You are on page 1of 10

Mathematical and Computer Modelling 53 (2011) 1670–1679

Contents lists available at ScienceDirect

Mathematical and Computer Modelling
journal homepage: www.elsevier.com/locate/mcm

Supplier selection in make-to-order environment with risks
Tadeusz Sawik ∗
AGH University of Science & Technology, Department of Operations Research and Information Technology, Al.Mickiewicza 30, 30-059 Kraków, Poland

article

info

Article history:
Received 14 April 2009
Received in revised form 7 June 2010
Accepted 21 December 2010
Keywords:
Supplier selection
Order allocation
Supply chain risk management
Conditional value-at-risk
Mixed integer programming

abstract
The problem of allocation of orders for parts among part suppliers in a customer driven
supply chain with operational risk is formulated as a stochastic single- or bi-objective
mixed integer program. Given a set of customer orders for products, the decision maker
needs to decide from which supplier to purchase parts required for each customer order to
minimize total cost and to mitigate the impact of delay risk. The selection of suppliers and
the allocation of orders is based on price and quality of purchased parts and reliability of on
time delivery. To control the risk of delayed supplies, the two popular percentile measures
of risk are applied: value-at-risk and conditional value-at-risk. The proposed approach
is capable of optimizing the supply portfolio by calculating value-at-risk of cost per part
and minimizing mean worst-case cost per part simultaneously. Numerical examples are
presented and some computational results are reported.
© 2010 Elsevier Ltd. All rights reserved.

1. Introduction
An important issue in supply chain risk management is how to best allocate the orders for parts among various part
suppliers to fulfill customer orders for products at low cost and to mitigate the impact of risk. The selection of supply portfolio
in the presence of supply chain delay risk, i.e., the supplier selection and order allocation under uncertain quality of supplied
materials and reliability of on-time delivery, is based on price, quality (defect rate) and reliability (on-time delivery rate)
criteria that may conflict with each other. Furthermore, to reduce the fixed ordering costs, the number of suppliers and the
total number of orders should be minimized. However, to mitigate the impact of delay risk the selection of more suppliers
sometimes may divert the risk of unreliable supplies.
In spite of the importance of supplier selection and order allocation problems, the decision making is not sufficiently
addressed in the literature (for a recent review, see [1], in particular for the make-to-order manufacturing environment,
e.g. [2–6]. The vast majority of the decision models are mathematical programming models with either a single objective,
e.g. [7,8] or multiple objectives, e.g. [9–13].
The models developed for supplier selection and order allocation can be either single-period models (e.g. [9,11]) that do
not consider inventory management or multi-period models (e.g. [8,12,14,15]) which consider the inventory management
by lot-sizing and scheduling of orders. Since common parts can be efficiently managed by material requirement planning
methods, this research is focused on custom parts that can be critical in make-to-order manufacturing. For customengineered products no inventory of custom parts can be kept on hand. Instead, the custom parts need to be requisitioned
with each customer order and hence the custom parts inventory need not to be considered.
When selecting suppliers and allocating orders to generate a supply portfolio, the producer faces uncertain costs and must
place orders with a set of suppliers with different quality and reliability. In stochastic supply settings, supplier selection
allows the producer to decide whether it should cooperate with low cost, yet risky suppliers over more expensive but
possibly more reliable suppliers. A common risk-neutral objective of minimizing expected cost is therefore influenced by

Tel.: +48 12 617 39 92; fax: +48 12 617 39 84.
E-mail address: ghsawik@cyf-kr.edu.pl.

0895-7177/$ – see front matter © 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.mcm.2010.12.039

. . differ in price and quality of offered parts and in reliability of on time delivery of parts.T. Denote by ci the capacity of supplier i ∈ I. the producer can be charged with a much higher penalty cost of delayed customer orders for products. Since different suppliers may deliver the ordered parts with different delays with respect to requested dates. this approach has been applied to solution of the newsvendor problem (e. quality loss cost. and discrete tolerances.or bi-objective static supplier selection and order allocation. value-at-risk (VaR) and conditional value-at-risk (CVaR) to be applied for managing the risk of supply delays. the portfolio approach presented in [5. VaR and CVaR have been widely used in financial engineering in the field of portfolio management (e. and let r be the largest acceptable average . .. For example. As a result. Similarly. the delivery date and the corresponding reliability of on-time delivery may randomly vary. CVaR allows the evaluation of worst-case costs and shaping of the resulting cost distribution through optimal supplier selection and order allocation decisions. Problem description In the supply chain under consideration various types of products are assembled by a single producer to satisfy customer orders.or multi-objective mixed integer program with the risk of defective or unreliable supplies controlled by the maximum number of delivery patterns (combinations of suppliers’ delivery dates) for which the average defect rate or late delivery rate can be unacceptable. Each order j ∈ J is described by the quantity dj of required custom parts and by requested delivery date. i. Uryasev [19] and Rockafellar and Uryasev [20. and by pij the purchasing price of part for customer order j ∈ J from supplier i ∈ I. In addition. [17] use the stochastic integer programming to model the relationship between manufacturing cost. . assembly yield. For each supplier. Feng et al. where each scenario s ∈ S is a unique combination of m delivery delays. in addition. This paper demonstrates that for a finite number of scenarios. The mixed integer programs for a single objective selection of supply portfolio to minimize either the expected cost per part or expected worst-case cost part are developed in Section 3. For example. using custom parts purchased from multiple suppliers (for notation used. . a different operational risk can be associated with each supply portfolio. in Sawik [6] the portfolio approach has been enhanced to consider a single-period supplier selection and order allocation in the presence of supply chain disruption risks.e. the quality of parts delivered by each supplier may randomly vary. the suppliers have different limited capacity and. which is the allocation of orders among the suppliers combined with the allocation of orders among the planning periods. m} be the set of m suppliers and J = {1. The problem is formulated as a single. [16]).6] has been enhanced to consider a single-period supplier selection and order allocation in the make-to-order environment in the presence of supply chain delay risk. Sawik / Mathematical and Computer Modelling 53 (2011) 1670–1679 1671 uncertainty and risk. Denote Ps as the probability that delivery scenario s is realized. In this paper. n} the set of n customer orders for the products. where the latter need not to be explicitly considered when selecting a supplier. Likewise. the expected delivery delay of supplier i. [18]). [22]) or recently to risk averse selection of orders. The ordered parts are dispatched to the producer after the completion time of their manufacturing to meet the requested delivery dates. one for each supplier and S is the index set of all scenarios. with δis = 0 for on-time delivery. where the approach is combined with a scenario-based method [23].g.g. In Sawik [5]. by oi the cost of ordering parts from supplier i ∈ I. The supply delays may result in the shortage of required parts and the corresponding delay penalty costs of delayed customer orders should be incorporated into the model.g. Each supplier can provide the producer with custom parts for all customer orders. number of days behind the requested delivery date) of supplier i in scenario s. However. however. whereas parts delivered late may be paid for at a reduced price. . define by ri = s∈S Ps δis . The trade-off (mean-risk) model for a bi-objective selection of supply portfolio is presented in Section 4. The proposed portfolio approach allows the two popular in financial engineering percentile measures of risk. Numerical examples and some computational results are provided in Section 5. In contrast to the dynamic portfolio. The proposed mixed integer programming models provide the decision maker with a simple tool for evaluating the relationship between expected and worst-case costs. that is for the allocation of orders for parts among the suppliers with no timing decisions. Although. caused by the shortage of required parts due to defective or delayed supplies. Clearly. In Section 2 a description of the supplier selection problem in a customer driven supply chain with risks is provided. the selection of optimal supply portfolio. chance-constrained programming models were developed by Kasilingam and Lee [7] to account for stochastic demand and by Wu and Olson [13] to consider expected losses from quality acceptance inspection or late delivery. When the suppliers are selected the risk of defective and unreliable (late) deliveries can be considered using past observations. 2. The paper is organized as follows. Let I = {1. Then. and final conclusions are made in the last section. . known ahead of time. the producer does not need to pay for ordered and defective parts. However. the largest acceptable average defect rate of supplies. research seldom considers uncertainty and risk (see. a portfolio approach is proposed for the problem of allocation of orders for custom parts among suppliers in make-to-order manufacturing. the supplier selection problem is stochastic in nature. Let δis ≥ 0 denote the random delivery delay (e. .21] introduced a new approach to select a portfolio with the reduced risk of high losses. new non-risk-neutral objectives of minimizing the number of outcomes that could occur above an acceptable cost level are observed in practice. let qi be the ∑expected defect rate of supplier i and q. The portfolio is optimized by calculating VaR and minimizing CVaR simultaneously. CVaR is used in conjunction with VaR and is applied for estimating the risk with non-symmetric cost distributions. The mixed integer programming models are proposed for a single. see Table 1).

where − xi = 1 i∈I and 0 ≤ xi ≤ 1 is the fraction of the total demand for parts ordered from supplier i (for definition of problem variables. . if an order for parts is placed on supplier i. where e′j is the per unit and per period penalty cost of delayed customer order j. It is assumed that eij is already reduced by a fraction of the purchasing price for parts delivered late (i. Minimization of expected costs In this subsection the selection of a risk-neutral supply portfolio is considered. . . s ∈ S = {1.e. . where D = j∈J dj is the total demand for parts and eij is the per unit and per period penalty cost of delayed customer order j caused by the late delivery of required custom parts from supplier i. m} j = customer order. 3. Nevertheless. eij = e′j − βij pij .g. i∈I j∈J s∈S Ps eij δis dj zij /D. for determining a static supply portfolio. n} s = delivery scenario. see Table 2). . the allocation of orders for parts among the suppliers is not combined with the allocation of orders among the planning periods. at time 0). Indices i = supplier. in 100α % of scenarios. h} Input parameters ci = capacity of supplier i dj = number of parts to be purchased for customer order j ∑ D= j∈J dj —total demand for parts δis = random delivery delay of supplier i in scenario s eij = per unit and per period penalty cost of delayed customer order j caused by the late delivery of parts provided by supplier i oi = cost of ordering parts from supplier i pij = price of part for customer order j purchased from supplier i Ps = probability of delivery scenario s qi = expected defect rate of supplier i q = the ∑ largest acceptable average defect rate of supplies ri = s∈S Ps δis —expected delivery delay of supplier i r = the largest acceptable average delivery delay of supplies α = confidence level Table 2 Problem variables. . . When deciding on a static supply portfolio it is assumed that the orders for all parts are simultaneously placed on selected suppliers (e.e.1672 T. The decision maker needs to decide from which supplier to purchase custom parts required for each customer order to achieve a minimum average cost per part of ordering. . i i i∈I ∑ ∑ ∑ ∑ ∑ ∑ i∈I j∈J pij dj zij /D. . . 3. .. and βij is the per unit and per period price reduction factor for parts for customer order j delivered late by supplier i). The supply portfolio should be checked against the largest acceptable average defect rate and delivery delay of supplies. . j ∈ J = {1. i. purchasing. i ∈ I = {1. . . the static portfolio should be checked over the time horizon against the risk of too low quality of purchased parts (a too high defect rate) and a too low reliability of supplies (a too high delay). Therefore. i. In risk-neutral operating conditions the ∑ overall quality of the supply portfolio can be measured by the expected cost per part of ordering. Ts = the amount by which cost of portfolio in scenario s exceeds VaR (tail cost) VaR = the targeted cost of portfolio based on the α -percentile of costs.e. . otherwise yi = 0 (supplier selection variable) zij = the fraction of parts required for customer order j ordered from supplier i (customer order allocation variable) delivery delay of supplies. given past observations. . . the outcome cannot exceed VaR (value-at-risk) xi = the fraction of total demand for parts ordered from supplier i (total demand allocation variable) yi = 1. The static supply portfolio is defined as (x1 . and delay penalty. Single objective selection of supply portfolio In this section two mixed integer programs are proposed for a single-period supplier selection and order allocation problem. where both quality and reliability are randomly varying over time..1. and each supplier delivers all the ordered parts at the earliest possible delivery date. xm ). o y / D. Sawik / Mathematical and Computer Modelling 53 (2011) 1670–1679 Table 1 Notation. purchasing and delay penalty and to mitigate the impact of risk of the unreliable supplies by minimizing the potential worst-case cost per part.

i ∈ I. the producer is not charged for purchasing the additional quantities of defective parts.T. (1) j∈J subject to 1. i. j ∈ J. The mixed integer program SP_E for selection supply portfolio in a risk-neutral supply chain environment is formulated below. However. Sawik / Mathematical and Computer Modelling 53 (2011) 1670–1679 1673 The purchase orders for parts are assumed to be inflated by the reject rates qi of defective parts. – the average defect rate of the portfolio cannot be greater than q.e. VaR is a decision variable based on the α -percentile of costs. 1]. (4) j∈J 2. Supplier selection and order allocation constraints: – for each customer order the demand for required parts must be fully met. i∈I zij ∈ [0.e. VaR does not account for properties of the cost distribution beyond the confidence level and hence does not explain the magnitude of the cost when the VaR limit . Portfolio selection constraints: – the portfolio definition constraint (note that the order allocation variable xi is an auxiliary variable determined by zij ). Non-negativity and integrality conditions xi ≥ 0 . 3. the following two popular percentile measures of risk are applied [18]. To control the supply chain risk. (1). in the objective function (1). In other words.. dj zij i∈I (5) j∈J − qi xi ≤ q (6) ri xi ≤ r . In the proposed model. we allow 100(1 − α)% of the outcomes to exceed VaR. the likelihood that a given portfolio’s cost will not exceed the amount defined as VaR).2. In other words. j∈J (2) i∈I − (1 + qi )dj zij ≤ ci yi . and the mean value of these outcomes is represented by CVaR. Model SP_E: Selection of supply portfolio to minimize expected cost per part Minimize Expected Cost per Part − E (x) =  D+ oi yi −− i∈I i∈I  pij dj zij j∈J D+ −−− s∈S i∈I Ps δis eij dj zij  D. However. (7) i∈I − i∈I 3. the outcome may exceed VaR. xi = −  D. in 100(1 − α)% of the scenarios.e. dj zij . where α ∈ (0. the corresponding purchasing cost is simply given by i∈I j∈J pij dj zij .. the portfolio will be optimized by minimizing expected cost per part E (x). are equal to (1+qi )dj zij for all i ∈ I . 1}. • Conditional Value-at-Risk (CVaR) at a 100α % confidence level is the expected cost of the portfolio in the worst 100(1−α)% of the cases. i∈I (3) j∈J yi ≤ − i ∈ I. Minimization of expected worst-case costs In this subsection the selection of a risk-averse supply portfolio is considered. VaR represents the maximum cost associated with a specified confidence level of outcomes (i. − zij = 1. since the producer does not ∑need ∑to pay for ordered and defective parts in the amount of qi dj zij . (8) (9) (10) Note that the demand for suppliers capacity in (3) is already inflated by the expected reject rates of selected suppliers. j ∈ J. 1) represents the confidence level for the cost distribution across all scenarios: • Value-at-Risk (VaR) at a 100α % confidence level is the targeted cost of the portfolio such that for 100α % of the scenarios. – for each selected supplier the total quantity of ordered parts cannot exceed its capacity. i∈I yi ∈ {0. i. – at least one customer order should be assigned to each selected supplier. However. – the average delivery delay of the portfolio cannot be greater than r. the outcome will not exceed VaR..

Define Ts as the tail cost for scenario s. When all custom parts of the same type that are required for a customer order should be purchased from a single supplier. VaR may be better for optimizing portfolios when good models for tails are not available.g. that is. s∈S (12) j∈J 4. subject to (1)–(13). Also. Bi-objective selection of supply portfolio In the single objective approach the supply portfolio is selected by minimizing either the mean cost per part. s ∈ S. When using CVaR to minimize worst-case costs. there may be portions of the nondominated set (near a weakly nondominated solution) that the weighted-sum approach is unable to compute. where the suppliers offer discounts based on quantity or business volume of ordered parts. e. The nondominated solution set of the bi-objective supply portfolio can be found by the parameterization on λ the weighted-sum program WSP presented below. Risk constraints: – the tail cost for scenario s is defined as the nonnegative amount by which cost per part in scenario s exceeds VaR. it has been assumed that each supplier is capable of manufacturing all required part types. Sawik / Mathematical and Computer Modelling 53 (2011) 1670–1679 is exceeded.g. (see. [23]) is formulated as the optimization of a weighted-sum of the expected cost and the CVaR as a risk measure. each supplier may only be prepared to manufacture a subset of part types and provide with the parts the corresponding subset of customer orders. a risk aversive decision maker wants to minimize the worst-case costs exceeding VaR. Model SP_CV: Selection of supply portfolio to minimize expected worst-case cost per part Minimize Expected Worst-Case Cost (CVaR) C (x) = VaR + (1 − α)−1 − Ps Ts (11) s∈S subject to 1. CVaR is always not less than VaR. where costs exceed VaR. We assume that the decision maker is willing to accept only portfolios for which the total probability of scenarios with costs greater than VaR is not greater than 1 − α . (1) or the mean worst-case cost per part. The mixed integer program SP_CV for selection supply portfolio to reduce the risk of high costs is formulated below.10]). 4. VaR is the acceptable cost level above which we want to minimize the number of outcomes or realizations and CVaR considers those portfolio outcomes. In the SP_CV model presented below. even if the complete parameterization on λ is attempted. the parts required for each customer order are assumed to be partially provided by one or more suppliers. Steuer [24] proved that for mixed integer programs. Model WSP: Bi-objective selection of supply portfolio to minimize weighted sum of mean cost and mean worst-case cost Minimize λE (x) + (1 − λ)C (x) (14) where 0 ≤ λ ≤ 1. the parameter α is fixed by the decision maker to control the risk of losses due to unreliable supplies. a large reduction in VaR may result in more scenarios with positive tail costs. . E (x). CVaR (conditional value-at-risk) focuses on the tail of the cost distribution.g. The trade-off model known in the literature as the mean-risk model (e. In the proposed models. (13) Note that as Ts is constrained to being positive. (11). The proposed models can be further enhanced for a different cost structure. In this section the two cost functions are considered simultaneously. By measuring CVaR. In the supply portfolio selection problem. e. Order assignment constraints: (2)–(4) 2. for a discount environment. Portfolio selection constraints: (5)–(7) 3. where tail cost is defined as the amount by which costs in scenario s exceed VaR. The mixed integer program WSP is based on SP_CV model with the addition of objective (1) of SP_E model. Ts ≥ −  oi yi D+ −− i∈I i∈I  pij dj zij j∈J D+ −− i∈I δis eij dj zij  D − VaR. [5. Since VaR and CVaR measure different parts of the cost distribution. on outcomes with the highest cost. On the other hand. then the corresponding continuous allocation variable zij should be redefined as a binary assignment variable denoting whether or not all parts required for order j are provided by supplier i.1674 T. In a more general setting. Non-negativity and integrality conditions: (8)–(10) Ts ≥ 0. [19–21]. C (x). The portfolio will be optimized by calculating VaR and minimizing CVaR simultaneously. In other words. In the proposed models CVaR is represented by an auxiliary function (11) introduced by Rockafellar and Uryasev [20]. the magnitude of the tail costs is considered to achieve a more accurate estimate of the risks of minimizing cost. and a bi-objective selection of supply portfolio is presented aimed at minimizing both objective functions to balance expected costs with the risk tolerance. However. otherwise CVaR may be preferred. the model tries to decrease VaR and hence positively impact the objective function.

Confidence level α 0. • α . was exponentially distributed.75.. was equal to 50. was equal to 0. number of binary variables. = 10. • δis ∈ {0. Nonz. mδ − δ max i∈I is • Ps .04 a CPU seconds for proving optimality on a PC Pentium IV.009 10 Var. 10}. was equal to 5000 for each supplier i.T. was uniformly distributed over [10. Bin. or 1% of all scenario outcomes. α . Var. .50.90 0.e. = 5083.of Suppliers CPUa 520 10 83 2060 21. 1. of suppliers selected a 201 21. the number of customer orders.e. delay penalty cost per period and per unit of customer order j caused by the late delivery of required custom parts provided by supplier i. In the computational experiments the confidence level.90. Cons. Ps = hm . Table 4 also presents the probability 1 − F (VaR) of outcomes with worst-case cost above VaR. were integers uniformly distributed over [100. the portfolio is more diversified and the total demand is allocated among all 10 suppliers.69 42.05. For example.41 20. It is observed that expected cost is greater for the SP_CV model as a result of risk considered in the supply portfolio selection decision. the largest acceptable average delivery delay of the portfolio.59 0. was equal to ⌈2 j∈J dj /m⌉ (⌈·⌉ denotes the smallest integer not less than ·). Computational examples In this section some computational examples are presented to illustrate possible applications of the proposed mixed integer programming approach for selection of supply portfolio under uncertainty and risk. = 5521. Solution results for the risk-neutral model SP_E and the risk-averse model SP_CV with different confidence levels are shown in Tables 3 and 4.s∈S (δis ) is the maximum δmax −∆ ∑ ∑ delay and ∆ = suppliers across all scenarios.67 33. which means that the focus is on minimizing the highest 50%.89 0. • pij . Nonz. . 10) are not selected at all. Nonz. The optimal supply portfolios for the SP_E model and for the SP_CV model with different confidence level α are ∑ shown in Fig.16 0. Fig. was equal to 10. RAM 1 GB /CPLEX 11. is set at five levels of 0. 1 demonstrates that for SP_E model most of the total demand is allocated among suppliers with a lower average unit price and the most expensive suppliers (i = 3. over 8. 6. . the closer are solution values of VaR and CVaR. • eij .17 0. 0.08. the total capacity of all suppliers was equal to the double total demand for parts. in particular for α = 0. the confidence level.32 39. respectively. a higher probability Ps is set for i∈I s∈S δis is the total delay of all ∑ scenarios s ∈ S with a smaller total delivery delay i∈I δis ). Sawik / Mathematical and Computer Modelling 53 (2011) 1670–1679 1675 Table 3 Solution results for SP_E model. the total demand for parts is D = j∈J dj = 137 000 parts. • qi . In contrast. was equal to 0. i.e. i. .70 30. where δmax = maxi∈I .99.. for the risk averse SP_CV model.69 36. 0.99 165 21.008 to 0. the higher is the confidence level α .1 for each supplier i. 1. Bin. Cons.75. 0. the probability that delivery scenario s is realized.236 10 139 21.83 25. was equal to 1.. the delivery delay (in days) of each supplier i in ∑ each scenario s. was equal to 20 for all suppliers i and orders j.50 32. ∑ For the test examples. 25%. number of constraints..90. (i.8 GHz. was equal to 5000. Bin. RAM 1 GB /CPLEX 11.50 0. for each supplier i the expected delay ri across all scenarios. Cons.95 or 0. which indicates that the impact of delay risks is mitigated by diversification of the supply portfolio.90. Table 4 Solution results for the SP_CV model. Expected Cost No.50.7% of the cost outcomes are above 30. was exponentially distributed. 5. • oi . 1. m.087 10 129 21. The following parameters have been used for the example problems: • • • • • h. and 0. the expected defect rate of each supplier i.17.499 10 CPU seconds for proving optimality on a PC Pentium IV. the largest acceptable average defect rate of the portfolio. • q. 0. In addition. for α = 0. = 2 562 060 CPUa Expected cost CVaR VaR 1 − F (VaR) No. • r. the unit price of parts required for each customer order j purchased from each supplier i. 5000]. n. 10%. 0. the number of suppliers. ranging from 0. the capacity of each supplier i. The size of the mixed integer programs is represented by the total number of variables. the cost of ordering parts from supplier i.5. For SP_CV model. and number of nonzero coefficients in the constraint matrix.95. .47 0. Note that VaR becomes smaller than the expected cost when α = 0. 1.15] and reduced by the factor (1 − qi ) to get a lower price for parts from the suppliers with a higher defect rate. as well as the average unit price j∈J pij /n are presented. Var. the number of delivery scenarios. ci .75 0.049 10 109 21. the numbers of required parts for each customer ∑order.99.. dj .95 0.8 GHz.9 the demand allocation is xi = 0. 0. The tables demonstrate that the number of selected suppliers is greater for the SP_CV model.64 26. 5%.49 6 0.65 29. costs per part.

25.58 12 0. For the bi-objective approach.90. For the selected nine levels of weight λ.10.67.5) 3 4 5 6 Supplier 7 SP_CV(alpha = 0. Supply portfolios for SP_E and SP_CV models. 0. a more risk averse decision making focuses on a smaller set of outcomes.50. 0. Fig. 0. The results obtained for the confidence level α = 0. 0. 0.18). 36. the model restricts the number of outcomes that exist above VaR.6 0. (21.15 0. CVaR) = (21.49. six different nondominated solutions were found: (Mean Cost. (21.2 0.85).99. 1.1676 T.52 0. (21.4 0. [24]. (21.5 11.51). 0.65.95 are presented in Table 5. 36.1 0.2 1 2 3 4 5 6 7 8 9 10 Supplier Expected Delay Average Unit Price Supply Portfolio for SP_E Model Fraction of Total Demand 0. Sawik / Mathematical and Computer Modelling 53 (2011) 1670–1679 Supplier Basic Characteristics 12.75. 36.9) 8 SP_CV(alpha = 0. When α increases.6 Unit Price Expected Delay 0.49.90 and 0.99) Fig. 2 shows the cost distribution for the optimal supply portfolio (model SP_CV) at the three confidence levels: 0.68. and the likelihood of having worst-case costs closer to VaR also is increasing.05 0 1 2 3 4 5 6 7 8 9 10 9 10 Supplier Supply Portfolios for SP_CV Model Fraction of Total Demand 0.54 11. 43. 36.99. for the mixed integer programs there may be portions of the nondominated set that the weighted-sum program is unable to compute. However. even if the complete parameterization on λ is attempted.56 0. the subsets of nondominated solutions were computed by parameterization on λ ∈ {0.05 0 1 2 SP_CV(alpha = 0.01.00} the weighted-sum program WSP. (21. As the value of α is increasing.5. The right-hand tail of the cost distribution is minimized with simultaneous increasing of the expected cost.1 0.69.15 0. 40. Note that solutions to single objective models SP_E and SP_CV are equivalent to the nondominated solutions of the weighted-sum program WSP with . 0.2 0.64). 1.00.50).53). 0.

00 Expected cost CVaR VaR No.48 10 21.67.95 is presented.T.53 32.1 0.10 6 λ = 1 and λ = 0.10 0.16.51 32.08 0.11 solver (with the default settings) on a laptop with Pentium IV processor running at 1.18 40. The results emphasize the effect of varying cost/risk preference of the decision maker.1 0. 2.69 36. When λ approaches 1. Cost distribution for different confidence level α in the SP_CV model.04 0.06 0. CVaR = 42.01 0.08 0.06 0. λ 0.99 1.69 36.02 0 1 5 9 13 17 21 25 29 33 Cost per Part 37 41 45 49 alpha = 0. respectively.59 10 21.51 10 21.04 0. the decision maker focuses on minimization of the expected cost and the number of selected suppliers decreases.14 0. CVaR = 26.50 32.25 0.70 Probability Mass Function 0.69 36. The computational experiments were performed using the AMPL programming language and the CPLEX v.50 32.1 0.99: Mean Cost = 21.68 36. .41 Probability Mass Function 0. CVaR = 33.06 0.17.04 0. The trade-off between the expected cost and the expected worst-case cost is shown in Fig.12 0. Table 5 Nondominated solutions for the weighted-sum program WSP:α = 0. Sawik / Mathematical and Computer Modelling 53 (2011) 1670–1679 1677 Probability Mass Function 0.95.12 Probability 0. Mean Cost = 21.64.89.47 10 21.39 7 21. where the convex efficient frontier of mean cost—the CVaR model with α = 0. of suppliers selected 21.02 0 1 5 9 13 17 21 25 29 33 Cost per Part 37 41 45 49 alpha = 0.08 0.50 32.02 0 1 5 9 13 17 21 25 29 33 Cost per Part 37 41 45 49 alpha = 0.49 40.49 43.14 Probability 0.49 10 21.00 0.69.50 32.52 10 21.14 0.75 0. VaR = 30.90 0.12 Probability 0.67 36.8 GHz and with 1 GB RAM.59 10 21.5: VaR = 20.9: Mean Cost = 21.64 32.69 36. VaR = 39. 3. The CPLEX solver was capable of finding proven optimal solutions within CPU seconds for all examples.32 Fig.85 35.65 36.50 0.

terrorist attacks. [4] T. labor strikes. The proposed portfolio approach is capable of optimizing the supply portfolio by calculating valueat-risk of cost per part and minimizing conditional value-at-risk of cost per part simultaneously. the business impact associated with disruption risks is much greater than that of the delay risk. in a non-risk-neutral environment the number of selected suppliers increases with the confidence level α . Sawik. References [1] N. While a single objective approach is capable of finding an optimal.N. M. using the CPLEX solver for mixed integer programming. Haouari. the selection of a dynamic. The limited computational experiments prove that the proposed approach based on mixed integer programming models provides the decision maker with a simple tool for evaluating the relationship between expected and worst-case costs. either risk-neutral supply portfolio that minimizes the expected cost per part or risk-averse supply portfolio that minimizes the expected worst-case cost per part. The trade-off model has been formulated as the optimization of a weighted-sum of the expected cost and the CVaR as a risk measure. 3. [8]).95. Single vs. focuses on varying cost/risk preference of the decision maker and provides her/him with a subset of nondominated portfolios. A framework for facilitating sourcing and allocation decisions for make-to-order items. Computers & Operations Research 34 (2007) 3516–3540. with the joint decisions: what. The experiments show that in a risk-neutral environment the more costly suppliers are rarely selected. For example. Akinc. [3] N. Conclusion The problem of optimal allocation of orders for parts among a set of approved suppliers in a customer driven supply chain under conditions of risk associated with uncertain reliability of suppliers has been modeled as a portfolio-type stochastic mixed integer program. This work has been partially supported by research grant of MNiSzW and by AGH. multiple objective supplier selection in a make-to-order environment. An important issue that needs to be further considered is the estimation of probabilities and the resulting costs associated with delay risks. 6. Aissaoui. the trade-off model: mean–risk. S. The supply portfolio is selected based on the expected cost per part of ordering. [2] U. On the other hand. Selecting a set of vendors in a manufacturing environment. Finally. Soni. Efficient frontier of Mean Cost—CVaR model WSP:α = 0. made over a time horizon (see. Ghosh. floods. . which indicates that the impact of delay risks is mitigated by diversification of the supply portfolio. Selection of supply portfolio under disruption risks. S. Hassini. multi-period supply portfolio should be considered. Sawik. purchasing and delay penalty and/or the expected worst-case cost per part. Sawik / Mathematical and Computer Modelling 53 (2011) 1670–1679 Fig. The optimal risk-neutral or risk averse supply portfolio can be found within CPU seconds for a limited number of scenarios considered. where and when to order.1678 T. Journal of Operations Management 11 (1993) 107–122. [6] T. Supplier selection and order lot sizing modeling: a review. Omega: The International Journal of Management Science 39 (2011) 194–208. Omega: The International Journal of Management 38 (2010) 203–212. [5] T. In most cases. Sawik. the delivery scenarios should combine the low impact supply chain delay risk with a high impact disruption risk referring to the major disruptions to normal activities caused by natural or man-made disasters such as earthquakes. in a more general setting. Acknowledgements The author is grateful to an anonymous reviewer for reading the manuscript very carefully and providing constructive comments which helped to substantially improve this paper. The two alternative objective functions were either considered separately or simultaneously in a bi-objective optimization problem. Decision Sciences 35 (2004) 237–259. Future research on supplier selection in customer driven supply chains with risks should also focus on various enhancements of the proposed approach. over a set of delivery scenarios. A cyclic versus flexible approach to materials ordering in make-to-order assembly. Murthy. E. Mathematical and Computer Modelling 42 (2005) 279–290.

Takano. Rockafellar. Conditional value-at-risk for general loss distributions. C. The Journal of Risk 2 (3) (2000) 21–41. Value-at-risk vs. Sawik / Mathematical and Computer Modelling 53 (2011) 1670–1679 [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] 1679 R. Wu. C. Basnet. Supply chain risk. 1996. S. Taaffe. J. New York. INFORMS 2008 (2008) 270–294. R. Sarykalin.L. O. J. S. Rockafellar. . Tutorials in Operations Research. and vendor selection.T. Selection of vendors a mixed-integer programming approach. Science. Gotoh. R. Omega: The International Journal of Management Science 37 (5) (2009) 996–1006. Z. Ustun. Uryasev. Kasilingam.G. D. Current. K.H. Wiley. International Journal of Production Economics 114 (2008) 646–655. Chahara. Demirtas. Tang. R. Wang. S. Computation and Application. K.A. Omega: The International Journal of Management Science 36 (2008) 76–90. Uryasev. Steuer.E. J. Wu. Supplier selection with multiple criteria in volume discount environments. The total cost of logistics in supplier selection under conditions of multiple sourcing. D. S.H. Supplier selection and supply quantity allocation of common and non-common parts with multiple criteria under multiple products. Leung. Ustun. Che. Xia. Conditional value-at-risk: optimization algorithms and applications. S. Financial Engineering News 14 (2) (2000). Uryasev. multiple criteria and capacity constraints.S.Y. An optimization model for concurrent selection of tolerances and suppliers. C.R. C. Risk averse demand selection with all-or-nothing orders. Omega: The International Journal of Management Science 36 (2008) 509–521. Olson. Computers and Operations Research 32 (2005) 1–14. Journal of Banking and Finance 26 (7) (2002) 1443–1471. Uryasev. An integrated multiobjective decision making process for supplier selection and order allocation. Demirtas. International Journal of Production Economics 103 (2006) 451–488. Wang. Computers and Industrial Engineering 55 (2008) 110–133. O.X.M. International Journal of Production Economics 73 (2001) 15–27.A.W. Newsvendor solutions via conditional value-at-risk minimization. Feng. E. C. Omega: The International Journal of Management Science 35 (2007) 494–504. A multiobjective approach to vendor selection. Z. Lee. simulation. C. J. S. Serraino. Ghodsypour. W. Computers and Industrial Engineering 31 (1996) 347–350. Computers and Industrial Engineering 40 (2001) 15–33. H. Wang. Y. J. O’Brien. Perspectives in supply chain risk management. Multiple Criteria Optimization: Theory. 38 (3–4) 203–212.S. G. Weber. Optimization of conditional value-at-risk. European Journal of Operational Research 179 (1) (2007) 80–96.P.T.S. An integrated multi-objective decision making process for multi-period lot sizing with supplier selection.A. E. European Journal of Operational Research 68 (1993) 173–184.T. Inventory lot-sizing with supplier selection. conditional value-at-risk in risk management and optimization.