We study a class of multiobjective bilevel programs with the weights of objectives being uncertain and assumed to belong to convex and compact set. To the best of our knowledge, there is no study about this class of problems. We use a worst-case weighted approach to solve this class of problems. Our “worst-case weighted multiobjective bilevel programs” model supposes that each player (leader or follower) has a set of weights to their objectives and wishes to minimize their maximum weighted sum objective where the maximization is with respect to the set of weights. This new model gives rise to a new Pareto optimum concept, which we call “robust-weighted Pareto optimum”; for the worst-case weighted multiobjective optimization with the weight set of each player given as a polytope, we show that a robust-weighted Pareto optimum can be obtained by solving mathematical programing with equilibrium constraints (MPEC). For an application, we illustrate the usefulness of the worst-case weighted multiobjective optimization to a supply chain risk management under demand uncertainty. By the comparison with the existing weighted approach, we show that our method is more robust and can be more efficiently applied to real-world problems.
Shanghai Institutions of Higher LearningTP2014043National Natural Science Foundation of China71401107715710551. Introduction
Multiobjective bilevel programs are a generalization of the scalar criterion bilevel programs [1] and used to model situations where decision-makers at both upper and lower levels, called players, take actions by considering their individual multiple objectives [2–4]. Many papers have been published in the last two decades about (multiobjective) bilevel programs because of their possible applications [1, 4, 5].
In this paper, we consider a special class of multiobjective bilevel programs with the weights of objectives being uncertain and assumed to belong to convex and compact set. To the best our knowledge, there is no study about this class of problems and the existing approaches in multiobjective bilevel programs can not solve this class of problems. Therefore, new methods are necessary to be proposed. In this paper, we present a robust-weighted approach to solve this class of multiobjective bilevel programs. In our model, we suppose that players at both upper and lower levels are risk-averse and a player uses the robust optimization approach to manage the weight uncertainty, assuming that the other player is a robust optimizer as well. Note that the robust optimization approach is not concerning problem data or the other players’ strategies but concerning the weights to the objectives. We note that this model is closely related to robust optimization [6, 7]. It can be seen as a natural extension of the robust modeling technique by replacing a decision variable with a general deterministic function [6, 7]. However, our model is also different from the studies in [6, 7] that we study the bilevel problem, while the model in [6, 7] is single level.
Multiobjective optimization has been extensively studied and a number of different approaches have also been proposed [8]. In multiobjective problems, since there are several competing objectives to be considered and it is not possible to simultaneously optimize all objectives, a commonly accepted approach for coping with this setting is the weighted approach by assigning a nonnegative weight by considering the importance of the corresponding objective function. And then the decision-maker can make a decision by optimizing a weighted sum objective [9–12]. A weighted Pareto optimal point can be obtained. In the setting of multiobjective models including input from multiple experts, a decision can often be obtained only after considering different opinions from different experts [11].
However current approaches to model-based multiexpert multiobjective decision-making have several shortcomings. How to choose a solution among (possibly infinite) generated Pareto optimal solutions? There is no guidance. In multiexpert optimization, the relative weights given by experts can differ significantly as experts with differing opinions often assign different relative weights to objectives. As shown in applications, the weights are not known in advance and the player has to choose them. Ambiguity often exists in the choice of the weights to objectives, as it is not easy to decide relative weights for each objective. In addition, as shown in multiobjective optimization in the literature, relative weights given by the same decision-maker may rely on the elicitation methods [12, 13]. In the bilevel setting considered in this paper, it is more difficult for the decision-maker at the upper level to exactly assume the weights of the decision-maker’s objectives at the lower level. Therefore, in this paper, we focus on a special class of multiobjective bilevel programs with the weights of objectives being uncertain and assumed to belong to convex and compact set. There is no study about this class of problems and there no methods in multiobjective bilevel programs can solve this class of problems. It is necessary to provide a new approach to cope with these issues.
Hence, our motivation to utilize a worst-case weighted approach is that it provides an alternative way to deal with the weights ambiguity. Furthermore, if each player in multiobjective bilevel problems chooses the worst-case weighted approach, then we show that the computation for robust-weighted Pareto optimum, with the choice of polytope weight set for every player, is reformulated as a solution to mathematical programing with equilibrium constraints (MPEC) which can be solved by the existing methods (e.g., the sequential quadratic programing (SQP) methods).
Though multiobjective bilevel programs have not attracted much attention in the literature, there are some interesting potential applications. One example is the multicriteria Stackelberg competition of a supply chain containing a manufacturer who supplies a set of products to a risk-averse retailer satisfying uncertain consumer demand. In our model, the manufacturer decides on the quantity for each product so as to maximize the profit and minimize the cost simultaneously by forecasting the order quantity from the retailer and the wholesale prices resulting from market clearing conditions. The retailer also decides their wholesale market order quantity for each product in order to simultaneously maximize the mean profit and minimize the standard deviation of the profit.
Since the possible applications, multiobjective bilevel programs have attracted some attention. For example, Yin [14] considers a multiobjective bilevel model for transportation planning and management problems, where genetic algorithms are proposed to solve the resulting model. Deb and Sinha [15] present evolutionary algorithms for solving multiobjective optimization problems. Pieume et al. [16] develop two methods for solving bilevel linear multiobjective optimization problems. Eichfelder [17] gives a solution method for solving nonlinear multiobjective bilevel problems based on a scalarization approach and the sensitivity analysis of adaptive parameters. To reduce traffic congestion as well as to improve workforce productivity, a bilevel multiobjective model is proposed for an urban logistics metropolis [18]. A discrete approach is proposed to solve the resulting model. The convergence result and numerical tests are also provided. Some special multiobjective bilevel programs with linear objectives are also studied extensively. For example, a linear bilevel optimization problem with multiple objectives at the upper level is studied in [19], where the original problem reduces to solving a series of linear bilevel problems with a single objective function at each level; a linear bilevel optimization problem with multiple objectives at the lower level is studied in [20], where the original problem is reformulated as an optimization problem over a nonconvex region given by a union of faces of the polyhedron defined by all constraints.
Our study is different from the above papers in that we focus on a special class of multiobjective bilevel nonlinear programs—the weights of objectives are uncertain and assumed to belong to convex and compact set. To deal with this uncertainty, a robust optimization approach is used. Then we reformulated the original problem as a robust bilevel nonlinear optimization problem. To the best of our knowledge, this is the first paper to consider these special multiobjective bilevel problems.
We feel that the primary contributions of this paper are as follows. We deal with a special class of multiobjective bilevel programs, where the weights of objectives are uncertain and assumed to belong to convex and compact set. To the best of our knowledge, this is the first paper to consider such multiobjective bilevel programs. In our model, we suppose that players at both upper and lower levels are risk-averse and a player uses the robust optimization approach to manage the weight uncertainty, assuming that the other player is a robust optimizer as well. As we know that, there is no method to solve such problem. We propose a worst-case weighted approach for solving such multiobjective bilevel programs, extending the notion of robust-weighted multiobjective optimization models [21] to multiobjective bilevel problems. We show that a robust-weighted Pareto optimal point can be calculated by solving MPEC when the weight sets chosen by the players are polytopes. We note that the MPEC has been extensively studied and can be solved by the sequential quadratic programing (SQP) method [22]. We demonstrate the usefulness of the worst-case weighted method in a bilevel multiobjective competition problem within supply chain. We note that, compared with the existing weighted approach [9], our method is more “robust" in that there are different solutions for choosing different weights by using weighted approach [9], but there is a unique solution by using our approach.
Throughout this paper we use the following notations. R is the set of real numbers; R+ denotes the set of nonnegative real numbers; and R++ is the set of strictly positive real numbers. Let (1)Rm=R×⋯×R︸m;R+m=R+×⋯×R+︸m;R++m=R++×⋯×R++︸m.For any u,v∈Rm, denote(2)v≥uresp.v≤u⟺v-u∈R+mresp.u-v∈R+m⟺vj-uj≥0resp.vj-uj≤0,j∈I;u≻vresp.v≺u⟺v-u∈R++mresp.u-v∈R++m⟺vj-uj>0resp.vj-uj<0,j∈I;v⪰uresp.v⪯u⟺v-u∈R+mresp.u-v∈R+m,u≠v⟺vj-uj≥0resp.vj-uj≤0,j∈I,u≠v,where I:={1,…,m} is the index set. Given any vector function h:Rn→Rm, by h∈C1(Rn,Rm), we indicate that h is a continuously differentiable function from Rn to Rm and we use ∇h(x)∈Rn×m to denote the gradient of the function h at x; for simplicity, in this paper we use superscript U (resp., L) to denote decision-makers in the upper level (resp., decision-makers in the lower level) parameters, decision variables, cost functions, and weights of payoff functions. We note that the different mechanisms (e.g., bilevel, equilibrium, multiobjective optimization, weighting, and Pareto) are used in this paper. To specially understand the links between all these components, we provide a figure (see Figure 1) to illustrate this.
Illustration of links about the different mechanisms used.
In Section 2, we review some basic concepts in multiobjective bilevel programs and then introduce a new concept of the robust-weighted Pareto optimum. We show that a robust-weighted Pareto optimal solution is also a weighted Pareto optimum. In Section 3, we show that when the weight sets are given as polytopes, the robust-weighted Pareto optimum can be cast as MPEC which can be efficiently solved by SQP methods. In Section 4, we show the efficiency of our approach by an example for analyzing a multiobjective bilevel competition problem within supply chain. Finally, we conclude this paper in Section 5.
2. Preliminaries
In this paper, we consider the following multiobjective bilevel optimization:(3)minxi∈Rni,i=U,LFUxU,xLs.t.GUxU,xL≥0,HUxU,xL=0,xL∈SLxU,where SLxU signifies the set of solutions of the lower-level problem(4)minxL∈RnLFLxL,xUs.t.GLxL,xU≥0,HLxL,xU=0and Fi:Rni×Rnj→Rmi, Gi:Rni×Rnj→Rpi, and Hi:Rni×Rnj→Rqi, i,j∈{U,L}, i≠j. In what follows, we assume that all functions considered in the bilevel problem (3) are continuously differentiable and mi≥2, i=U,L. Since we consider the case of a multiobjective optimization (mL≥2) at the lower level, given xU, the solution set SL(xU) denotes the Pareto optimal set or weak Pareto optimal set; that is, given any xU and x¯L∈SL(xU), there is no feasible point xL to (4) such that FL(xL,xU)⪯FL(x¯L,xU) or FL(xL,xU)≺FL(x¯L,xU).
Definition 1.
(xU,xL) is said to be feasible for problem (3) if GU(xU,xL)≥0, HU(xU,xL)=0, and xL∈SL(xU). (x¯U,x¯L) is said to be locally Pareto optimal (resp., weak Pareto) for problem (3), if it is feasible for (3) and there exists no other feasible point (xU,xL) in the neighborhood of (x¯U,x¯L) such that FU(xU,xL)⪯FU(x¯U,x¯L) (resp., FU(xU,xL)≺FU(x¯U,x¯L)).
In the multiobjective optimization literature, the weighted approach is one of the most often used methods. However the weighted approach is also criticized for several shortcomings. For example, as shown in applications, weights are not known in advance and the modeler or decision-maker has to choose them while it is often difficult to make a reasonable decision for choosing appropriate weights for objectives. In many situations there may be uncertainty in the weights provided by the decision-maker. Therefore there is a need for providing a novel methodology to cope with these issues. The robust-weighted approach is one possible choice as it provides an alternative to solve the weight uncertainty in the multiobjective bilevel model. Specifically, we assume that players in both levels use the robust-weighted approach to address the weight uncertainty; that is, both the leader and the follower use a robust-weighted approach to minimize their weighted payoff functions.
Definition 2.
Let Wi∈Wri≔{wi∈R+mi∣wi1=1} be a nonempty closed convex set of weights, i=U,L. (xU,xL) is called a robust-weighted optimal solution of (3), if it is an optimum to the following robust optimization problem:(5)minxi∈Rni,i=U,LmaxwU∈WUwUTFUxU,xLs.t.GUxU,xL≥0,HUxU,xL=0,xL∈SwLxU,where SwL(xU) signifies the set of solutions of the lower-level problem:(6)minxL∈RnLmaxwL∈WLwLTFLxL,xUs.t.GLxL,xU≥0,HLxL,xU=0.
Different from weighted Pareto concepts, the robust-weighted optimum proposed in this paper seeks to find an optimal solution that is robust and feasible for the worst-case weight within the family of weights. It directly follows from Theorem 2.2 of Hu and Mehrotra [21] and Definitions 1 and 2 that the following theorem holds.
Theorem 3.
Let Wi⊂Wri be defined as in Definition 2, i=U,L. if x¯:=(x¯U,x¯L) is a robust-weighted optimum of (5) with the combination W≔(WU,WL) of weight sets, then we have the following results:
x¯ is a weak Pareto optimum of (3).
if all of the weights in WU are positive, then x¯ is a Pareto optimum of (3).
if x¯ is the unique optimal solution of (5), then x¯ is a Pareto optimum of (3).
3. MPEC Reformulations for (5)
Now that a robust-weighted optimum of (5) is a (weak) Pareto optimum of (3), our next step is to discuss how to realize such a solution. To do this, we assume that the weight sets Wi, i=U,L, are given as polyhedral regions. Then the problem of computing a robust-weighted optimal point could be cast as MPEC which has been extensively studied and can be solved by the solver PATH [23].
We suppose that the uncertain weights can be described by a fixed reference point and a perturbation region around the point; that is, the weight sets WU and WL can be given as follows: let w¯i∈Rmi be the reference point of wi and Ci∈Rbi×mi be a coefficient matrix used to construct a perturbation region around w¯i, i=U,L, and then define the perturbation region around w¯i as follows:(7)Wi=wi∈Rmi∣wi=w¯i+CiTvi,vi∈Vi⊂Rbi,i=U,L,where Vi,i=U,L, are two uncertain sets which belong to(8)V~i=vi∈Rbi∣w¯i+CiTvi≥0,eiTw¯i+CiTvi=1,i=U,L,and here ei≔(1,…,1)∈Rmi, i=U,L. Note that the above method for defining uncertain parameters has been widely used in the robust optimization literature [12, 13].
Define(9)Vi≔vi∈V~i∣Aivi=γi,Bivi≥ci,i=U,L,where Ai∈Rli×bi, γi∈Rli, ci∈Rki, and Bi∈Rki×bi, i=U,L. We now present a reformulation for problem (5) with Wi given as (7)–(9). Next, we show that the computation for a robust-weighted Nash equilibrium point of MBG can be casted as solving MPEC.
Lemma 4.
Suppose the weight sets Wi, i=U,L, are given as (7)–(9), and then the robust-weighted bilevel optimization problem (5) is equivalent to the following bilevel optimization problem:(10)minxi,zji,i=U,L;j=1,2,3,4w¯UTz3U+z4U-γUTz1U-cUTz2Us.t.CUz3U+AUTz1U+BUTz2U2=0FUxU,xL-z3U-z4Ue≤0z1U∈RlU,z2U∈R+kU,z3U∈RmU,z4U∈RGUxU,xL≥0,HUxU,xL=0xL,z1L,z2L,z3L,z4L∈SwLxU,where SwL(xU) signifies the set of solutions of the lower-level problem:(11)minxL,zjL,j=1,2,3,4w¯LTz3L+z4L-γLTz1L-cLTz2Ls.t.CLz3L+ALTz1L+BLTz2L2=0FLxL,xU-z3L-z4Le≤0GLxL,xU≥0,HLxL,xU=0z1L∈RlL,z2L∈R+kL,z3L∈RmL,z4L∈R.
Proof.
This lemma can be proven as follows. Given the leader’s strategy xU, it is easy to show that the follower’s optimization problem (6) can be rewritten as(12)minxL,tts.t.maxwL∈WLwLTFLxL,xU≤t,GLxL,xU≥0,HLxL,xU=0.For any given (xU,xL), the left-hand side in the first inequality constraint in the above problem is equivalent to(13)maxvLCLFLxL,xUTvL+w¯LTFLxL,xU,s.t.w¯L+CLTvL≥0,eTw¯L+CLTvL=1,ALvL=γL,BLvL≥cL.The corresponding dual problem is(14)minz1L,z2L,z3L,z4Lw¯LTz3L+z4L-γLTz1L-cLTz2L,s.t.CLz3L+ALTz1L+BLTz2L=0FLxL,xU-z3L-z4Le≤0z1L∈RlL,z2L∈R+kL,z3L∈RmL,z4L∈R,where “0” and “e” are the zero and unit vectors with appropriate dimensions, respectively.
Therefore it follows from (12)–(14) and the strong duality theorem that the assertion of this theorem is true.
The above lemma shows the equivalence between the bilevel problems (5) and (10); that is, if (xU,xL) is a robust-weighted optimal solution to (5), then there is (zji)i=U,L;j=1,2,3,4 such that (xi,zji)i=U,L;j=1,2,3,4 is an optimal solution to (10); otherwise if (xi,zji)i=U,L;j=1,2,3,4 is an optimal solution to (10), then (xU,xL) is a robust-weighted optimal solution to (5). So, in what follows, we focus on the bilevel problem (10).
Given xU, it is expected that a solution (xL,zjL)j=1,2,3,4 of the lower-level problem (11) is a stationary point of (11); that is, it solves the following Karush-Kuhn-Tucker (KKT) system:(15)∇xLΓ=∇xLFLxL,xUλ2L-∇xLGLxL,xUλ4L+∇xLHLxL,xUλ5L=0;∇z1LΓ=-γL+ALλ1L=0;∇z2LΓ=-cL+BLλ1L-λ3L=0;∇z3LΓ=w¯L+CLTλ1L-λ2L=0;∇z4LΓ=1-eTλ2L=0;HLxL,xU=0;FLxL,xU-z3L-z4Le≤0,λ2L≥0;λ2LTFLxL,xU-z3L-z4Le=0;z2L≥0,GLxL,xU≥0,λ3L≥0,λ4L≥0;λ3LTz2L=0,λ4LTGLxL,xU=0,where Γ is the Lagrangian of (11) defined as(16)ΓxL,xU,z1L,z2L,z3L,z4L,λ1L,λ2L,λ3L,λ4L,λ5L≔w¯LTz3L+z4L-γLTz1L-cLTz2L2+λ1LTCLz3L+ALTz1L+BLTz2L+λ2LTFLxL,xU-z3L-z4Le-λ3LTz2L-λ4LTGLxL,xU+λ5LHLxL,xU.
Therefore replacing the lower level optimization problem (11) with (15) derives MPEC:(17)minxi,zji,i=U,L;j=1,2,3,4,λkL,k=1,…,5w¯UTz3U+z4U-γUTz1U-cUTz2Us.t.CUz3U+AUTz1U+BUTz2U2=0FUxU,xL-z3U-z4Ue≤0z1U∈RlU,z2U∈R+kU,z3U∈RmU,z4U∈RGUxU,xL≥0,HUxU,xL=0xL,z1L,…,z4L,λ1L,…,λ5Lsolves15.
This problem can be efficiently solved by the solver PATH [23]. We note that the PATH algorithm is originally described in [24]. This algorithm computes a stationary solution of MPEC (since it solves an equation related to the necessary optimality conditions). If for any given xU functions FL(·,xU) and HL(·,xU) are convex in xL and GL(·,xU) is concave in xL, then under some constraint qualification an optimal solution of the bilevel optimization problem (10) is also a stationary solution of the MPEC (17) [25].
Theorem 5.
Suppose the weight sets Wi, i=U,L, are given as (7)–(9). If for any given xU functions FL(·,xU) and HL(·,xU) are convex in xL and GL(·,xU) is concave in xL, then under some constraint qualification an optimal solution of the bilevel optimization problem (5) is also a stationary solution of the MPEC (17).
Proof.
The proof for this theorem follows directly from Lemma 4, the above discussions, and Proposition 5 [25].
4. An Application4.1. Model Description
In this section, we model the multicriteria competition in a supply chain of M manufacturers who cooperatively decide on volume to supply P products to N risk-averse retailers who also cooperatively decide on volume to satisfy uncertain consumer demand. Manufacturers have two objectives (i.e., to maximize their profit and to minimize their cost of products supplied simultaneously) by choosing the supply quantity for each of P products in the anticipation of the order quantities of the retailers and also anticipating of the wholesale prices resulting from the market clearing conditions. Retailers have also two objectives (i.e., to maximize the expected utility from retail sales and to minimize the risk expressed as the standard deviation of the profit) by deciding the wholesale market order quantity for each of P products.
The retail demand satisfies the following stochastic linear inverse demand function: (18)p^=a^-B^x^ξ,where p^=(p1,…,pN)T∈RNP is the aggregate price vector, pj is the vector of the jth retailer prices with pj=(pj1,…,pjP), pjk is the price of the jth retailer for the kth product, and ξ=∑l=1rλlξl with ξl as a positive scalar random variable with mean μl and finite standard deviation σl, λl≥0, and ∑l=1rλl=1. The vector a^=(a1,…,aN)T∈RNP with aj=(aj1,…,ajP) gives the prices that the consumers would be willing to pay if the retail market supply was 0 and ξ is a finite number. The matrix B^∈RNP×NP is the positive definite matrix of inverse demand sensitivities which can be expressed as (19)H1G12⋯G1NG21H2⋯G2N⋮⋮⋱⋮GN1GN2⋯HN,with Hj∈RP×P, j=1,…,N, Gjκ∈RP×P, and j,κ=1,…,N, j≠κ. The vector x^=(x1,…,xN)T∈RNP is the aggregate retailer order vector, where xj is the vector of the jth retailer order vector xj=(xj1,…,xjP).
Retailers make their order quantity decision as follows:(20)max-∑j=1NxjTπ+xjTaj-B^j·x^Eξ,min∑j=1NxjTaj-B^j·x^σξ,s.t.xj≥0,j=1,…,N,where π=(π1,…,πP) is the vector of wholesale prices and B^j·=(Gj1,…,Hj,…,GjN), for jth retailer whose first objective is to maximize their mean profit and second objective is to minimize their standard deviation for the stochastic demand. We assume that the weight vector to the two objectives xjTπ+xjT(aj-B^j·x^)E[ξ] and xjT(aj-B^j·x^)σ(ξ) in (20) is wjL:=(wj1L,wj2L)∈R++2.
Manufacturers choose their supply quantity to maximize their profit in the wholesale market, anticipating the market clearing wholesale market price as well as the retailer order quantities. Mathematically, manufacturers face the following decision problem:(21)min∑i=1Mwi1UciTyi-wi2UπTyi,s.t.∑k=1Myk=∑j=1Nxjv,yi≥0,i=1,…,M,v∈RP,where ci∈RP is the ith manufacturer unit cost, yi∈RP is the supply quantity of the ith manufacturer and xj(π):RP→RP is the jth retailer order quantity for a given wholesale price π, wiU:=(wi1U,wi2U)∈R++2 is the weight vector to manufacturer i’s profit and cost, and the first constraint is the wholesale market clearing conditions. We assume that all the uncertainty exists in the retail market and that the wholesale market is deterministic except for the weight choice. The reason for this is that the demand uncertainty in the retail demand function is realized when all wholesale market related decisions are made in our model. We also assume that the weight vector wiU is uncertain and belongs to a known convex and compact set. By this assumption, we can utilize the proposed robust-weighted approach.
4.2. Numerical Results
We now detail the numerical tests for finding a robust-weighted manufacturer-retailer equilibrium by using the numerical solver PATH [23]. We present numerical results for the robust-weighted manufacturer-retailer equilibrium of models in the decentralized supply chain with finite weights and general polyhedral weight regions, respectively. Suppose M=2, N=2, and P=2, that is, two manufacturers, two retailers, and two products. We assume that E[ξ]=1 and σ(ξ)=0.5; that is, the mean and the standard deviation of the stochastic variable ξ are 1 and 0.5, respectively. We give the following assumption for the parameters in supply chain:(22)a1≔911;a2≔710;c1≔0.60.6;c2≔0.50.5Hj=Gjκ≔10.50.51,j,κ=1,2,j≠κ.
Analytical Hierarchy Process (for brief AHP) [26] is used to assign weights for the two objectives in models (20) and (21). We assume that retailer j is unsure about the relative importance of f1j(x^):=xjTπ-xjT(aj-B^j·x^)E[ξ] and f2j(x^):=xjT(aj-B^j·x^)σ(ξ) for any j=1,2. Consequently, we first consider that there are finite weights to quantify the relative importance for the two objective functions. Then we generalize the finite weights to the weight set given by a polytope. We first suppose that retailers allow three possible scale indices for the two objective functions as f11/f21=0.7, 1, or 1.5 and f12/f22=0.8, 1.28, or 1.69. Similar to the assumption about the retailers’ weights, we also assume that manufacturer i is also unsure about the relative importance of g1i:=-πTyi and g2i:=ciTyi, i=1,2. As such, manufacturers allow for two possible scale indices to the two objective functions as g11/g21=0.3 or 1 and g22/g22=0.5 or 1.2.
The numerical results are reported in Tables 1–4, where (x^,y^,π) (manufacturer-retailer equilibriums) are the weighted Nash equilibrium by using the weighted approach [9], and here the vector y^=(y1,…,yM)T∈RMP is the aggregate manufacturer supply vector. The robust-weighted Nash equilibrium (RWNE) with finite weights is reported in the last row of Table 4. There are different weighted manufacturer-retailer equilibriums for a different choice of the weights. The robust-weighted retailer equilibrium is computed by using the weight region generated as the convex hull of the nine weight vectors. It is easy to see that the robust-weighted approach presents a unique equilibrium which mitigates the conflict with different weighted manufacturer-retailer equilibria for a different choice of the weights in Tables 1–4. We note that, compared with the existing weighted approach [9], our method is more “robust" in that there are different solutions for choosing different weights by using weighted approach [9], but there is a unique solution by using our approach.
Numerical test results with finite weights for g11/g21=0.3 and g12/g22=0.5.
f11/f21
f12/f22
x^
y^
π
0.7
0.8
(0.583, 0.527, 0, 1.654)
(0, 0, 0.583, 2.181)
(1.851, 2.203)
0.7
1.28
(0, 0, 0.393, 1.893)
(0, 0, 0.393, 1.893)
(2.633, 3.547)
0.7
1.69
(0, 0, 0.430, 1.930)
(0, 0, 0.430, 1.930)
(2.964, 4.021)
1
1.28
(0.833, 0.923, 0, 1.554)
(0, 0, 0.833, 2.476)
(2.817, 3.384)
1
0.8
(0.833, 1.833, 0, 0)
(0, 0, 0.833, 1.833)
(2.750, 3.250)
1
1.69
(0.833, 0.513, 0, 2.000)
(0, 0, 0.833, 2.513)
(2.910, 3.570)
1.5
1.69
(0.917, 1.385, 0, 1.170)
(0, 0, 0.917, 2.555)
(3.465, 4.096)
1.5
1.28
(0.917, 1.801, 0, 0.717)
(0, 0, 0.917, 2.518)
(3.338, 3.843)
1.5
0.8
(0.917, 1.917, 0, 0)
(0, 0, 0.917, 1.917)
(3.500, 4.167)
Numerical test results with finite weights for g11/g21=0.3 and g12/g22=1.2.
f11/f21
f12/f22
x^
y^
π
0.7
0.8
(0.769, 0.664, 0, 1.751)
(0, 0, 0.769, 2.415)
(1.692, 2.044)
0.7
1.28
(0.278, 0, 0.814, 1.980)
(0, 0, 1.092, 1.980)
(1.897, 3.099)
0.7
1.69
(0, 0, 0.505, 2.005)
(0, 0, 0.505, 2.005)
(2.805, 3.862)
1
1.28
(0.939, 0.995, 0, 1.622)
(0, 0, 0.939, 2.617)
(2.658, 3.625)
1
0.8
(0.939, 1.939, 0, 0)
(0, 0, 0.939, 1.939)
(2.591, 3.091)
1
1.69
(0.939, 0.597, 0, 2.045)
(0, 0, 0.939, 2.642)
(2.751, 3.411)
1.5
1.69
(0.996, 1.429, 0, 1.241)
(0, 0, 0.996, 2.670)
(3.305, 3.937)
1.5
1.28
(0.996, 1.833, 0, 0.812)
(0, 0, 0.996, 2.654)
(3.179, 3.684)
1.5
0.8
(0.966, 1.996, 0, 0)
(0, 0, 0.996, 1.996)
(3.341, 4.008)
Numerical test results with finite weights for g11/g21=1 and g12/g22=1.2.
f11/f21
f12/f22
x^
y^
π
0.7
0.8
(0.924, 0.778, 0, 1.832)
(0, 0, 0.924, 2.610)
(1.560, 1.911)
0.7
1.28
(0.515, 0, 0.768, 2.053)
(0, 0, 1.283, 2.053)
(1.764, 2.967)
0.7
1.69
(0.255, 0, 1.036, 2.068)
(0, 0, 1.291, 2.068)
(1.834, 3.310)
1
1.28
(1.028, 1.055, 0, 1.697)
(0, 0, 1.028, 2.733)
(2.525, 3.092)
1
0.8
(1.028, 2.028, 0, 0)
(0, 0, 1.028, 2.028)
(2.458, 2.958)
1
1.69
(1.028, 0.667, 0, 2.082)
(0, 0, 1.028, 2.749)
(2.618, 3.278)
1.5
1.69
(1.062, 1.466, 0, 1.300)
(0, 0, 1.062, 2.766)
(3.173, 3.804)
1.5
1.28
(1.062, 1.860, 0, 0.891)
(0, 0, 1.062, 2.751)
(3.046, 3.551)
1.5
0.8
(1.063, 2.062, 0, 0)
(0, 0, 1.063, 2.062)
(3.208, 3.875)
Numerical test results with finite weights for g11/g21=1 and g12/g22=0.5.
f11/f21
f12/f22
x^
y^
π
0.7
0.8
(0.817, 0.699, 0, 1.776)
(0.817, 2.475, 0, 0)
(1.651, 2.003)
0.7
1.28
(0.351, 0, 0.800, 2.003)
(1.151, 2.003, 0, 0)
(1.856, 3.059)
0.7
1.69
(0, 0, 0.525, 2.025)
(0.525, 2.025, 0, 0)
(2.764, 3.821)
1
1.28
(0.967, 1.013, 0, 1.639)
(0.967, 2.653, 0, 0)
(2.617, 3.184)
1
0.8
(0, 2.867, 0, 1.167)
(0, 3.033, 0, 0)
(3.100, 2.550)
1
1.69
(0.967, 0.619, 0, 2.056)
(0.967, 2.675, 0, 0)
(2.710, 3.370)
1.5
1.69
(1.017, 1.440, 0, 1.259)
(1.017, 2.700, 0, 0)
(3.265, 3.896)
1.5
1.28
(0, 2.650, 0, 0.836)
(0, 3.186, 0, 0)
(4.859, 3.643)
1.5
0.8
(1.017, 2.017, 0, 0)
(1.017, 2.017, 0, 0)
(3.300, 3.967)
(RWNE)
(0.583, 0.527, 0, 1.654)
(0, 0, 0.583, 2.181)
(1.851, 2.203)
The managerial impact of the proposed method is as follows. The proposed approach overcomes several shortcomings for the current approaches to the multicriteria competition in a supply chain. There is no clear guidance to a decision-maker (manufacturers or retailers) to choose a solution among (possibly a very large number of) generated Pareto optimal solutions. However, there is a unique solution by using our approach. In the multicriteria competition in a supply chain, since different decision-maker (manufacturers or retailers) may have different opinions, the relative weights given by different decision-maker can differ significantly. Furthermore, there may be ambiguity in the weights given by a decision-maker (manufacturer or retailer), because it is often difficult to give relative weights for each objective.
For the general polytope weight region, we suppose the reference point w¯j and matrices CiU and CjL are defined as follows: w¯iU:=(1/2,1/2), i=1,2, w¯jL:=(1/3,2/3), j=1,2, and(23)CiU≔1212,i=1,2CjL≔1011,j=1,2.
The perturbation subsets are defined as(24)ViU≔viU∈R2∣eTviU≤1-δ,i=1;VjL≔vjL∈R2∣eTvjL≤1-δ,j=1,2,where δ is a parameter defined in 0,1. With the above data the robust-weighted manufacturer-retailer equilibrium with different δ can be calculated by utilizing the method developed in Section 5. The stability results are shown in Table 5 which contains small data errors due to the small changes in δ.
Numerical test results with general polytope weight regions.
δ
x^
y^
π
0
(0.778, 1.444, 0.778, 1.444)
(0, 0, 1.556, 2.889)
(2.750, 3.250)
0.007
(0.777, 1.444, 0.777, 1.444)
(0, 0, 1.544, 2.887)
(2.748, 3.248)
0.02
(0.776, 1.442, 0.776, 1.442)
(0, 0, 1.541, 2.884)
(2.745, 3.245)
0.05
(0.772, 1.439, 0.772, 1.439)
(0, 0, 1.544, 2.878)
(2.737, 3.237)
0.1
(0.767, 1.433, 0.767, 1.433)
(0, 0, 1.533, 2.867)
(2.725, 3.225)
0.5
(0.722, 1.389, 0.722, 1.389)
(0, 0, 1.444, 2.778)
(2.625, 3.125)
5. Conclusion
This paper introduces a robust-weighted approach to multicriteria bilevel problems. The computation of the robust-weighted Pareto optimal point, with the polyhedral weight sets, can be equivalently transformed into solving MPEC. As an application, the new approach is used to analyze supply chain multicriteria bilevel competition. The numerical tests show that our method is more robust than existing approaches and a unique robust-weighted Pareto optimum can always be obtained for general polyhedral weight sets.
Conflicts of Interest
The authors declare that they have no conflicts of interest.
Acknowledgments
The work is supported by The Program for Professor of Special Appointment (Eastern Scholar) at Shanghai Institutions of Higher Learning (no. TP2014043). The work is also supported by Natural Scientific Foundation of China (no. 71401107; no. 71571055).
DempeS.Annotated bibliography on bilevel programming and mathematical programs with equilibrium constraints200352333335910.1080/0233193031000149894MR19963212-s2.0-0042737841BonnelH.Optimality conditions for the semivectorial bilevel optimization problem200623447467MR2265749BonnelH.MorganJ.Semivectorial bilevel optimization problem: penalty approach2006131336538210.1007/s10957-006-9150-4MR22835382-s2.0-33845640514LiuB.WanZ.ChenJ.WangG.Optimality conditions for pessimistic semivectorial bilevel programming problems20142014, article no. 4110.1186/1029-242x-2014-41MR33476722-s2.0-84899815153ZhengY.WanZ. P.A solution method for semivectorial bilevel programming problem via penalty method2011371-220721910.1007/s12190-010-0430-7MR28315342-s2.0-80052614687Ben-TalA.GoryashkoA.GuslitzerE.NemirovskiA.Adjustable robust solutions of uncertain linear programs200499235137610.1007/s10107-003-0454-yMR20390452-s2.0-14844327381JeyakumarV.LeeG. M.LiG.Characterizing robust solution sets of convex programs under data uncertainty2015164240743510.1007/s10957-014-0564-0MR32979702-s2.0-84939877970EhrgottM.20052ndBerlin, GermanySpringerMR2143243WangS. Y.Existence of a Pareto equilibrium199379237338410.1007/bf00940586MR12521422-s2.0-0027696346YuJ.YuanG. X.The study of Pareto equilibria for multiobjective games by fixed point and Ky Fan minimax inequality methods1998359172410.1016/s0898-1221(98)00053-4MR1625036TsiporkovaE.BoevaV.Multi-step ranking of alternatives in a multi-criteria and multi-expert decision making environment2006176182673269710.1016/j.ins.2005.11.010ZBL1102.686552-s2.0-33745368723SchoemakerP. J. H.WaidC. C.An experimental comparison of different approaches to determining weights in additive utility models198228218219610.1287/mnsc.28.2.1822-s2.0-0020089312WeberM.BorcherdingK.Behavioral influences on weight judgments in multiattribute decision making199367111210.1016/0377-2217(93)90318-h2-s2.0-0027593910YinY.Multi-objective bilevel optimization for transportation planning and management problems20023619310510.1002/atr.56703601062-s2.0-0036885370DebK.SinhaA.2009SpringerPieumeC. O.MarcotteP.FotsoL. P.SiarryP.Solving bilevel linear multiobjective programming problems20111421421910.4236/ajor.2011.14024EichfelderG.Multiobjective bilevel optimization2010123241944910.1007/s10107-008-0259-0MR25959202-s2.0-77949540007JiY.QuS. J.YuZ. S.Bi-level multi-objective optimization model for last mile delivery using a discrete approach201610.1080/10236198.2016.1210607CalveteH. I.GaléC.Linear bilevel programs with multiple objectives at the upper level2010234495095910.1016/j.cam.2008.12.010MR26095502-s2.0-77950339458CalveteH. I.GaléC.On linear bilevel problems with multiple objectives at the lower level2011391334010.1016/j.omega.2010.02.0022-s2.0-77953687861HuJ.MehrotraS.Robust and stochastically weighted multiobjective optimization models and reformulations201260493695310.1287/opre.1120.1071MR2979432ZBL1342.900742-s2.0-84866383103JianJ.-B.LiJ.-L.MoX.-D.A strongly and superlinearly convergent SQP algorithm for optimization problems with linear complementarity constraints2006541174610.1007/s00245-005-0848-8MR22276222-s2.0-33744726245FerrisM. C.MunsonT. S.Complementarity problems in GAMS and the PATH solver200024216518810.1016/s0165-1889(98)00092-xMR1727862DierkseS. P.1994Madison, Wis, USAUniversity of Wisconsin-MadisonHuX.RalphD.Using EPECs to model bilevel games in restructured electricity markets with locational prices200755580982710.1287/opre.1070.0431MR23609502-s2.0-38549176844SaatyT. L.The modern science of multicriteria decision making and its practical applications: the AHP/ANP approach20136151101111810.1287/opre.2013.1197MR31385592-s2.0-84887223649