Homotopy Interior-Point Method for a General Multiobjective Programming Problem
X. Zhao,1,2S. G. Zhang,1and Q. H. Liu3
Academic Editor: Yongkun Li
Received26 Oct 2011
Revised10 Feb 2012
Accepted12 Feb 2012
Published29 Apr 2012
Abstract
We present a combined homotopy interior-point method for a general multiobjective
programming problem. For solving the KKT points of the multiobjective programming problem,
the homotopy equation is constructed. We prove the existence and convergence of a smooth homotopy path
from almost any initial interior point to a solution of the KKT system under some basic assumptions.
1. Introduction
We consider the following multiobjective programming problem:
where , and .
Since Kellogg et al. [1] and Smale [2] have published their remarkable papers, more and more attention has been paid to the homotopy method. As a globally convergent method, the homotopy method now becomes an important tool for numerically solving complementary problem and nonlinear mathematical programming problem [3β5].
Among most interior-point methods for solving mathematical programming, one of the main ideas is numerically tracing the center path generated by the optimal solution set of the logarithmic barrier function. Usually, the strict convexity of the logarithmic barrier function or boundedness of the solution set is needed [6, 7]. Lin et al. [8] presented a new interior-point method, combined homotopy interior-point method (CHIP method), for convex nonlinear programming. Subsequently, Lin et al. [9] generalized the CHIP method to convex multi-objective programming with only inequality constraints. Recently, Song and Yao [10] generalized the combined homotopy interior-point method to the general multiobjective programming problem under the so-called normal cone condition. In that paper, they proved the existence of the homotopy path under the following assumptions:(A1) is nonempty and bounded;(A2)for all , the vectors are linearly independent;(A3)for all , ,
whereββ.
In [10], the combined homotopy method was given as follows:
where , and . However, the solution simply yields for all . In fact, from the last equation, we have . According to this, we know that for all of . That is, the method given in [10] just solves the single-objective programming problem.
The purpose of this paper is to generalize the combined homotopy interior-point method for a general multiobjective programming problem (MOP) under quasinorm cone condition that weakens the assumptions more than the ones in [10] and constructs a new homotopy equation which is much different and simpler the that one given in [9].
The paper is organized as follows. In Section 2, we recall some notations and preliminaries results. In Section 3, we construct a new combined homotopy mapping and prove the existence and convergence of a smooth homotopy path from almost any interior initial point to a KKT point of MOP under some assumptions. In Section 4, numerical results are given.
2. Some Definitions and Properties
Let and denote the nonnegative and positive orthants of , respectively. For any two vectors and in , we use the following conventions:
Suppose that , , and are three times continuously differentiable functions. Let
and let
denote the active index set at a given point.
Definition 2.1. A point is said to be an efficient solution to multiobjective programming problem (MOP), if there is no such that holds.
Definition 2.2. Let be an open set, and let be a smooth mapping. If Rangeβ for all , then is a regular value and is a regular point.
Definition 2.3. Let . For any is said to be positive linear independent with respect to and if
implies that
Lemma 2.4 (parametric form of the Sard Theorem on a smooth manifold; see [11]). Let and be smooth manifolds of dimensions and . Respectively, let be a map, where . If is a regular value of , then for almost all , 0 is a regular value of .
Lemma 2.5 (inverse image theorem; see [12]). If 0 is a regular value of the mapping , then consists of some smooth manifolds.
Lemma 2.6 (classification theorem of one-dimensional manifold; see [12]). A one-dimensional smooth manifold is diffeomorphic to a unit circle or a unit interval. The following three basic assumptions are commonly used in this paper: (C1) is nonempty and bounded;(C2)for any , there exists a positive linear independent map with respect to and such that, is positive linear independent with respect to and ;(C3)for any , there exists a positive linear independent map with respect to and , such that
(quasinormal cone condition).
Remark 2.7. If satisfies assumptions (A1)β(A3), then it necessarily satisfies assumptions (C1)β(C3). In fact, if we choose , then it is easy to get the result. Clearly, if satisfies assumptions (C1)β(C3), then it does not necessarily satisfies assumptions (A1)β(A3).
3. Main Results
Let . We say that is a KKT point of (MOP) if there exists , such thatwhere , , .
Meanwhile, the KKT system of MOP is (3.1a)β(3.1c).
For a convex multiobjective programming problem, the solution of the MOP can be obtained from the KKT system, and for a non-convex multi-objective programming problem, it is significant when we get a solution of the KKT system.
To solve the KKT system (3.1a)β(3.1c), βwe construct a homotopy equation as follows:
where , , , , , and .
For a given , rewrite as . The zero-point set of is
Theorem 3.1. Suppose that is defined as in (3.2) and let , and be three times continuously differentiable functions. In addition, let assumptions (C1)β(C3) βhold, and let be two times continuously differentiable function. Then, for almost all initial points , 0 is a regular value of and consists of some smooth curves. Among them, a smooth curve, say , is starting from .
Proof. Denote the Jacobi matrix of by . For any and , we have . Now, we consider the submatrix of . For any ,
where . We obtain that
That is, 0 is a regular value of . By the parametric form of the Sard theorem, for almost all , 0 is a regular value of . By the inverse image theorem, consists of some smooth curves. Since , there must be a smooth curve, denoted by , starting from .
Theorem 3.2. Let assumptions (C1)-(C2) βhold. For a given , if 0 is a regular value of , then the projection of the smooth curve on the component is bounded.
Proof. Suppose that the conclusion does not hold. Since is bounded, there exists a sequence such that
From the last equality of (3.2), we have
If we assume , this hypothesis implies that
Since , , it follows that the second part in the left-hand side of some equations in (3.8) tends to infinity as . But the other two parts are bounded. This is impossible. Thus, the component is bounded.
Theorem 3.3. Let , and be three times continuously differentiable functions. In addition, let assumptions (C1)β(C3) hold, and let be two times continuously differentiable function. Then, for almost all of , contains a smooth curve , which starts from . As , the limit set ββof is nonempty and every point in is a solution of the KKT system (3.1a)β(3.1c).
Proof. From the homotopy equation (3.2), it is easy to see that . βBy Theorem 3.1, for almost all ,ββ0 is a regular value of and contains of a smooth curve starting from . By the classification theorem of one-dimensional smooth manifolds, is diffeomorphic to a unit circle or the unit interval . Noticing that
By assumption (C2), we know that . And because , we obtain that ββis nonsingular. Therefore, the smooth curve , which starts from , is diffeomorphic to . Let be a limit point of . Only three cases are possible:(a),
(b),
(c). Because has a unique solution , case (c) will not happen. In case (b), because and are bounded sets and by assumption (C2), for any , there exists a positive linear independent map with respect to and such that is positive linear independent with respect to and . From the first equality of (3.2), we get that the component of is bounded. If case (b) holds, then there exists a sequence such that
Because and are bounded and by Theorem 3.2, there exists a sequence (denoted also by ) such that
From the third equality of (3.2), we have
Hence, the active index set is nonempty. From the homotopy equation (3.2), it follows that
(i)When , rewrite (3.14) as Because , are bounded and by assumption , when , we observe that
Using and , we have from (3.16) that
It is easy to see that the right-hand side of (3.17) is bounded. By assumption (C2) and (3.17), we get
where . Then, we have
which contradicts assumption . (ii)When , rewrite (3.14) as We know that, since and are bounded as , the right-hand side of (3.20) is bounded. But by assumptions (C2) and , if , then the left-hand side of (3.20) is infinite, this is a contradiction. As a conclusion, (a) is the only possible case, and is a solution of the KKT system (3.1a)β(3.1c). Let be the arc-length of . We can parameterize with respect to .
Theorem 3.4. The homotopy path is determined by the following initial-value problem for the ordinary differential equation:
The component of the solution point , for , is a solution of the KKT system.
4. Algorithm and Example
The following algorithm describes the numerical computation for Pareto optimal solutions.
Algorithm 4.1 ((MOP) Euler-Newton method). Step 1. Give an initial point , an initial step length , and three small positive numbers . Let .Step 2. Compute the direction of the predictor step:(a)compute a unit tangent vector ββof at ;(b)determine the direction of the predictor step. If the sign of the determinant ββis , take . If the sign of the determinant is , take .Step 3. Compute a corrector point
where
is the Moore-Penrose inverse of . If , let and go to Step 4. If , let and go to Step 4. If , let and go to Step 3.Step 4. If and , let and go to Step 2. If and , let ,ββgo to Step 3, and recompute ββfor the initial point . If , let ,ββgo to Step 3, and recompute for the initial point . If and , then stop.
Remark 4.2. In Algorithm 4.1, the arc length parameter is not computed explicitly. The tangent vector at a point on has two opposite directions, one (the positive direction) makes increase and the other (the negative direction) makes decrease. The negative direction will lead us back to the initial point, so we must go along the positive direction. The criterion in Step 2 (b) of Algorithm 4.1 that determines the positive direction is based on a basic theory of homotopy method, that is, the positive direction at any point on keeps the sign of the determinant invariant.
Example 4.3. We have
The results are shown in Table 1.
Example 4.4. We have
The results are shown in Table 2.
Acknowledgments
This work was supported by the National Natural Science Foundation of China (Grant no. 10771020) and the Jilin Province Natural Science Foundation (Grant no. 20101597).
References
R. B. Kellogg, T. Y. Li, and J. Yorke, βA constructive proof of the Brouwer fixed-point theorem and computational results,β SIAM Journal on Numerical Analysis, vol. 13, no. 4, pp. 473β483, 1976.
S. Smale, βA convergent process of price adjustment and global newton methods,β Journal of Mathematical Economics, vol. 3, no. 2, pp. 107β120, 1976.
S. N. Chow, J. Mallet-Paret, and J. A. Yorke, βFinding zeroes of maps: homotopy methods that are constructive with probability one,β Mathematics of Computation, vol. 32, no. 143, pp. 887β899, 1978.
R. D. C. Monteiro and I. Adler, βAn extension of Karmarkar type algorithm to a class of convex separable programming problems with global linear rate of convergence,β Mathematics of Operations Research, vol. 15, no. 3, pp. 408β422, 1990.
Z. Lin, B. Yu, and G. Feng, βA combined homotopy interior point method for convex nonlinear programming,β Applied Mathematics and Computation, vol. 84, no. 2-3, pp. 193β211, 1997.
Z. H. Lin, D. L. Zhu, and Z. P. Sheng, βFinding a minimal efficient solution of a convex multiobjective program,β Journal of Optimization Theory and Applications, vol. 118, no. 3, pp. 587β600, 2003.
W. Song and G. M. Yao, βHomotopy method for a general multiobjective programming problem,β Journal of Optimization Theory and Applications, vol. 138, no. 1, pp. 139β153, 2008.