z

Fr'

THE OPERATIONS RESEARCH PROGRAM z

0z
DEPARTMENT OF INDUSTRIAL ENGINEERING
STATE UNIVERSITY OF NEW YORK AT BUFFALO rc
N
O
SUNY AT BUFFALO n

MULTILEVEL LINEAR PROGRAMMING

by
Wayne F. Bialas
Mark H. Karwan

Research Report No. 78-1

May 1978

Department of Industrial Engineering

State University of New York at Buffalo
Buffalo, New York 14260
 *
MULTILEVEL LINEAR PROGRAMMING

Wayne F. Bialas t

Mark H. Karwan

Technical Report No. 78-1

*Presented at the Joint National Meeting of the Operations Research

Society of America and The Institute of Management Sciences,
May 1-3, 1978.

tAssistant Professor, Department of Industrial Engineering, State

University of New York at Buffalo, Amherst, N. Y. 14260.
+Assistant Professor, Department of Industrial Engineering, State
University of New York at Buffalo, Amherst, N. Y. 14260.
 Abstract

The general multilevel programming problem is a set of nested

optimization problems over a single feasible region. Control over the

decision variables is partitioned among the levels, but a decision variable

may impact the objective functions of several, if not all, levels. The

two-level linear resource control problem is a special case where the

level-two planner controls the effective resource space for the level-

one planner. This produces a solution structure with the feasible region

viewed by level two as a nonconvex subset of the overall feasible region.

An algorithm to solve this problem is proposed.

•
Introduction

Many planning situations require the analysis of several objectives.

Multiobjective optimization techniques have been developed to permit a more

faithful analysis of the tradeoffs among competing goals, and assist a planner

in reaching an acceptable compromise. Such methods assume that all objectives

are those of a single planner, impacting directly on his state of well-being.

Multiobjective optimization fails to recognize that many objectives are ordered

within an administrative or other hierarchical structure. For example, the

energy policies set forth by the Federal government affect the objectives and

options, and hence the strategies, of state officials. This process continues

within a hierarchy including local governments, planning agencies, and basic

economic units such as firms and households. Each unit in the hierarchy wishes

to maximize its individual benefits in view of the partial exogenous control

exercised at other levels. In the example, actions at the state level affect

the benefits sought by the Federal government which can, in turn, constitutionally

exercise first, but only partial control over the state.

Multilevel optimization techniques partition control over the decision

variables among the levels, and analyze objectives at all levels of the

planning process. A planner at one level of the hierarchy may have his

objective function determined, in part, by variables controlled at other

levels. However, his control instruments may allow him to influence the

policies at other levels and thereby improve his own objective function.

The problem of multilevel linear programming was offered by Candler

and Norton [2]. Unfortunately, their general problem was imprecisely defined

and they did not realize that the effective feasible regions for higher

1
levels were nonconvex sets. This causes the Candler and Norton algorithm

to fail since it assumes convexity. Although the general problem formulation

and solution technique proposed by them is incorrect, the basic premise of

their problem has far-reaching applications. Some particular problems are

amenable to this framework:

(i) Control of oil imports: With the desire to limit oil imports, the
Federal government can impose import quotas and duties. This action
will affect the consumption of oil by economic units which are maximizing
profits. The decisions made by these units would then impact on the
objectives of the Federal government including a change in consumption
patterns and levels of oil imports.

(ii) Floodplain planning: Seeking to reduce flood risk, government can

implement floodplain zoning programs and subsidize flood insurance.
These policies will influence the land activity patterns based on the
budgets and objectives of land users.

(iii)Job incentiveprograms: As industry seeks to maximize productivity, it

may implement profit-sharing and other job incentive programs. Such
programs, in turn, affect the goals of employees whose decisions impact
on industrial productivity.

Similar work in this general area has been conducted on Stackelberg

games [1,3,7].Within the broad definition of such games, a static Stackelberg

game with fixed leaders and a continuous decision space could be defined to

encompass multilevel optimization problems. However, current methodology

does not consider the activity space of one player to be a function of the

strategies of other players. Such an extension of Stackelberg games would

require the payoff function of one level to have discontinuities dependent on

the decisions of other players. This formulation is, at best, unwieldly,

and perhaps intractable.

The term "multilevel" has been frequently used to describe approaches

to similar important planning problems. However, in the context of the

multilevel programming problem defined in the next section, these previous

approaches are primarily decomposition techniques applied to single level

problems [4,5,6]. For example, Haimes, Foley and Yu [5] employ lagrangian

duality to decompose and efficiently solve a large model for the control of

water quality with a single overall system objective of a central planning

agency. The dual variables are then interpreted, as with Dantzig-Wolfe

decomposition, to determine prices (taxes) to be charged to each subproblem

(polluter) for violating pollution standards.

The flexibility of the multilevel approach can answer questions

regarding the assignment of control over variables to various levels. For

example, in some cases, coalitions of levels can improve the objective functions

of all levels. Furthermore, because of the structure of some problems, a single

level could exercise complete control over all levels even though controlling

only a proper subset of the variables. This methodology can assess the value

of controlling a particular subset of variables, and with this information,

policy makers could determine what control should be relinquished or maintained

over certain variables.

The next section will provide the formal, general definition of a multilevel

programming problem. Then the discussion will focus on a special case of this

general problem, the two-level linear resource control problem, and its

mathematical structure. This will include a characterization of the solution

set and some general algorithmi c approaches to solve the problem. Finally, a

specific algorithm for finding local optimal solutions will be presented.

General Definition of Multilevel Programming Problems

n
Let the decision variable space (Euclidean n-space) R x = (x ,x ,...,x )
1 2 n

be partitioned among r levels,

nk • 4 x k k k k,
R = (x 1 ' 2 "-- ,xn k=1,...,r ,
1
k
r n
where nk = n. Denote the maximization of a function f(x) over R by
k=1 n n
k k+1 k+2 r nk +l k+2 x
varying only x E R k given fixed x, x in R ^ R

by max f(x)
k k+1 k+2
x ix ,x ,...x .

The general multilevel programming problem can then be defined as

max f (x)
r
r
x
st: max f (x)
r-1
r-1 1 r
x Ix

st: max f (x)

r-2
xr-2 r-1 r
ix ,x
r-1
(P ) 4
r-2
(P ) 4
st: max f (x)
11 2 1
x ix ,x- ,...,x

st:xESCR

This establishes a collection of nested mathematical programming problems

{P 1 ,.. ,P r }. The feasible region, S = S l , is defined as the level-one feasible

1 n 2 3 r
region. The solutions to P in R for each fixed x ,x ,..., x form a set
1
2 1
S = {x E S :f (x) = max f (x) } called the level-two feasible region over
1 1
-r
xi 1X 2 ,X 3 ,..., x
2 3 r
which f (x) is then maximized by varying x for fixed x ,...,x .
2
In general, the level-k feasible region is defined recursively as

^ Sk-1 :
S = {t x
x E - fk-1 (x)= max fk-1 (x) }
k-1 ^k ^r
x ix ,...,x
k-1 r
Note that x is a function of x lc ,...,x . Furthermore, the problem at

each level can be written as

max f (x)
k
k k+1
x ix ,...,x
( Pk )
x E Sk

k+1 r r
which is a function of x , and (P ): max f (x) defines the entire
x€S r r
problem.

4
The Two-Level Linear Resource Control Problem

The two-level linear resource control problem is the multilevel programming

problem, of the form

2
max c x
x
1
st: max c x
2
(P ) 4 x
(P l 2
A x + A x <b
1 2 -
x > 0 .

2
Here, level 2 controls x which, in turn, varies the resource space of level
l 2
one by restricting A x < b - A x .
1 2

Ill-Defined Problems
k
Care must be taken when P results in alternative optimal solutions for
k+1r
fixed x . Although not affecting the value of the level-k objective

function, f k (x), these solutions can have drastically varying impact on the

objectives of other levels. Therefore, control over the choice among alternative

optimal solutions may have to be delegated to other levels or an incentive

scheme may be required to induce the level-k planner to choose a solution

desirable to level k+1. If no such scheme is employed, the problem may be

ill defined.

Consider the following example of a two-level linear resource control

problem:

5
 max + 1 x2
1 2 x
x
2 max x
2 + x
st: 3
x i , x3 2
1x

2 x + x = 4
2 3
( P )1 1 x > 1
(P ) 2
- x + 2x < 2
1 2
x + x < 4
1 2

For x = 2, the unique level-one solution is (x 11 x 2, x 3 ) = (2,2,2) with value 4.

2
The corresponding level-two solution value is 3. However, for x 2 = 1, there

exists a set of alternate optimal solutions to P 1 X = f(x ,x x )
1 2' 3
0 < x < 3, x = 1, x = 3} . The corresponding level-two objective for
1 2 3
1 1
this set of solutions ranges continuously from to 3 For a unique solution
2'
to be returned to level two for fixed x
2 and to induce level one to return
x = 3 for x = 1, a side payment to the level-one objective from level two may
1 2
be employed. For the example, the level-one objective function,

max x + x + E(x + 1 -x ) with E > 0 sufficiently small, is a perturbation

2 3 1 2

which accomplishes this. Given this side payment scheme, x 2 = 1 is the optimal

decision for level two.

Nonconvexity

In the two-level linear resource control problem,

r 2
S 1 = l x > 0 : A x l + A2 x < b ,
1

is a convex set. However,

2 1 l
S = Ix f S : c x = max c 1 xi1
x11i x2

6
 2
need not be. Therefore, P , which can be written as
2
max c x
x(S 2

involves the optimization of a linear function over a nonconvex region.

2
Consider the geometric example in Figure 1. The set S for the problem in
1
Figure 1 is a subset of the edges of the boundary of S . In problems of higher
2 1
dimension, S is composed of edges and faces of the boundary of S . Consider

the following three dimensional example:

max x + x
2 3
x
2
st: max x
3
xi ,x3 x 2
x + x - x > 2
1 2 3
x - x + x < 2
1 2 3
x + x + x < 6
1 2 3
x + x + x > 2
1 2 3
x < 1 ,
3 -
2
where S is the hatched region shown in Figure 2.

Relationships to Multiobjective Programming

Optimal solutions to the multilevel programming problem may not be

Pareto-optimal. While cooperation might improve the objective functions at

every level, the order and independence with which decisions are made prevent

such cooperation. This rules out any algorithmic approach which seeks only

Pareto-optimal solutions and is one of the main distinguishing characteristics

between multiobjective and multilevel programming.

For an example of this behavior, consider Figure 1. Both levels have

higher objective function values at point (a). However, for x fixed at x

2 2'
2
level one will choose x l = xl (point (b)), thus point (a) is not in S This

leads to the best choice of x to be x = 0 with the optimal solution at point (c).
2 2

7
A Sufficient Condition for Complete Control

Consider the two-level linear resource control problem. Given any basis

B C A for the set of constraints Ax = b, one can write the equivalent set of

constraints on x:

Bx + Nx = b
B N
-1 -1
Or xB = B b - B N xN .

When x N is fixed, x is uniquely determined. Thus to have complete control of

B
the solution, the level-two planner need only control the complete set of

nonbasic variables corresponding to any basis.

2
Further Characterization of S and P2

The following theorem and its corollaries help to characterize both S2

2
and the optimal solution for P in the two-level linear resource control

problem.

Theorem I

Suppose S= {x : Ax = b, x > 0} is bounded. Let S 2 = {x =(x 1 ,x 2 ) E S i :

11 1 1,
c x = max c x /. Then the following hold
x l 1X 2

2 1
(i) S C S

(ii) Let {yd r be any r points of S i , such that

t=1
2
x = A y with A > 0, X = 1. Then A > 0 implies y ( S2
t t S t - t t t
t

2 1 2
Proof: (i) s S by the definition of S .
1
(ii) (By contradiction) Let y1 , y ..' y f S with
2" r
r
1 2 2
x = (x ,x ) = X A y , and
t t S
t=1
r
> 0 A l > 0, X A = 1.
at - t
t=1

8
 k optimal
solution

Figure 1
2
Example of NOnconvexity of S
Figure 2

EXample of S 2in Three Dimensions

10
 1 2 -1 2 _2
Suppose y
= (Y1' Y1 ) f S
Then there exists y1
1
such that = Y 1 Y1 ) (
1 I 1 1
with c 17 1 > c y l .

Using (i), Y l E S . Therefore,'

=AY + XAyc
1 1 t=2 t t
- ,
2
since S1 is convex. Noting that x = x2 and A. >
1
1 1 1 1 1 1
CX c y < cA Y + cAy
= c 1A 1y1 t t t 1 1 1 ttt
t=2 t=2
1 2 2
Given, x = (x ,x ) S and A l > 0, we have established an x with the

following properties:
sl
(a) x=
- 2 2
(b) x =
1 1 < c1-1
x .
(c) c x
2 2 1 1
This contradicts the definition of S since x t S should have maximized c x
2 2
for the fixed value of x . Therefore Al > 0 implies y S . Since the
1
choice of y l 'among the y's was arbitrary, we have proven A t > 0 implies yt c S
Any point which positively contributes in any convex combination
2
forming a point in S , must also be in S2. Since this is true of any point,

including y which are extreme points of S 1 , the following corollary results:

Corollary 1. If x is an extreme point of S 2 , then x is an

1
extreme point of S .

Proof: (By contradiction) Let x be an extreme point of S te . Suppose x is

not an extreme point of S1 . Then there exist extreme points y ,...,y r c S1, and

A > > 0, Such that

1
t=1

x y
t=1
From Theorem 1, this implies .,Yr t S and hence x cannot be an extreme
2
point of S a contradiction.

11


2
Recalling that P may be formulated as max c x and noting the
x(S 2

correspondence of extreme points in S2

S and S 1 , the following result is derived.

Corollary 2. An optimal solution to the two-level linear resource

control problem (if one exists) occurs at an extreme point of the constraint

set of all variables (S ).

Proof: The two-level linear resource control problem can be written as

2
max c x
2
x(S
2
Since c x is linear,if a solution exists, one must occur at an extreme
2
point of'S (alternative optimal solutions at nonextreme points may exist),

By Corollary 1, this must be an extreme point of S 1 .

This result justifies extreme point search procedures as a basis for

algorithmic approaches to solving the two-level linear resource control problem.

Algorithmic Approaches

It has been shown that an optimal solution to the two-level linear

resource control problem occurs at an extreme point of the level-twO feasible

, 2
region, S 2 . Let LS ] denote the convex hull of S 2 . Since the sets of extreme
2
points for S and [S 2 ] are identical, the problem

2
(P2 ) max c x

st: x ( [S 2 ]
is an equivalent formulation for the two-level linear resource control problem.

This suggests a search for cutting plane procedures to approximate the convex
2
hull of S as a direction for future research..

12
 Any desirable algorithm for the two level linear resource control problem

should exhibit some particular properties.

^1 ^2
Consider the solution, x = (x ,x ) to the following problem:
2
max c x
6; )
st: Ax=b, x>0

In (P), the level two planner is given full control over all variables. Now
2 ^2
fix x = x and solve the following problem with solution x to determine if

x f S2 : 1
max c x
1 2
st: = (b-A x )
A1 x
1
x > 0
2 is an optimal solution to the overall problem. For
If x = x, then x t S

example, note that in Figure 1, the vector c 2 could be changed to produce a

9
solution to (P) at any extreme point of S 2 .. The set $ does not vary with

changes in the second level objective, and, hence quite different choices of c 2

can produce an optimal solution after solving (P). For the example shown in
2
Figure 1, two particular choices of c which lead to such a condition for the
1 2
level-one objective shown are both c2= c and c = -c 1 . Thus both highly

complementary and highly conflicting objectives (as well as many inbetween)

may lead to solutions after solving the two linear programming problems (P) and

(P). Any reasonable algorithm should have the ability to easily solve any

problems for which x e S 2 .

An Algorithm to Find Local Optimal Solutions

Consider the following portion of abounded simplex tableau to be employed

in the proposed algorithm:

13
 2 2 2
xl RHS
.._
2 2
r 22 2
r r r
1 k
x Y1
B Y 11 Y lk
1
x -b2
1-;
B Y21 Y22 — - — '''''' " Y2k
. 2 .. .
_
xB Yml Y m2 b
m Ymk m

2
The variables, x 2 x
k represent the nonbasic level-two variables which are
at nonzero values, and r i2 ,...,rk2 represent the reduced costs of these variables

with respect to the level-two objective function. In terms of the present

basisBCA,b=B_ b- y yx 2 where (y ,Y ,t
= Y, = B
-1
(a ,a
j lj 2j mj l j 2j
j= 1
and xB
x denotes the ith basic variable.

Assume that S is boUnded with no degeneracy and no alternative optimal

solutions exist for P 1 for any feasible x 2 .

The following algorithm guarantees a local optimal solUtion.

Step 1. Solve the following problem with optimal solution x = (x ,x and

^1 ^2

optimal tableau T via the simplex method:

2
Step 2. Set x2= x and solve the following problem via bounded simplex

(k=u=x 2 ) beginning with tableau T:

1
max c x
st: Ax=b
2 ^2
x =x
x >0

Let the optimal solution be x. If x = x, stop; x is a global optimal

solution. Otherwise, go to step 3a with current tableau T and relax

2 ^2
the constraints x = x .

14


Step 3a. If all nonbasic variables are equal to zero, go to step 4 with current

tableau T. Otherwise go to step 3b.

Step 3b. If b. > 0 for all i, go to step 3c. Otherwise, without loss of

generality, consider b t = 0. Choose such that 1 < j < k and

yQj
y 0. Bring x2 into the basis via a degenerate pivot. Go to
3
3
step 3a.

Step 3c. Consider any nonbasic variable which is at a strictly positive value,
2 2 2 2
say x.. If r. < 0, increase xj until it enters the basis. If r, %P. 0,
3 -
decrease x 2 until either it reaches zero or it must enter the basis.

Go to step 3a.

Step 4. Beginning with tableau T solve the following problem via a modified

simplex procedure:
2
max c X

st: Ax=b
x>0

The modification is as follows. Given a candidate to enter the basis

2 x will increase) only allow it to enter if the
(one for which c
2.
resulting basic solution, x, will be contained in S This is

determined by obtaining the solution x to the following problem:

1
max c x
2
st: A i < (b-A x )
1x 2
1 2 2
x > 0, x = x
1
via dual simplex on repeated applications of step 4. If x = x

then enter the candidate into the basis. Repeat step 4 until no more

candidates exist which satisfy the above mqdification, then stop,

15
Validation and Convergence

The algorithm begins by finding the maximum of the second level objective

over the entire feasible region, S 1 . A check in step 2 is then made to determine

if the resulting solution is in S 2 . If so, the algorithm terminates with a

global optimal solution and has solved what was previously termed an easy

problem. If termination does not occur in step 2, the resulting solution from

step 2 is by definition contained in S 2 . Since the bounded simplex algorithm

was employed, a number of nonbasic level-two variables may be at nonzero values

2
corresponding to appropriate components of x. Degeneracy may also have been
2
introduced by fixing the components of x from step 1.
2
is indeed contained in 5 . Since bounded simplex was employed, a number of

nonbasic level-two variables may be at nonzero values corresponding to

appropriate components of x 2 . Degeneracy may also have been introduced by

fixing the components of x 2 from step 1.

The purpose of step 3 is to relax the constraint x = x 2 and to move to an

extreme point x° which satisfies x o 2 o

S and cx > cx. If a right hand side,

, from the current tableau is equal to zero then step 3b is entered to

perform a degenerate pivot. Some nonbasic variable, x 2 , j=1,2,...,k, is then

brought into the basis at its current positive level and x B becomes nonbasic

at its current value of zero. Thus the number of basic variables at level zero

is reduced by one. This is repeated until no degeneracy is present. Note that

such a pivot is always possible, that is, yk 0 for some j=1,...,k. Suppose
that y 2, = 0 for all j=1,...,k. Then repeated applications of step 3c would

result in a degenerate extreme point of the original feasible region, S 1 , since

xB will remain zero no matter how x 2l ,...,x are varied. This contradicts the

original nondegeneracy assumption.

16
 2
If all b.> 0 but there are still nonbasic level-two variables, x2 . .. ,x k'
2
at nonzero values,. then step 3c is entered. Any variable x j , j=1,...,k, is

chosen to be increased or decreased depending on its reduced level-two cost,

2 2
r.. Since there are no explicit upper bounds on x 3 increase is limited

by a current basic variable reaching zero. The original problem is bounded,so

2 is decreased, again a current basic variable may reach
this must occur. If x
2 itself will become zero. In either case, the number of nonzero
zero or else x
J
nonbasic variables is decreased by one.
2
The points generated in step 3c can be shown to be contained in S which

is assumed when step 4 is entered. Recall that 1;1 0 for all i as a result

of step 3b. Thus there exists two scalars, 0 1 > 0 and 0 2 > 0 such that any
increase or decrease in x, by an amount less than or equal to 01 and 02

respectively results in a feasible solution (i.e., a point in S 1 ). This implies

that the current solution, which is in S 2 , is a convex combination of two

feasible points resulting from a strict increase and a strict decrease in x 23 .

By Theorem I, such points must also be in S 2 . Thus each point resulting from

step 3c must be contained in S 2 .

Step 4 is entered when an extreme point of S has been obtained. A modified

2
simplex method is used to take steps in S along which the level two objective

increases. This is accomplished by using the normal simplex rules with

objective c 2x along with a check that no basis change results in leaving S 2 .

The algorithm terminates with an extreme point solution in S 2 which has the

property that all adjacent extreme points either lead to a decrease in c 2 x or

are not in S 2 . Thus a local optimal solution is obtained.

Convergence of the algorithm is established by noting the following

facts:

17

(i) The feasible region defined by S Ax = b, x > 0t is bounded

and each basis is nondegenerate.

(ii) Steps 1, 2 and 4 are finite since the simplex, bounded simplex and

dual simplex procedures are finite under fact (i).

(iii) Each application of step 3b strictly decreases the number of basic

variables at level zero and also the number of nonzero nonbasic

variables.

(iv) Each application of step 3c reduces the number of nonzero nonbasic

variables by one.

Conclusions

Multilevel mathematical prograMming problems, if carefully defined, can

serve as useful tools in modelling structured economic units. Such models can

predict the inefficiencies of non-Pareto-optimal decisions and identify the

seats of true control within hierarchical organizations.

This paper has proposed a general mathematical structure for such problems,

and specifically characterized the two-level linear resource control problem.

For this problem, Theorem I illustrates a key property of the nonconvex feasible

region viewed by level two. As a foundation, it justifies extreme point solution

techniques and obviates the need for methods to establish the convex hull of the

level two feasible region. Towards this goal, this paper has offered an adjacent

extreme point method which can find local, and sometimes global, optimal solutions

to the two-level linear resource control problem.

Acknowledgement

The authors wish to expreSs their sincere gratitude to Professbr Daniel P.

Loucks for introducing them to this problem

18
Figure Captions

2
FigUre 1. Example of Nonconvexity of S
2
Figure 2. Example of S in Three Dimensions

19
References

1. Basar, T., "On the Relative Leadership Property of Stackelberg

Strategies," Journal of Optimization Theory and Applications.
Volume II, No.6, 1973, pp.655-661.

Candler, W. and R. Norton; Multi-Level Programming, Unpublished

Research Memorandum, DRC, World Bank, Washington, D.C. August 1976.

3. Cruz, J.B., "Stackelberg Strategies for Multilevel Systems," in

Directions in Decentralized Control, Many Person Optimization
and Large-Scale Systems, Y.C. Ho and S.K. Mitter, Eds., New York,
Plenum Press, 1976, pp.139-147.

4. Goreaux, L.M. and A.S. Manne, Multi-Level Planning: Case Studies in

Mexico, North-Holland, Amsterdam, 1973.

5. Haimes, Y.Y., J. Foley and W. Yu, "Computational Results for Water

Pollution Taxation Using Multilevel Approach," Water-Resources
Bulletin, Volume 8, No.4, August 1972, pp.761-771.

6. Haimes, Y.Y., W.A. Hall and H.T. Freedman, Multiobjective Optimization

in Water Resources Systems, Elsevier, Amsterdam, 1975.

7. Simaan, M. and J.B. Craig, "On the Stackelberg Strategy in Nonzero-Sum

Games," Journal of Optimization Theory and Applications, Volume 11,
No,5, 1973, pp.533-555.

20