https://orcid.org/0000-0003-2951-0613
Figures
Abstract
This paper studies the flexible large-scale supplier selection and order allocation problem with various quantity discounts, i.e., no discount, all-unit discount, incremental discount, and carload discount. It fills a literature gap that models usually formulate one or seldom two types because of the modeling and solution difficulty. All suppliers offering the same discount are far from reality, especially when the number of suppliers is large. The proposed model is a variant of the NP-hard knapsack problem. The greedy algorithm, which solves the fractional knapsack problem optimally, is applied to cope with the challenge. Three greedy algorithms are developed using a problem property and two sorted lists. Simulations show the average optimality gaps are 0.1026%, 0.0547%, and 0.0234% and the model can be solved in centiseconds, densiseconds, and seconds for supplier numbers 1000, 10000, and 100000. This allows the full use of data in the big data era.
Citation: Wang Q (2023) Flexible supplier selection and order allocation in the big data era with various quantity discounts. PLoS ONE 18(3): e0283585. https://doi.org/10.1371/journal.pone.0283585
Editor: Ashkan Memari, Bond University, AUSTRALIA
Received: August 18, 2022; Accepted: March 12, 2023; Published: March 27, 2023
Copyright: © 2023 Qing Wang. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All data files are available from Figshare (http://doi.org/10.6084/m9.figshare.20508816).
Funding: The author received no specific funding for this work.
Competing interests: The authors have declared that no competing interests exist.
Introduction
For decades, supplier selection has been an extensive research topic [1–3]. Often, it combines with order allocation if one supplier cannot supply all the quantities [4, 5]. It associates with quantity discounts when suppliers offer discount schemes. Today’s markets are changing, and customer tastes vary and are difficult to predict. Manufacturers should meet unique needs of color, shape, material, decoration, and others. Flexibility in choosing suppliers is necessary to meet specific requirements. Quick decisions are desired to save time, speed up order response, and improve the user experience.
In the era of big data, massive amounts of data are collected, stored, transferred, and processed [6–8]. Leveraging big data helps companies to improve the quality and agility of decision-making and brings great business value. With big data, suppliers can be selected from a much larger pool. More alternatives are provided to the decision-makers, which results in higher profits or lower costs. Large-scale models are developed. For NP-hard models, such as knapsack problems, conventional algorithms suffer from the "curse of dimensionality". They take long solution times or are unable to solve within a reasonable time limit–one or several hours.
The fractional knapsack problem formulates the original supplier selection and order allocation problem. The greedy algorithm is efficient and solves it to optimality. The knapsack formulation becomes more complex when setup costs and various quantity discounts are involved. Whether the greedy algorithm can still provide good solutions becomes questionable.
This paper studies flexible supplier selection and order allocation in the big data era with setup and four quantity discount types. A two-layer framework makes use of big data and screens technical requirements. The proposed nonlinear integer programming model is a variant of the knapsack problem. Three greedy algorithms are progressively developed by leveraging an underlying mechanism and utilizing two sorted lists. The lists use actual unit costs to balance the setup costs and different quantity discounts. The main contributions of this paper are twofold:
It proposes a nonlinear integer programming model that formulates four common types of quantity discounts. It fills a gap in the literature where models consider a limited number of discount types. These models are not suitable for practical applications.
The exact algorithms encounter the “curse of dimensionality” for solving the proposed model. To fully exploit the big data, this paper develops three greedy algorithms capable of solving very large-scale problems (up to 100000 suppliers) in seconds, with average optimality gaps of 0.1026%, 0.0547%, and 0.0234%.
The remaining paper is organized as follows. Section 2 reviews the relevant literature. Section 3 gives the formulation of the model. Section 4 presents the greedy algorithms for solving the large-scale model. A numerical study is provided in section 5, and computational experiments are conducted in section 6. Section 7 concludes the paper.
Literature review
Flexible supplier selection
Flexible decisions are desired in supply chains to increase resilience and agility. [9] evaluated supply, manufacturing, and logistics flexibility with the developed stochastic programming model. [10] developed a flexible dynamic sustainable procurement model to cope with the uncertainties of global supply chains. [11] modeled flexible sustainable supplier selection decisions in multitier supply chains.
Supplier selection with quantity discount
A literature study on supplier selection with quantity discounts shows that most models have one type of quantity discount [12–14]. A few have two [15, 16]. It is not realistic, especially when the number of suppliers is large.
Different quantity discount schemes bear different cost structures. The no discount cost is linear, the incremental discount cost is concave, the all-units discount cost is discontinuous, and the carload discount cost is a convex polyline [17, 18]. Models formed with one or two types often cannot describe those with more. Supplier selection models are NP-hard [19]. As the number of discounts increases, modeling complexity and solution difficulty increase. Models with various discounts and their efficient solutions are demanded in the era of big data.
Knapsack problem
Knapsack problems are extensively studied combinatorial optimization problems with a wide range of applications [20, 21] and many variants of forms. Knapsack problems are NP-hard. [22, 23] reviewed variants of the knapsack problems and solution algorithms and heuristics, structured in two papers. 0–1 knapsack is the most traditional form. In the 0–1 knapsack, objects cannot be partitioned, i.e., either selected or not selected. In the fractional knapsack, part of an object can be selected. The greedy algorithm chooses greedily the part with the highest value, which can lead to an optimal solution [21]. Though it can solve the fractional knapsack problem perfectly, it is not an optimal algorithm for other knapsack problems.
Curse of dimensionality
When applying exact solution algorithms like dynamic programming or branch-and-bound to solve knapsack or other combinatorial optimization problems, the “curse of dimensionality” will occur. The “curse of dimensionality”, introduced by Bellman for dynamic programming, indicates the exponential growth of hypervolume as a function of dimensionality [24, 25]. It leads to an exponential growth of the solution time as the dimensions of the problem increase. For large problem sizes, it may not be possible to solve within a reasonable time limit (one or several hours).
[26] proposed a branch-and-bound algorithm for solving an integer quadratic multi-knapsack problem. Table 1 in the paper shows that when the problem dimensions m and n are increased from 100 to 2000, the solution time of the four exact algorithms increases rapidly from less than/greater than one second to more than the time limit of 3 hours. [27] developed a Lagrangian-based branch-and-cut algorithm for the interval min-max regret generalized assignment problem. Table 3 in the paper shows that when the dimension m increases to 10 and n increases to 80, the problem soon becomes unsolvable by the exact algorithms—Benders-like decomposition, basic branch-and-cut, and Lagrangian-based branch-and-cut, within a time limit of 3600 seconds. [28] proposed an exact branch-and-bound algorithm for the quadratic combinatorial optimization problem. Table 4 in the paper shows the very fast increase in the computation time for three solvers and two algorithms, from a few seconds to several hours or over the time limit, when the number of variables increases from 180 to 420.
Greedy algorithm
The greedy algorithm is a solution scheme that greedily makes the best choice at the current stage. It can solve the fractional knapsack problem to optimality. Greedy algorithms obtain solutions efficiently, but the quality doesn’t guarantee. [29] proposed greedy algorithms to solve two-dimensional knapsack problems with binary weights. [30] developed greedy algorithms for linear and quadratic knapsack problems. Other than knapsack problems, greedy algorithms apply in many areas. [31] presented a greedy algorithm for the maximization of sequence submodular functions. [32] developed a greedy policy for the dynamic information provision model. [33] applied greedy algorithms to solve the single-demand facility location problem.
Flexible supplier selection and order allocation in the big data era with various quantity discounts
The two-layer framework for flexible supplier selection and order allocation in the big data era is shown in Fig 1.
Layer 1: Supplier screening
When a customer order arrives, its technical requirements are identified. These requirements may include specific technologies, delivery lead times, delivery reliability, quality standards, services to provide, materials to use, the durability of the product, sustainability issues, etc. The requirements are converted into codes to match the data stored in the database. For instance, technology is stored in the database by its code Tk. The order requires technologies T3 and T10. Suppliers must possess the technologies to fulfill the order. Data like the percentage of late delivery and average lead time are recorded and updated after the order fulfillment. In case of urgent orders, suppliers with short average lead times satisfy. Suppliers are screened to obtain a list of suppliers that meet all the requirements.
Layer 2: Supplier selection and order allocation
The list of suppliers is grouped by the pricing schemes and numbered. The supplier data form the model parameters, for instance, Si, cij. Combining them with the model stored in the system’s model library produces the mathematical programming model to be solved. A nonlinear integer programming model is proposed in this paper. Three algorithms, SGDQG, IGDQD, and RGDQD, are proposed for solving the model efficiently. The proposed model and the algorithms are elaborated below.
After order fulfillment, the performance data are used to update the database.
Model formulation
Four common pricing schemes are formulated, namely no discount, all-units discount, incremental discount, and carload discount. The cost structure for each discount type is shown in Figs 2–5. The horizontal axis of the figures denotes the order quantity and the vertical axis denotes the cost. The indices, parameters, and decision variables used in the figures and models are listed in Tables 1–3. First, the model for each type is formulated separately.
No discount case. The model for the no-discount case is formulated as follows.
The objective function (1) minimizes the total cost, which is composed of the setup cost and the unit cost. Constraint (2) ensures that the total order quantity is equal to Q. Constraint (3) restricts the order quantity xi of supplier i to be between 0 and the supplier capacity bi1 if supplier i is selected, or 0 if supplier i is not selected. Constraints (4) and (5) ensure that decision variables xi and yi are integer and binary variables, respectively.
All-units discount case. The model for the all-units discount case is as follows.
The objective function (6) minimizes the summation of the setup costs and the unit costs. Constraint (7) equals the summation of quantities xij to the total order quantity Q. Constraint (8) denotes if interval j for supplier i is selected, i.e., aij =1, the purchase amount xij must be between the interval boundaries, otherwise, xij = 0. Constraint (9) denotes if supplier i is selected, one and only one discount interval for the supplier must be selected. Constraint (10)-(12) restrict aij, yi and xij to be integer and binary variables.
Incremental discount case. The model for the incremental discount case is as follows.
The objective function (13) minimizes the summation of the setup costs and the unit costs for the incremental discount. The constraints are the same as in the all-units discount case.
Carload discount case. The model for the carload discount case is as follows.
The objective function (14) minimizes the summation of the setup costs and the unit costs for the carload discount case, the odd and even intervals have different cost structures as shown in Fig 5, and are calculated separately. The constraints are the same as in the all-units discount case.
Supplier selection and order allocation model with setup and various quantity discounts
The above models are integrated into one model that incorporates all four types of pricing schemes for supplier selection.
Problem (P) is a nonlinear integer programming model. It can be perceived as some combination of 0–1 knapsack, multiple choice knapsack, fractional knapsack, and knapsack with setup problems, and is thus NP-hard. Problem (P) has decision variables and
constraints.
Solution methodology
Solution methodologies are sought for to solve problem (P) efficiently. The following proposition shows a property of the solution of (P).
Proposition 1: There exists an optimal solution to problem (P) in which at most one quantity value xi or xij lies strictly between two breakpoints.
Proof: Denote xi or xij by qi, suppose there are two suppliers d and e that get non-breakpoint orders, and
. Suppose cd≤ce. Then switching a partial order
from supplier e to supplier d can reduce the total cost by at least (ce−cd)δ. After switching, either supplier d or supplier e get breakpoint order.
With Proposition 1, since the amount Q is an integer, and all boundary values are integers, the only non-boundary value must be an integer. The integer constraints of xi’s and xij’s, i.e. (22) and (23), can be relaxed. From Proposition 1, the complexity of (P) may be reduced. A direct way of solving (P) is to iterate through all fractional intervals and compare the objective values. Let f denote the index of the supplier with the non-boundary interval. Problem (P) can be transformed into problem (Pf).
(24)
(25)
(26)
(27)
(28)
(29)
(30)
(31)
(32)
(33)
cij in objective function (15) is changed to
such that
and
. By substituting the equality constraints (28), (29), (32) and (33) into the objective function (24), the problem sizes of (Pf) are reduced to
or
decision variables and
or
constraints, but the problem needs to be solved
times as f iterates through the intervals.
Though problem sizes reduce from (P) to (Pf), (Pf) is a variant of the knapsack problem, and is thus NP-hard. The complexity needs to be further reduced.
Large-scale problem
In the big data era, great vendor information may be utilized, and large-scale problem (P) needs to be solved. However, exact solution algorithms suffer from the “curse of dimensionality”. This section seeks efficient solution methodologies for the large-scale problem (P).
As in the “Literature review” section, the greedy algorithm solves the fractional knapsack problem to optimality and is applied to solve other knapsack problems. Whether it can be a good algorithm for (P) will be investigated. To design a greedy algorithm for (P), how to make the greedy choice is the key issue. The fractional knapsack greedily selects the unit revenue. (P) aims to find the lowest total cost, dividing by the order quantity Q, which is the same as finding the lowest average cost. Filling the knapsack greedily with the lowest actual cost per unit accounting for all the costs becomes a good choice.
For the fractional knapsack, the unit revenues can be calculated and sorted. The unit revenue doesn’t change for a single object, no matter how many quantities of the object are selected. This makes the greedy selection possible. But with setup costs and different discounts, the actual cost per unit varies with quantity, making the sorting tedious.
Proposition 1 shows that in an optimal solution to (P), all but one of the quantity values are boundary values. The actual unit costs of boundary values can be calculated and sorted. Greedy choices can be made based on these costs. This forms List 1 below. The ordering of actual unit costs for the remaining quantity forms List 2. The proposed greedy algorithms in the following are based on these two sorted lists.
Greedy algorithms
The actual unit cost is defined as the average unit cost of ordering a single item, considering all the costs incurred. The calculation for a particular xi or xij differs for different quantity discount types, and is given in Eqs (34)–(38).
For incremental discount,
(36)
For carload discount, for quantities between odd intervals,
(37)
for quantities between even intervals,
(38)
List 1 and List 2 are obtained as follows.
List 1: Rank ascendingly the actual unit costs of boundary quantity bij’s for all supplier i and interval j, and mark the respective supplier no. i, interval no. j, and boundary quantity bij.
List 2: Rank ascendingly the actual unit costs of a certain quantity ΔQ for all supplier i and interval j, and mark the associated supplier no. i, interval no. j, where the suppliers no. is not selected in List 1. If the quantity is in the range of an interval j, i.e., bi,j−1<ΔQ<bij, use the quantity ΔQ to calculate the actual unit cost for interval j; If the quantity exceeds the range of an interval j, i.e., bij≤ΔQ, use the interval’s boundary value bij to calculate the actual unit cost for interval j; Otherwise if the quantity is below the range of interval j, i.e., bi,j−1≥ΔQ, the actual unit cost for interval j is assigned a large number, e.g., 1000, and interval j will not appear in the rank of List 2. Mark the associated quantity of each ranked item with the quantity to calculate the actual unit cost.
In List 2, interval j’s whose upper boundary bij less than ΔQ are also taken into account, since some combination of intervals up in the list may give better results.
First, the simple greedy algorithm is proposed. The procedure is as follows.
Algorithm SGDQD: Simple greedy algorithm for supplier selection with different quantity discounts
Step 1. Obtain List 1 with no repetitive supplier i until adding the next item will exceed Q. Obtain the sum of boundary quantity bij’s in the list as Q′. Calculate ΔQ = Q−Q′.
Step 2. Obtain List 2 with one item for the quantity ΔQ. Let q be the associated quantity, calculate ΔQ = ΔQ−q.
Step 3. Repeat Step 2 until ΔQ = 0.
SGDQD is simple, and a spreadsheet can be used to find solutions. Place the supplier data in the cells of the spreadsheet and use "function" to get sorted lists—List 1 and List 2. Regardless of the size of the supplier set, SGDQD only needs to get the top “n” suppliers of List 1 and the topmost supplier of List 2. Solvers need to solve a mathematical programming problem with large dimensions, which is complex and time-consuming. SGDQD can attain the optimal solution for some problems, but improvements are possible. More calculations and comparisons are performed. The following improved greedy algorithm are obtained.
Algorithm IGDQD: Improved greedy algorithm for supplier selection with different quantity discounts
Step 1. Obtain List 1 with no repetitive supplier i until adding the next item will exceed Q. Obtain the sum of boundary quantity bij’s in the list as Q′. Calculate ΔQ = Q−Q′.
Step 2. Obtain List 2 for the quantity ΔQ until the associated quantity of an item equals ΔQ. Form trajectories for the items in List 2. The last item is an end of trajectory. If a single item is obtained, the associated trajectory ends.
Step 3. For items except the last item, let q be the associated quantity, calculate ΔQ = ΔQ−q, and repeat Step 2.
Step 4. Calculate the total cost for each trajectory in the obtained tree and select the minimum. The associated trajectory gives the optimal solution.
Step 1 in IGDQD is the same as in SGDQD, i.e., List 1 is used the same to fill Q. More calculations are performed on List 2 to obtain the combination with the lowest cost for the quantity gap of Step 1. An example of the calculations is given using randomly generated Data20. Table 4 shows List 1 for Q = 300000 without allowing duplicate supplier numbers. The quantity gap of Step 1 is 24900. Then, as shown in Table 5, List 2 is obtained for the amount 24900. The trajectories are formed as shown in Fig 6. List 2’s for the amounts 3100, 10300, and 5700 are obtained to form the rest of the trajectories in Fig 6. In this case, the uppermost trajectory gives the minimum total cost.
In rare cases, List 2 has many small quantities. Many replicated trajectories are present. To further simplify, combinations of the quantities are calculated and compared.
IGDQD improves the solutions to some extent by exploring List 2 more. In fact, List 1 can also be explored to obtain better solutions. The refined greedy algorithm improves IGDQD by performing calculations on both List 1 and List 2.
Algorithm RGDQD: Refined greedy algorithm for supplier selection with different quantity discounts
Step 1. Obtain List 1. Select and add boundary quantity bij’s in List 1 by order with no repetitive supplier no. allowed until adding the next item will exceed Q. The last selected quantity has the rank no. l in List 1. Obtain the quantities selected whose rank no. is less than l−d1+1, enumerate and select the quantities in List 1 from the rank no. l−d1+1 to l+d2 with no repetitive supplier no. i such that Q is not exceeded. Obtain the sum of boundary quantity bij’s for each combination of selected items as Q′. Calculate ΔQ = Q′−Q.
Step 2. Obtain items in List 2 of quantity ΔQ in order until the associated quantity of an item equals ΔQ. Form trajectories for the items in List 2. The last item is an end of the trajectory. If a single item is obtained, the associated trajectory ends.
Step 3. For items except the last item, let q be the associated quantity, calculate ΔQ = ΔQ−q, and repeat Step 2.
Step 4. Calculate the total cost for each trajectory in the obtained tree and select the minimum. The associated selections along the trajectory form the solution for a combination in Step 1.
Step 5. Compare the minimums in Step 4 for all combinations in Step 1 and obtain the minimum solution.
Step 2–4 in RGDQD are the same as step 2–4 in IGDQD that the operations on List 2 are identical. Step 1 enumerates different combinations in List 1. As d1 increases to l, total enumeration of l+d2 quantities is performed. As d1 and d2 decrease to 0, Step 1 becomes the Step 1 in IGDQD. The larger the parameters d1 and d2, the more combinations and the better the result. The extreme case is d1 = l and d2 = N−l so that all suppliers are considered. However, as d1 and d2 increase, the number of enumerated combinations also increases. The selection of d1 and d2 takes into account the average number of suppliers in an order for a particular Q, and the tradeoff between optimality and simplicity of the computations.
Numerical study
A numerical example is presented for illustrative purposes. An order of 52 products arrives. After screening the technical requirements, 7 suppliers are available for selection. These suppliers are grouped by discount types, numbered 1 to 7, and denoted by S1 to S7. S1 offers no discount, S2, S3, and S4 offer the all-units discount, S5 and S6 with incremental discount, and S7 with carload discount. For simplicity and without loss of generality, all suppliers have a setup cost of $10 and a capacity of 20 units. Except for S1, the rest suppliers have 2 discount intervals. All boundary quantities of the first interval are 10 units. The unit costs are $2.3, $2.2, $1.9, $2.4, $2.1, $2.3, and $2.3 for S1 to S7. The discounted costs are $1.7, $1.8, $2.2, $1.9, $2, and $2.2 for S2 to S7. The actual unit cost is denoted by “auc”.
First, SGDQD is applied. List 1 is obtained: S2, 20 units, auc 2.2; S3, 20 units, auc 2.3. ΔQ = 52-20-20 = 12. One item list 2 of 12 is obtained: S7, 10 units, auc 2.84. ΔQ = 12–10 = 2. One item list 2 of 2 is obtained: S5, 2 units, auc 7.1. ΔQ = 2–2 = 0 and the algorithm terminates. The solution is: S2, 20 units; S3 20 units; S7 10 units; S5 2 units. The cost is $132.6. Then, IGDQD is applied. List 1 is the same. List 2 of 12 is obtained: S7, 10 units, auc 2.84; S5, 12 units, auc 2.9. Two trajectories are formed. The upper is the same as SGDQD. The costs of $132.6 and $124.8 are compared. The solution is S2, 20 units; S3 20 units; S5 12 units; cost $124.8. RGDQD is applied with d1 = 1 and d2 = 1. The enumerations are: S2, 20 units, S3, 20 units; S2, 20 units, S5, 20 units. The first combination is the same as in IGDQD. For the second combination, obtain List 2 of 12: S3, 12 units, auc 2.63. The cost is $125.6. The first combination gives lower cost and the solution is the same as IGDQD. The solutions by IGDQD and RGDQD are optimal.
Tests instances
This section conducts computational experiments to show the performance of the proposed algorithms in solving the large-scale model (P).
Random data generation
Supplier data are retrieved from the database and after supplier screening, 30% of suppliers offer no discount, 30% all-units discount, 30% incremental discount, and 10% carload discounts. The supplier data in the study are generated randomly as shown in Table 6.
In Table 6, the discount rate α is generated as follows. The range of discount rate is [0.88, 0.99]. This range is divided by Mi−1. Uniform dividing points are obtained which are rounded to two decimal points, e.g., for Mi = 4, the dividing points are 0.92 and 0.95. α values are randomly generated uniformly between the dividing points, e.g., for Mi = 4, discount rates α’s are generated uniformly between [0.95,0.99], [0.92,0.95], and [0.88,0.92].
Solution quality of the proposed algorithms
20 instances of data of 1000 suppliers are randomly generated. Using the randomly generated data, 120 problems are formulated with different Q’s, i.e., Q = 3000, 15000, 75000, 150000, 300000, 600000. SGDQD, IGDQD, and RGDQD are applied to solve the 120 problems. RGDQD is applied with parameters d1 set to 1 and d2 set to 2.
The solutions are compared with the optimal solutions to evaluate the solution quality of the algorithms. Tables 7–12 give the solutions. The optimal solutions are in bold.
The optimality gap is calculated as follows,
Figs 7–12 show the improvements in solution quality with narrowing optimality gaps solved by SGDQD, IGDQD, and RGDQD. Tables 13–15 show the relevant statistics.
Tables 7–12 and Figs 7–12 show that the solutions obtained by the proposed greedy algorithms are optimal or very close to optimal. In the statistics of Tables 13–15, 59.2%, 70%, and 80% of the solutions are optimal solutions solved by SGDQD, IGDQD, and RGDQD respectively. The average optimality gaps are only 0.001026, 0.000547, and 0.000234 for SGDQD, IGDQD, and RGDQD respectively, with standard deviations of 0.002363, 0.001569, and 0.000833 respectively. The solution quality improves from SGDQD to IGDQD and from IGDQD to RGDQD.
In the simulations, d1 and d2 in RGDQD are set to small values, 1 and 2. Results show that by simple enumeration, where both d1 and d2 bear small values, improvements in solution quality are obvious. Increasing these values gives better results. For instance, if d1 is set to 7 and d2 is set to 6, 90% of the solutions will be optimal.
Further improvements.
Further improvements of RGDQD are possible through some more observations. For instance, it is often observed that for the carload discount, although an even interval may have a smaller actual unit cost than the next odd interval, the quantity of the odd interval constitutes a larger proportion of Q. Choosing it often yields better results. Thus, the following step may be added to Step 1 of RGDQD:
Step 1+. If both an even interval and its next odd interval of carload discount are present above the order l+d2 in List 1, and the even interval is above the odd interval, then the even interval can be replaced by its next odd interval and obtain the combinations.
The effect of adding Step 1+ to RGDQD is tested for Q = 600000. The resulting algorithm is termed RGDQD’. Fig 13 shows the improvements in solution quality solved by RGDQD’ compared with the original RGDQD.
The efficiency of the proposed algorithms
To test the efficiency of the proposed algorithms, the number of suppliers N is set to 1000, 10000, and 100000. 10 instances of each are generated. The order quantity Q is set to 10000, 100000, and 1000000. The algorithms are coded in Python and the simulation is conducted on a personal computer, Intel core i5-7200U, 2.5GHz 2.71GHz, 8GB RAM. The solution time is the average of 10 runs with generated instances for each combination of the two parameters. The average solution time of the three proposed algorithms is listed in Table 16.
Table 16 shows that with N = 1000, (P) can be solved in centiseconds, and with N = 10000, (P) can be solved in less than half a second. Even with a very large dimension–up to 100000 suppliers, (P) can still be solved in several seconds. Table 16 also shows that the running time increases slowly with the increase of the order quantity Q.
Comparison with the genetic algorithm
The genetic algorithm (GA) has been an excellent and widely applied solution mechanism for combinatorial optimization problems [14, 22, 23]. Compared to traditional algorithms, GA tends to solve problems more efficiently. Although GA provides a general evolution mechanism, the design of GA is problem specific. [14] applied GA to solve the proposed NP-hard mixed integer nonlinear programming model for the multiple sourcing problem with supplier failure risk and quantity discount. The problem in this study has only 10 suppliers. [34] designed a hybrid genetic algorithm to solve the two-dimensional single large object placement problem and showed its good performance in terms of solution time and quality compared to other algorithms. In [35], a variable-grouping based GA for large-scale integer programming was proposed that outperformed the standard GA. Quadratic knapsack problems with less than 400 variables were solved with a solution time of up to 9726 seconds.
To compare with the proposed greedy algorithms, attempts of applying GA to solve the large-scale (P) are made with Geatpy2—the genetic and evolutionary algorithm toolbox for Python with high performance. Geatpy2 is a GA toolbox that provides GA templates with adjustable operators and delivers leading solution performance. Details of the toolbox can be found at http://geatpy.com/ and resources can be found in the geatpy-dev/geatpy directory on GitHub. For all the trials, feasible solutions are not able to be found. Applying GA to solve large dimensional nonlinear integer programming problems remains a challenge. The proper design of GA to solve (P) becomes another interesting research topic.
The proposed greedy algorithms provide a very good solution substitute when other algorithms are unable to find optimal or even feasible solutions for the large-scale (P).
Sensitivity analysis
Sensitivity analysis is performed to study the impact of parameters on solution quality and the final solution.
Sensitivity analysis of the order quantity Q on solution quality.
A sensitivity analysis of Q on solution quality is performed. Holding other values constant, Tables 13–15 show that the optimality gaps generally decrease as Q increases. But the percentage of optimal solutions also decreases with increasing Q.
For very small Q, i.e., Q = 3000, all solutions are optimal. Small or large Q corresponds to whether the average number of suppliers in the optimal solution is small or large, refer to Fig 14. For Q = 3000, the optimal solution has only one supplier. (P) becomes a supplier selection problem with no order allocation. Table 13 shows that when SGDQD selects only one supplier, it obtains optimal solutions for all cases. The other two algorithms are improvements of SGDQD and thus both obtain optimal solutions.
As Q increases to 15000, it is still quite small and there are 1.4 suppliers in the optimal solution on average. Only 60% of the solutions achieve optimality for SGDQD. All the solutions are optimal if only one supplier is selected. If more suppliers are involved, all the solutions obtained are nonoptimal. The average optimality gap of these solutions is 0.5247%, compared to 0.4769%, 0.3718%, 0.0887%, and 0.0400% for other order quantities. With a small Q, the selection of non-optimal suppliers contributes larger to the solution. This can be tackled by IGDQD. 95% of the solutions are optimal for IGDQD and RGDQD. With more calculations performed, optimal solutions are obtained generally. The only nonoptimal solution has a small optimality gap.
With the further increase of Q, the solution quality generally improves. The trend becomes more pronounced when the order quantity becomes large. The reason is that as more suppliers join to fill Q, a larger percentage of suppliers are the same as in the optimal solutions. It results in smaller optimality gaps. The improvements by IGDQD and RGDQD are not as large for Q = 15000.
Sensitivity analysis of parameters on the final solution.
By changing the parameters in the numerical example, the selected suppliers may change. Table 17 shows the change of parameters and the corresponding optimal solutions.
Analyzing the solutions of the computational tests, it is found that intervals with low unit costs are selected. Add the same amount to the randomly generated setup costs Si, and keep the other data constant, the selected suppliers’ intervals change for large amounts. It means that not only the unit costs take effect. By varying the scale of the other parameters, solution analyses show that actual unit costs are most important in the selection. The capacity of the supplier, the number of discount intervals, the discount rate, the setup cost alone, or the unit cost alone is not.
Managerial insights
The results lead to the following managerial insights.
First, from the results in the “Tests instances” section, the proposed algorithms can assist managers in making timely decisions in the face of a huge supplier set. Supplier big data can be fully utilized. For very large-scale model solving, efficient algorithms are demanded.
Second, flexible supplier selection decisions are necessary to obtain the lowest cost allocation. The supplier screening process obtains different suppliers. The selected suppliers change with model parameter changes. In the big data era, where data are dynamically recorded, fixing the set of suppliers may not be a good choice.
Third, the order quantity affects the solution quality of the algorithms. When the average number of suppliers in the order are 1, the solutions by SGDQD are generally optimal. When the number is between 1 and 2, the solutions by IGDQD are generally optimal. If the order quantity is very small, SGDQD is the suitable algorithm with the greatest efficiency. If the quantity is small, IGDQD may be the choice. RGDQD gives the best solution quality. If the company receives orders with large amounts, RGDQD can be chosen for selection.
Last, using the actual unit costs to make greedy selections, the average optimality gaps of proposed greedy algorithms are only 0.1026%, 0.0547%, and 0.0234%. Moreover, sensitivity analysis shows that the actual unit costs play a major role in selecting the optimal suppliers. Therefore, if the company purchases internationally, which implies high setup costs (including transportation costs), the setup costs become more important in selecting suppliers. It is essential to consider not only the unit costs but also the suppliers’ distance and mode of transportation. If the suppliers are near the location of the company and setup costs are low, the company may select suppliers with low unit costs.
Conclusion
In this paper, greedy algorithms are proposed to solve the formulated large-scale model with four quantity discount types for flexible supplier selection and order allocation in the big data era. Three greedy algorithms—SGDQD, IGDQD, and RGDQD are developed incrementally from the previous algorithm, utilizing more information and obtaining better performance. RGDQD is not yet the optimal algorithm, though it approaches optimality by increasing its parameters d1 and d2. The run time of the algorithm increases with the increase of the parameters. d1 and d2 in the study are set to small values– 1 and 2. For future research, the tradeoff between optimality and run time may be studied to find the best d1, d2 assignment for real applications. Existing models in the literature may be reformulated to take into account more quantity discounts. Efficient methodologies for solving large-scale models should be investigated in the era of big data.
References
- 1. Chai J, Ngai EWT. Decision-making techniques in supplier selection: Recent accomplishments and what lies ahead. Expert Syst Appl. 2020;140: 112903.
- 2. Ho W, Xu X, Dey PK. Multi-criteria decision making approaches for supplier evaluation and selection: A literature review. Eur J Oper Res. 2010;202: 16–24.
- 3. Wetzstein A, Hartmann E, Benton jr.WC , Hohenstein N-O. A systematic assessment of supplier selection literature–State-of-the-art and future scope. Int J Prod Econ. 2016;182: 304–323.
- 4. Hamdan S, Cheaitou A. Dynamic green supplier selection and order allocation with quantity discounts and varying supplier availability. Comput Ind Eng. 2017;110: 573–589.
- 5. Hosseini S, Morshedlou N, Ivanov D, Sarder MD, Barker K, Khaled AA. Resilient supplier selection and optimal order allocation under disruption risks. Int J Prod Econ. 2019;213: 124–137.
- 6. Kaffash S, Nguyen AT, Zhu J. Big data algorithms and applications in intelligent transportation system: A review and bibliometric analysis. Int J Prod Econ. 2021;231: 107868.
- 7. Sestino A, Prete MI, Piper L, Guido G. Internet of Things and Big Data as enablers for business digitalization strategies. Technovation. 2020; 102173.
- 8. Yang L, Jiang A, Zhang J. Optimal timing of big data application in a two-period decision model with new product sales. Comput Ind Eng. 2021;160: 107550.
- 9. Almeida JF de F, Conceição SV, Pinto LR, de Camargo RS, Júnior G de M. Flexibility evaluation of multiechelon supply chains. Deng Y, editor. PLOS ONE. 2018;13: e0194050. pmid:29584755
- 10. Kaur H, Singh SP. Flexible dynamic sustainable procurement model. Ann Oper Res. 2019;273: 651–691.
- 11. Cui L, Wu H, Dai J. Modelling flexible decisions about sustainable supplier selection in multitier sustainable supply chain management. Int J Prod Res. 2021; 1–22.
- 12. Ghaniabadi M, Mazinani A. Dynamic lot sizing with multiple suppliers, backlogging and quantity discounts. Comput Ind Eng. 2017;110: 67–74.
- 13. Jackson JE, Munson CL. Common replenishment cycle order policies for multiple products with capacity expansion opportunities and quantity discounts. Int J Prod Econ. 2019;218: 170–184.
- 14. Meena PL, Sarmah SP. Multiple sourcing under supplier failure risk and quantity discount: A genetic algorithm approach. Transp Res Part E Logist Transp Rev. 2013;50: 84–97.
- 15. Cui L, Deng J, Wang L, Xu M, Zhang Y. A novel locust swarm algorithm for the joint replenishment problem considering multiple discounts simultaneously. Knowl-Based Syst. 2016;111: 51–62.
- 16. Alfares HK, Turnadi R. Lot sizing and supplier selection with multiple items, multiple periods, quantity discounts, and backordering. Comput Ind Eng. 2018;116: 59–71.
- 17. Burke GJ, Carrillo J, Vakharia AJ. Heuristics for sourcing from multiple suppliers with alternative quantity discounts. Eur J Oper Res. 2008;186: 317–329.
- 18. Jucker JV, Rosenblatt MJ. Single-period inventory models with demand uncertainty and quantity discounts: Behavioral implications and a new solution procedure. Nav Res Logist Q. 1985;32: 537–550.
- 19. Rafigh P, Akbari AA, Bidhandi HM, Kashan AH. A sustainable supply chain network considering lot sizing with quantity discounts under disruption risks: centralized and decentralized models. J Comb Optim. 2022;44: 1387–1432. pmid:36062162
- 20. Changdar C, Pal RK, Mahapatra GS, Khan A. A genetic algorithm based approach to solve multi-resource multi-objective knapsack problem for vegetable wholesalers in fuzzy environment. Oper Res. 2020;20: 1321–1352.
- 21. Jacko P. Resource capacity allocation to stochastic dynamic competitors: knapsack problem for perishable items and index-knapsack heuristic. Ann Oper Res. 2016;241: 83–107.
- 22. Cacchiani V, Iori M, Locatelli A, Martello S. Knapsack problems—An overview of recent advances. Part I: Single knapsack problems. Comput Oper Res. 2022;143: 105692.
- 23. Cacchiani V, Iori M, Locatelli A, Martello S. Knapsack problems—An overview of recent advances. Part II: Multiple, multidimensional, and quadratic knapsack problems. Comput Oper Res. 2022;143: 105693.
- 24.
Bellman R. Dynamic Programming. Princeton: Princeton University Press; 1957.
- 25.
Bellman R. Adaptive control processes: a guided tour. Princeton: Princeton University Press; 1961.
- 26. Quadri D, Soutif E, Tolla P. Exact solution method to solve large scale integer quadratic multidimensional knapsack problems. J Comb Optim. 2009;17: 157–167.
- 27. Wu W, Iori M, Martello S, Yagiura M. Exact and heuristic algorithms for the interval min-max regret generalized assignment problem. Comput Ind Eng. 2018;125: 98–110.
- 28. Buchheim C, Traversi E. Quadratic Combinatorial Optimization Using Separable Underestimators. Inf J Comput. 2018;30: 424–437.
- 29. Gorski J, Paquete L, Pedrosa F. Greedy algorithms for a class of knapsack problems with binary weights. Comput Oper Res. 2012;39: 498–511.
- 30. Glover F. Advanced greedy algorithms and surrogate constraint methods for linear and quadratic knapsack and covering problems. Eur J Oper Res. 2013;230: 212–225.
- 31. Bernardini S, Fagnani F, Piacentini C. A unifying look at sequence submodularity. Artif Intell. 2021;297: 103486.
- 32. Renault J, Solan E, Vieille N. Optimal dynamic information provision. Games Econ Behav. 2017;104: 329–349.
- 33. Cheung S-S, Williamson DP. Greedy algorithms for the single-demand facility location problem. Oper Res Lett. 2017;45: 452–455.
- 34. Hadjiconstantinou E, Iori M. A hybrid genetic algorithm for the two-dimensional single large object placement problem. Eur J Oper Res. 2007;183: 1150–1166.
- 35. Hua Z, Huang F. A variable-grouping based genetic algorithm for large-scale integer programming. Inf Sci. 2006;176: 2869–2885.