du, R., A. Banerjee, and S. L. Kim. 2013. “Coordination of Two-echelon Supply Chains Using Wholesale Price Discount and Credit
Option.” International Journal of Production Economics 143: 327–334.
El-Ghaoui, L., and H. Lebret. 1997. “Robust Solution to Least-squares Problems with Uncertain Data.” SIAM Journal on Matrix Analysis
and Applications 18: 1035–1064.
El-Ghaoui, L., F. Oustry, and H. Lebret. 1998. “Robust Solutions to Uncertain Semidefinite Programs.” SIAM Journal on Optimization
9 (1): 33–52.
Elmaghraby, W., and P. Keskinocak. 2003. “Dynamic Pricing in the Presence of Inventory Considerations: Research Overwiev, Current
Practices, and Future Directions.” Management Science 49 (10): 1287–1309.
Eppen, G. D., and R. K. Martin. 1987. “Solving Multi-item Capacitated Lot-sizing Problems Using Variable Redefinition.” Operations
Research 35 (6): 832–848.
Fisher, M., J. Hammond, W. Obermeyer, and A. Raman. 1994. “Making Supply Meet Demand in an Uncertain World.” Harvard Business
Review(May-June):83–94.
Gabrel, V., C. Murat, and A. Thiele. 2014. “Recent Advances in Robust Optimization: An Overview.” European Journal of Operational
Research 235: 471–483.
Gorissen, B. L., I. Yanıkoğlu, and D. den Hertog. 2015. “A Practical Guide to Robust Optimization.” Omega 53: 124–137.
Downloaded by [EP- IPSWICH] at 00:55 21 November 2017

Higle, J. L., and K. G. Kempf. 2011. “Production Planning under Supply and Demand Uncertainty: A Stochastic Programming Approach.”
In Stochastic Programming, edited by G. Infanger, 297–314. New York, USA: Springer Science+Business Media.
Huang, X., S.-M. Choi, W.-K. Ching, Tak-Kuen Siu, and Min Huang. 2011. “On Supply Chain Coordination for False Failure Returns: A
Quantity Discount Contract Approach.” International Journal of Production Economics 133 (2): 634–644.
Iancu, D. A., M. Sharma, and M. Sviridenko. 2013. “Supermodularity and Affine Policies in Dynamic Robust Optimization.” Operations
Research 61: 941–956.
Jeuland, A., and S. M. Shugan. 1983. “Managing Channel Profits.” Marketing Science 2 (3): 239–272.
Lal, R., and R. Staelin. 1984. “An Approach for Developing an Optimal Quantity Discount Pricing Policy.” Management Science 30 (12):
1524–1539.
Lee, H. L., and M. J. Rosenblatt. 1986. “A Generalized Quantity Discount Pricing Model to Increase Supplier’s Profits.” Management
Science 32: 1177–1185.
Li, J., and L. Liu. 2006. “Supply Chain Coordination with Quantity Discount Policy.” International Journal of Production Economics 101
(1): 89–99.
Phillips, R. L. 2005. Pricing and Revenue Optimization. Stanford, CA: Stanford Business Books.
Pishvaee, M. S., M. Rabbani, and S. A. Torabi. 2011. “A Robust Optimization Approach to Closed-loop Supply Chain Network Design
under Uncertainty.” Applied Mathematical Modelling 135: 637–649.
Qi, X., J. F. Bard, and G. Yu. 2004. “Supply Chain Coordination with Demand Disruptions.” Omega 32 (4): 301–312.
Qin, Y., H. Tang, and C. Guo. 2007. “Channel Coordination and Volume Discounts with Price-sensitive Demand.” International Journal
of Production Economics 105 (1): 43–53.
Shapiro, A. 1994. “Quantitative stability in Stochastic Programming.” Mathematical Programming 67: 99–108.
Shapiro, A., and A. Nemirovski. 2005. On Complexity of Stochastic Programming Problems, 111–146. Boston, MA: Springer, US.
Shapiro, A., and A. Philpott. 2007. A Tutorial on Stochastic Programming. Discussion Paper. Atlanta GA USA: Georgia Institute of
Technology.
Solyalı, O., J. F. Cordeau, and G. Laporte. 2012. “Robust Inventory Routing under Demand Uncertainty.” Transportation Science 3:
327–740.
Soyster, A. L. 1973. “Convex Programming with Set-inclusive Constraints and Applications to Inexact Linear Programming.” Operations
Research 21 (5): 1154–1157.
Soyster, A. L. 1974. “A Duality Theory for Convex Programming with Set-inclusive Constraints.” Operations Research 22 (4): 892–898.
Tsai, J.-F. 2007. “An Optimization Approach for Supply Chain Management Models with Quantity Discount Policy.” European Journal
of Operational Research 177 (2): 982–994.
Tsay, A. A., and N. Agrawal. 2000. “Channel Dynamics under Price and Service Competition.” Manufacturing & Service Operations
Management 2 (4): 372–391.
Tunc, H., O. A. Kilic, S. A. Tarim, and B. Eksioglu. 2013. “A Simple Approach for Assessing the Cost of System Nervousness.” International
Journal of Production Economics 141 (2): 619–625.
Vargas, V. 2009. “An Optimal Solution for the Stochastic Version of the Wagner–Whitin Dynamic Lot-size Model.” European Journal of
Operational Research 198 (2): 447–451.
Wagner, H. M., and T. M. Whitin. 1958. “Dynamic Version of the Economic Lot Size Model.” Management Science 5 (1): 89–96.
Weng, Z. K. 1995. “Channel Coordination and Quantity Discounts.” Management Science 41: 1509–1522.
Wiesemann, W., D. Kuhn, and M. Sim. 2014. “Distributionally Robust Convex Optimization.” Operations Research 62: 1358–1376.
Xie, J., D. Zhou, J. C. Wei, and X. Zhao. 2010. “Price Discount based on Early Order Commitment in a Single Manufacturer-multiple
Retailer Supply Chain.” European Journal of Operational Research 200 (2): 368–376.
 International Journal of Production Research 6821

Zokaee, S., A. Jabbarzadeh, B. Fahimnia, and S. J. Sadjadi. 2014. “Robust Supply Chain Network Design: An Optimization Model with
Real World Application.” Annals of Operations Research 1–30.

Appendix 1. The final robust counterparts for the case of the budget uncertainty
In order to obtain the final robust counterparts, we reformulate the models from Section 3 in the way described in Step 3 of Section 3.2.

A.1 The final robust counterpart for the model presented in Section 3.2

minimize τ
subject to


T 
t 
T 
I 
T 
T 
I
h0 α0v + s0 S0t + si Ritk − ci
t=1 v=1 t=1 i=1 t=1 k=t i=1
⎡

I 
T T u 
T
+ d̄iu ⎣h 0 (T + 1 − v)βivu − h 0 (T + 1 − v)Rivk
Downloaded by [EP- IPSWICH] at 00:55 21 November 2017

i=1 u=1 v=u v=1 k=u

⎤

u 
T 
I 
T 
I
+ hi (u − t)Ritk + pi ⎦ + wi1u + i z i1 ≤ τ
t=1 k=u i=1 u=1 i=1

⎡

T 
u 
T
vui11 + vui12 = ⎣h 0 (T + 1 − v)βivu − h 0 (T + 1 − v)Rivk
v=u v=1 k=u
⎤

u 
T
+ hi (u − t)Ritk + pi ⎦ εiu d̄iu i > 0, u = 1, . . . T
t=1 k=u

wi1u ≥ vui11 , wi1u ≥ −vui11 , z i1 ≥ vui12 i > 0, u = 1, . . . T


I 
t 
I 
t 
I
M S0t ≥ α0t + βitu d̄iu + wit2u + i z it2 ∀t
i=1 u=1 i=1 u=1 i=1

vuit21 + vuit22 = βitu εiu d̄iu i > 0, ∀t, u = 1, . . . , t

vuit21 + vuit22 = 0 i > 0, ∀t, u = t + 1, . . . T
wit2u ≥ vuit21 , wit2u ≥ −vuit21 , z it2 ≥ vuit22 i > 0, ∀t, u = 1, . . . , t


T
Ri1k = 1 i >0
k=1


t−1 
T
Rikt−1 = Ritk i > 0 ,t > 2
k=1 k=t


t 
I 
T min
 (u,t) 
T 
I 
T 
I
α0v +
−
d̄iu (βiu Rivk ) − wit3u − i z it3 ≥ 0 ∀t
v=1 i=1 u=1 v=1 k=u i=1 u=1 i=1
⎛ ⎞
min
 (u,t) 
T

−
vuit31 + vuit32 = − ⎝βiu Rivk ⎠ εiu d̄iu i > 0, ∀t, u = 1, . . . , T
v=1 k=u

wit3u ≥ vuit31 , wit3u ≥ −vuit31 , z it3 ≥ vuit32 i > 0, ∀t, u = 1, . . . , T


I 
t 
I 
t 
I
α0t + βitu d̄iu − wit4u − i z it4 ≥ 0 ∀t
i=1 u=1 i=1 u=1 i=1
 6822 V. Buhayenko and D. den Hertog

vuit41 + vuit42 = −βitu εiu d̄iu i > 0, ∀t, u = 1, . . . , t

vuit41 + vuit42 = 0 i > 0, ∀t, u = t + 1, . . . T

wit4u ≥ vuit41 , wit4u ≥ −vuit41 , z it4 ≥ vuit42 i > 0, ∀t, u = 1, . . . , t

S0t ∈ {0, 1}, Rivk ∈ {0, 1} ∀i, t, k.

A.2 The final robust counterpart for the model given in Section 3.3

minimise ci

subject to
⎡ ⎤

T 
t 
T 
T 
T
hi αiv + si Sit + ⎣ pi − h i (T + 1 − u) + h i (T + 1 − v)βivu ⎦ d̄iu
t=1 v=1 t=1 u=1 v=u
Downloaded by [EP- IPSWICH] at 00:55 21 November 2017

T
+ wi1u + i z i1 ≤ ci
u=1
⎡ ⎤

T
vui11 + vui12 = ⎣ pi − h i (T + 1 − u) + h i (T + 1 − v)βivu ⎦ εiu d̄iu u = 1, . . . T
v=u

wi1u ≥ vui11 , wi1u ≥ −vui11 , z i1 ≥ vui12 u = 1, . . . T


t 
t
M Sit ≥ αit + βitu d̄iu + wit2u + i z it2 ∀t
u=1 u=1

vuit21 + vuit22 = βitu εiu d̄iu ∀t, u = 1, . . . t

vuit21 + vuit22 = 0 ∀t, u = t + 1, . . . T

wit2u ≥ vuit21 , wit2u ≥ −vuit21 , z it2 ≥ vuit22 ∀t, u = 1, . . . t

 t 

t 
t  
t
αiv + d̄iu βivu − 1 − wit3u − i z it3 ≥ 0 ∀t
v=1 u=1 v=u u=1
 t 

vuit31 + vuit32 = − βivu − 1 εiu d̄iu ∀t, u = 1, . . . t
v=u

vuit31 + vuit32 = 0 ∀t, u = t + 1, . . . T

wit3u ≥ vuit31 , wit3u ≥ −vuit31 , z it3 ≥ vuit32 ∀t, u = 1, . . . t


t 
t
αit + βitu d̄iu − wit4u − i z it4 ≥ 0 ∀t
u=1 u=1

vuit41 + vuit42 = −βitu εiu d̄iu ∀t, u = 1, . . . t

vuit41 + vuit42 = 0 ∀t, u = t + 1, . . . T

wit4u ≥ vuit41 , wit4u ≥ −vuit41 , z it4 ≥ vuit42 ∀t, u = 1, . . . t

Sit ∈ {0, 1} ∀t.

 International Journal of Production Research 6823

A.3 The final robust counterpart for the model presented in Section 3.4

maximise pit ∗
t ∗ ∈W
subject to

T 
T
αit ∗ −1 + si +
[ p ∗ + (u − t ∗ )h p ∗ ] − β ∗ ) +
d̄iu (βit ∗ −1u + n iu wit ∗ 1u + i z it ∗ 1 ≤ αik ∗ (t ∗ , k ∗ ) ∈ W
it i it ik u
u=1 u=1
∗ ∗
[ p ∗ + (u − t ∗ )h p ∗ ] − β ∗ )ε d̄
vuit 11 + vuit 12 = (βit ∗ −1u + n iu it i it ik u iu iu t ∗ ∈ W, u = 1, . . . , T
∗ ∗ ∗
wit ∗ 1u ≥ vuit 11 , wit ∗ 1u ≥ −vuit 11 , z it ∗ 1 ≥ vuit 12 t ∗ ∈ W, u = 1, . . . , T


T 
T
αik ≤ αit−1 + si +
[ p + (u − t)h p ] − β ) −
d̄iu (βit−1u + n iu wit2u − i z it2 ∀t, k
it i it iku
u=1 u=1

[ p + (u − t)h p ] − β )ε d̄
vuit21 + vuit22 = −(βit−1u + n iu ∀t, u = 1, . . . , T
it i it iku iu iu

wit2u ≥ vuit21 , wit2u ≥ −vuit21 , z it2 ≥ vuit22 ∀t, u = 1, . . . , T

Downloaded by [EP- IPSWICH] at 00:55 21 November 2017

pit ∗ ≤ pi ∀t ∗
pit = pi ∀t = t ∗ .
Copyright of International Journal of Production Research is the property of Taylor & Francis
Ltd and its content may not be copied or emailed to multiple sites or posted to a listserv
without the copyright holder's express written permission. However, users may print,
download, or email articles for individual use.