Sie sind auf Seite 1von 19

Compulers

them. Engng,
Printed in Great Britain.

Vol. I?. No. 9, pp. 909-927,


All

rights

1993

0098-1354/93

$6.00 + 0.00

Copyright
0 1993Pcrgamon
PressLtd

reserved

SYMBOLIC
INTEGRATION
OF LOGIC IN
MIXED-INTEGER
LINEAR
PROGRAMMING
TECHNIQUES
FOR PROCESS SYNTHESIS
R. RAMAN and I. E. GROSSMANN?
Departmentof ChemicalEngineering,CarnegieMellon University,Pittsburgh,PA 15213, U.S.A.
(Received

10 February 1992; final revision received 18 December


received for publicarion 7 January 1993)

1992;

Abstract-This
paper deals with the branch and bound solution of synthesis problems that are modeled
as mixed-integer
linear programming
(MILP)
problems.
Logic relations between potential units in a

superstructure
are consideredthroughsymbolicintegrationwithinthe numericalbasedbranchand bound
scheme. The objective of this integration is to reduce the number of nodes that must be enumerated by
using the logic to decide on the branching of variables, and to determine by symbolic inference whether
additional variables can be fixed at each node. Two different strategies for performing the integration are
proposed that use the disjunctive and conjuctive normal form representations of the logic, respectively.
The paper also addresses the question of how to systematically generate the logic for process flowsheet
superstructures.
Computational
results are presented to compare
the performance
of the proposed
methods and a variant that includes violated logic inequalities in the model with the cases when all logic
inequalities are included in or excluded from the model.

Discrete decisions constitute an important element in


process design and synthesis problems (e.g. deciding
what units to integrate in a flowsheet). Mathematical
programming approaches make use of 0-l binary
variables to model these decisions. Together with
continuous variables, one can model design and
synthesis problems through an objective function and
a set of constraints representing material, heat balances and specifications. When these involve nonlinearities, a mixed-integer nonlinear program (MINLP)
results. Depending on the assumptions and the superstructure representation, the synthesis problem can
often be posed as a mixed-integer linear programming (MILP) problem. The modeling step is crucial
to the success of the synthesis method and a discussion on this topic can be found in Kocis and Grossmann (1989) and Grossmann (1990).
Another major step involves the solution of the
MI(N)LP optimization model. The most common
solution technique for the MILP problem is the well
known LP-based branch and bound algorithm (see
Beale, 1977; Nemhauser and Wolsey, 1988) which is
implemented in most computer codes (e.g. LINDO,
ZOOM,
MPSX,
SCICONIC).
Recently these
methods have been improved by the use of preprocessing techniques that can reduce dimensionality,
and cutting planes that can reduce the integrality gap

tTo

whom all correspondenceshould be addressed.


909

of the MILP (e.g. see Van Roy and Wolsey, 1987). An


alternative method for MILP is the Benders decomposition method (Benders, 1962) which also requires the branch and bound method for solving the
master problem. Similarly, the Generalized Benders
Decomposition (GBD) (Geoffrion, 1972) and the
Outer Approximation (OA) (Duran and Grossmann,
1986) algorithms for MINLP problems require that
the corresponding MILP master problem be solved
by branch and bound. Finally, the solution of the
MINLP can also be performed directly with a branch
and bound method where NLP subproblems are
solved at each node of the tree (see Beale, 1977;
Gupta, 1980; Nabar and Schrage, 1990). From the
above, it is clear that the branch and bound method
plays a central role in the solution of the mixedinteger optimization problems.
Since in many synthesis problems the number of
&-I variables that is required can be rather large, the
potential size of the tree that must be examined by a
branch and bound method can become a major
bottleneck in the computations. On the other hand,
knowledge about the synthesis problem in the form
of heuristics and logic of relations among units can
potentially provide information about the design
space and make the problem easier to solve. Raman
and Grossmann (1991a) showed how both logic
relations and heuristics expressed in the form of
propositional logic can be represented in terms of
linear inequalities involving O-1 variables. Based on
this representation, Raman and Grossmann (1992)

910

R.

RAMAN

and 1. E. C~ROSMANN

address specifically synthesis problems in which the


proposed the integration of logic and heuristic knowlsuperstructures have significantly fewer design
edge using a quantitative framework. They conconfigurations than the total number of 0-l combisidered the addition of the constraints for the logic
nations. Preliminary work and results with the aprelations among units in the MI(N)LP to decrease the
preach have been presented in Raman and
relaxation gap (difference between integer solution
Grossmann (199lb). For the case of superstructures
and relaxed continuous problem). Constraints for
that do not exhibit this property, the logic will be
heuristic rules and logic relations were used to imconverted into the CNF form. The paper will also
prove the search in the master problem of MI(N)LP
address the question of how to systematically generalgorithms. The computational results showed that
ate the logic for process flowsheet superstructures.
the addition of logic constraints in MILP problems
Computational results and comparisons are also
often produces significant reductions in compupresented.
tational time. This is due to the reduction in the
number of nodes that have to be enumerated in the
THE BRANCH AND BOUND PROCEDURE
branch and bound tree despite the increased size of
the MILP. Lien and Wahl(l991) proposed the use of
Consider the MILP formulation (Pl) for a given
unit resolution to generate cutting planes from the
synthesis problem, where x and y are the vectors of
logic inequalities. The basic idea is to add few resolcontinuous and binary variables, respectively,
ution cuts to the MILP in addition to the original
logic inequalities. Although the authors did not deZ,, = min uTx + bTy,
velop a procedure for deciding which resolution cuts
s.t.
Cx+Dy<d,
to select, they obtained order of magnitude improvements in solution times using ZOOM on test probEy < e,
lems related to the MILP model for heat integrated
x 30,
y E {O,ltP.
(Pl)
distillation sequencing with discretized temperatures
by Andrecovich and Westerberg (1985).
The mixed-integer constraints represent mass and
While the above improvements in mixed-integer
energy balances as well as logical conditions; the pure
optimization are encouraging, a basic question that
integer constraints correspond to the logical specificastill remains is how to best represent and incorporate
tions.
logic and heuristics in optimization. In contrast to the
The linear programming relaxation of (Pl) is given
quantitative approaches, another possibility is to
by:
represent logic propositions in terms of Boolean
Z,, = min aTx + bTy,
variables and perform symbolic inference on them.
The two extreme ways of representing propositional
s.t.
Cx+Dy<d,
logic are the conjunctive (CNF) and the disjunctive
Ey=Ge,
(DNF) normal forms, respectively. Propositional
logic based inference techniques involve application
x20,
O<y<l.
(P2)
of Boolean algebra for the DNF case. For the case
Z, is a lower bound on Z,,, where commonly some
of CNF representations, techniques include unit
resolution (Clocksin and Mellish, 1984). Here, Horn
of the y-variables take fractional values in the solution to (P2). The branch and bound procedure is a
clause systems are particularly easy to tackle since the
inference for this case can be shown to be linear in
rigorous tree enumeration method for finding the
global optimum of the MILP. Its building blocks rely
time (Chandru and Hooker, 1989).
In this paper we propose methods for the symbolic
on a certain set of procedures or rules, some of which
integration of logic relations between potential units
are heuristic in nature. In particular, the major
in a superstructure to aid the branch and bound
components of the branch and bound procedure are
search in MILP synthesis models. The objective of
the following.
this integration is to reduce the number of nodes that
Upper bound. Any feasible solution provides an
must be enumerated, by using the logic to decide on
upper bound to the solution of a minimization probthe branching of variables and to determine by
lem. For instance, the pivot and complement method
symbolic inference whether additional variables can
of Balas and Martin (1980) is a general heuristic
be fixed at each node. A unique feature of this
procedure that can be used to generate an initial
approach is that it does not require additional conupper bound. Alternatively, heuristics or the qualitatstraints in the MILP. Two different strategies for
ive MILP model reported in Raman and Grossmann
performing the integration are proposed. In one
(1991 a) can be used to generate an initial configurstrategy the logic is converted to the DNF form to
ation in a synthesis problem.

Symbolic integrationof logic


Lower bound. This can be obtained by solving the
linear relaxation to problem (P2) at the root node,
and at successive nodes with a subset of &I variables
fixed at the discrete values. Whenever the former
problem exhibits integer solutions the search can be
terminated. Since this is seldom the case, the latter
subproblems must be considered. Whenever their
solutions is infeasible or exceeds the current upper
bound, their corresponding node in the tree fathomed.
Branching rule. At any node, the search space has
to be further divided into subspaces in order to
continue with the search. The most common rule for
this purpose is to fix to 0 or I one of the binary
variables that has taken Fractional values in the
relaxed problem. The variable upon which to perform
this branching is chosen by methods that range from
the simple one that chooses the variable that is closest
to 0.5, to the more sophisticated ones that rely on
information concerning penalties for causing any
variable to be integral (see Gartinkel and Nemhauser,
1972).
Sub-problem
selection rule. This rule selects the
subproblem to solve among all the available nodes in
the tree which are still open for examination. This
may involve just a simple depth-first or breadth-first
enumeration or a combination of both.
Logical testing. Certain logical tests can be superimposed to make decisions on the other binary
variables due to the fixing of the chosen binary
variable. This may involve the development of cutting
planes to improve the lower bounds or else may
involve fixing other binary variables simultaneously
(chaining). The typical example here are special ordered sets of type 1 (SOSl), whereby fixing one of the
variables to 1 in the constraint Xvi = 1, also fixes the
remaining variables to zero.
The efficiency of the branch and bound algorithm
can be determined from the number of nodes that
must be enumerated in the tree. This depends heavily
on the relaxation gap (the difference between the
integer solution and the lower bound at the root
node), the upper bound, the branching rule and the
subproblem selection. Since the last three rely on
heuristics that are general in nature, the branch and
bound method does not take advantage of the underlying logic structure that characterizes a specific

911

problem as will be shown with a simple example in


the next section.
MOTIVATING

To provide some insight as to why conventional


branch and bound methods may be inefficient in
optimizing
highly constrained superstructures, consider the simple example in Fig. 1. There are 3
potential processes that can be selected for manufacturing a maximum of 10 ton h- of chemical C. The
maximum capacities of these processes are 30,30 and
50 ton h-, respectively. Since there is the specification that processes 1 and 2 cannot be selected simultaneously, there are only 3 feasible configurations:
(a) select processes 1 and 3;
(b) select processes 2 and 3;
(c) do not select any of the three processes.
The number of these alternatives is clearly smaller
than the 23 = 8 combinations implied by the three O-l
variables representing each process. The MILP model
for the optima1 selection of processes that minimizes
the total cost is assumed to be given as follows:
min Z = l4y, + 12~5 f 10~5 + 3x,
+2.8x,

s.t.

x3 - 0.9x, = 0,
x5 - 0.85x, = 0,
x,-x,-xX,=0,
x, - 0.75x, = 0
x, - 3oy, c 0,
x, - 3oy2 < 0,
x, - 5oy, Q 0,
YI +_rz<
Xl.%,

. . . , x7 3 0;

Y,,Y2,Y3=@1-

1,

x, < 10;
(1)

If a standard branch and bound method is applied,


such as in an earlier version of the program LINDO,
the enumeration tree that results is shown in Fig. 2.

x6

1. Superstructure

- 9x, + 2x,,

x, - x* - x, = 0,

Fig.

EXAMPLES

for selection of processes.

R. RAMAN

912

and I. E. GIWSMANN

Fig. 2. Tree with conventional

Here the relaxed


objective

LP yields a noninteger

I3.9608

indicating

solution

potential

for

a common

heuristic).

expanded
y, = 1.
(-

LP it is the one closest

by

Since

12.1482)

selected

Two

imposing
the

former

node

integer

objective

point

Therefore,

and

an

Z = 3.629
than

[alternative

the current

solution
(a)].

upper

(3) with an

bound

(maximum

logical constraints
LP relaxation.

the

information

MILP

y3 = 1

these

arc

with

OUTLINE

greater

solutions

are

and bound

continues

by backtracking

the additional
order

Since y, is the only

it is used to create nodes

y, = 0 with Z = 12, and node 8 withy,


[alternative

search terminates

= 1. The latter

to

to confirm

the branch and


that the solution

y, = 0, y, = 0, y, = 0, Z = 0, is the global
For

this simple

method
root

had to examine

node)

feasible

example

despite

alternatives

the

the branch

optimum.
and

a total of 8 nodes
fact

that

there

in the superstructure!

bound

the branch

of

there

the

While

and

bound

is no

explicit
logic

that can effectively

limit

In the next sections

this objective.

One could

OF PROPOSED

systematic

APPROACH

the performance
the proposed

logic relations

of the branch

approach

utilizes

that define the connec-

the various units in the superstructure

provide

more

information

Instead of processing

1992),

in the model

to

the

them in the form of


(Raman

and Gross-

they will be used in symbolic

procedure

proposed.

in

search

form.

to generate these logic relations

Approaches

for

performing

the

integration

for the cases when the logic is converted

into

and

DNF

CNF

representations

next.

Both these approaches

that

potentially

earliest.

(plus the

are only

mann,
is first

Since at this point all

the open nodes have been fathomed,


bound

7 with

(b)] which is discarded

the upper bound.

problem

algorithm,

linear constraints

as it exceeds

the existence

in the next section).

is that

In order to improve

procedure.

yields Z = 0.7059

one could also add extra

relate

we will provide the basis of a logical inference scheme

next with an objective


variable,

30, 30, 50) of the

the nodes that must be examined.

to the root node where node 2 with yr = 1 is examined


noninteger

of the

in the tree search on the underlying

of the above

with Z = 14 for

both

of - 7.2941.

capacities

will improve
problem

tivity among
and bound

that

that can accomplish

discarded.
The branch

MILP

the value

Alternatively,

schemes

the

instance,

could be reduced to yield a tighter

constraints

method,

of

formulation

For

is selected next

for

As

be-

is noninteger.

variable

upper bounds

these

is

node (4) yields

which

yielding a solution
integer

is now

now

the

above.

units (these are described

y, = 0 yields a

(c)] which

The second

(-4.3703)

y,

improve

given

logical

objective

since its value

the constraint

y3, the remaining

for the branching

next.

variable

for the first node

the upper bound.


objective

y2 = 0 and

a lower

1 is examined

Z = 0 [i.e. alternative

a lower

Y3 = 0,

yields

as the branching

feasible

1 and 2 are then

the constraints

closest to 0.5. Imposing

comes

nodes

since

to 0.5 (this is

and bound.

course

model

profit.

The variable y, is selected first for the branching


in the relaxed

of

with

branch

leads

Furthermore,

to
both

branch
an

integer

algorithms

many of the binary variables


based on the available

are presented
on the variable
solution

the

try to fix as

as possible at each node

logic. The next sections discuss

each of these issues in detail.

Symbolic integration of logic


FORMAL

GENERATION
OF LOGIC
A SUPERSTRUCTURE

913

IN

As a first step in the generation of the logic, the


superstructure is represented by a directed graph. The
nodes and arcs in the graph are further classified into
various subgroups depending on the tasks that they
perform in the superstructure:
1. Mixer: mixes two or more streams. No other
unit operation is involved.
2. Splitter: splits a stream into multiple streams.
No other unit operation is involved.
3. Unit: These include units that perform a change
in composition and pressure and temperature in
the output streams. Examples include compressors, reactors, distillation columns, absorption
towers and membrane separators.
4. Sources/sinks: These provide the inlets and outlets to the process flowsheet: feed, product and
byproduct streams.
Recycle streams are handled by removing the connectivity with the main process path at the end of the
loop. This prevents one from obtaining circular relationships, without losing the connectivity of the
flowsheet. The proposed procedure to develop the
logical relationships in the model is described below.
1. Associate Boolean variables with every node in
Boolean variables Y will bc used to represent the existence of all the units (U,) in the
superstructure and 2 to represent the splitters, mixers, sources and sinks.
2. Develop
reiationships
between
Boolean
variables. This depends on the class of nodes to which
the variables belong. The logic associated with the
various kinds of nodes (see Fig. 3a-c) is as follows:
the graph.

(a) mixer:
zffl =z- YavYb,
Zm 2j Yc.

(2)

If Ya v Yb =E-Zm also holds, then the first relation


can be rewritten as Ya v Yb 9 Zm. Similarly, if
Yc * Zm holds, then the second relation can be
rewritten as Yc 0 Zm:
(b) splitter:
Zs *

Ya,

zs =+ Yb v Yc.

(3)

If Ya S- Zs holds, then the first relation can be


rewritten as Zs 0 Ya. Similarly, if Zs =+ Yb v Yc
holds, then the second relation can bc rewritten as
ZsoYbvYc:
(c) Unit u E U,:
Yu*

YaAYbh

...

Yu*YlAY2A-~.AYm.
CACE
,7,%-E

nYn,

(4)

(b)
Ya
Ybo
Yn

Yl

yu

.. . . .

c?Ym

(cl
Fig. 3. Nodes in the graph of a superstructure: (a) mixer;
(b) splitter; and (c) Un component.

3. User specifications. Apart from the logic associated at every node, user specifications like limits on
the selection (e.g. at most one reactor) and availability of feed streams also need to be considered.
In case an input or output stream is known to
exist, set the corresponding
Boolean variable to
TRUE and simplify the logical relationships. The
above method generates all the logic describing the
connectivity between the various units in terms of Y
and Z variables. This approach, however, requires
the use of the variables Z which are often not
included in the quantitative model (usually only
Y-variables are used). The number of variables Z
can be decreased by applying the following procedure.
Reduction

of logical relations

Check for valid Exclusive OR relations involving Z variables from user specifications. Each
Exclusive OR can lead to the removal of one
variable from the logic.
Check if relations with no implications occur in
other expressions to establish their truth and
substitute for Z variables accordingly.
Eliminate equivalence relations with one or
more variables and occurring only once in
another expression. Repeat until no further
reduction is possible.
If the MINLP model does not require assigning O-l
variables to mixers and splitters, the logic is further
simplified by removing all the Z variables so as to

R. RAMAN

914

and 1. E. C~ROSSMANN

Fig. 4. Flowsheet superstructure.

express it only in terms of Boolean variables for the


process units (Y).
EXAMPLE:

GENERATION

OF LOGIC

To illustrate the procedure described above, consider the superstructure in Fig. 4 in which there are
eight possible units that can be used for the manufacture of the final product. We will use Z to denote
Boolean variables for mixers and splitters, and Y for
process units to show how the logic reduces in the
case when O-1 variables are only assigned to process
units. The graphical representation of the superstructure is shown in Fig. 5. The application of the
proposed procedure is as follows:
Eighteen Boolean variables are required to
mode1 the graph representation of the superstructure. Eight variables (Y 1, Y2, . . . , Y8)
model the units in the graph while 15 variables
(Zl,Z2,...,
Z15) model the splitters, mixers,
sources and sinks.
Expressing the relationships between variables
at every node, the following logical relationships
are obtained,
ZloZ2
220

280
Ylv

Y2

Y4eZ9

Y4

YlvY2023

Z9oY6vY7

23024

Y6v

z4oz5vz7

z100211

Z5.=-Z6vY3

Zll

270

213 9

Y5vZ8

z13oz15v

Y3v

Y5

Y7oZlO

oz12
Y8

YS oz14.

(5)

3. The only user specification is the availability of


the feed stream associated with the variable Zl
for the source. Setting Z 1 to TRUE in the logic
simplifies the expressions to:
22

280

Z2oYlvY2

Y4-oZ9

YlvY2ctZ3

Z90

23024

Y6v

z4oz5vz7

ZlO9Zll

Z5026vZ3

Zll

270

213 ++ Y8

Y5vZ8

z13*z15v

Y3v Y5

Y4

Y6v

Y7

Y7oZlO

oz12

Y8 0 214.

(6)

The logic in (6) is reduced by the procedure


proposed above. There are no exclusive OR
relationships in the model. Since 22 is true,
Y6

Fig. 5. Graphical representation of superstructure.

Symbolic integration of logic

Y 1 v Y2 and therefore 23, 24, 25 v 27 are true.


The relationships can now be expressed as:
Yl

280

Y2

Z5vZ7
ZSoZ6v
279

Y4

Y4oZ9
290

Y3

Y6v

Y7

915

operators. There are two ways of transforming these


propositions in A in order to perform the inference in
the branch and bound method. In the first case, the
logic propositions are converted into the CNF form
(see Clocksin and Mellish, 1981; Raman and Grossmann, 1991a):

Y6v Y7oZlO

YSvZ8

z13oz15v

ZlO~Zll

Y3v Y5

Zll9212
2130

Y8

Y8 -+Z14.

(7)

Substituting for Z5, 27, 28, Z9, ZlO, Zll, 212,


Z 13, Z 14 from the equivalence relations in which
they occur, the relationships reduce to:
Y40

YlvY2
Z6v

Y3v

Y6v

Y7

Y5vZ8

Y8oZlSv

Y3v

Y5

(8)

The binary variables required to model the system are


(Yl, Y2, _ . , Y8) and (26,Z8,215).
All other variables are either not required or else can be expressed
in terms of the remaining ones. Finally, if no Z
variables are desired to model the problem but only
Y variables are used, as is commonly the case, the
logic relations reduce to:
Y40

Yl v Y2
Y3vY5==-

Y6v

Y7.

Y8

(9)

Note that in the above, 26 v Y3 v Y5 v Y8 has been


eliminated because when 26 and 28 are absent, the
logical expression no longer applies. However, for the
implication Y8 - Z15 v Y3 v Y5, only the one way
implication Y3 v Y5 = Y8 applies.
REPRESENTATIONS

FOR LOGICAL

INFERENCE

Let the logic relations of the superstructure be


given by a conjunctions of q different clauses:

A = (L, AL,

A . . . AL,},

(10)

where Li is a logical proposition expressed in terms


of implications, OR, EXCLUSIVE OR and AND
L,:

where Pi and z are subsets of the Boolean variables


that correspond to some of the O-l variables, and s
is the number of disjunctive c1auses.t
In the second representation, the logic propositions
are converted into the DNF form (see Clocksin and
Mellish, 1981):

Y, v Y, =a Yj

Lz:

Y, *

L,:

(1 Y, v 7 Yz)

Y, v Y2

(process 1 or process 2 imply process 3),


(process 3 implies process 1 or process 2),
(do not select process 1 or do not select process 2).

disjunctiveclause,also known as a disjunct,is defined


here as a disjunction of literals.
$A conjunctive clause, also known as a conjunct, is defined

?A

here as a conjunctionof literals.

where Yi is a Boolean variablerepresentingthe existence


of unit i and 1 Y, its nonexistence; Qj and i are the
index sets of the Boolean variables which correspond
to a patiition of the &l variables yi. i = 1, . . .p. Each
conjunctive clause% separated by a disjunction represents the assignment of units in a feasible conf&uration, where it is assumed that each Boolean variable
has a one-to-one correspondence with the O-l binary
variables of the MILP model. Therefore, r represents
the number of alternatives in the superstructure.
Thus, the DNF representation is useful in cases when
the number of feasible alternatives is significantly
smaller than the total number of O-1 combinations.
To illustrate the CNF and DNF representations in
(11) and (12), consider the small example problem
shown in Fig. 1. The following propositional logic
expressions apply:

Removing the implications in L, and L, , and using


DeMorgans theorem in L, , the corresponding CNF
representation for the logic is:
q=(lYlv

Y3)n(-tYZv

Y3)

h(lY3vYlvY2)A(lY1vlY2).

(13)

R.

916

RAMAN

and I. E.

Distributing the OR over the AND operators, the


corresponding DNF representation is obtained and is
as follows:

Note that the disjunction in (14) represents the


three alternatives (a), (b) and (c) that were presented
earlier in the paper.
INTEGRATION

OF LOGIC
BOUND

WITHIN

BRANCH

AND

SEARCH

In this section we will propose a scheme for


incorporating the logical relationships in the DNF
and CNF forms in the branch and bound method. As
mentioned earlier, the constraints for the logical
relationships amoug units can be added to the MILP
model to tighten the LP relaxation. However, if the
number of constraints is large, this can lead to a
significant increase in the solution time of each LP
subproblem. Therefore, it is proposed in this work
not to add these constraints to the MILP model (PI),
but simply keep those logical constraints that are
specifications of the original model (e.g. do not select
both processes I and 2 in Fig. 1). The DNF and CNF
representations of logic will be analyzed and solved
symbolically for the branching rule and logical chaining as discussed below. The detailed description of the
proposed procedures will be presented in the next
section and will be followed by a small illustrative
example.
Branching rule. Instead of relying only on general
heuristics {e.g. variable with value closest to 0.5) or
penalties, we will make use of the logic in the DNF
and CNF forms that applies to each node.
The basic idea in the case of the DNF form is as
follows. In order to prune branches as soon as
possible, it is advisable to search among branches
that quickly lead to integer solutions. In order to do
this with the DNF representation, the variable that
has the highest priority for branching is the one that
occurs the least number of times in nonnegated or
negated form in the disjunctive representation of the
logic at the current node of the tree. The reasoning
behind the argument is that the variable that occurs
the fewest number of times in nonnegated or negated
form has the lowest number of feasible solutions
among the different alternatives implied by the disjunctive logic. Also, since such a variable has the
potential of eliminating a larger number of clauses
from the logic by making them false (i.e. infeasible),
this increases the possibility of solving the Boolean
equation for all or some of the remaining variables
which can then be fixed at that node.

GROSSMANN

For the case of the CNF


representation, the
Jeroslow and Wang (1990) heuristic, developed for
the satisfiability problem, has proved to be the most
effective for a similar reason as in the DNF case. The
variable that occurs in the largest number of short
clauses is chosen as the variable to branch on. The
motivation here is that using unit resolution it is more
likely for one to be able to fix other variables in a
short clause; furthermore, the greater number of
short clauses the branching variable appears in, a
greater amount of variables are likely to be fixed at
the resulting node.
Logical chaining. Both approaches attempt to fix
as many of the other binary variables as possible by
analyzing the available logic. The CNF based approach utilizes unit resolution (Clocksin and Mellish,
1981) for the purpose of making inferences, while the
DNF
approach
simplifies the current Boolean
equation, and searches for variables that occur with
the same sign in all the clauses and setting them to
that value at the node.
FORMAL
I.

PROCEDURE

Disjunctive

FOR PROPOSED

normal form

STRATEGY

approach

to state the procedure


formally, consider a node n that is under examination
in the tree. Let N,, be the subset of&l
variables with
fixed discrete value, and M, the subset of O-l variables that are treated as continuous (free). That is
N,, = {i 1y, is fixed at node n }, while M, = (i 1yi is free
at node n}_ Furthermore, consider the partition of
the
set
of
fixed
variables
N.
that
such
~,,=(ioN,lYi=l}
and F,={ioN,Iy,=O}
(of
course at the root node of the tree T,, = 0, F = 0).
We can then associate the Boolean variables Y, for
i E T,, and 1 Y, for i E F.,. Since these Boolean variables are fixed at node n, one can combine them with
the logic at node n expressed in DNF with the
following equation:
Branching

rule. In order

where Qmjand Q,, j = 1. r, are a partition of the set


M, of free variables. In order to select the variable to
be branched on among those belonging to the set M.,
define uj and E to identify the negated and nonnegated terms as:
a =

1
0

if
if

iEf&
i E Q,

pj=

1 -a/,

(161

where it is assumed that all Boolean variables occur


in each clause. When only a subset of the Boolean

Symbolic

integration

variables is used, the above procedure can be easily


extended. The index 4 of the binary variable with the
fewest number of nonnegated or negated terms is
then given by:

q=arg

min [ 2: -"(i

CCtj

+&M)]}.

917

Then the following Boolean variables can be calculated,

chaining.

-[

(18)

LL,,(,

(Y,) ,&

,Ep,

(Yi>

P:

Define 7; and 7:
negated terms as:

If

q Eiz,,,.

W;+

approach

(1 Yi)

idL,)]~

where

IEq

,Ep

(1 Y,)

VJ,AF

(1 Y,) .

(20)

to identify the nonnegated and

yf = 1 if i E Pj\T,

= FALSE,

and

where

Qm+?,=0a,

(21)

i2n+tj=0zr-

From equations (15) and (18) the Boolean equation


for node n + 1, will then be given by:
j.+P+1=

it

*
I-

vJE^,

(1 Y,)A yq
n

,=

1.r

i~Qn+l,

Y8,e~+1, (1 Y;)

Define A,, the length of the clause, as:

(22)
To measure the largest number of occurrences in
short clauses, we define, as in Jeroslow and Wang

I>

(1990):

y: = c y;: 2-4

(19)
Finally, the above Boolean equation can be solved
defining:
~m+l=(,fi,
and

iES+.

Branching rule. Consider a node n that is under


examination in the tree. Let IV,,= {i ( yi is fixed at
node n } and M, = (i 1y, is free at node n }. Furthermore, consider the partition of the N, such that
T,,={ioN,(y,=
1) and F,={i~iV.(y~=tJ}.
Then
the logic at node n can be expressed in the CNF
as:

WY+ =

Y,=FALSE,

2. Conjuncrive normal form

We then proceed to simplify equation (18) as follows:

6q EQ,,

i E R+;

This then implies that in addition to fixing y4 = 1 at


node n + 1, we can also fix y;= 1, i E R+,
and
provided these sets are nonempty. The
y,=O,iES+
case when the variable y, is set to zero (1 Y,) at node
n + 1 is similar to the above steps.

i=l....,r.

If

Y;= TRUE,

(17)

We will consider that having


selected the binary variable y,, we will expand at node
n the two nodes for y4 = O(l Y,) and y1 = 1 (Y,).
However, before solving the corresponding LP subproblems at the two new corresponding nodes, we
will perform a logical test to determine whether
additional O-l variables can be fixed. For this we
solve the Boolean equation in (15) by adding the
conjunction for the new fixed Boolean variable. Consider the case when this variable is Y, (this is equivalent to fixing y, = 1). We then first compute for each
of the r clauses in (15) the expression:
Logical

of logic

~~,,le.+.+0>

and
y, = c y; 2-*I.
I

(23)

The index q of the binary variable Y, that is to be


chosen for branching is then given by:
max (yr), max(yr)

I>.

(24)

R. RAMAN and 1. E.

918

Logical chaining. Assuming we branch on Y,, the


resulting CNF with s clauses, j = 1,2, . . . , s, is of the
form:

GROSSMANN

If mp = 0,

Set Yi = FALSE

i 4 F,, U FIX,,

FIX, = FIX,,
FIX, = FIX,U fi),

To determine which additional O-1 variables can be


Iixed when branching on Y4, unit resolution
(Clocksin and Mellish, 1981) can be performed and
the steps are as follows:
0. Set FIX,= {q}, FIX,,= 0.
1. Set j = 1, VARFIXED
= FALSE.
the
clause
2. If
(TeUFIXp)nPj#QI,
I!+ = TRUE since at least one of the nonnlgated Boolean variables is true. Go to Step 4.
3. To determine the number of unknown Boolean
variables, in the jth clause, set
W-Z;= IP,] - ]F,n P,] - 1FIXJ-I Z_,I
and

b. If rni + ml, > 1 (i.e. two or more unknown variables), go to Step 4.


C. If rni + m, = 0, all binary variables in the conjunct are fixed.
4. Ifj -z s, j =j + 1, go to Step 2. Otherwise go to
Step 5.
{1,2,. . .,p},
then all binary
5. If FIX,UFIX,=
variables are fixed. Go to Step 6. Note that one
still needs to solve the LP at the node.
If (T,UFIX,)fl(F,UFIX,)
# 0, there is a contradiction in the logic. The node is fathomed.
stop.
If VARFIXED
= FALSE, then no variable has
been tixed during the pass. Go to Step 6. Else
go to Step 1.
6. The following Boolean variables (and the associated 0-l variables) can be fixed:
Yi= FALSE, i E FIX,.
Y,=TRUE, i E FIX,;
The case when the variable Y, is set to zero (1
node n + 1 is similar to the above steps.
EXAMPLE

mi, = ]Pi] - ]TnflPj] - ]FIX,nPjl.


a. If m, + rni = 1 (i.e. only one unknown Boolean
variable), then:
VARFIXED
If m, = 0,

= TRUE,

Set Yi = TRUE

i 4 T,,UFIX,,,

FIX, = FIX,U (i},


FIX, = FIX,,

Y,) at

REVISITED

In order to illustrate the procedure suggested in the


previous section with the DNF, consider the example
in Fig. 1. The relaxed LP in this case is also
Z = - 13.9608 (see node 0 in tree of Fig. 6). Consider
now the disjunction for the three alternative designs
at the root node [see equation (14)]:
W=(Y,A1Y*AY3)v(lY1AYznY,)
v (1

Y, A 1

Fig. 6. Branchand bound tree with DNF method.

Y, A 1

Y,).

Symbolic integration
USitIgequation (16), we then have:
3
for Y,, c ai, = 1;

of logic

919

y, = 1 at node 3), we then have by analyzing IV* with


(18) the following:
Y2~

Wf=(l

j-1

Y,),

Y,)A(Y,A

W; = FALSE,

for Y,, $

x4= 1;

W:=(lY,AY,)A(lY,AlY,)=FALSE,

j-1

which leads to
Y,=TRUE.

W,=(TY,AY,)A(Y~)=-

for 1 Y,, i j3{= 2;


1-I
for iY2,

for i

5 BJz=2;
j= I

We can then fix y, = 1 in addition to fixing yt = 1;


this then leads to an integer point in node 3 with
objective Z = 3.6296 which is discarded as it exceeds
the upper bound. Finally, we consider the case when
y, = 0 (i.e. 1 Y,). This then leads to:
W;=(~Y,A~Y,)A(Y,AY~)=FALSE,

Y,, i pi = 1.
j- 1

Since. Y,, Y, and 1 Y, occur the least number


of times, we can choose any of the three binary
variables for branching. Say we select yz since
its fractional value is closest to 0.5. Consider first
the case Y, ( yz = 1). From equation (18) we then
have:
IV; = Yz A (Y, h 1 Y, A Y,) = FALSE,
IV:=

Y2A(lY,A

W;=

Y,A(~Y,A~Y~A~Y~)=FALSE.

Y,A Y,),

which leads to IV = Y2 A (1 Y, A Ys)==-Y, = FALSE,


Y ,=TRUE,
since R = {3}, S = (I}. Thus fixing
y2 = 1 means we can also fix y, = 0 and ys = 1. This
then leads to an integer point whose objective is
Z = 0.7058 which is an upper bound (see node 1 in
Fig. 6).
We consider next the case 1 Y, (i.e. fixing y, = 0).
We then have:

IV; = FALSE,

which leads to
Y,=FALSE.

W4=(1YzAlY,)A(lY3)=-

We can then fix yJ = 0 in addition to fixing y2 = 0


and y, = 0; this then leads to the global optimum in
node 4 with objective Z = 0. Since all nodes are
fathomed at this point the search is terminated. Thus,
it can be seen that in this simple example only 4 nodes
had to be enumerated as opposed to the 8 that were
needed with a conventional branch and bound
method. The search tree is shown in Fig. 6.
The approach for the CNF form follows similarly
and will only be illustrated at the root node. Consider
the CNF form at the root node [see equation (13)
which has four clauses]:
yO=(1Ylv

Y3)~(1Y2v

~(1Y3v

Ylv

Y3)

Y2)~(1Ylv1Y2).

From equations (21) and (22), we have:


wf=lY,A(Y,AlY,AY,),

w$ = 1

Yz A (1

ro 0 17
Y, A -I Y, A 1

rl 0 07

Y3 ),

which leads to

A, = 2,

Az=2,

A,=3,

A,=2.

Hence from equation (23), we have:


In this case we cannot fix either y, or y3 since
R = (0},
S = (0).
We then compute the corresponding LP with yz = 0 which yields Z = - 12. I482
as shown in node 2 of Fig. 6. In this node we can
select either yr or y3 for branching since both occur
in nonnegated and negated form once. Since y, is
closest to 0.5, we select Y,. For the case Yl (fixing

r: = 0.125,
y: = 0.5,

r: = 0.125,
7: = 0.5,

r: = 0.5,

y: = 0.125.

Since 1 Y,, 1 Y, and Ys all have the largest


Jeroslow-Wang value, we can choose any of the three
binary variables for branching. Say we select 1 Y2
(i.e. Y, = FALSE) since the fractional value of yz is

R.

920

closest

to 0.5.

comes

from

The

CNF

form

with

and I. E. GR~~.WANN

RAMAN

1 Y2

then

be-

the procedure
= {2},

outlined

Since

Pi

variable,

and

contain

Pi

variables

hand,

we now branch

CNF

form

V2=

From

more

on

than

Step 3, Pi
FIX,

{1,2,3),

meaning

Fig. 7. Note

tation

The most
and CNF

for

such

Generating

Kocis

and

but only 198

problems

the DNF

can

be done

and hence.

structure.

The

the

hand,

CNF

based

form

approach,

can be readily applied

of

represenreasonably

of the superon the other

to problems

number of feasible alternatives


However,

Y3)

with large

and is reasonably

= (l}
that

representation.

for DNF

is generally

easy

Hence

FIX,,UFIX,

y2 = 1 implies

=
fixing

form,

it also

but the order in which

these

changes.

DNF

than the CNF


that

important

PROPOSED

difference

APPROACHES

between

the DNF

is related to the class of prob-

lems that each of them

are more

adapted

for. The

Fig. 7. Branch and bound

to

the size of the

tree is usually smaller


a worst case

on the maximum

be

to

examined

number of

with

the

DNF

approach:
Theorem

1. The number

of nodes to be examined

by branch and bound with the disjunctive


is not more

OF THE

need

once the

easier in the DNF

based tree. Furthermore,

bound can be established

rule

as will

the procedure

case. Therefore,

based branch and bound

nodes

the branching

results. Secondly,
is much

it

than on

than for CNF

variable has been chosen,

case than in the CNF

have only one term. So

are obtained,

on the DNF

Firstly,

stronger

check for fixed variables

that as in Fig. 6 for DNF

approaches

the CNF

branching

which leads to fixing Y3 =


fixing

inferences

be shown in the numerical

P:=lYI.

and Pi

once the representations

is easier to perform

Pi=Y3,

only 4 nodes,

DlSCUSSION

by

[i.e. 198 clauses in (12)] which

due to the constrained

= (2):

Yl = FALSE.

are examined

considered

of

For example,

has 13 t%l variables

efficiently

y 1 = 0 and y3 = 1. The entire search tree is shown in


involves

(1989)

the

of combinations

the system.

unknown

one

Y, = TRUE),

Y3)r\(lY2v

P;=TRUE,

and

superstructure

Grossmann

where

is significantly

to generate.

Y,,-,[(-,Ylv

= {2,3},

the HDA

the binary variables.

is possible,

Y2 (i.e.

P;=lYlvY3,

FIX,

solutions

to the number

feasible configurations

becomes:

TRUE

design

is better suited

problems

is much smaller than the 213 = 8192 combinations

can be fixed. If on the other

Step 2, with FIX,

Applying

Step 2:

P:=TRUE.

no further inference

no additional

in the previous section,

P; = TRUE,

Yl,

P;=1Y3v

feasible

integer

binary variables that model

we get from

PI = 1 Yl v Y3,

of

based approach

of mixed

smaller compared

h(lY3vY1vY2)A(lY1vlY2)].

with FIX,

representation

number

I/=(~Y,)A[(~Y~vY~)A(~Y~vY~)

Using

DNF

for the class

(25):

clauses
The

1, where

logic in (12)

r is the number

of

in the disjunction.

proof

interesting
applied

than 2r -

for the theorem


to

note

that

to the motivating

tree with CNF

method.

is in Appendix

if

the

example

above

B. It is

theorem

is

which has a dis-

Symbolic integrationof logic


junction with 3 clauses, a maximum of 5 nodes (4 plus
the root node) is predicted which is in fact what was
obtained with the tree in Fig. 6.
COMPUTER

IMPLEMENTATION

The proposed branch and bound schemes were


initially tested with the computer codes ZOOM 2.2
(Singal et al., 1987) and SCICONIC
2.1 (1990)
through CAMS 2.25 (Brooke et al., 1988) by externally generating the sequence of LP subproblems
since these codes do not provide facilities for interfacing routines for branching or fixing variables.
An automated implementation was developed
using the Optimization Subroutine Library (OSL)
Version. 2 (IBM, 1991) which is a collection of
high-performance mathematical subroutines that can
be used to solve optimization problems on the RISC
System/6000 workstations. All the modifications to
the default branch and bound search procedure were
coded as FORTRAN
77 user-exit subroutines to
OSL. The branching variable was chosen by the
logic-based schemes proposed in earlier sections. Any
ensuing tie was broken by simply branching on the
tied variable that is closest to 0.5 in the relaxed
solution. It must be noted that more sophisticated
numerical-based branching rules could also be used

OSL

921

instead. The branching rule was incorporated in the


OSL user-exit EKKBRNU
subroutine which is invoked before the solution of the LP subproblems at
each node. The proposed procedures to identify other
variables to be fixed at a node apart from the
branching variable were incorporated in the user-exit
EKKEVNU
subroutine. These variables are then
fixed by changing their bounds in OSL. All other
components of the branch and bound search corresponded to the default options provided by OSL. The
DNF or CNF representations of the logic are inputs
to the system along with the model equations which
are specified through CAMS 2.25.
NUMERICAL

RESULTS

In order to demonstrate the effect of the proposed


procedure on the branch and bound search, two sets
of problems have been solved. The first one involves
the MILP model for the synthesis of separation
systems with no heat integration (Andrecovich and
Westerberg, 1985) and this has been solved for problems involving 4,5 and 6 components. The second set
of problems correspond to heat integrated distillation
sequences in which the temperatures are treated as
continuous variables as opposed to the model by
Andrecovich and Westerberg (1985). A description of

Proposed scheme

Fig. 8. Computer implementation of proposed algorithm.

R. RAMAN and I.

922

(a)

E.

GROSSMANN

Superstructure
Y6=+Y3hY9
Y7aY3AY8
YS_Y2VY7
Y9aYSvY6
YIQ*YtvY4

Yl *Y4vYS
Y2aYgAYIO
Y3e.Y6vY7
Y4aYIAYIO
YS=,YlAY9
(b)

Logic

reIatIons

Fig. 9. Superstructure
for Qcomponent separation.

this model is given in Appendix A. It should also be


noted that all problems were solved to global optimality (OPTCR = 0 in GAMS 2.25), and with the
default settings for the other parameters of the MILP
codes.
The superstructure for the 4-component system
without heat integration is shown in Fig. 9 along with
the logical constraints that were derived with the
Tabk

1. Four component sytxhesis problem

otixittal
for&&ion
Si2.C
Constraints
Variables
Billll~

procedure described in this paper. The problems were


solved with the original MILP formulation (Pl), the
formulation with logical constraints and with the
proposed method described in this paper. The branch
and bound codes used were ZOOM 2.2 (an academic
code), SCICONIC 2.11 and OSL Version 2 (both
commercial codes). The results for the problems are
shown in Tables l-3.
Note that the formu-

27
31
IO

Formulation
with logic

DNF based
approach

CNF based
approach

49

21
31
10

2-l
31
IO

31
IO

LP daxation

3601.6

3625.8

3601.6

3601.6

MILP solution

3625.8

3625.8

3625.8

3625.8

ZOOM
No. of nodes
No. of iterations
CPU time@)-

33
123
0.15

87
0.14

6
54
0.09

98
59
0.1

SCICONIC
No. of nodes
No. of iterations
CPU time(sy

17
45
0.13

32
0.08

38
0.06

8
39
0.08

12
26
0.54

0
21
0.32

8
20
0.6

IO
43
0.9

OSL

No.
of nodes
No. of iterations
CPU time(s)*

On an IBM R6WO/530.

Symbolic
Table

2.

Five

OrigilMll

formulation
Size
Constraints

integration
component

synthesis

Variables
Binary

61
20
4262.7

4278.

MILP

4304.I
44
577
0.86

SCJCONIC
No. of nodes
No. of iterations
CPU

CNF

based

approach
51

51
20
61

20
61

:A

4262.7

4262.7

4304.1

4304.I

4304.1

16
302
0.5

4
118
0.2

13
221
0.35

26

time(s)

problem

90

LP r-&ration
solution

923

DNF
based
approach

Formulation
with logic

51

ZOOM
No. of nodes
No. of iterations
CPU time(s~

of logic

13

79
0.19

0:::

60
0.09

68
0.13

52
179
1.23

2
48
0.49

11
55
0.74

14
79
1.33

OSL
No. of nodes
No. of iterations
CPU
time(s)
On

lations

with

explicit

an IBM

R6000/530

constraints

for

the

logic

is from 12 to 8 and 10 nodes with the DNF or CNF


approaches, respectively. The time savings with the
reduction of nodes was not realized with OSL due to
the small size of the problem and the overhead of the

have

that the
original MILP model, but produce a tighter LP
relaxation.
For the case of OSL, the CPU times
reported include the time for the symbolic manipulations of the logic; this is not the case for ZOOM and
SCICONIC.
The 4-component problem only involves 5 feasible
alternatives. As can be seen in Table 1, the addition
of the logic constraints causes the problem to be
solved as a relaxed LP and decreases the time despite
the increase in the size of the problem. On the other
hand, use of the propositional logic maintains the size
of the problem to that in the original problem, and
decreases the number of nodes from 33 in ZOOM and
17 for SCICONIC to 6 using the DNF and 8 using
the CNF based approaches. With OSL the reduction
almost

twice

the

number

of

constraints

Table

3.

symbolic

analysis.

In the 5-component problem, there are only 14


feasible 4-column sequences although 20 binary variables are required for modeling the superstructure. As
can

be seen

from

Table

2, the proposed

Six component

synthesis

shows

search

probkm
DNF
based
aooroach

CNF
based
anoroach

156
106
35

86
106
35

86
106
35

18.114.7

18,161.6

18,114.7

l8,l

18.170.3

18.170.3

18,170.3

18.170.3

Original
formulation

Formulation
with Ionic

SiZC
Constraints
Variables
Binary
LP

86
106
35

relaxation

MlLP

solution

ZOOM
No.
No.

of nodes
of iterations

CPU

time(s)=

SCICONIC
No. of nodes
No. of iterations
CPU time(s)
OSL
No. of nodes
No. of iterations
CPU time(s)
On

an IBM

R6000/530

55
1050

2.5
61

141
386
3.46

branch

and

considerable improvement over


the original model. The addition of the logic inequalities does reduce the time with both ZOOM and
SCICONIC (42 and 22%). However, the proposed
DNF based approach yields even more substantial
time savings with both codes (78 and 5g%). This is
mainly due to the fact that the proposed method only
requires 4 nodes as opposed to 44 with ZOOM and
bound

16
787

2.01
7

14.7

IO
292

12
326

0.64

0.72

IO

12

I62

98

0.31

0.14

0.16

8
219
1.18

18
84
1.06

26
154
3.09

loo

R.

924

OMAN

and I. E. GROSSMANN

Table 4. Synthesis of heat integrated distillation columns


Original
formulation

Formulation
with logic

DNF based
approach

258
291
100

473
291
100

258
291
100

1117.72
1900.58

1117.72
1900.58

1117.7
1900.58

could not solve


> 2000

78
20.78

185
8.5

> 100,000
> 2000

18
4.8

I .6

74
540
8.37

20
238
2.76

Size
Constraints
Variables
Binary
LP relaxation
MILP solution
ZOOM
No.

of nodes
CPU time(s)

SCICONIC
No. of nodes
CPU time(s)
OSL
No. of nodes
No. of iterations
CPU time(s)

> 100,000
> I ,ooo,ooo
15000

29

=On an IBM R6000/530

26 with SCICONIC. The CNF based approach, too,


yields time savings of 60 and 32% with ZOOM and
SCICONIC,
respectively, and requires I3 nodes.
With OSL, the original algorithm requires 52 nodes
while the DNF and CNF based algorithms solve the
problem in 11 and 14 nodes, respectively. It should be
noted that with OSL the number of nodes with the
DNF and CNF approaches is not the same as in
ZOOM and SCICONIC presumably because different subproblem selection rules were applied. In the
case of the ZOOM and SCICONIC implementations,
a depth first strategy was used until feasible integer
solutions were obtained; the selection of the subsequent node was based on the best lower bound.
In the 6-component problem, 35 &l variables are
used to model a superstructure that contains only 42
S-column sequences. Here, as shown in Table 3, only
IO nodes were required with the DNF based approach and 12 with the CNF based approach, as
opposed to the 55 in ZOOM and the 61 in SCI-

CONIC. Also, the addition of the logic constraints


required 16 and 7 nodes, respectively. Furthermore,
the proposed method was able to reduce the time
requirements from the original model by about 74
and 68 and 72% and 65% with the DNF and CNF
based approaches, respectively. With OSL, the original algorithm requires 141 nodes while the DNF and
CNF based algorithms both require only 18 and 26
nodes. respectively.
The second type of problem also involves the
synthesis of a 4-component separation system but in
addition considers heat integration by allowing for
heat matches between reboilers and condensers of the
various separation columns in order to reduce the
utility cost for operation. The number of binary
variables required to model this system increases
sharply from the previous case. For a 4-component
system, 100 binary variables are needed-10 to model
the existence of the distillation columns and 90 to
model the potential matches between the reboilers

Table 5. Effect of addition of violated inequalities at root node

Four components
Logic inequalities
No. of nodes
No. of iterations
CPU time(s)
Five components
Logic inequalities
No. of nodes
No. of iterations
CPU time(s)*
Six components
Logic inequalilies
No. of nodes
No. of iterations
CPU time(s)
Heat integrated distillation
Logic inequalities
No. of nodes
No.

of iterations

CPU time(sr

On an IBM R600/530

Original
formulation

Formulation
with lopic

Original model with


violated ineaualities

DNF based
aooroach

12
26
0.54

22
0
21
0.32

8
0
16
0.29

8
20
0.6

I6
0.29

0
52
179
1.23

39
2
48
0.49

6
32
101
0.85

0
11
55
0.74

6
4
49
0.46

0
141
386
3.46

70
8
219
I.IR

II
60
292
I.91

0
I8
84
I .06

II
5
67
0.7

215

74
540

> 100,000
> 1.ooo.olM

20
238

17
264

>

100,ooo
> 1 ,ooo,ooo

>5000

8.37

>5000

2.76

DNF with
violated ineaualities

8
0

2.62

Symbolic integration
and condensers

of the various

this problem
for

with

columns.

the DNF

this example

based

involved

320 clauses

tives),

and with the formulation

shown

in Table

bound

the

and

problem

even

hour

after

solves

ZOOM,
OSL!

OSL

were

While

not

I million

able

the problem

(Differences

the branching
ties.) Using
reductions

in

although

with

based

approach.

relaxation

and

over

185 nodes

of nodes

also

as with

with

are due to

in the case of

very

the

It is interesting

at the root node

based

with logic inequalities


are

as large

the

significant,

symbolic

to note

that the LP

was not improved

in

in this

inequalities,

solution

of

one might

ositions

only

a subset

MILP.

This

for instance

(1991)

addition

each
this

In this paper,
logic

relaxation
branch
bound

tree.

To

we

have

scheme,
hybrid

the

scheme

successively

LP

in

all satisfied.

performed

symbolically

scheme

are

given

with

in

Table

with logic constraints,


violated
proach.

As seen from

be

nodes

the

most

was

stopping

the DNF
5 (last

in

all

cases.

The

with the DNF

case was marginal

It is interesting

to note that only

be as efficient

(third

as handling

column)

and

with only
DNF

ap-

scheme
required

and proved

example.

inequalities

then

formulation

however,
violated

they

was

column)

a few logic inequalities,

efficient

con-

of symbolic

and bound

with

node

to

difference
in the last
adding

the

did not prove

to

all the logic inequalities

in

and

effective

hy-

Numerical

and

form,

models

this method

reductions

rules

well

as

variables

at

as

is a trade-off

separation

can greatly

form

problem

constraints.
a

in which

are included

vs

This

hybrid
only

solving

solving

trade-off
scheme

violated

stage are included

that

reduce

the most

as inequalities,

in part
has

were

the case when

a tighter

inequalities

the
and

distillation

of magnitude
with

a problem
can

syn-

in the branch
producing

of two orders

between

motivating

to the case when no

As for the comparison

all logic constraints


MILP

shown,

in the case of a heat integrated

achieved.

at

has been

for

of nodes to be enumerated
search with the DNF

and

inferences

of O-l

that compared

prop-

a branch

logical
As

by

logic

results on a small

on MILP

is included,

ation

these

subsets

among

generated

branching

fixing

thesis have shown

with

CNF

proposed.

for

be

Converting

has been

provides

logic

can

performs

a mechanism
example

that logic of relations

that

been

larger

with

fewer

be resolved
proposed

at the LP

in the symbolic

there

but

branch

relaxand

bound.
authors gratefully acknowledge
financial support from the Engineering Design Research
Center Carnegie Mellon University and from the Eastman
Chemicals Division.
Acknowledgements-The

REFERENCES

that

logic.

symbolic

DNF

solved

for the hybrid

formulation

scheme

the

when

these results the hybrid

than the symbolic

such
with

and bound

and a pure

only had to consider


fewer

LP

formulation,

original

inequalities

of

and

inequalities

results with OSL

with the original

lower

branch

the examples

the relaxed

branch

symbolic
the

effectiveness

all

step and

The

the

in the

scheme

in the LP

tighten

all the logic

at each

The computational
compared

the

solved

in which

vioIated

to

subproblems

by adding

were

within

order

study

were

and

cuts in

a hybrid

that are violated

incorporated

bound

of

in the

by Lien

few resolution

we have considered

bound

to

constraints.

inequalities

are
and

all logic

constraints

was proposed

procedure.

method

problem,

approach

with

as an alternative

who incorporated

to the logic

in which

integration

of the logic

branch

superstructure

into

bound

savings;

the MILP

consider

include
Wahl

SCHEME

the symbolic

the direct

systematic

number

In comparing

for

This paper has shown


units

bound

with

schemes

extensive
firm

schemes.

each node.

DNF

case.

HYBRID

more
to draw

the results do show the potential

integration
brid

Although

is required

CONCLUSIONS

the

and 20 nodes with

in the number

CPU-time

not

solve

the DNF

rules used by each code

the formulation

are
and

problems.

experience

ZOOM,

to

iterations

RISC/6000,

29 nodes with SCICONIC

three

clusions,

altema-

branch

last

925

improvement

formulation.

on the IBM

approach

symbolic

a considerable

original

SCICONIC
CPU

shows

(i.e.

the

numerical

which

with inequalities

4. The proposed

search

over

The result for

approach,

of logic

Andrecovich M. J. and A. W. Westerbcrg, An MILP


formulation for heat-integrated distillation sequences synthesis. AIChE JI 31. 1461-1474 (19851.
Balas E. and C. H. Martin, Pivot and complement-a
heuristic for &i programming. Mgtnt Sci. 26, 86-96
(1980).
Beak E. M. I., Integer Programming. The State of the Art
in Numerical Analysis (D. Jacobs, Ed.), pp. 408-8.
Academic Press, London (1977).
Benders J. F., Partitioning procedures for solving mixedvariables programming problems. Namer. Math.
4,
238-252

(1962).

Brooke A.,
Users

D.

Guide.

Kendrick and A. Meeraus, CAMS-A


Scientific Press, Redwood City, CA

The

(1988).
Chandru V. and J. N. Hooker. Extended horn sets in
propositional logic. Working P&r 88-89-39. Graduate
School of Industrial Administration, Carnegie Mellon
University, Pittsburgh (1989).
Clocksin W. F. and C. S. Mellish. Programming in Prolog.
Springer-Verlag, New York (198 I).

R. FLAMAN and 1. E. GROSSMANN

926

Duran M. A. and 1. E. Grossmann, An outer-approximation algorithm for a class of mixed-integer nonlinear


programs. Marls. Program. 36, 307-339 (1986).
Garftnkel R. S. and G. L. Nemhauser, Integer Programming.
Wiley, New York (1972).
Geoffrion A. M., Generalized Benders Decomposition.
J. Optimization

Theory

Applic.

10, 237-260

Model

for

the Synthesis
Distillation

s.t.

C [pk + flk
*

Sequences

Fk + (CS)QSTEAM,
T CONDL
- 330

330

F,-

1
,P&

t,F;=Oo,

m = I,..

Qk-K,F,=O.

k = l,...,COL,

Fk-

k = I ),...)

u,y,=SO,

.,NI,

COL,

T REEL= T,,w,~+DT,,~

k = l,...,COL,

c
QEXkj+QCWQCW,=Q,,
, EcoL,/l
c
QEX,, + QSTEAMt
JEcoLVl

k=l....,

COL,

= Q,.
k=I,....COL,

QEK,

T ReB.j -

UP,jZ,j
TCOND~

< 0,
-

k,j

EMAT

= 1,

. . . , COL,

+ UB,,.(l - 2&j) 2 0,
k,j=l,...,

kEcoL

COL,

yk<NCOMP-11,

Fk3 TEBL 1Q,, QEX,, WW,,

QSTEAM,,

T COND.j, & 2 0. Yk. zij = {W).

where COL is the number of units in the superstructure. and


is the number of intermediate streams and NCOMP is
the number of components in the feed stream. The sets,
parameters and variables are as follows:

NI

Sets

psm=

{all columns which produce a given intermediate m as distillate or bottoms),


FS, = (all columns having an intermediate product
m as fead},
FS, = {all columns with the original feed stream as
feed}.

Parameters

~1~= cost coefficient for fixed charge of column


k,
CS = unit cost of steam for the reboilers,
CCW = unit cost of cooling water for the condensers.
e, = split fraction relating distillate or bottoms
Rows to feed flows,
K* = heat load coefficient for column k,
DT,,,,
= Temperature difference in reboiler and
condenser of column k.
EMAT = minimum exchanger approach tempcra-

ture,

of Heat-lnregrated

The following model assumes distillation columns with


sharp splits which can operate at arbitrary pressure levels.
These columns are represented through temperatures in the
condensers and reboilers whose difference is assumed to be
a constant for each column. These temperatures are also
bounded by the temperatures of the utility streams. All
potential heat exchanges between condensers and reboilers
are considered.
The selection of the sequence, temperature levels and heat
exchanges is given by the following MILP:
min

,EFS,

(1972).

Grossmann I. E.. MINLP optimization strategies and algorithms for process synthesis. Proc. of FOCAPD
Meeting
(Edited by Siirola et al.), pp. 105-132 (1990).
Gupta 0. K., Branch and bound experiments in nonlinear
integer programming. Ph.D. Thesis, Purdue University
(1980).
IBM, O&5-Guide
and Reference (Release 2). Kingston, NY
(1991).
Jeroslow R. E. and J. Wang, Solving propositional satisfiability problems. Ann. Math. AI 1, 167-187 (1990).
Kocis G.
R. and I. E. Grossmann.
A modeling/
decomposition strategy for MINLP optimization of process flowshccts. Computers
ckem. Engng 13, 797-819
(1989).
Lien K. M. and P. E. Wahl, If you cant beat them, join
them. Combine artificial intelligence and operations research techniques in chemical process systems design.
Proc. PSE91,
Vol. IV, pp. 1.1-15, Montebello, Canada
(1991).
Nabar S. and L. Schrage, Modeling and solving nonlinear
integer programming problems. Presented at Annuul
AIChE Meeting,
Chicago (1990).
Nernhauser G. L. and L. A. Wolscy, Integer and Combimtorial Optimization. Wiley-Interscience, New York (1988).
Raman R. and I. E. Grossmann, Relation between MILP
modeling and logical inference for chemical process synthesis. Computers them. Engng 15, 73-84 (199la).
Raman R. and I. E. Grossmann, Logical inference in branch
and bound techniques for process synthesis. Proc. of
PSE91,
Vol. I, pp. 10.1-15,
Montebello,
Canada
(199lb).
Raman R. and I. E. Grossmann, Integration of logic and
heuristic knowledge in MINLP optimization for process
synthesis. Computers &em. Engng 16, 155-171 (1992).
Scicon Ltd, SCICON
V/M
User Guide Ver.2.11,
Milton
Keynes, U.K. (1990).
Singal J., R. E. Marsten and T. Morin, Fixed order branch
and bound methods for mixed-integer programming: the
ZOOM
system. Working paper, Management Information Science Department, The University of Arizona,
Tucson (1984).
Van Roy T. J. and L. A. Wolsey, Solving mixed integer
programs by automatic reformulation. Opers Res. 35,
45-57 (1987).
APPENDIX

c
F,=Fror,
LEf&

+ (CCW)

1- v,u

_Yk),

WWkl.

U, = upper bound for flowrate of stream k.


UP0 = upper bound on heat exchange between
r&oiler of i and condenser 1.
US, = upper bound on TREe., - T,,;;,. j - EM AT,
Variables
F* = feed flowrate to column k,

pk = variable fixed cost for column k.


Q* = utility requirement for column k.
QEX, = heat exchange between t&oiler of column
i and condenser of column i.
QCW, = cold water requirement for wlumn k,
- temperature of the reboiler of column -,
iT _-_..
UEILi_
TCOND,/ = temperature of the condenser of wlumn_j.
y* = existence of wlumn k.
Z,=
existence of heat exchange between reboiler of column i and condenser of
wlumn i.

Symbolic integration of logic

APPENDIX

Proof of Theorem
Theorem I. The number of nodes to be examined by
branch and bound with the disjunctive logic in (12) is not
more than 2r - 1, where r is the number of clauses in the
disjunction.
h-00/: Consider any node in a binary tree representing all
combinations of the binary variables, Ui = TRUE, FALSE,
i = I,
. , m and satisfying the disjunction in f2o as given in
rootnoda

leaf node
Fig. 81.

A branch and bound tree.

927

(12). Since, the DNF representation is valid, there are only


two possibilities:
(a) The subset of Boolean variables at that node does not
falsify more than r - 2 clauses. This implies that this
is an intermediate node that requires further branching.
(b) The Boolean variables at that node falsify r - 1
clauses. This implies that the remaining clause can be
set to TRUE fixing all remaining Boolean variables
which then defines a terminal node.
Since each TRUE value of a clause gives rise to a different
terminal node, the maximum number of leaf nodes in the
branch and bound tree is r, where r is the number of clauses
in the DNF. Consider the search tree shown in Fig. Bl that
involves r leaf nodes and one root node. Let the number of
intermediate nodes be p. Since the structure is a tree, the
total number of edges in the tree is (r + p + I - 1) = (r + p).
But since each intermediate node is connected to 3 edges, the
root node to 2 edges and the each leaf node to only 1 edge,
the total number of edges in the tree is also given by
(3p + 2 + r)/2. Equating the two expressions, we get
p = r - 2. Therefore, the total number of nodes in the tree
=r+p+l=Zr-1.

Das könnte Ihnen auch gefallen