See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/265668231

Accelerating Benders Decomposition: Algorithmic Enhancement and Model

Selection Criteria

Article  in  Operations Research · June 1981

DOI: 10.1287/opre.29.3.464

CITATIONS READS

404 1,029

2 authors, including:

Richard Wong

45 PUBLICATIONS   3,861 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Aggregation and Disaggregation in Operations Research View project

All content following this page was uploaded by Richard Wong on 13 September 2016.

The user has requested enhancement of the downloaded file.

 - ; : :::::-:: 0i:·
f~~~~~~~~

:-::: - : "-: . : . .:~-;-- :-.,: ,: - :i;

IV C-
;~~~ -)i :-: ,, S:- ~

.,.,! ,~ O -. ,. -: , :-: ,. ~- ... : . -::: - ;- ... : - :-~: .;-.

. F 0. Xf : - f 0:00t:f... '
-::-,}0

-·· .;.; .·' -··1 ;·

i.·-. I; .i.:· r:.;

r .·: r· r
:r
L·:·- i
:
i i I :- · :·:: ·:::··

:....
.. ·' ··
: I:·.-r'·.- - -LI
-: · --
I·'·-
-- ·- ·;· :;-:; ·· .
:-·· ·.
:'·· :·
i.. ··
....-
·-··-

0;- F. Tc :

OFT_.
, NO .
 ACCELERATING BENDERS DECOMPOSITION:
ALGORITHMIC ENHANCEMENTS AND
MODEL SELECTION CRITERIA

by

Thomas L. Magnanti

and

Richard T. Wong

OR 093-79 October 1979

This research was supported by the U.S. Department of Transportation

under Contract DOT-TSC-1058, Transportation Advanced Research Program
(TARP).
I
 Benders decomposition is an algorithm for mixed integer programming

that has been applied successively to a variety of applications. Florian,

Guerin, and Bushel [11 have used the algorithm to schedule the movement

of railway engines, Richardson [33] has applied the algorithm to airline

routing, and Geoffrion and Graves [16] have had great success applying the

algorithm to design industrial distribution systems. Fisher and Jaikumar

[9] have recently discussed the advantages of using the algorithm for ve-

hicle routing problems. These contributions demonstrate the potential for

using Benders decomposition to solve specially structured mixed integer

programs.

Unfortunately, Benders procedure has not been uniformly successful in

all applications. Geoffrion and Graves 16], among others, have noted

that reformulating a mixed integer program can have a profound effect upon

the efficiency of the algorithm. In earlier computational experience with

a class of network design problems [38], we have observed that straight-

forward implementation of Benders algorithm often converged very slowly,

requiring the solution of an exorbitant number of integer programming

problems. This experience raises a number of questions. How does problem

formulation effect the performance of Benders algorithm? Is there a "best"

formulation for a given problem? For those applications where the algo-

rithm does not work well, is there a mechanism for improving its conver-

gence properties? The resolution of these questions would enhance pros-

pects for applying Benders decomposition and has prompted the study re-

ported in this paper.

In the first section, we review the essential concepts of Benders

decomposition. In section 2, we describe an acceleration technique for

 -2-

reducing the number of iterations of Benders procedure. We accomplish

this by choosing judiciously from the possible cuts that could be gene-

ratad at any iteration to obtain "strong" or "pareto-optimal" cuts. This

sa.ction process involves a choice, made by solving a linear program,

-rm the multiple optimal solutions of another linear program.

The results given in sections 1 and 2 apply to a broader minimax set-

fng that also includes Dantzig-Wolfe decomposition for linear and non-

.near programs, and related "cutting plane" type algorithms that arise in

-source directive and price directive decomposition. For the sake of

;enerality, we cast our development of these two sections in the broader

setting of cutting plane, or relaxation, algorithms which includes, as a

special case, Benders procedure applied to general mixed integer programs.

In section 3, we show how this general methodology might be spe-

icialized for particular classes of applications. We consider Benders

decomposition applied to facility location problems on networks, deve-

loping a very efficient algorithm for generating pareto-optimal cuts

that exploits the underlying structure of these models. Since the linear

program for generating cuts in these applications is the dual of a network

optimization problem, multiple optimal solutions and, hence multiple

choices for Benders cuts, will be commonplace, thus providing an excellent

opportunity for applying our proposed methodology for selecting "good"

cuts.

Section 4 describes a related approach for accelerating Benders de-

composition via the "proper" formulation of mixed integer programs. Im-

proved formulations can also allow the generation of stronger cuts for

Benders decomposition. We discuss criteria for comparing various mixed

 -3-

integer formulations of a problem.

The final section summarizes our results and suggests possibilities

for further investigation.

1. BENDERS DECOMPOSITION AND MINIMAX OPTIMIZATION

Minimax Problems

Two of the most widely-used strategies for solving large scale opti-

mization problems are resource directive decomposition and Lagrangian

relaxation. Several papers in the mathematical programming literature

(see, for example, Geoffrion [13] and [14], and Magnanti [22]) point out

the central importance and unifying nature of these solution techniques.

The techniques are not only applied directly; their use is, at times, com-

bined with other approaches as when Lagrangian relaxation is embedded

within the framework of branch and bound for solving integer programming

problems (Fisher and Shapiro [10], Geoffrion [15]).

Since Benders algorithm, the focus of our analysis, is but one mani-

festation of resource directive decomposition, we shall consider a broader,

but somewhat more abstract, minimax setting that captures the essence of

both the resource directive and Lagrangian relaxation approaches. We

study the optimization problem

v = Min Max{f(u) + yg(u)} (1)

ycY UU

where Y and U are given subsets of R and R , f is a real valued function

defined on U and g(u) is an m-dimensional vector for any u U. Note that

we are restricting the objective function f(u) + yg(u) to be linear-affine

in the outer minimizing variable y for each choice of the inner maximizing
 -4-

variable u.

Benders decomposition leads to this minimax setting by considering

the following mixed integer program:

Minimize cx + dy

subject to: Ax + Dy = b (2)

x > 0, y Y .

In this formulation, x is an n-vector of continuous variables, y is a

k-vector of discrete variables, and Y is a subset of the integer points in

k-dimensions. The matrices A and D and vectors c, d, and b have dimensions

compatible with those of x and y.

We can formulate the mixed integer program in the equivalent form:

Minimize Minimize {cx + dy}

y Y x > 0

Ax = b - Dy

For any fixed value of y, the inner minimization is a linear program. If

it is feasible and has an optimal solution for all y Y , then dualizing

gives the equivalent formulation

Minimize Maximize {ub - uDy + dy}

y Y u CU

which is a special case of (1) in which U = {u Rm : uA < c}, f(u) = ub,

1
These assumptions can be relaxed quite easily, but with added complica-

tions that cloud our main development. See Garfinkel and Nemhauser [12]

or Lasdon 20] for a review of the algorithm in full generality.

 -5-

and g(u) = d - uD. This reformulation is typical of the resource direc-

tive philosophy of solving parametrically in terms of complicating vari-

ables, like the integer variables y of a mixed integer program.

The minimax problem (1) also arises when dualizing the constraints

g(u) > 0 of the optimization problem

Maximize f(u)

subject to: g(u) > 0 (4)

u U .

The resulting optimization problem is the Lagrangian dual, a form of the

minimax problem in which Y is the nonnegative orthant, or, more generally,

the convex subset of the nonnegative orthant for which the maximization

problem over U is finite valued.

Solving minimax problems by relaxation

For any given y Y, let v(y) denote the value of the maximization

problem in (1); that is,

v = Min v(y)
yEY

where

v(y) = Max {f(u) + yg(u)} . (5)

uEU

Since v(y) is defined as the pointwise maximum of linear-affine functions,

it is convex, though generally nondifferentiable. Consequently, whenever

the set Y is convex, the minimax problem can be viewed as a convex program.

There has been a great flourish of activity recently in modifying and ex-

tending algorithms of differentiable optimization to solve this class of

 -6-

problems (see Dem'yanov and Malazemov 7 ], Lemarechal [21], Mifflin [27],

Wolfe 37], and the references that they cite). An alternative solution

strategy that applies even when Y is not convex is a relaxation approach.

Rewrite (4) as

Minimize z

subject to: z > f(u) + yg(u) for all u U (6)

y Y, z R

and form a relaxation

Minimize z

subject to: z > f(uj ) + yg(uj )

(j = 1,2,. . .,K) (7)

y Y, z R

where each u
j is an element of U. The solution y ,
K zK of this "master

problem" (7) is optimal in (6) if it satisfies all of the constraints of

that problem; that is, if v(y ) < z . If, on the other hand, v(y ) > z ,

and u solves the "subproblem" (5) when y y , then we add

K +
(uK+ =
z > f(u ) + yg(uK+)

as a new constraint, or cut as it is usually called, to the master problem

(7). The algorithm continues in this way, alternately solving the master

problem and subproblem.

When applied to problems (3) and (4), this algorithm is known, res-

pectively, as Benders Decomposition and Dantzig-Wolfe Decomposition or

2
As before, to simplify our discussion we assume that this problem always

has at least one optimal solution.

 -7-

generalized programming. The master problem is an integer program with

one continuous variable when Benders algorithm is applied to mixed integer

programs; it is a linear program when Dantzig-Wolfe decomposition is ap-

plied to nonlinear programs. The convergence properties of the relaxation

algorithm are well-known, although usually stated in the context of par-

ticular instances of the algorithm, (see, for example, Benders [3 ],

Dantzig [5, CH.24 ], and Magnanti et al. [23]). If the subproblem is a

linear program, then the point u j in (7) can be chosen as extreme points

of U and the algorithm terminates after a finite number of iterations. If

the set U is compact and the functions f and g are continuous, then any

limit point y E Y, if one exists, to the sequence {y }K> is optimal in

K>l
(1). -Neither of these convergence properties depends upon structural pro-

perties of Y. Nevertheless, the structure of Y does determine whether or

not the master problem (7) can be solved efficiently.

2. ACCELERATING THE RELAXATION ALGORITHM

A major computational bottleneck in applying Benders Decomposition is

that the master problem, which must be solved repeatedly, is an integer

program. Even when the master problem is a linear program as in the ap-

plication of Dantzig-Wolfe Decomposition, the relaxation algorithm has not

generally performed well due to its poor convergence properties (Orchard-

Hays 31], Wolfe [36]). There are several possibilities for improvement:

(i) making a good selection of initial cuts,

i.e., values of the u j , for the master problem;

(ii) modifying the master problem to alter the choice

of yK at each step;

(iii) formulating the problem "properly"; and

 -8-

(iv) if there are choices, selecting good cuts to

add to the master problem at each step.

In a number of studies of mixed integer programs, Mevert [26] found

that the initial selection of cuts can have a profound effect upon the

performance of Benders algorithm. Geoffrion and Graves 16] have reported

similar experience with facility location problems.

There have been several proposals to alter the master problem for

Dantzig-Wolfe Decomposition. Nemhauser and Widhelm [28] (see also O'Neill

and Widhelm [30]) show that scaling the constraints of the master problem

to find the "geometrically centered" value of y at each step, can be bene-

ficial. Marsten, Hogan, and Blankenship [25], (see also Marsten [24]),

have had success in restricting the solution to the master problem at each

step to lie within a box centered about the previous solution. Hollaway

19] shows how to select among multiple optima of the master problem to

obtain better convergence.

Several researchers have illustrated the importance of problem for-

mulation. Two different formulations of the same problem might be iden-

tical in terms of feasible solutions, but might be distinguishable in

other ways. For example, they might have different linear programming or

Lagrangian relaxations, one being preferred to the other when used in con-

junction with algorithms like branch and bound or Benders decomposition.

Recent studies by Cornuejols, Fisher, and Nemhauser [4 and Geoffrion and

McBride [17] provide theoretical insight and computational experience con-

cerning the role of model formulation in Lagrangian relaxation. Davis and

Ray [6], Beale and Tomlin [2], and Williams [35], in the context of

linear programming relaxation for branch and bound, and Geoffrion and

Graves [16], in the context of Benders decomposition applied to facility

 -9-

location models, show that proper model formulation can generally improve

the computational efficiency of these procedures. Section 4 explores fur-

ther the question of proper model formulation for mixed integer programs

in the context of Benders decomposition.

In many instances, as when Benders decomposition is applied to net-

work optimization problems, the selection of good cuts at each iteration

becomes an issue. Recall from section 1 that any solution to the sub-

problem (5) at each iteration of a relaxation algorithm such as Benders

decomposition defines a cut. In network applications where the matrix

A of the mixed integer program (2) models network flow structure, multiple

optimal solutions to the subproblem are the norm; equivalently, degenerate

solutions to the inner maximization problem in (3)

Minimize {dy + cx : Ax = b - Dy, x > 0}

are to be expected because the shortest route, transshipment, and other

network optimization problems are reknowned for their degeneracy. In this

section, and the following one, we introduce methods and algorithms for

choosing from the alternative optima to (5) at each iteration, a solution

that generates a cut that is in some sense "best".

First, we must formalize some definitions. We say that the cut (or

constraint), z > f(u ) + yg(u ) in the minimax problem (1) dominates or is

stronger than the cut, z > f(u) + yg(u), if f(u1 ) + yg(u ) > f(u) + yg(u)

for all y E Y with a strict inequality for at least one point y Y. We

call a cut pareto optimal if no cut dominates it. Since a cut is deter-

mined by the vector u U, we shall also say that u dominates (is

stronger) than u if the associated cut is stronger, and we say that u is

pareto optimal if the corresponding cut is pareto optimal. Let us call

 -10-

any point y contained in the relative interior of Y , a core

oint of Y.

The following theorem provides a method for choosing from among the

alternate optimal solutions to the subproblem (5) to generate pareto-

optimal cuts.

Theorem 1: Let y be a core point of Y, i.e., y ri(Yc), let u() de-

note the set of optimal solutions to the optimization problem

Max {f(u) + g(u)} (8)

uEU

and et u solve the problem:

Max {f(u) + y g(u)}

uU(y) (9)

Then u is pareto optimal.

Proof: Suppose to the contrary that u is not pareto optimal; that is,

there is a u U that dominates u . We first note that since

©
f(u) + yg(u) > f(u© ) + yg(u ) for all y Y, (10)

it is true that

f(u) + wg(u) > f(u0 ) + wg(u) for all w E Y (11)

To establish the last inequality, recall that any point w Y can be ex-

pressed as a convex combination of a finite number of points in Y, i.e.,

w = {h y y £ Y}, where X > 0 for all y Y, at most a finite number of

the X are positive, and {Xy : y Y} = 1.

Also, note from the inequality (10) with y = Y, that u must be an

optimal solution to the optimization problem (8), that is, u U(y). But
 -11-

then (10) and (9) imply that

f(u) + y g(u) = f(u0) + y g(u )

. (12)

Since u dominates u ,

f(u© ) + yg(u ) < f(u) + yg(u) (13)

0 C
c)
for at least one point y £ Y. Also, since y ri(Y , there exists (see

(34, Theorem 6.4]) a scalar e > 1 such that w - Oy + (1 - 8)y belongs to

YC. Multiplying equation (12) by 8 and multiplying inequality (13) by

(1 - e), which is negative and reverses the inequality, and adding gives:

f(u0) + wg(u ) > f(u) + wg(u)

But this inequality contradicts (11), showing that our supposition that u

is not pareto optimal is untenable. This completes the proof.

= k
When f(u) ub, g(u) = (d - uD), and U = {u £ R : uA < c}, as in

Benders Decomposition for mixed integer programs, problem (8) is a linear

program. In this case, U(y) is the set of points in U satisfying the li-

near equation, u(b - Dy) = v(y) - dy, where v(y) is the optimal value of

the subproblem (5). Therefore, to find a pareto optimal point among all

the alternate optimal solutions to problem (8), we solve problem (9), which

is the linear program:

Maximize {dyo + u(b - Dyo)}

subject to: u(b - Dy) = v(y) - dy (14)

and uA < c .

We should note that varying the core point y might conceivably gene-

rate different pareto optimal cuts. Also, any implementation of a strong

 -12-

cut version of Benders algorithm has the option of generating pareto op-

timal cuts at every iteration, or possibly, of generating these cuts only

periodically. The tradeoff will depend upon the computational burden of

solving problem (9) as compared to the number of iterations that it saves.

In many instances, it is easy to specify a core point y for imple-

menting the pareto optimal cut algorithm. If, for example,

Y = {y R : y > 0 and integer}, then any point y > 0 will suffice; if

k h a vc o 0)
Y = {y R : yj = 0 or 1 for j = 1,2,. . .,k}, then any vector y with

0 < y. < 1 for j = 1,2,. . .,k suffices; and if

k k
Y = {y R : y < p, y > 0 and integer},
j=l

as in the inequality version of the p-median problem, then any point y

k
with y > 0 and y < p suffices. In particular, if p > k/2, then
j=l
y = (1/2,1/2,. . .1/2) is a core point.

One particular version of the preceding theorem merits special mention.

Suppose that U is a product of sets U = U U x . .. x U and that f

and g are additively separable over the sets UJ; that is,

f(u) = E f(u
j=l

and

g(u) = gj(u(),
j=l

where u = (u(1),u(2)'. . .,u(J)) is a partition of u with u(j) U j. The

notation u(j) distinguishes this vector from the component u. of u U.

Then, for any y Y, the subproblem (5) separates as:

 -13-

J
v(y) = v (y)
j=l j

where, for each j,

vj(y) = Max f (u ) + yg(u.) . (15)

j
J ~ U i (j) ()
(j)

Since, for any u(j) U,

f(u)) + ygj(u(j) < v ()

j (J)) j i) - j

the vector u belongs to U(i), meaning that the sum over j of the lefthand

sides of these expressions equals the sum of the righthand sides if, and

only if,

f (u() + gj(u ) = vj(9)

] (j) i(j) j

for all j. That is, choosing u to be one of the alternate optimal solu-

tions to (8) is equivalent to u(j) being an alternate optimal solution to

(15) when y = . Consequently, finding a pareto optimal cut decomposes

into independent subproblems, as recorded formally in the following corol-

lary stated in terms of the notation just introduced.

Corollary 1: Let y be a core point of Y, and, for each i = 1,2,. . .,J,

let u (Y) denote the set of optimal solutions to the optimization problem

Max .{f (u ) + gj(u

aU t s
(j)

and et u () solve the problem:

(i)
 -14-

Max {f(u ) + ygj(u)}

~U~(j)U
()j)

hen =(1) u ( 2)' u ()) is pareto optimal for (1).

The separability of f and g in this discussion has historically been a

major motivation for considering resource directive decomposition and

Lagrangian relaxation. Whenever the constraint matrix A of the variables

x in problem (2) is block diagonal, this separability property applies.

In this case, problem (14) decomposes into several linear programs, one for

each subvector u(j) of u. In Lagrangian relaxation, dualizing induces

separability in the subproblem, from the "complicating constraints"

J
g(x)-= gj(uj)) >
j=l (J) -

of the original problem formulation (4) if U=U 1 xU 2 x...xUJ is separable.

3. ACCELERATING BENDERS METHOD FOR NETWORK OPTIMIZATION

Although solving the linear program (14) always generates pareto

optimal cuts whenever Benders method is applied to mixed integer programs,

it might be possible to generate strong cuts more efficiently in certain

situations. In particular, when the Benders subproblem involves network

optimization, special purpose network algorithms might be preferred to

the general purpose methodology.

In this section, we describe special network algorithms for gene-

rating strong cuts for facility location problems. We begin by considering

a facility location problem formulated as the following mixed integer

program:
 -15-

n m m
v = Min E E c..x.. + E d.y.
i=l j=l z] i. j=l

m
subject to: x.. > 1 (i = 1,2,. . .,n) (16)
j=l

x.. < y (1 < i < n)

13 J -

x.. > 0 (1 < j < m)

yj = 0 or 1

yEY

where m = number of potential facilities

n = number of customers

and Y = set of feasible values for y c (0,1)

If yj = 1, we construct facility j and incur a fixed cost of d.. If

x.. = 1, customer i receives service at facility j. The first constraint

requires that each customer be serviced by some facility. The second con-

straint states that no customer can be serviced at a facility unless that

facility is constructed. The references cited at the outset of section 2

and our discussion in section 4 suggest reasons for choosing this parti-

cular form of the problem formulation instead of an equivalent formulation

with constraints x.ij < nyj for all j in place of the constraints xij < yj

for all i and j.

m
If Y = {Yl y. = p}, n = m, and c.. = 0 for all j, then (16) becomes
j=l 33
the well-known uncapacitated plant location problem. The set Y might in-

corporate a number of additional conditions imposed upon the configuration

of open (i.e., yj = 1) facilities. Among these might be contingency

 -16-

constraints such as "location i is opened only if location j is opened,"

multiple choice constraints such as "open at most two of the locations i,

j, and k," and other conditions of this nature.

Suppose we fix y = y Y; then (16) reduces to the following pure

linear programming subproblem:

n m
v(y) = Min Z Z C ..x..
i=l j=l

m
subject to: O x.. >1 (i = 12,. . .,n) (17)
j=l 1]-

0 < x.. < 1 j 0

- 1] -

O < x.. < O j C

- 1 -

(1 < i < n)

where

= {jlyj = 1}, the set of open facilities, and

C = {jj = 0}, the set of cosed facilities.

3

The linear program dual of this problem is:

n m
v (y) = Max £ [X. - z y.r. 4
i= 1 j=l 1
i=l

subject to: X - . . < . . (1 < i < n) (18)

i 1] - 13

X. > (1 < j < m)

1 -

Tr . > .
13 -

Any solution to this problem determines a cut of the form:

 -17-

n m m
v > (X. - Z r. .y.) + Z d.y. (19)
i=l I j=l 3] 3 j=l ]

(Note that we have appended the term Zdjyj to the right hand side of the

cut. This term was omitted from the objective function of the subproblem (17)

because it is a constant for any given choice of the configuration vari-

ables yj.)

Careful inspection of the linear program (17) reveals that, for most

problems, it will have a degenerate optimal basis. This implies that it

usually will be possible to derive more than one Benders cut. We next des-

cribe a procedure for generating pareto-optimal cuts.

Note that for any choice of y Y, the linear programs (17) and (19)

decompose into separate subproblems, one for each index i = 1,2,. . . ,n.

Also, the "natural solution" (see Balinski [ 1)

= c min {c..; j O}
i ij(i) i 3

7ij= 0 if j 0 (20)

and 1.. = max (O,X. - c..) if j C ,

to the linear programming dual problem (18) has the property that the op-

timal value of the ith subproblem is v. (y) = .. Consequently, corollary 1

1 I

with u (i) = (Xi, ilT i2- . . in) implies that solving for each i the

3
The optimal dual variables have a convenient interpretation in terms of

the facility location problem. X. is the cost of servicing customer i when

1

= y. 7..is the reduction in the cost of servicing customer i when fa-

cility j is opened and Yi = Yi for all i j.

_ Ysl·-ll·llll---------YI·IIC IXI-^-I-II·LIIIIIIIY--UL (_I_

 -18-

the subproblem
m
Max (X. - y.T. .]
j=l

m
subject to: X.1 - z yr.. = X. (21)
J13 1

X. -. < c.. (j = 1,2,...,m)

1 13 - 13

T.. >O
13 -
X.1 ->0

provides a pareto-optimal vector with components \i and 7.ij for i = 1,2,.

.,n and j = 1,2,. . .,m. Here, as before, y denotes the current value of

the integer variables and y belongs to the core of y, i.e., y £ ri(Y)

Our first objective is to show that, for each i, the subproblem (21)

is piecewise linear as a function of X.. Note that, since the equality

constraint of this problem reads,

X. - 'g T..
I j sO ] 1 ij(i)

and since
i ij(i) - ij(i)

and r.. > 0 for all j,

13 -

it must be true that 7r.. = 0 for all j j(i), j £ 0 and

13

I _ I _1_1__
 -19-

ij(i) = - ci = . - X

Also, if we substitute for .i in the objective function of (21) from the

equality constraint, the objective becomes:

m
Max X + Z (j y.)rij
y-

Consider any index j C. Since yj = 0, the coefficient

_ 0
j =yj - Yj

of 7.. is nonpositive. Thus, an optimal choice of i.. satisfying the

13 13

two constraints

. -r.. < c.. and Tr.. > 0

1 13-13 13 -

is T.. = max {0,X - c.}

.
.z ] 1 13

Collecting these results, we see that the optimal value of problem

(21) as a function of the variable . is:

1

A+c (A. - .) + Z . max {O,X. - c.

As an aid to optimizing (21), we note the following upper and lower

bounds on i.:

. < . < L.

where, by definition, L. = min {c.. : j s 0 and j j(i)}. The lower

1 13

bound is simply a consequence of the equality constraint of problem (21),

because each y > 0 and each .ij > 0. The upper bound is a consequence of
3- 1 3-
our previous observation that, for all j j(i) and j E 0, V.. = 0 and,
13
 -20-

therefore, the constraint. - .. < c.. becomes . < c...

]1]- 1 1-- 1]
Now, since the function (27) is piecewise linear and concave in X.,

we can minimize it by considering the linear segments of the curve in the

interval X. < X < L. in order from left to right until the slope of any
1 - 1 - 1

segment becomes nonpositive. Formally,

(1) Start with X. = X.

1 1

(2) Let T = {j C : cij < iX.and let s = Eij(i) + {. j T}

1- 1 i(i) j

s is the slope of the function (25) to the right of f..

(3) If s < 0, then stop; X :is optimal. If s > 0 and T = C, then

stop, X. = L. is optimal.

(4) Let ck = min {c.i : j C and j T. If Li < cik, set i = L.

and stop. Otherwise, increase Xi to cik. Repeat steps (2)-(4).

Once the optimal value of Xi is found using this algorithm for each-i,

the remaining variables .ij can be set using the rules given above. Then,

by virtue of corollary 1, the cut obtained by substituting these values in

(19) is pareto optimal.

The above algorithm should be very efficient. For each customer i,

at most m (m = # of possible facilities) steps must be executed. So in the

worst case, the number of steps required by this procedure is bounded by

(# of customers) (# of possible facilities).

We might emphasize that this algorithm determines a pareto optimal cut

for any given point y in the core of Y. Also, the algorithm applies to

any of the possible modeling variations that we might capture in Y, such as

the contingency and configuration constraints mentioned at the beginning of

this section.
 -21-

This algorithm can also be extended to more complex models like the

capacitated facility location problem 17,18]. The basic concepts of such

a procedure are similar to the algorithm that we have just described and

will not be given here.

It is interesting to note that the pareto-optimal algorithm specified

above is similar to a dual ascent procedure proposed by Erlenkotter [ 8,

pp. 997-998]. Both procedures essentially give a set of rules for in-

creasing the A.. One distinguishing feature of our algorithm is that it

solves problem (21) exactly, whereas Erlenkotter's procedure gives an ap-

proximate solution to the dual of the linear programming relaxation of

problem (16).

To conclude this section, we note that it is possible to generate cuts

stronger than the natural cuts defined by setting Ai = .i and i..= r.. in

(19), but without assurance of pareto-optimality. For details, see

Balinski [1 ], Nemhauser and Wolsey 29], and Wong [38].

4. A MODEL SELECTION CRITERION FOR BENDERS DECOMPOSITION

Selecting the "proper" model formulation is another important factor

that effects the computational performance of Benders decomposition applied

to network design and other mixed integer programming models. This section

discusses a criterion for distinguishing between different but "equivalent"

formulations of the same mixed integer programming problem to identify

which formulation is preferred in the context of Benders decomposition.

Many network optimization problems have several "natural" mixed in-

teger formulations. For example, as we noted in section 3, various varia-

tions of the facility location problem can be stated in several possible

ways as mixed integer programs. We demonstrate in this section why some

formulations lead to such pronounced improvements over others in the

 -22-

performance of Benders decomposition (see also references [4 1, [6 ], and

[15], cited earlier).

To illustrate the role of model selection, we consider an example of

Benders decomposition applied to the p-median facility location problem

[4 i. The p-median problem can be formulated as:

N N
(P) Minimize Z Z d..x..
i=l j=l 1 1]

N
subject to: x.. = 1 V. (23)

x.. < Yi V(i,j) (24)

N
Z y. = p (25)
i=l 1

x. . > 0 and Yi integer V(i,j) . (26)

N is the number of nodes in the problem and p is the number of facilities

to be located. Yi indicates whether a facility is located at node i and

x.. indicates whether customer j is serviced at node i.

1]
As we noted in section 3, an equivalent formulation is:

N N
(Q) Minimize Z Z d..x..
i=l j=1 13

subject to: (23), (25), (26), and

N
x.i < Ny (27)
j=l

Note that (27) represents an aggregation of the constraints in (24). Con-

sequently, although P and Q are equivalent mathematical descriptions,

if we relax the integrality constraint on the yi, the feasible region for
 -23-

Q is a proper subset of the feasible region for P .

Let us examine the following p-median problem represented in figure

5.1:

Figure 5.1 p-median Example

N is equal to 4, p is equal to 2, -and all d.. are 100.

13
The application of Benders decomposition to this example formulated

as Q yields the following set of Benders cuts:

400
z > 200 - y1 - 400y 2 + 0y 3 + 0y4

z > 200 - 400y1 + Oy 2 - 400y 3 + Oy4

z > 200 - 400y l + OY 2 + Oy3 - 400y 4

z > 200 + Oy 1 - 400y 2 - 400y 3 + Oy4

400
z > 200 + Oy 1 - 400y 2 + Oy3 - y4

z > 200 + Oy 1 + 0y 2 - 400y 3 - 400y 4

This set of cuts has the property that every single one must be generated

in order for Benders algorithm to converge.

Applying Benders decomposition to our example formulated as P

requires the single cut:

z > 400 - 100y1 - 100y 2 - 100y 3 - 100y 4

We can generalize this example in the following way: let p = N/2

and let d.. = 100 for all i j and d.. = 0 for all i = j. With this
1] 1]
 -24-

class of examples, we have problems where the Q formulation requires

N/2 cuts, an exponential number of cuts with respect to N, for Benders

algorithm to converge. For these same problems, the P formulation, in

every case, requires only one Benders cut for convergence! This example

dramatically illustrates the importance of intelligent model formulation

for Benders decomposition.

Now we present a formal framework for comparing model formulations for

Benders decomposition. This framework is then utilized to prove our main

results.

Suppose we have two mixed integer programs P and Q that are

represented as:

(P) Minimize [vP(y)] where [vP(y)l = Minimum cx + dy

yY
subject to: Ax + By= b

x > 0

and

(Q) Minimize [vQ(y)] where [vQ(y)] = Minimum hw + dy

y Y
subject to: Dw + Gy = g

w > 0

x,w, and y are column vectors of problem variables; b and g are column

vectors; c, d, and h are row vectors; A, B, D, and G are appropriately

dimensioned matrices. The set Y is a set of integer valued vectors that

captures the integer constraints of the problem. We assume that the set

Y is finite.

We will say that P and Q are equivalent mixed integer programming

representations of the same problem if vP(y) = vQ(y) for all y E Y .

---
 -25-

That is, the two models have the same integer variables and may have dif-

ferent continuous variables and constraints, but always give the same ob-

jective function value for any feasible assignment of the integer variables.

We will say that the two formulations are identical if v (y) = vQ(y) for

all y belonging to the convex hull of Y.

In the context of Benders decomposition, another possible interpre-

tation of equivalence is that vP (y) and vQ(y) represent the linear pro-

gramming subproblems when Benders decomposition is applied to P and Q.

So the two models are equivalent if their respective Benders subproblems

always have the same optimal value.

We evaluate these two models by comparing the cuts generated from

the application of Benders decomposition to these models. Following the

derivation of Benders decomposition given in section 1, we can rewrite P

and Q as

Minimize z

subject to: z > (b - By) + dy 7 £

y C Y

where is the set of points in the polyhedron A < c; and

Minimize z

subject to: z > y(g - Gy) + dy y r

Y Y£ Y
ys

As in earlier sections, we assume that the linear programming subproblems

v (y) and v(y) are feasible and have optimal solutions for all y Y.

These constraints can be relaxed, but with added complications that do not

enrich the development in an essential way.

 -26-

where r is the set of points of the polyhedron yD < h.

The inequalities z > (b - By) + dy and z > y(g - Ey) + dy will be

referred to as the Benders cuts for P and Q, respectively. We remark that

our definition of Benders cuts, in which a cut can be generated from any point

in the subproblem dual feasible region, produces a larger set of possible cuts

than the usual definition which restricts the cuts to those corresponding to

the extreme points of the subproblem dual feasible region. With this limited

definition of Benders cuts, the results of this section need not always be valid.

To compare equivalent model formulations, we adapt the concept of a pareto

optimal cut, introduced in section 2, by saying that a Benders cut (or

constraint)

z > (b - By) + dy

for P dominates a Benders cut

z > y(g - Gy) + dy

for Q if

7(b - By) + dy > y(g - Gy) + dy

for all y Y with a strict inequality for at least one point y Y.

A cut z > y(g - Gy) + dy for Q will be called unmatched with respect to the

formulation P if there is no cut for P that is equal to it (in the sense that

two cuts are equal if their right-hand sides are equal for all y Y) or

dominates it.

A formulation Q is superior to an equivalent formulation P if Q has at

least one Benders cut that is unmatched with respect to P, but P does not have

any cuts that are unmatched with respect to Q.

In a very loose sense, Q is superior to P if they are equivalent formu-

lations and the set of Benders cuts for P is a proper subset of the Benders

cuts for Q.
 -27-

With these definitions we can now prove several properties concerning

model formulation and the strength of Benders cuts.

Lemma 1 Let P and Q be equivalent formulations of a mixed integer

progrcning problem. Q has a Benders cut that is unmatched with respect

to P if, and only if, there is a y belonging to the convex hullZZ yc of

Y that satisfies v Q (y ° ) > vP (y).

Proof: ( -) Let z > y (g - Gy) + dy be a Benders cut that is unmatched

with respect to P. Since we are assuming that the set Y is finite, the

definition of an unmatched cut implies:

Max [min (b - By) + dy - y (g - Gy) - dy] < O .

7TA<c yEY

Now observe that the above inequality still holds if we replace the set Y
c
by Y . Using linear programming duality theory, we can reverse the order

of the max and min operation to obtain

Min [ max IT(b - By) + dy - y (g - Gy) - dy] < 0

yEYC 7TA<c

Linear programming duality theory, when applied to the inner maxi-

mization, allows us to rewrite the above expression as

Min cx + dy- [y (g- Gy) + dy] < 0

subject to: Ax + By = b

x > 0, y Yc

This implies that there is a y E Y satisfying

MMin cx+ dyo = v P(y°) < (g - Gy) + dyo

subject to: Ax = b - By

x> .

Another application of linear programming duality theory, in this case

to Q, gives us:
 -28-

v (y) < y (g - GyO) + dyO < Min hw + dy °

subject to: Dw = g - Gy

w> 0

or

vP (yO) < y) .

(<-) The reverse implication has essentially the same proof with all

the steps reversed. Explicit details will not be given here. E

This lemma leads to the following theorem about preferred formulations:

Theorem 2: Let P and Q be equivalent formulations of a mixed integer

progranuming problem. Q is superior to P if, and onZy if, vQ(y) > vP (y)

for all y YC with a strict inequality for at least one y yC .

Proof: (-) If v(y) > vP (y) for all y E YC, Lemma 1 says that P

does not have any Benders cuts that are unmatched with respect to Q. But because

there is a y Y such that v(y) > v(yO) Lemma 1 implies that Q has a cut that

is unmatched in P. So Q satisfies the definition of being superior to P.

(-----) If Q is superior to P, then P, by definition of superior,

does not have any cuts that are unmatched with respect to Q. Lemma 1

then tells us that vQ(y) > vP (y) for all y Y. The definition of

superior also states that Q has a cut that is unmatched with respect to

P and using lemma 1 we can say that there exists a y Y such that

y( o .)
vQ(yO) > v

The implications of Theorem 2 may become more apparent when inter-

preted in another way. Let the reZlaxed primal problem for any formu-

lation of a mixed integer program be defined by replacing Y by its

convex hull Y . Theorem 2 states that for a formulation of a mixed

integer programming problem, the smallest possible feasible region (or

 -29-

the "tightest" possible constraint set) for its relaxed primal problem

is preferred for generating strong Benders cuts. For any formulation P,

a smaller feasible region for its relaxed primal problem will result in.

larger values of the function v (y) which Lemma 1 and Theorem 2 indicate

is desirable.

As a concrete example, the p-median problem discussed at the beginning

of this section has two formulations P and Q. They differ only in that

P has constraints of the form x.. < Yi for all (i,j), whereas Q has

constraints of the form x < 4y for all i.

j=l -
Since the later set of constraints is an aggregation of the former

constraints, the feasible region for the relaxed primal problem of P

is no larger than that for Q. So vP(y) > vQ(y) for all y yC. A straight-

)
forward computation shows that vP(y = 200 > vQ(y0 ) = 0 for yo = (2,2,)

So the formulation P is superior to Q for this example.

A general consequence of Theorem 2 is that for any mixed integer pro-

gramming formulation, the convex hull of its feasible region will be a

model formulation that is "optimal" in terms of generating Benders cuts

since it has a relaxed primal problem whose feasible region is the smallest.

In order to formalize this observation for any formulation P of a mixed

integer program as in (2), let C(P) denote the mixed integer program whose

feasible region is the convex hull of the feasible region for P.

Theorem 3: Given any formulation P of a mixed integer program,

vC() (y) > vQ(y) for all y yC and for all equivalent formulations Q

of this problem.
* cy *
Proof: Let y Y be arbitrary and let x be an optimal solution to
* C(P) * * *
C(P) when y = y ; that is, V (y ) = cx + dy . By definition of
 -30-

* *
convex hull (y ,x ) is a convex combination with weights X. of a finite
1

number of points (y ,x ) that are feasible in P. Linearity of the ob-

jective function cx + dy implies that cx + dy = X (cx + dy ). Since

(y ,x ) is feasible in P, vP(y ) cxi + dy. Therefore,

vC(P) C(y *>) >(

v v (yi). But since P and Q are equivalent formulations

vC() (Y*) > Z .vQ(yl) and by convexity of vQ (y), the right-hand side of

this last expression is no smaller than vQ(y ). Consequently,

v (y) > v(y ) for all y £ Y. n

Combining this theorem with Theorem 2 establishes

Corollary 2: Given any two equivalent formulations P and Q of a mixed

integer program, the convex huZZll formulation C(P) of P is either superior

or identical to Q.

Another interesting property of the convex hull formulation of a

problem is that when Benders algorithm is applied to it, only one cut is

necessary for it to converge. More formally, let us suppose that the

constraints of the following problem define the convex hull of the mixed

integer program P in (2):

C(P)
v = min cx + dy

subject to: Rx + Qy = q

x > O,y Y

Then we have

Theorem 4: For any formulation of a mixed integer program, the convex

huZZ formulation C(P) requires only one Benders cut for convergence.
C(P)
Proof: v min min cx + dy
yEY x>O

subject to: Rx + Qy = q
 -31-

Since C(P) is the convex hull formulation, we can substitute YC for Y

without affecting the optimal solution value. Then applying linear pro-
C(p)
gramming duality theory (and again assuming that v (y) is feasible for
C(p)
all y E Y) we have v = min max u(q - Qy) + dy. Another application
EYC uR<c C(P) *
of linear programming duality theory yields v = min u (q - Qy) + dy for
s* ysyc
some u- satisfying uR<c. Since the last objective function is a linear function

of y, we can substitute Y for its convex hull and write:

C(P)
v = min z

subject to: z > u (q - Qy) + dy

y Y .

*jk~~~~~~ ~C(P) C(P) S

Let y be a solution of this problem. Then v v (y ) So the

single Benders cut generated by u is sufficient to solve the convex hull

formulation C(P). O

Although we have shown that a reduced feasible region for the relaxed

primal problem of a formulation is desirable, there are other issues that

must be considered in selecting a model for use with Benders decomposition.

First, there remains the difficulty of constructing alternative models

for mixed integer programming problems. Although, in principle, the convex

hull formulation of a problem requires only a single Benders cut for con-

vergence, in general, it will be very difficult to determine this cut by

building such a model. There is no efficient procedure known for generating

the constraints representing the convex hull of a set of points. Padberg

and Hong [32] have recently had success generating such constraints itera-

tively for traveling salesman problems. Finding efficient methods for

 -32-

generating alternative models appears to be an area for future research.

The formulation of network optimization problems is potentially one

rich application area for the results of this section. Network problems

usually have several evident "natural" formulations. The facility location

problem discussed in section 3, the multicommodity distribution system

problem solved by Geoffrion and Graves [16], and the capacitated plant

location problem described by Guignard and Spielberg [18], are all net-

work examples that have several easily derived formulations. For these

problems, since the alternative formulations usually have the same problem

variables, we can compare them by inspecting the size of the feasible

region for their respective relaxed primal problems. Due to the compara-

tively simple constraint sets of network problem formulations, it may also

be possible to derive additional constraints from the current ones. In

such a situation, these new constraints could be evaluated by testing if

they reduce the size of the feasible region for the modified primal problem.

Another issue that should be considered is the difficulty of solving

the Benders (linear programming) subproblems. Adding constraints to a

formulation strengthens the Benders cuts that can be derived, but also com-

plicates the solution of the linear subproblems. So there is a trade-off

between the quality of Benders cuts available and the time needed to solve

the Benders subproblems.

Finally, a related issue is that adding constraints to a formulation

can cause the linear programming subproblems to become degenerate since we

are adding constraints to a linear program while keeping the number of

variables constant. Thus, there may be a choice as to which cut to generate

at each iteration of Benders algorithm.

 -33-

So by "tightening" the formulation of a problem we can get stronger

Benders cuts, but these stronger cuts may have to be distinguished from

other weaker cuts. The methodology described in section 2 should be

useful in such a situation.

5. CONCLUSIONS AND FUTURE WORK

Recent computational successes have underscored the promise of Benders

decomposition. However, the straightforward application of this solution

strategy frequently leads to computational excesses. This paper has des-

cribed new results and methodology for accelerating the convergence of

Benders decomposition. This approach is also applicable to a broader class

of relaxation algorithms for minimax problems such as Dantzig-Wolfe decom-

position for the Lagrangian dual of a nonlinear problem. The adaptation

of this technique in section 3 to Benders decomposition applied to facility

location models yielded an efficient special purpose algorithm. Compu-

tational work applying these techniques to facility location and network

design problems is currently in progress. We hope to report these results

in a future paper.

Section 4 discussed the relationship between the proper mathematical

formulation of mixed integer programming models such as the facility location

problem and the computational performance of Benders decomposition. We

presented a criterion for selecting among alternate model formulations for

use with Benders decomposition. Suggestions were also made for modifying

model formulations in order to improve the performance of Benders procedure.

A potentially fruitful avenue for future research would be to construct new

formulations for mixed integer programs based upon our results and to per-

form computational tests evaluating our criteria for selecting among alter-

native model formulations.

(_I _____
 -34-

REFERENCES

1. Balinski, M. L., "Integer Programming: Methods, Uses, Computation,"

Man. Sci., Vol. 12, pp. 253-313, 1965.

2. Beale, E. M. L. and Tomlin, J. A., "An Integer Programming Approach

to a Class of Combinatorial Problems," Math. Prog., Vol. 3,
pp. 339-344, 1972.

3. Benders, J. F., "Partitioning Procedures for Solving Mixed Variables

Programming Problems," Twhm. Math., Vol. 4, pp. 238-252, 1962.

4. Cornuejols, G., Fisher, M. L., and Nemhauser, G. L., "Location of

Bank Accounts to Optimize Float: An Analytic Study of Exact
and Approximate Algorithms," Man. Sci., Vol. 23, pp. 789-810, 1977.

5. Dantzig, G. B., Linear Programming and Extensions, Princeton University

Press, Princeton, NJ, 1973.

6. Davis, P. S. and Ray, T. L., "A Branch-Bound Algorithm for Capacitated

Facilities Location Problems," Nav. Res. Log. Q., Vol. 16, pp. 331-
344, 1969.

7. Dem'yanov, V. F. and Malozemov, V. N., Introduction to Minimax,

(Translated from Russian by D. Louvish), John Wiley & Sons,
New York, 1974.

8. Erlenkotter, D., "A Dual Based Procedure for Uncapacitated Facility

Location," Opns. Res., Vol. 26, pp. 992-1009, 1978.

9. Fisher, M. L. and Jaikumar, R., "A Decomposition Algorithm for Large

Scale Vehicle Routing," Report No. 78-11-05, Department of
Decision Sciences, University of Pennsylvania, 1978.

10. Fisher, M. L. and Shapiro, J. F., "Constructive Duality in Integer

Programming," SIAM JournaZ on Applied Math, Vol. 27, pp. 31-52,
1974.

11. Florian, M. G., Guerin, G., and Bushel, G., "The Engine Scheduling
Problem in a Railway Network," INFOR JournaZ, Vol. 14, pp. 121-
138, 1976.

12. Garfinkel, R. and Nemhauser, G., Integer Programming, Wiley, New York,
1972.

13. Geoffrion, A. M., "Elements of Large Scale Mathematical Programming,

Parts I and II," Man. Sci., Vol. 16, pp. 652-691, 1970.

14. Geoffrion, A. M., "Primal Resource-Directive Approaches for Optimizing

Nonlinear Decomposable Systems," Opns. Res. Vol. 18, pp. 375-
403, 1970.
 -35-

15. Geoffrion, A. M., "Lagrangian Relaxation and Its Uses in Integer

Programming," Math. Prog. Study 2, pp. 82-114, 1974.

16. Geoffrion, A. M. and Graves, G., "Multicommodity Distribution

System Design by Benders Decomposition," Man. Sci., Vol. 5,
pp. 822-844, 1974.

17. Geoffrion, A. M. and McBride, R., "Lagrangian Relaxation Applied

to Facility Location Problems," AIE Transactions, Vol. 10,
pp. 40-47, 1979.

18. Guignard, M. and Spielberg, K., "Search Techniques With Adaptive

Features for Certain Mixed Integer Programming Problems,"
Proceedings IFIPS Congress, Edinburgh, 1968.

19. Holloway, C., "A Generalized Approach to Dantzig-Wolfe Decomposition

for Concave Programs," Opns. Res., Vol. 21, pp. 210-220, 1973.

20. Lasdon, L. S., Optimization Theory for Large Systems, The Macmillan
Company, New York, 1970.

21. Lemarechal, C., "An Extension of Davidon's Methods to Non-Differentiable

Problems," Math. Prog. Study 3, pp. 95-109, 1975.

22. Magnanti, T. L., "Optimization for Sparse Systems," in Sparse Matrix

Computations, (J. R. Bunch and D. J. Rose, eds.), pp. 147-179,
Academic Press, New York, 1976.

23. Magnanti, T. L., Shapiro, J. F., and Wagner, M. W., "Generalized

Linear Programming Solves the Dual," Man. Sci., Vol. 22, pp. 1194-1203,
1976.

24. Marsten, R. E., "The Use of the BOXSTEP Method in Discrete Optimization,"
iMath. Prog. Study 3, pp. 127-144, 1975.

25. Marsten, R. E., Hogan, W. W., and Blankenship, J. W., "The BOXSTEP
Method for Large-Scale Optimization," Opns. Res., Vol. 23,
pp. 389-405, 1975.

26. Mevert, P., "Fixed Charge Network Flow Problems: Applications and
Methods of Solution." Presented at Large Scale and Hierarchical
Systems Workshop, Brussels, May, 1977.

27. Mifflin, R., "An Algorithm for Constrained Optimization with Semi-
smooth Functions," International Inst. for Appl. Sys. Analysis,
Laxenberg, Austria, 1977.

28. Nemhauser, G. L. and Widhelm, W. B., "A Modified Linear Program

for Columnar Methods in Mathematical Programming," Opns. Res.,
Vol. 19, pp. 1051-1060, 1971.
 -36-

29. Nemhauser, G. L. and Wolsey, L. A., "Maximizing Submodular Set

Functions: Formulations and Analysis of Algorithms," Technical
Report No. 398, School of Operations Research and Industrial
Engineering, Cornell University, 1978.

30. O'Neill, R. P., and Widhelm, W. B., "Computational Experience with

Normed and Nonnormed Column-Generation Procedures in Nonlinear
Programming," Opns. Res., Vol. 23, pp. 372-382, 1972.

31. Orchard-Hays, W., Advanced linear Programming Computing Techniques,

McGraw-Hill, New York, 1968.

32. Padberg, M. and Hong, S., On the Symmetric TraveZing Salesman ProbZlem:
A ComputationaZ Study, Report #77-89, New York University, 1977.

33. Richardson, R., "An Optimization Approach to Routing Aircraft,"

Trans. Sci., Vol. 10, pp. 52-71, 1976.

34. Rockafellar, R. T., Convex AnaZysis, Princeton University Press,

Princeton, NJ, 1970.

35. Williams, H. P., "Experiments in the Formulation of Integer Programming

Problems," Math. Prog. Study 2, pp. 180-197, 1974.

36. Wolfe, P., "Convergence Theory in Nonlinear Programming," in

(J. Abadie, ed.), Integer and NonZinear Programming, North Holland,
Amsterdam, 1970.

37. Wolfe, P., "A method of Conjugate Subgradients for Minimizing Non-
differentiable Functions," Math Prog. Study 3, pp. 145-173, 1975.

38. Wong, R., "Accelerating Benders Decomposition For Network Design,"

Ph.D. Thesis, Department of Elec. Eng. and Comp. Sci., Mass. Inst.
of Tech., January, 1978.

View publication stats I -