Vous êtes sur la page 1sur 47

Introduction Reformulation Relaxation Decomposition based Methods Application Result

Application of Decomposition Based Optimization


Methods to MINLP problems
Application to MINLP problems
Pratik Patil
10302018
under guidance of
Ravindra Gudi
&
Mani Bhushan
July 20, 2012
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Outline
1 Introduction
2 General Reformulation of MINLP
3 Relxation techniques for MINLP
4 Decomposition based Methods
5 Application
6 Result and Conclusions
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Mixed Integer Nonlinear Program (MINLP)
_

_
min
x,y
f (x, y)
subject to g(x, y) 0;
h(x, y) = 0;
x X, y Yinteger
Desired Properties:
f , g, h are smooth (convex) functions
X, Y are polyhedral sets, e.g. Y = {y [0, 1]
p
| Ay b}
y Y integer hard problem
f , c not convex Very hard problem
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Application of Decopmposition
Solving MINLP is hard due to combined integrality and
nonlinearity
Consider integer variable y as complicating/interacting
variable and apply model coordination and linearization on it
(Model Coordination)
This seperates MINLP into 2 subproblems:
NLP subproblem with y = y
k
for that iteration, commonly
called primal problem.
MILP subproblem which has its nonlinearities removed,
commonly called master problem.
Various methods dier on formation of master problem.
Another type of decomposition is to separate sparse MINLP
into smaller subproblems. (Goal Coordination)
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Primal Problem
Primal problem is similar for all decomposition based methods
If Primal is infeasible for given y
k
, primal feasibility problem is
solved
Primal P(y
k
)
min
x
f (x, y
k
)
s.t. h(x, y
k
) = 0
g(x, y
k
) 0
x X R
n
Primal Feasibility-1
min
xX
p

i =1

i
s.t. h(x, y
k
) = 0
g(x, y
k
)
i
,

i
0
i = 1, 2, . . . , p
Primal Feasibility-2
min
x

i I

w
i
g
+
i
(x, y
k
)
s.t. g
i
(x, y
k
) 0
h(x, y
k
) = 0
x X i I
Primal povides UBD and cut information
Feasibility problem is used to exclude feasible point through
Master problem
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Block-Separable Reformulation
The block-structure h
i
(x) =

p
k=1
h
k
i
(x
J
k
) inuences the
quality and computation of a relaxation
small blocks: fast computation of underestimators and cuts
large blocks: better relaxation (smaller duality gaps)
Many MINLP problems have a natural sparse structure which
can be reformulated to block separable.
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Splitting-schemes
Sparse MINLP can be reformulated to block-separable with almost
arbitrary block-sizes by
1 Partition the sparsity structure:
E
sparse
=
_
(i , j ) V
2
|

2
h
l
(x)
x
i
x
j
= 0 for some l {0, . . . , m} and x [x, x]
_
and V = 1, . . . , n (Vertex set)
into blocks J
1
, . . . , J
p
.
2 For each adjacent node set
R
k
= {i
p
_
l =k+1
J
l
| (i , j ) E
s
parse, j J
k
},
add new variables y
k
R
|R
k
|
and copy-constraints x
R
k
= y
k
where k = 1, . . . , p.
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Extended block-separable reformulation
By replacing block-separable constraints
p

k=1
h
i ,k
(x
I
k
) 0
by
p

k=1
t
i ,k
0, g
i ,k
(x
I
k
, t
i ,k
) := h
i ,k
(x
I
k
) t
i ,k
0, k = 1, . . . , p.
we obtain a problem with linear coupling constraints
min c
T
x + c
0
s.t. Ax + b 0
g
i ,k
(x
J
k
) 0, i M
k
, k = 1, . . . , p (P
ext
)
x [x
I
k
, x
I
k
]
(easier for generation and use of cuts)
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Convex relaxation and Lagrangian relaxation
Convex relaxation of (P
ext
):
(C) min{c
T
x + c
0
| x conv(G X) H}
Lagrangian relaxation of (P
ext
):
(D) max

min
x
{c
T
x + c
0
+
T
(Ax + b) | x conv(G X)}
Duality gap val(P
ext
) - val(D) smaller if blocks are larger
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
1 Dual methods: Lagrangian Decomposition
Solve (D) approximately by a subgradient method
2 Cutting plane methods:
Solve (C) approximately by generating supporting hyperplanes
3 Column generation methods: Simplicial Decomposition
Solve (C) approximately by generating extreme points and
extreme rays.
Role of block-separation/Decomposition:
Subgradients, supporting hyperplanes and extreme points are
computed by solving small NLPs.
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Convex underestimator
Replacement of nonlinear functions h
i
by a convex underestimator

h
i
in (P) yields a nonlinear convex relaxation:
min

h
0
(x)
s.t.

h
i
(x) 0, i = 1, . . . , m
x [x, x]
-underestimators

f (x) = f (x) +, Diag(x x)(x x)


and 0, x [x, x]
Polynomial underestimators
1 Generate underestimator, q(x) f (x), x [x, x] by
sampling
2 Set q(x) = q(x) +, Diag(x x)(x x)
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Valid Cut addition
solve separation problem (small NLP):

k
= min f
k
(x
J
k
;
)
s.t. g
k
i
(x
J
k
) 0,
x
J
k
[x, x],
and optimal variable value is x
J
k
The valid cuts added then are
Linearization cuts: g
i ,k
( x
J
k
)
T
(x
J
k
x
J
k
) 0, i A
k
Lagrange cuts: L
K
(x
J
k
; ) D
k
( ) where
D
k
( ) = min
xG
k
L
K
(x; )
Level cuts: f (x) f ( x) where x =
p
k=1
x
J
k
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Methods employed
MINLP problem
Generalized Benders decomposition
Outer Approximation
Extended cutting plane method
Generalized cross decomposition
LP-NLP based Branch & Bound
Simplicial Decomposition
Lagrangian Decomposition
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Generalized Benders Decomposition
Classical BD was for MILP, adapted to MINLP by Georion
Has only one continuous variable, others are projected out
Applied to following class MINLP problems:
min
x,y
f (x, y)
h(x, y) = 0
g(x, y) 0
x X R
n
y Y {0, 1}
q
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Under following conditions:
C1 : X is nonempty, convex and the functions, f (x, y), g(x, y)
are convex for each y {0, 1}
q
and the functions h(x, y) are
linear for each y {0, 1}
q
.
C2 : The set
Z = {z R
p
| h(x, y) = 0, g(x, y) 0 for some x X}
is closed for each y Y.
C3 : For each xed y Y V where
V = { y | h(x, y) = 0 &g(x, y) 0 for some x X},
one of the following condition holds:
1 the resulting problem has nite solution and has an optimal
multiplier vector for equalities and inequalities.
2 the resulting problem is unbounded, that is its objective
function goes to
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Generalized Benders Decomposition
Primal Problem for
k
th
iteration:
min
x
f (x, y
k
)
s.t. h(x, y
k
) = 0
g(x, y
k
) 0
x X R
n
Master Problem for k
th
iteration:
min
yY,
B

B
s.t.
B
(y;
k
,
k
), k = 1, 2, . . . , K
0

(y;

l
,

l
), l = 1, 2, . . . ,
where, (y;
k
,
k
) = min
xX
L(x, y, , )

(y;

l
,

l
) = min
xX

L(x, y,

, )
Here L(x, y,
k
,
k
) = f (x
k
, y
k
) +
k
T
h(x, y
k
) +
k
T
g(x, y
k
),
is Lagrangian of primal problem.
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Generalized Benders Decomposition - Algorithm
General Procedure
Primal problem provides
upper bound and Master
problem provides lower
bound
Primal and master are
solved iteratively
Optimal solution involves
UBD LBD , 0
is small constant.

Primal (NLP) Subproblem P(y
k
)
Master (MILP) Subproblem
MILP
Infeasible?
yes
No
Stop
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Outer Approximation
Uses rst order linearization instead of dual projection
Applied to following types of MINLP problem
Simple OA
min
x,y
c
T
y + f (x)
s.t g(x) + B(y) 0
x X, y Y
OA + Equality
Relaxation
min
x,y
c
T
y + f (x)
s.t h(x) = 0
g(x) + B(y) 0
x X
y Y
Generalized OA
min
x,y
f (x, y)
s.t. g(x, y) 0
x X R
n
y Y = {0, 1}
q
where
x X = {x | x R
n
, A
1
x a
1
} R
n
y Y = {x | x {0, 1}
q
, A
2
y a
2
}
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Outer Approximation- General Procedure I
General Procedure of OA:
let (x
j
, y
j
)solveNLP(y
j
)
linearize nonlinear f , g, h
about (x
j
, y
j
)
new objective variable
f(x.y)
MINLP(P) MILP(M)


f(x)
(M)
_

_
min
x,y
c
T
y +
OA
s.t. f (x
k
) +f (x
k
)(x x
k
)
0 g(x
k
) +g(x
k
)(x x
k
)
0 T
k
[h(x
k
) +h(x
k
)
T
(x x
k
)]
x X, y Y
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Outer Approximation - Master Problem
For non-convex f , g, h augmented penalty variant of OA is
used :
min
x,y
c
T
y +
OA
+

k
w
o
k
s
o
k
+

i ,k
w
p
i ,k
p
i ,k
+

i ,k
w
q
i ,k
q
i ,k
s.t. + s
o
k
f (x
k
) +f (x
k
)(x x
k
)
p
k
g(x
k
) +g(x
k
)(x x
k
)
q
k
T
k
[h(x
k
) +h(x
k
)
T
(x x
k
)]
x X = {x | x R
n
, A
1
x a
1
} R
n
y Y = {x | x {0, 1}
q
, A
1
x a
2
}
s
o
k
, p
i ,k
, q
i ,k
0, k = 1, 2, ..., K
Master problem: lower bound (underestimate of f,g), primal
problem: upper bound
Stop if LBD UBD
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Generalized Cross Decomposition
Consists of two phases:
Primal and Dual subproblem phase
Master problem phase
Condition on MINLP are same as GBD with extra condition
that
min
x
L(x, y, )
can be performed independently
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Generalized Cross Decomposition
Phase-1
Phase-1 contains iterative solution of Primal and Dual
problems
Primal problem is equivalent to OA/GBD
Dual problem is given by:
min
x,y
a
1
.f ( x, y ) +
k
T
1
g
1
( x, y ) + a
2
.
s.t. g
2
( x, y ) a
2
.
x X R
n
y Y = { 0, 1 }
q
where a
1
= 1; a
2
= 0 for feasible primal
a
1
= 0, a
2
= 1 for infeasible primal
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Generalized Cross Decomposition I
Phase-2
Two master problems
Primal Master Problem
min
y,
c

c
s.t.
c
q
1
_
min
xX
(q
3
( x,
1
,
2
)), y,
1
,
2
_
;
0 q
2
_
min
xX
(q
4
(x,
1
,
2
)), y,
1
,
2
_
;
y Y
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Generalized Cross Decomposition II
Phase-2
Relaxed Lagrange Master Problem
max

1
,
c

c
s.t.
c
h
k
(
1
), k = 1, 2, . . . , K
0 h
l
(
1
), l = 1, . . . , L

1
0
Pratik Patil
Decomposition Techniques for MINLP
Generalized Cross Decomposition - owchart





y = y
1
Primal
feasibility
CTDU
Dual
Feasibility
Primal
Subproblem
CTP
OT
CTD
Dual
Subproblem
OT
Relaxed Primal
Master Problem
Relaxed Lagrange
Master Problem
Stop
Stop
Infeasible feasible
No
Yes
y
k
No
Yes
No
Yes
Yes
y
k
No
Yes
x
k
, y
k

1
k
,
2
k


1
l
,
2
l

x
k
, y
k
y
k

1
k
,
2
k

Introduction Reformulation Relaxation Decomposition based Methods Application Result
LP-NLP based Branch & Bound
Avoids re-solving MILP master problems
Consider MILP branch and bound
Nonlinear function linearized and
continuous relaxation applied
Initial continuous solution branches
to start BB tree

Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
LP-NLP based Branch & Bound
Interrupt MILP, when y
k
Z
I
is
found
Solve Primal NLP (y
k
) get x
k

Integer
Feasible
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
LP-NLP based Branch & Bound
Linearize f , c about (x
k
, y
k
)
add linearization to the tree

Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
LP-NLP based Branch & Bound
Continue MILP tree-search until
LBD UBD

Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Extended Cutting Plane Method (ECP)
Solves only MILP master problem in OA by linearization (cutting
plane) around feasible points
min
Z
c
T
Z
S.t. l
(k)
j
0, j = 1, . . . , L
k
,
AZ a, BZ = b,
Z X Y, 1
where l (Z) = g
i
(z
k
) + .g
i
(z
k
)
T
(z z
k
)
No primal (NLP) subproblem, instead Kelley cutting plane
method used
Slow nonlinear convergence
Each y
k
needs is checked against 3 conditions for optimality
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Extended Cutting Plane Method (ECP) I
1 C1: If l
k
J
(Z
j
)

(atZ=Z
k
)
g
j
(Z
k
) Local Underestimator,
j = 1, . . . , L
k
. If not then
k
j
is updated as: ( > 1)
=
_
.
k
j
, l
k
j
(z
k
) > g
j
(z
k
),

k
j
, otherwise.
2 Case 2: If Z
k
is feasible then if,

k
j

g(Z
j
)

h
Feasible Underestimators j = 1, . . . , L
k
is tolerance, otherwise is updated as: ( )
=
_
.
k
j
,
k
j
<
g(Z
j
)

h
,

k
j
, otherwise.
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Extended Cutting Plane Method (ECP) II
3 Case 3: If the MILP problem is infeasible, do next iteration
with

k+1
j
= .
k
j
, j = 1, . . . , L
k
Pratik Patil
Decomposition Techniques for MINLP

L1 = 0, k= 0
k = k+1
Solve MILP
Feasible
Solution?
Calculate
gk = maxi{gi(zk)}
Local
Underestimators
Feasible Solution
to Main
Prbolem?
Feasible Solution
to Main
Prbolem?
Optimal Point
(xk,yk)
Update
According to C3
Update
According to C1
Update
According to C2
Add linearization
Lk+1 = Lk + 1
gk 0
yes
No
yes
yes
yes
yes
No
No
No
No
yes
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Simplicial Decomposition
MINLP(P)
MINLP (P
split
)
NLP P
split
(x, y)
min P
split
(x, y)
min P
k
(x
k
, y)
k = 1, . . . , p
convergence
min P(

x
i

i
+
0
x
0
)
s.t.

i
= 1
MIP
min P( x, y)
tolerance
criteria
yes, x
opt
, y
opt
Block-separation
y = y
smaller subproblems
x =
p
k
x
k
x
no
x
new
yes, x
objective value
no, new y
Figure 1: Employed algorithm of Simplicial decomposition for MINLP
1
Pratik Patil
Decomposition Techniques for MINLP
Lagrangian Decomposition
[allowframebreaks]
Algorithm 1: Algorithm for Lagrangian Decomposition
Input: Block-separable MINLP (P
split
)
Initialize with y = y
1
, y
1
Y feasible set of binary variable ;
while UBD
main
LBD
main
do
Set v = 1,
(v1)
down
= and Initialize duals

(v)
=
0
,
(v)
=
0
;
while
||
(v+1)

(v1)
||
||
(v)
||
< and
||
(v+1)

(v1)
||
||
(v)
||
< do
Solve relaxed primal-split problem (??) and get
minimizer x
(v)
and objective function value
(v)
;
Update lower bound of subproblem
(v)
down

(v)
if
(v) >
(v1)
down
;
Update dual variables as per subsections ??- ?? ;
Update upper bound of subproblem D(
(v)
) < Upper
bound for
(v)
;
end
Update LBD
main

(v)
if
(v)
> LBD
main
;
Solve MIP (P(x
(v)
)), with x = x
(v)
and subject to y Y;
Update UBD
main
P(x
(v)
) if UBD
main
< P(x
(v)
);
Fix binary variables, y = y
k
, where y
k
is solution of
P(x
(v)
);
end
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Application to Process Synthesis Problems
Heat Exchanger Network Synthesis Problem
Cyclic Scheduling Problem
Feedtray location problem
Synthesizing process ow sheet
Non-sharp separation in distillation
Multiproduct batch plant
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Heat exchanger network syntthesis I
Table: Heat exchanger network synthesis results
n |B| m p max nonl. var. conv.
Example-1 50 9 62 10 28 no
Example-2 50 9 62 10 28 no
Example-3 135 31 222 62 no
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Heat exchanger network syntthesis II
GBD OA-AP LP/NLP B&B simplicial Lagrangian
Decomposition Decomposition
Example-1
Iterations 7 5 21 25 23
of NLP-MIP
Total CPU 6.445 8.486 13.832 17.359 16.87
time (sec.)
Annual cost 16062.49 15499.7 15499.7 167602 16235
Example-2
Iterations 9 5 27 30 22
of NLP-MIP
Total CPU 7.482 9.76 15.91 18.548 17.356
time (sec)
Annual cost 40296 37526 39759 44586 42653
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Heat exchanger network syntthesis III
Example-3
GBD OA-AP GCD LP/NLP B & B
Iterations 11 8 10 30
of NLP-MIP
Total CPU 12.244 22.37 30.717 36.4611
time (sec)
Annual cost 585618 601102 602043 586784
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Cyclic scheduling problem I
Table: Cyclic scheduling results
n |B| m p max nonl. var. conv.
38 12 20 5 21 no
GBD OA-AP LP-NLP based B&B Simplicial Lagrangian SBB
Decompn. Decompn. (Global)
Iterations of 12 9 25 13 11
NLP-MIP
Total CPU 12.49 18.14 30.12 20.81 11.16
time (sec)
Total Prot 128833 128015 146821.38 136358.02 140104.6 153650
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Feedtray location problem I
Table: Feedtray location problem results
n |B| m p max nonl. var. conv.
88 36 284 11 17 no
GBD OA-ER-AP Simplicial Lagrangian SBB
Decomposition Decomposition (Global)
Iterations of 10 7 13 10
NLP-MIP
Reux ratio (r ) 0.9210 0.9899 0.9056 0.9103 0.99
Top product rate (P1) 66.134 69.57 65.367 64.42 69.61
Bottom product rate (P2) 33.86 30.41 34.63 35.58 30.38
Feedtray number 11 11 12 12 11
Total CPU 19.78 32.24 35.6 22.1536
time (sec)
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Synthesizing process ow sheets I
Table: Synthesis of process system - Application of various methods
n |B| m p max nonl. var. conv.
Prob.-1 7 3 6 3 2 yes
Prob.-2 2 5 14 5 2 yes
Prob.-3 18 8 23 8 2 yes
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Synthesizing process ow sheets II
Number of iterations Optimal
GBD OA-AP LP-NLP B & B Simplicial Lagrangian Solution
Decomposition Decomposition
Prob.-1 5 4 4 8 6 6.01
Prob.-2 7 4 6 13 9 73.04
Prob.-3 9 7 9 20 16 68.01
Total CPU time (sec)
GBD OA-AP LP-NLP B & B Simplicial Lagrangian
Decomposition Decomposition
Prob.-1 0.49 0.7987 1.2838 0.882 0.5488
Prob.-2 1.77 2.8851 4.78 3.328 1.9834
Prob.-3 7.27 11.85 16.94 12,72 8.977
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Non-sharp separation in distillation
Table: Nonsharp separation results
n |B| m p max nonl. var. conv.
51 2 43 15 5 no
GBD Lagrangian Simplicial OA-AP
Decomposition Decomposition
Iterations of 8 10 13 12
LP-MIP
Total CPU 0.846 1.996 3.156 7.9
time (sec)
Objective:Total annual cost = 156700
with column I and II in series as starting point
Pratik Patil
Decomposition Techniques for MINLP
Introduction Reformulation Relaxation Decomposition based Methods Application Result
Multiproduct Batch Plant
Table: Batch plant problems-results
n |B| m p max nonl. var. conv.
Problem-1 20 9 20 5 2 no
Problem-2 47 24 73 10 2 no
GBD OA LP/NLP B&B Lagrangian
Decomposition
Problem-1
Iterations of 5 3 10 7
NLP-MIP
Total CPU 4.573 3.26 8.856 3.76
time (sec)
Optimal Investment cost: = 167427.66
Problem-2
Iterations of 8 5 15 11
NLP-MIP
Total CPU 9.22 6.93 18.24 7.82
time (sec)
Optimal investment cost: = 285506.7
Pratik Patil
Decomposition Techniques for MINLP
General Results
CPU time Time taken by solver excluding compilation time
GBD < Block-separable techniques < OA-AP <
LP-NLP based B&B
Depends upon both computational intensity of
master problem and speed of convergence
Number of iterations Shows eectiveness of master problem in
forcing convergence
OA-AP < GBD < Lagrangian Decomposition <
Simplicial Decomposition < LP/NLP based B&
B
Optimal objective value Convergence to Global minima of
nonconvex problem.
None of the methods could guarantee the Global
minima
OA-AP shows good tendency towards obtaining
global solution to nonconvex problem.
Conclusions
Some heuristics can be suggested for application decomposition to
MINLP based upon results of solved problems
MINLPs in which discrete variables are present in large
portion and involved in nonlinearity should be solved with
Global B&B search.
Duality based decomposition techniques should be used for
problems involving high number of nonlinear constraints.
OA based techniques are suitable for problems where cost of
solving nonlinear is less. Augmented penalty should be used in
case of noncovex problems.
Application of block-separable tactics should be favoured
when-
Problem structure is naturally sparse
smaller number of nonlinear complicating constraints
Nonlinearity in complicating constraints is univariate or
separated according to blocks.

Vous aimerez peut-être aussi