HanLecture2 ME

Multiobjective Linear Programming Biobjective LPs and Parametric Simplex Multiobjective Simplex Method
International Doctoral School Algorithmic Decision Theory: MCDA and MOO

Lecture 2: Multiobjective Linear Programming
Matthias Ehrgott
Department of Engineering Science, The University of Auckland, New Zealand Laboratoire dInformatique de Nantes Atlantique, CNRS, Universit e de Nantes, France
MCDA and MOO, Han sur Lesse, September 17 21 2007
Matthias Ehrgott
MOLP I
Overview
Multiobjective Linear Programming Formulation and Example Solving MOLPs by Weighted Sums Biobjective LPs and Parametric Simplex The Parametric Simplex Algorithm Biobjective Linear Programmes: Example Multiobjective Simplex Method A Multiobjective Simplex Algorithm Multiobjective Simplex: Examples
Matthias Ehrgott
MOLP I
Formulation and Example Solving MOLPs by Weighted Sums
Overview
Matthias Ehrgott
MOLP I
Variables x Rn Objective function Cx where C Rpn Constraints Ax = b where A Rmn and b Rm min {Cx : Ax = b , x X = {x Rn : Ax = b , x is the feasible set in decision space Y = {Cx : x X } is the feasible set in objective space 0} 0} (1)
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Example min 3x 1 + x 2 x1 2x2 subject to x2 3x 1 x 2 x 3 1 1 2
3 6 0 3 6
C=
A=
0 1 1 0 3 1 0 1
b=
Matthias Ehrgott
MOLP I
Example min 3x 1 + x 2 x1 2x2 subject to x2 3x 1 x 2 x 3 1 1 2
3 6 0 3 6
C=
A=
0 1 1 0 3 1 0 1
b=
Matthias Ehrgott
MOLP I
x2 X x1
0 1
Matthias Ehrgott
x3
x4
2 3 4
MOLP I
Feasible set in decision space
0 0
10
11
12
13
14
Cx
-1 -2
Cx4 Y
-3
-4
-5
-6
Cx2
-7
-8
-9
Cx3
Feasible set in objective space
Matthias Ehrgott MOLP I
-10
Denition Let x X be a feasible solution of the MOLP (1) and let y = Cx . x is called weakly ecient if there is no x X such that Cx < C x ; y = Cx is called weakly nondominated. x is called ecient if there is no x X such that Cx C x ; y = Cx is called nondominated. x is called properly ecient if it is ecient and if there exists a real number M > 0 such that for all i and x with ciT x < ciT x and there is an index j and M > 0 such that cjT x > cjT x ciT x ciT x M. cjT x cjT x
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
0 0
10
11
12
13
14
Cx1 Cx4 Y
-1
-2
-3
-4
-5
-6
Cx2
-7
-8
-9
Cx3
Nondominated set
-10
x2 X x1
0 1
Matthias Ehrgott
x3
x4
2 3 4
Ecient set
MOLP I
Overview
Matthias Ehrgott
MOLP I
Let 1 , . . . , p LP ()
0 and consider
p
min
k =1
T k c k x = min T Cx
subject to Ax x
= b 0
with some vector 0 (Why not = 0 or 0?) LP () is a linear programme that can be solved by the Simplex method If > 0 then optimal solution of LP () is properly ecient If 0 then optimal solution of LP () is weakly ecient Converse also true, because Y convex
Matthias Ehrgott
MOLP I
Let 1 , . . . , p LP ()
0 and consider
p
min
k =1
subject to Ax x
= b 0
Matthias Ehrgott
MOLP I
Let 1 , . . . , p LP ()
0 and consider
p
min
k =1
subject to Ax x
= b 0
Matthias Ehrgott
MOLP I
Let 1 , . . . , p LP ()
0 and consider
p
min
k =1
subject to Ax x
= b 0
Matthias Ehrgott
MOLP I
Let 1 , . . . , p LP ()
0 and consider
p
min
k =1
subject to Ax x
= b 0
Matthias Ehrgott
MOLP I
Illustration in objective space

2
. . . . . . . . . . . . . . . . . . .. . . . .
y2 2 4 6 8 10 12 14 y
0 Cx 2 4 6 8
10
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . 1 . .. . .. . 1 . . . . . . . . . . . . . .. ... . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . .. .. . . . . . 4 . . . . . . .. . . . . . . . . . . . . . . .. . .. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . .. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. . ... . . . . . . .. . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . 2 T . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 T 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 . . . . . . . . . . . . . . . . . .
Cx
Y = CX
Cx
{( ) Cx = 2 } Cx
{( ) Cx = }
{( ) Cx = 3 }
3 T
1 = (2, 1), 2 = (1, 3), 3 = (1, 1)
Matthias Ehrgott
MOLP I
Illustration in objective space

2
. . . . . . . . . . . . . . . . . . .. . . . .
y2 2 4 6 8 10 12 14 y
0 Cx 2 4 6 8
10
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . 1 . .. . .. . 1 . . . . . . . . . . . . . .. ... . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . .. .. . . . . . 4 . . . . . . .. . . . . . . . . . . . . . . .. . .. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . .. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. . ... . . . . . . .. . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . 2 T . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 T 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 . . . . . . . . . . . . . . . . . .
Cx
Y = CX
Cx
{( ) Cx = 2 } Cx
{( ) Cx = }
{( ) Cx = 3 }
3 T
1 = (2, 1), 2 = (1, 3), 3 = (1, 1)
Matthias Ehrgott
MOLP I
y Rp satisfying T y = dene a straight line (hyperplane) Since y = Cx and T Cx is minimised, we push the line towards the origin (left and down) When the line only touches Y nondominated points are found Nondominated points YN are on the boundary of Y Y is convex polyhedron and has nite number of facets. YN consists of nitely many facets of Y . The normal of the facet can serve as weight vector
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Question: Can all ecient solutions be found using weighted sums? If x X is ecient, does there exist > 0 such that x is optimal solution to min{T Cx : Ax = b , x 0}?
Lemma A feasible solution x 0 X is ecient if and only if the linear programme max eT z subject to Ax = b Cx + Iz = Cx 0 x, z 0,
(2)
where e T = (1, . . . , 1) Rp and I is the p p identity matrix, has an optimal solution ( x, z ) with z = 0.
(2)
(2)
Proof. LP is always feasible with x = x 0 , z = 0 (and value 0) Let ( x, z ) be optimal solution If z = 0 then z = Cx 0 C x = 0 Cx 0 = C x There is no x X such that Cx Cx 0 because (x , Cx 0 Cx ) would be better solution x 0 ecient If x 0 ecient there is no x X with Cx Cx 0 there is no z with z = Cx 0 Cx 0 max e T z 0 max e T z = 0
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Lemma A feasible solution x 0 X is ecient if and only if the linear programme min u T b + w T Cx 0 subject to uT A + w T C w u
0 e Rm
(3)
has an optimal solution ( u, w ) with u T b + w T Cx 0 = 0. Proof. The LP (3) is the dual of the LP (2)
Matthias Ehrgott
MOLP I
Lemma A feasible solution x 0 X is ecient if and only if the linear programme min u T b + w T Cx 0 subject to uT A + w T C w u
0 e Rm
(3)
has an optimal solution ( u, w ) with u T b + w T Cx 0 = 0. Proof. The LP (3) is the dual of the LP (2)
Matthias Ehrgott
MOLP I
Theorem A feasible solution x 0 X is an ecient solution of the MOLP (1) if and only if there exists a Rp > such that T Cx 0 T Cx for all x X . Note: We already know that optimal solutions of weighted sum problems are ecient (4)
Matthias Ehrgott
MOLP I
Theorem A feasible solution x 0 X is an ecient solution of the MOLP (1) if and only if there exists a Rp > such that T Cx 0 T Cx for all x X . Note: We already know that optimal solutions of weighted sum problems are ecient (4)
Matthias Ehrgott
MOLP I
Proof. Let x 0 XE By Lemma 4 LP (3) has an optimal solution ( u, w ) such that u T b = w T Cx 0 u is also an optimal solution of the LP min u T b : u T A which is (3) with w = w xed There is an optimal solution of the dual of (6) max w T Cx : Ax = b , x 0 (7) w TC , (6) (5)
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Proof. By weak duality u T b w T Cx for all feasible solutions u of (6) and for all feasible solutions x of (7) We already know that u T b = w T Cx 0 from (5) x 0 is an optimal solution of (7) Note that (7) is equivalent to min w T Cx : Ax = b , x with w 0
e > 0 from the constraints in (3)
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
The Parametric Simplex Algorithm Biobjective Linear Programmes: Example
Overview
Matthias Ehrgott
MOLP I
Modication of the Simplex algorithm for LPs with two objectives min subject to (c 1 )T x , (c 2 )T x Ax x
= b 0
(8)
We can nd all ecient solutions by solving the parametric LP min 1 (c 1 )T x + 2 (c 2 )T x : Ax = b , x for all = (1 , 2 ) > 0 0
Matthias Ehrgott
MOLP I
Modication of the Simplex algorithm for LPs with two objectives min subject to (c 1 )T x , (c 2 )T x Ax x
= b 0
(8)
We can nd all ecient solutions by solving the parametric LP min 1 (c 1 )T x + 2 (c 2 )T x : Ax = b , x for all = (1 , 2 ) > 0 0
Matthias Ehrgott
MOLP I
We can divide the objective by 1 + 2 without changing the optima, i.e. 1 = 1 /(1 + 2 ), 2 = 2 /(1 + 2 and 1 + 2 = 1 or 2 = 1 1 LPs with one parameter 0 1 and parametric objective
c () := c 1 + (1 )c 2 min c ()T x : Ax = b , x 0 (9)
Matthias Ehrgott
MOLP I
We can divide the objective by 1 + 2 without changing the optima, i.e. 1 = 1 /(1 + 2 ), 2 = 2 /(1 + 2 and 1 + 2 = 1 or 2 = 1 1 LPs with one parameter 0 1 and parametric objective
c () := c 1 + (1 )c 2 min c ()T x : Ax = b , x 0 (9)
Matthias Ehrgott
MOLP I
Let B be a feasible basis

T B 1 N Recall reduced cost c N = cN cB
Reduced cost for the parametric LP c () = c 1 + (1 ) c2 is an optimal basis of (9) for some Suppose B ) 0 c ( (10)
Matthias Ehrgott
MOLP I

Matthias Ehrgott
MOLP I

Matthias Ehrgott
MOLP I

Matthias Ehrgott
MOLP I

Matthias Ehrgott
MOLP I
Case 1: c 2
From (10) c () 0 for all < is optimal basis for all 0 B Case 2: There is at least one i N with c i2 < 0 such that c there is < ()i = 0 c i1 + (1 ) ci2 = 0 ( ci1 c i2 ) + c i2 = 0 =
c i2 1 c i c i2
is not optimal Below this value B
Matthias Ehrgott
MOLP I
Case 1: c 2
c i2 1 c i c i2
Matthias Ehrgott
MOLP I
Case 1: c 2
c i2 1 c i c i2
Matthias Ehrgott
MOLP I
Case 1: c 2
c i2 1 c i c i2
Matthias Ehrgott
MOLP I
Case 1: c 2
c i2 1 c i c i2
Matthias Ehrgott
MOLP I
Case 1: c 2
c i2 1 c i c i2
Matthias Ehrgott
MOLP I
Case 1: c 2
c i2 1 c i c i2
Matthias Ehrgott
MOLP I
Case 1: c 2
c i2 1 c i c i2
Matthias Ehrgott
MOLP I
Case 1: c 2
c i2 1 c i c i2
Matthias Ehrgott
MOLP I
I = {i N : c i2 < 0, c i1
0} c i2 . c i1 c i2 (11)
:= max
i I
] is optimal for all [ , B As soon as < new bases become optimal Entering variable xs has to be chosen where the maximum in (11) is attained for i = s
Matthias Ehrgott
MOLP I
I = {i N : c i2 < 0, c i1
0} c i2 . c i1 c i2 (11)
:= max
i I
Matthias Ehrgott
MOLP I
I = {i N : c i2 < 0, c i1
0} c i2 . c i1 c i2 (11)
:= max
i I
Matthias Ehrgott
MOLP I
I = {i N : c i2 < 0, c i1
0} c i2 . c i1 c i2 (11)
:= max
i I
Matthias Ehrgott
MOLP I
I = {i N : c i2 < 0, c i1
0} c i2 . c i1 c i2 (11)
:= max
i I
Matthias Ehrgott
MOLP I
Algorithm (Parametric Simplex for biobjective LPs) Input: Data A, b , C for a biobjective LP. Phase I: Solve the auxiliary LP for Phase I using the Simplex algorithm. If the optimal value is positive, STOP, X = . Otherwise let B be an optimal basis. Phase II: Solve the LP (9) for = 1 starting from basis B found in Phase I and b. . Compute A yielding an optimal basis B Phase III: While I = {i N : c i2 < 0, c i1 0} = . := max
c i2 . i1 c i2 i I c
n o c i2 . s argmax i I : c 1 2 i c i n o j b r argmin j B : A , Ajs > 0 .

js
and b. Let B := (B \ {r }) {s } and update A End while. Output: Sequence of -values and sequence of optimal BFSs.
Matthias Ehrgott
MOLP I
Overview
Matthias Ehrgott
MOLP I
Example min subject to 3x 1 + x 2 x1 2x2 x2 3x 1 x 2 x 3 6 0
LP () min (4 1)x1 + (3 2)x2 subject to x2 + x3 = 3 3x 1 x2 + x4 = 6 x 0.
Matthias Ehrgott
MOLP I
Use Simplex tableaus showing reduced cost vectors c 1 and c 2 Optimal basis for = 1 is B = {3, 4}, optimal basic feasible solution x = (0, 0, 3, 6) Start with Phase 3
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Iteration 1: c 1 c 2 x3 x4 3 -1 0 3 1 -2 1 -1 0 0 1 0 0 0 0 1 0 0 3 6
= 1, c () = (3, 1, 0, 0), B 1 = {3, 4}, x 1 = (0, 0, 3, 6) 1 2 I = {1, 2}, = max 3+1 , 1+2 =2 3 s = 2, r = 3
Matthias Ehrgott
MOLP I
Iteration 2 c 1 c 2 x2 x4 3 -1 0 3 0 0 1 0 -1 2 1 1 0 0 0 1 -3 6 3 9
= 2/3, c () = (5/3, 0, 0, 0), B 2 = {2, 4}, x 2 = (0, 3, 0, 9) 1 I = {1}, = max 3+1 =1 4 s = 1, r = 4
Matthias Ehrgott
MOLP I
Iteration 3 c 1 c 2 x2 x1 0 0 0 1 0 0 1 0 -2 7/3 1 1/3 -1 1/3 0 1/3 -12 9 3 3
= 1/4, c () = (0, 0, 5/4, 0), B 3 = {1, 2}, x 3 = (3, 3, 0, 0) I=
Matthias Ehrgott
MOLP I
Weight values 1 = 1, 2 = 2/3, 3 = 1/4, 4 = 0 Basic feasible solutions x 1 , x 2 , x 3 In each iteration c () can be calculated with the previous and current c 1 and c 2 . Basis B 1 = (3, 4) and BFS x 1 = (0, 0, 3, 6) are optimal for [2/3, 1]. Basis B 2 = (2, 4) and BFS x 2 = (0, 3, 0, 9) are optimal for [1/4, 2/3], and Basis B 3 = (1, 2) and BFS x 3 = (3, 3, 0, 0) are optimal for [0, 1/4]. Objective vectors for basic feasible solutions: Cx 1 = (0, 0), Cx 2 = (3, 6), and Cx 3 = (12, 9)
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Values = 2/3 and = 1/4 correspond to weight vectors (2/3, 1/3) and (1/4, 3/4) Contour lines for weighted sum objectives in decision are parallel to ecient edges 2 (3x1 + x2 ) + 3 1 (3x1 + x2 ) + 4
4
1 5 (x1 2x2 ) = x1 3 3 3 5 (x1 2x2 ) = x2 4 4
x2 X x1
0 1
x3
Feasible set in decision space and ecient set
x4
2 3 4
Matthias Ehrgott
MOLP I
4
1 5 (x1 2x2 ) = x1 3 3 3 5 (x1 2x2 ) = x2 4 4
x2 X x1
0 1
x3
x4
2 3 4
Matthias Ehrgott
MOLP I
4
1 5 (x1 2x2 ) = x1 3 3 3 5 (x1 2x2 ) = x2 4 4
x2 X x1
0 1
x3
x4
2 3 4
Matthias Ehrgott
MOLP I
Weight vectors (2/3, 1/3) and (1/4, 3/4) are normal to nondominated edges
0 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
Cx1 Cx4 Y
-1
-2
-3
-4
-5
-6
Cx2
-7
-8
-9
Cx3
Objective space and nondominated set
-10
Algorithm nds all nondominated extreme points in objective space and one ecient bfs for each of those Algorithm does not nd all ecient solutions just as Simplex algorithm does not nd all optimal solutions of an LP Example min (x1 , x2 )T subject to 0 xi 1 Ecient set: {x R3 : x1 = x2 = 0, 0 i = 1, 2, 3 x3 1}
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
A Multiobjective Simplex Algorithm Multiobjective Simplex: Examples
Overview
Matthias Ehrgott
MOLP I
min{Cx : Ax = b , x
0} 1 Let B be a basis and C = C CB A B A and R = CN How to calculate critical if p > 2? 3 1 N = At B1 : C , = 2/3, = (2/3, 1/3)T and 1 2 N = (5/3, 0)T T C 3 1 N = At B2 : C , = 1/4, = (1/4, 3/4)T and 1 2 N = (0, 5/4)T T C Find Rp , > 0 such that T R 0 (optimality) and T r j = 0 (alternative optimum) for some column r j of R
Matthias Ehrgott
MOLP I
min{Cx : Ax = b , x
Matthias Ehrgott
MOLP I
min{Cx : Ax = b , x
Matthias Ehrgott
MOLP I
min{Cx : Ax = b , x
Matthias Ehrgott
MOLP I
min{Cx : Ax = b , x
Matthias Ehrgott
MOLP I
min{Cx : Ax = b , x
Matthias Ehrgott
MOLP I
Lemma If XE = then X has an ecient basic feasible solution. Proof. There is some > 0 such that minx X T Cx has an optimal solution Thus LP () has an optimal basic feasible solution solution, which is an ecient solution of the MOLP
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Denition 1 A feasible basis B is called ecient basis if B is an optimal basis of LP() for some Rp >. are called adjacent if one can be obtained 2 Two bases B and B from the other by a single pivot step.
3
Let B be an ecient basis. Variable xj , j N is called ecient nonbasic variable at B if there exists a Rp > such that T R 0 and T r j = 0, where r j is the column of R corresponding to variable xj . Let B be an ecient basis and let xj be an ecient nonbasic variable. Then a feasible pivot from B with xj entering the basis (even with negative pivot element) is called an ecient pivot with respect to B and xj
Matthias Ehrgott
MOLP I
3
Matthias Ehrgott
MOLP I
3
Matthias Ehrgott
MOLP I
3
Matthias Ehrgott
MOLP I
No ecient basis is optimal for all p objectives at the same time Therefore R always contains positive and negative entries Proposition Let B be an ecient basis. There exists an ecient nonbasic variable at B .
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
It is not possible to dene ecient nonbasic variables by the existence of a column in R with positive and negative entries Example R= 3 2 2 1
T r 2 = 0 requires 2 = 21 T r 1 0 requires 1 0, an impossibility for > 0
Matthias Ehrgott
MOLP I
Lemma Let B be an ecient basis and xj be an ecient nonbasic variable. Then any ecient pivot from B leads to an adjacent ecient basis . B Proof. xj ecient entering variable at basis B
T there is Rp > with R
0 and T r j = 0
xj is nonbasic variable with reduced cost 0 in LP() Reduced costs of LP() do not change after a pivot with xj entering be the resulting basis with feasible pivot and xj entering Let B , B is an optimal basis Because T R 0 and T r j = 0 at B for LP() and therefore an adjacent ecient basis
0 and T r j = 0
0 and T r j = 0
0 and T r j = 0
0 and T r j = 0
0 and T r j = 0
0 and T r j = 0
How to identify ecient nonbasic variables? Theorem Let B be an ecient basis and let xj be a nonbasic variable. Variable xj is an ecient nonbasic variable if and only if the LP max et v j subject to Rz r + Iv z , , v has an optimal value of 0. (12) is always feasible with (z , , v ) = 0 = 0 0 (12)
Matthias Ehrgott
MOLP I
How to identify ecient nonbasic variables? Theorem Let B be an ecient basis and let xj be a nonbasic variable. Variable xj is an ecient nonbasic variable if and only if the LP max et v j subject to Rz r + Iv z , , v has an optimal value of 0. (12) is always feasible with (z , , v ) = 0 = 0 0 (12)
Matthias Ehrgott
MOLP I
Proof. By denition xj is an ecient nonbasic variable if the LP min subject to 0T = 0 RT 0 j T (r ) = 0 I e
(13)
has an optimal objective value of 0, i.e. if it is feasible (13) is equivalent to min subject to 0T = 0 RT 0 j T (r ) 0 I e
(14)
Matthias Ehrgott
MOLP I
Proof. By denition xj is an ecient nonbasic variable if the LP min subject to 0T = 0 RT 0 j T (r ) = 0 I e
(13)
has an optimal objective value of 0, i.e. if it is feasible (13) is equivalent to min subject to 0T = 0 RT 0 j T (r ) 0 I e
(14)
Matthias Ehrgott
MOLP I
Proof. The dual of (14) is max subject to rj eT v + Iv z , , v
Rz
= 0 0.
(15)
Matthias Ehrgott
MOLP I
Need to show: ALL ecient bases can be reached by ecient pivots Denition are called connected if one can be Two ecient bases B and B obtained from the other by performing only ecient pivots. Theorem All ecient bases are connected.
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Matthias Ehrgott
MOLP I
Proof. two ecient bases B and B Rp , > such that B and B are optimal bases for LP() and LP() Parametric LP ( [0, 1]) with objective function T C + (1 )T C c () = is rst basis (for = 1) Assume B for LP() After several pivots get an optimal basis B p + (1 ) R> for all [0, 1] all bases are Since = optimal for LP( ) for some Rp > , i.e. ecient If B = B , done by ecient pivots: they are Otherwise obtain B from B alternative optima for LP ()
(16)
(16)
(16)
(16)
(16)
(16)
(16)
(16)
3 cases
X = , infeasibility X = but XE = , no ecient solutions X = , X E =
result in three phase multiobjective Simplex algorithm Phase I: Solve min{e T z : Ax + Iz = b , x 0, z 0} If optimal value is nonzero, X = Otherwise nd bfs of Ax = b , x 0 from optimal solution Phase II: Find ecient bfs by solving appropriate LP () Note: LP () can be unbounded even if XE = Solve min{u T b + w T Cx 0 : u T A + w T C 0, w e } If unbounded then XE = Otherwise nd optimal w and solve min{w Cx : Ax = b , x 0} Optimal bfs x 1 exists and is ecient bfs for MOLP Phase III: Starting from x 1 nd all ecient bfs by ecient pivots, even with negative pivot elements
3 cases
3 cases
3 cases
Algorithm (Multicriteria Simplex Algorithm.) Input: Data A, b , C of an MOLP. Initialization: Set L1 := , L2 := . Phase I: Solve the LP min{e T z : Ax + Iz = b , x , z 0}. If the optimal value of this LP is nonzero, STOP, X = . Otherwise let x 0 be a basic feasible solution of the MOLP. Phase II: Solve the LP min{u T b + w T Cx 0 : u T A + w T C 0, w e }. If the problem is infeasible, STOP, XE = . Otherwise let ( u, w ) be an optimal solution. Find an optimal basis B of the LP min{w T Cx : Ax = b , x L1 := {B}, L2 := .
0}.
Matthias Ehrgott
MOLP I
Algorithm
Phase III: While L1 = Choose B in L1 , set L1 := L1 \ {B}, L2 := L2 {B}. b, and R according to B. Compute A, EN := N . For all j N . Solve the LP max{e T v : Ry r j + Iv = 0; y , , v 0}. If this LP is unbounded EN := EN \ {j }. End for For all j EN . For all i B. If B = (B \ {i }) {j } is feasible and B L1 L2 then L1 := L1 B . End for. End for. End while. Output: L2 .
Example There can be exponentially many ecient bfs min xi min xi subject to xi xi i i i i = 1, . . . , n = 1, . . . , n = 1, . . . , n = 1, . . . , n.
1 1
n variables, m = 2n constraints, p = 2n objective functions all 2n extreme points of the feasible set are ecient
Matthias Ehrgott
MOLP I
1 1
Matthias Ehrgott
MOLP I
1 1
Matthias Ehrgott
MOLP I
1 1
Matthias Ehrgott
MOLP I
Overview
Matthias Ehrgott
MOLP I
Example min subject to 3x 1 + x 2 x1 2x2 x2 3x 1 x 2 x 3 6 0
LP () min (4 1)x1 + (3 2)x2 subject to x2 + x3 = 3 3x 1 x2 + x4 = 6 x 0.
Matthias Ehrgott
MOLP I
Phase I: MOLP is feasible x 0 = (0, 0) Phase II: Optimal weight w = (1, 1) Phase II: First ecient solution x 2 = (0, 3) Phase III: Ecient entering variables s 1 , x 2 Phase III: Ecient solutions x 1 = (0, 0), x 3 = (3, 3) Phase III: No more ecient entering variables
x2 X x1
0 1
x3
x4
2 3 4
Matthias Ehrgott
MOLP I
x2 X x1
0 1
x3
x4
2 3 4
Matthias Ehrgott
MOLP I
x2 X x1
0 1
x3
x4
2 3 4
Matthias Ehrgott
MOLP I
x2 X x1
0 1
x3
x4
2 3 4
Matthias Ehrgott
MOLP I
x2 X x1
0 1
x3
x4
2 3 4
Matthias Ehrgott
MOLP I
x2 X x1
0 1
x3
x4
2 3 4
Matthias Ehrgott
MOLP I
Example min x1 2x2 min x1 +2x3 min x1 x3 subject to x1 +x2 1 x2 2 x1 x2 +x3 4. Slack variables x4 , x5 , x6 introduced to write the constraints in equality form Ax = b
Matthias Ehrgott
MOLP I
Phase I: B = {4, 5, 6} is a basis with bfs x 0 = (0, 0, 0, 1, 2, 4) Phase II:

min subject to u1 + 02u2 + 4u3 1 1 0 1 T @ 0 1 0 0 u 1 1 1 0 0 1 0 1 0 0 1 0 A + w T @ 1 1 1 2 0 0 0 0 0 2 0 0 1 0 0 1 0 0 A 0 w 0 e
w = (1, 1, 1) min{x1 2x2 + x3 : Ax = b , x 0} B 1 = {2, 5, 6}, x 1 = (0, 1, 0, 0, 1, 3) is ecient bfs, L1 = {{2, 5, 6}}
Matthias Ehrgott
MOLP I
Phase I: B = {4, 5, 6} is a basis with bfs x 0 = (0, 0, 0, 1, 2, 4) Phase II:

min subject to u1 + 02u2 + 4u3 1 1 0 1 T @ 0 1 0 0 u 1 1 1 0 0 1 0 1 0 0 1 0 A + w T @ 1 1 1 2 0 0 0 0 0 2 0 0 1 0 0 1 0 0 A 0 w 0 e
w = (1, 1, 1) min{x1 2x2 + x3 : Ax = b , x 0} B 1 = {2, 5, 6}, x 1 = (0, 1, 0, 0, 1, 3) is ecient bfs, L1 = {{2, 5, 6}}
Matthias Ehrgott
MOLP I
Phase III Iteration 1: B 1 = {2, 5, 6}, L1 = , L2 = {{2, 5, 6}} c 1 c 2 c 3 x2 x5 x6 1 -1 1 1 -1 2 0 0 0 1 0 0 0 2 -1 0 0 1 2 0 0 1 -1 1 0 0 0 0 1 0 0 0 0 0 0 1 2 0 0 1 1 5
EN := {1, 3, 4}
Matthias Ehrgott
MOLP I
Check x1 1 1 -1 1 Check x3 1 1 -1 1 1 2 -1 0 0 0 0 0 0 0 0 2 0 1 0 0 2 0 -2 0 1 0 -1 0 1 0 0 1 LP has optimal solution, x3 is ecient 1 2 -1 0 0 0 0 0 0 0 0 2 -1 1 0 0 2 0 1 0 1 0 -1 0 -1 0 0 1 LP has optimal solution, x1 is ecient
Matthias Ehrgott
MOLP I
Check x1 1 1 -1 1 Check x3 1 1 -1 1 1 2 -1 0 0 0 0 0 0 0 0 2 0 1 0 0 2 0 -2 0 1 0 -1 0 1 0 0 1 LP has optimal solution, x3 is ecient 1 2 -1 0 0 0 0 0 0 0 0 2 -1 1 0 0 2 0 1 0 1 0 -1 0 -1 0 0 1 LP has optimal solution, x1 is ecient
Matthias Ehrgott
MOLP I
Check x4 1 1 -1 1 1 0 2 -1 LP 2 -2 0 0 0 0 0 0 0 2 -2 0 0 0 0 is unbounded, 1 0 0 0 1 0 0 0 1 x4 is not ecient
EN = {1, 3} Feasible pivot x1 enters and x2 leaves: basis B 2 = {1, 5, 6} Feasible pivot x3 enters and x6 leaves: basis B 3 = {2, 3, 5} L1 := {{1, 5, 6}, {2, 3, 5}}
Matthias Ehrgott
MOLP I
Check x4 1 1 -1 1 1 0 2 -1 LP 2 -2 0 0 0 0 0 0 0 2 -2 0 0 0 0 is unbounded, 1 0 0 0 1 0 0 0 1 x4 is not ecient
EN = {1, 3} Feasible pivot x1 enters and x2 leaves: basis B 2 = {1, 5, 6} Feasible pivot x3 enters and x6 leaves: basis B 3 = {2, 3, 5} L1 := {{1, 5, 6}, {2, 3, 5}}
Matthias Ehrgott
MOLP I
Iteration 2: B 2 = {1, 5, 6} with BFS x 2 = (1, 0, 0, 0, 2, 3) L1 = {{2, 3, 5}}, L2 = {{2, 5, 6}, {2, 3, 5}} c 1 c 2 c 3 x2 x5 x6 EN = {2, 3, 4} 0 0 0 1 0 0 -1 1 -1 1 1 -2 0 2 -1 0 0 1 1 1 -1 1 0 -1 0 0 0 0 1 0 0 0 0 0 0 1 1 1 -1 1 2 3
Matthias Ehrgott
MOLP I
Check x2 : Leads back to B 1 = (2, 5, 6) Check x3 : -1 1 1 1 1 -1 -1 0 -2 1 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 -1 0 1 2 -1 -1 x3 not ecient Check x4 -1 1 -1 0 1 2 -1 -1 x4 not ecient EN =
1 1 1 -1
-1 -1 -1 1
0 1 0 0
0 0 1 0
0 0 0 1
0 0 0 0
Matthias Ehrgott
MOLP I
1 1 1 -1
-1 -1 -1 1
0 1 0 0
0 0 1 0
0 0 0 1
0 0 0 0
Matthias Ehrgott
MOLP I
1 1 1 -1
-1 -1 -1 1
0 1 0 0
0 0 1 0
0 0 0 1
0 0 0 0
Matthias Ehrgott
MOLP I
1 1 1 -1
-1 -1 -1 1
0 1 0 0
0 0 1 0
0 0 0 1
0 0 0 0
Matthias Ehrgott
MOLP I
Iteration 3 B 3 = {2, 3, 5} with bfs x 3 = (0, 1, 5, 0, 1, 0) L1 = , L2 = {{2, 5, 6}, {1, 5, 6}, {2, 3, 5}} c 1 c 2 c 3 x2 x5 x3 EN = {1, 4, 6} 1 -5 3 1 -1 2 0 0 0 1 0 0 0 0 0 0 0 1 2 -2 1 1 -1 1 0 0 0 0 1 0 0 -2 1 0 0 1 2 -10 5 1 1 5
Matthias Ehrgott
MOLP I
Check x1 -1 1 -5 3 Check x4 -1 1 -5 3 1 2 -2 1 -1 0 -2 1 -1 -2 2 -1 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1 2 -2 1 -1 0 -2 1 1 -1 5 -3 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0
x4 is not ecient
x4 is not ecient Check x6 : Leads back to B 1

Check x1 -1 1 -5 3 Check x4 -1 1 -5 3 1 2 -2 1 -1 0 -2 1 -1 -2 2 -1 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1 2 -2 1 -1 0 -2 1 1 -1 5 -3 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0
x4 is not ecient

Check x1 -1 1 -5 3 Check x4 -1 1 -5 3 1 2 -2 1 -1 0 -2 1 -1 -2 2 -1 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1 2 -2 1 -1 0 -2 1 1 -1 5 -3 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0
x4 is not ecient

Iteration 4: L1 = , STOP Output: List of ecient bases B 1 = {2, 5, 6}, B 2 = {1, 5, 6}, B 3 = {2, 3, 5}
x {2, 5, 6}
. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
x2 {1, 5, 6}
x {2, 3, 5}
.. .. . . . . . . . . . . . . . . . . . . . . 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . 2 . . . . . . . . . . . . . . . . . . . . . . . .. . .. . . . . 5. . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . 3 . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . 1 . . . . . . . .. . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . .. . . . . . .. . . . . . . . . . . . . . .. . . . 1 . . . .. .. . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . 2 . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . 0 . . . . .. . . . . . . . . . . . . 1 . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 . . . . . . . . . . . . . . . . . . . . .
. x3 ...... . . . . . . .
6 5
4 x 3 2 1
XE x0 = (0, 0, 0) x1 = (0, 1, 0) x2 = (1, 0, 0) x3 = (0, 1, 5) x4 = (1, 0, 3) x5 = (0, 0, 4)
Matthias Ehrgott
MOLP I

HanLecture2 ME

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

HanLecture2 ME

Uploaded by

Copyright:

Available Formats

Multiobjective Linear Programming Biobjective LPs and Parametric Simplex Multiobjective Simplex Method

International Doctoral School Algorithmic Decision Theory: MCDA and MOO

MCDA and MOO, Han sur Lesse, September 17 21 2007

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Example min 3x 1 + x 2 x1 2x2 subject to x2 3x 1 x 2 x 3 1 1 2

Formulation and Example Solving MOLPs by Weighted Sums

Example min 3x 1 + x 2 x1 2x2 subject to x2 3x 1 x 2 x 3 1 1 2

Formulation and Example Solving MOLPs by Weighted Sums

Feasible set in decision space

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Illustration in objective space

1 = (2, 1), 2 = (1, 3), 3 = (1, 1)

Formulation and Example Solving MOLPs by Weighted Sums

Illustration in objective space

1 = (2, 1), 2 = (1, 3), 3 = (1, 1)

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

Formulation and Example Solving MOLPs by Weighted Sums

e > 0 from the constraints in (3)

Formulation and Example Solving MOLPs by Weighted Sums

e > 0 from the constraints in (3)

Formulation and Example Solving MOLPs by Weighted Sums

e > 0 from the constraints in (3)

Formulation and Example Solving MOLPs by Weighted Sums