Houyuan Jiang, Daniel Ralph, copyright 1997.
Matlab code accompanied the paper: Jiang, H., Ralph D. QPECgen, a MATLAB generator for mathematical programs with Computational Optimization and Applications 13 (1999), 25–59.
Python implementation coded by Patricia Gillett, 2013-2016.
This code generates random test problems of MPEC with quadratic objective functions and affine variational inequality constraints, and certain special cases.
The MPEC problem is defined as:
min f(x,y) s.t. (x,y) in Z y in S(x), S(x) solves AVI(F(x,y), C(x)) F(x,y) is linear with respect to both x and y C(x) is polyhedral in y-space
- x: n dimensional first level variable.
- y: m dimensional second level variable.
- P: P=[Px Pxy^T; Pxy Py] -- Hessian of the objective function.
- c, d: coefficient vectors associated with x and y, respectively.
- A, a: A is an l by (m+n) matrix, a is an l dimensional vector matrix. Used in the upper level constraints A*[x;y] + a <= 0 (models are described in paper in terms of G and H where A=[G, H])
- F, q, N, M: N is an m by n matrix, M is an m by m matrix, and q is an m by 1 vector. These define F linearly in terms of x and y: F=Nx+My+q.
- D, E, b: D is a p by n matrix, E is a p by m matrix, b is an m dimensional vector. Used in the lower level constraints for type 100 problems.
- u: m dimensional vector used in lower level constraints for type 200 problem.
AVI-QPEC (type 100)
min 0.5*[x;y]^T*P*[x;y] + [c;d]^T*[x;y] s.t. A*[x;y] + a <= 0 D*x + E*y + b <= 0 lambda >= 0 (D*x + E*y + b)^T lambda = 0 N*x + M*y + E^T*lambda + q = 0
BOX-MPEC (type 200)
For this case, let y=[y1;y2] where variables y1 have both upper and lower bounds and y2 variables only have lower bounds. Because there are no other lower level constraints, the case simplifies and there are no lambda variables.
min 0.5*[x;y]^T*P*[x;y] + [c;d]^T*[x;y] s.t. A*[x;y] + a <= 0 0 <= y1 <= u 0 <= y2 N2*x + M2*y + q2 >= 0 complements 0 <= y2 N1*x + M1*y + q1 complements 0 <= y1 <= u
PATRICIA'S ADDITION: SPECIAL BOX-MPEC (type 201)
It is convenient for us to have one more type where all second level variables have both lower and upper bounds and all first level variables are bounded above and below as well.
min 0.5*[x;y]^T*P*[x;y] + [c;d]^T*[x;y] s.t. A*[x;y] + a <= 0 0 <= x <= ux 0 <= y <= uy N1*x + M1*y + q1 complements 0 <= y <= uy
min 0.5*[x;y]^T*P*[x;y] + [c;d]^T*[x;y] s.t. A*[x;y] + a <= 0 0 <= y N*x + M*y + q >= 0 (N*x + M*y + q)^Ty = 0
GOOD-LCP (type 800)
The objective function is equivalent to sum((x-1)^2) + sum((y+2)^2) shifted by a constant. It is minimized by the point closest to (1 ... 1,-2 ... -2), which is the origin.
min x^Tx + y^Ty - 2*sum(x) + 4*sum(y) s.t. x <= y 0 <= y (y-x)^Ty = 0
This problem has multiple local minima. The objective function is equivalent to sum((x+1)^2) + sum((y-2)^2) shifted by a constant. It is minimized by the feasible point closest to (-1 ... -1, 2 ... 2), which is (-1 ... -1, 0 ... 0).
min x^Tx + y^Ty + 2*sum(x) - 4*sum(y) s.t. x <= y 0 <= y (y-x)^Ty = 0