; d
kt
2 Z 13
The objective function (8) is to maximize the prot. Con-
straints (9) are the capacity balancing equation ensuring
the capacity allocated to each type of manufacturing cell
not larger than that required by the orders. The remaining
constraints are the same as those in model I.
3.3. Inter-factories capacity planning model
The third model is formulated to solve an inter-factories,
two-party supply demand negotiation problem. A media-
tor considers both the excess capacity and excess orders
of two factories. The goal of the mediator is, from the per-
spective of a capacity purchaser through outsourcing, to
maximize the net prot by completing excess orders (d
0
kj
)
using excess capacity (N
0
kt
) of the factories. The model is
denoted as model III,
Maximize z
X
p
k1
X
v
t1
X
n
j1
P
j
C
tj
x
ktj
=1 I
k
14
Subject to
X
v
t1
a
tj
x
ktj
6 d
0
kj
; 8k; j 15
N
0
kt
X
n
j1
a
tj
x
ktj
Wu
kt
l
tj
P0; 8k; t 16
N
0
kt
; x
ktj
2 Z
17
The objective function (14) is to maximize the net prot.
The meaning of the other constraints is the same as the
ones in models I and II.
In this model, the two individual factories may bargain
on cost term (i.e., the price of resources, C
tj
) to reach a deal.
4. Solving the problem by ant algorithm
Due to the complexity of problems, this study proposes
an ant heuristic algorithm to nd an ecient resource port-
folio plan in which the resource investment decision, capi-
tal usage plan, resource conguration, and task allocation
are determined simultaneously. The algorithm not only
reduces the total cost of producing all orders but also
improves total prot in a factory level as well as the system
level. In many industrial situations, ant algorithm has been
shown to oer successful solution strategies for large and
complex problems of production systems. In the problem
addressed in the study, negotiating parties require to re-
examine capacity plans frequently on the base of updated
resource prices from its counter party. In this regards, the
ant algorithm provides excellent property in dealing with
changing parameters.
4.1. MAA algorithm
The proposed ant algorithm (called MAA) follows the
classical ACO algorithmic scheme and improves its e-
ciency by incorporating a constraint propagation proce-
dure for solving the problem, as follows:
Procedure: Modied Ant Algorithm (MAA)
Begin
Set parameters and initialize pheromone trails
Sort variables by the most constrained variable rule.
Repeat
For c from 1 to MaxCycle
For n from 1 to N
ants
A /
While jAj < jXj Do
Select a variable x
j
2 X that is not
assigned in A
Choose a value v 2 D(x
j
) with probability
P
A
(hx
j
,vi) using the repair mechanism to guarantee all
solutions are feasible;
A A [ {hx
j
,vi}
End While
End For
Update pheromone trails using the best ant of cycles
(the cycle best) {A
k
}
If (several cycles pass by) then reinforce
pheromone trails using the best ant trial (the global
best) {A
l
}
End For
Until max trials reached
End
2834 K.-J. Wang, M.-J. Chen / Expert Systems with Applications 36 (2009) 28312842
MAA ranks variables x
j
s by the most constrained vari-
able rule. The ant searching strategy begins with the repair
mechanism that allows an articial ant to construct a com-
plete non-violated assignment of values from D(x
j
)(the
domain of variable x
j
) to variables x
j
s. A cycle involves a
set of articial ants, each responsible for a solution of the
problem. When all the ants have completed their solution,
the cycle is over and the best assignment, A
k
(supposed it is
performed by the kth ant), is dened as the cycle best. Fur-
thermore, a trial contains several cycles. The best of all the
cycle bests, A
l
(made by the ant lth) is thus called the trial
best or global best.
When a new cycle best is less than the current cycle best,
the pheromone on path of the new cycle best is updated. A
cycle best is reset to a large number at the each end of trial.
When a new global best is updated after a lower cycle best is
found, the pheromone on the new global best path is also
updated. This new global best assignment is used to rein-
force the pheromone trails. The algorithm terminated when
the maximum number of trials is reached.
4.2. Pheromone and transition probability
MAA uses articial ants to construct a solution (by
building a path) on a graph that is specic to the problem.
In this study, the graph is constructed associated with the
decision variables, x
ktj
in the model.
More formally, consider an ant that has already visited a
set A of vertices, A = {hx
1
, v
1
i, . . ., hx
k
, v
k
i} which is a set of
variable-value pairs that corresponds to the simultaneous
assignment of values v
1
, . . ., v
k
to variables x
1
, . . ., x
k
,
respectively.
Let x
0
be the Start vertex, x
1
the rst assigned variable
and so forth. Supposed that the next selected variable to be
assigned is x
j
. Articial ants lay pheromone on each vari-
able-value pair hx
j
, vi in which the graph is G = (V, E)
where V = {hx
j
, vijx
j
2 X, v 2 D(x
j
)} and E = {(hx
j1
, mi,
hx
j
, vi) 2 V
2
, v 2 D(x
j
)}.
s(hx
j1
, mi, hx
j
, vi) is dened as the amount of phero-
mone that represents the learnt desirability of assigning
value m to variable x
j
and value m to variable x
j1
simulta-
neously. Formula (18) states the pheromone intensity of
the arc between the candidate vertex hx
j
, vi and the previ-
ous assigned vertex hx
j1
, mi. s
A
(hx
j
, vi) depends on the
pheromone laid on the edge between hx
j
, vi and hx
j1
, mi.
s
A
hx
j
; vi s
hx
j1
;mi2A
hx
j1
; mi; hx
j
; vi 18
Formula (19) denes the probability of transition that
an ant selects vertex hx
j
, vi,
p
A
hX
j
; vi
s
A
hX
j
; vi
a
g
A
hX
j
; vi
b
P
w2Dxj
s
A
hX
j
; wi
a
g
A
hX
j
; wi
b
19
a is a factor for weighting the pheromone, a > 0. The heu-
ristic term of [g
A
(hX
j
, vi)]
b
is included and dened as
1
Dxj
b
Dx
j
is the domain of variable x
j
). To set up a
low, initial production quantity can meet all the constraints
of the problem and nd an initial solution for further
improvement. Based on our experiments experiences, b is
set zero once an initial solution is found. Besides, in the
study b = 0 as the constraint of
P
v
t1
a
tj
x
ktj
d
kj
is applied
(model I) and b=1 as
P
v
t1
a
tj
x
ktj
6 d
kj
applied (models II
and III).
Formula (20) species the quantity of pheromone
deposited on edge (i, j) by the kth ant that has built the best
assignment, A
k
. TC(A
k
) is the total cost of kth ants assign-
ment. The objective function is to maximize total prot, so
Ds(A
k
, i, j) is proportional to the reciprocal of TC(A
k
),
DsA
k
; i; j
1
TCA
k
; if i; j f g 2 A
k
0; otherwise
(
20
At the end of each cycle, the algorithm checks whether
the current cycle best is less than the previous cycle best.
If so, formula (21) is applied to locally update the quantity
of pheromone on each edge (i, j) according to the best
assignment of the kth ant. q is the local evaporation rate,
0 6 q 6 1,
si; j 1 q si; j DsA
k
; i; j for all i; j 21
After several cycles in a trial, a global update is per-
formed. The global best of the trial is again used to inten-
sify pheromone, as indicated in formula (22), and c is the
global evaporation rate, 0 6 c 6 1. A
l
is the ant l with the
global best in the trial. The cycle best and global best are
used to enhance the magnitude of pheromone on the prom-
ising trails according to a double-phase elitist strategy that
helps articial ants always select the most promising arcs,
si; j 1 c si; j DsA
l
; i; j 22
4.3. Constraint propagation procedure
The proposed constraint propagation procedure
together with the ant algorithm xes the value domains
of the variables that have not yet been searched. Hence,
each articial ant walks in the search space of feasible solu-
tion regions. The procedure uses formulas (2), (9) and (15)
to conne values assigned to x
ktj
. After a cycle is ended,
each x
ktj
is assigned a value and N
kt
is computed using
the ceiling integer of the right hand side of formula (23),
N
kt
P
X
n
j1
a
tj
x
ktj
Wu
kt
l
tj
P0; 8k; t 23
5. Experiments
The parameters of the proposed MAA algorithm were
carefully investigated and tuned in a sensitivity analysis.
Major parameters include the evaporation rate (q, c), the
number of articial ant (N
ants
), the number of reinforce-
ment cycles (RF) and the number of cycles in a trial (C).
Table 1 presents the resulting setting of the parameters.
K.-J. Wang, M.-J. Chen / Expert Systems with Applications 36 (2009) 28312842 2835
This setting is determined by a 2
6
full factorial design of
experiments approach on the six parameters.
JAVA language is used to develop the code of the
proposed ant algorithm run on a P4 CPU with 256MB
RAM.
A case with ten types of machines, ten types of orders
and ten periods of production horizon problems are con-
sidered to verify the performance of proposed MAA
algorithm.
5.1. Case study of model I
The available operational time of each manufacturing
cell, W, is 1800 h for each period of time. The target utili-
zation of each kind of manufacturing cells is 100%. The
interest rate is 6% in all periods. The upper bound of bud-
get, CC, is 7 million USD. The initial number of manufac-
turing cells is N
0
= (4, 4, 4, 4, 4, 4, 3, 3, 3, 3) in which each
column represents the number of machine in a cell. The
unit price of selling capacity of remaining manufacturing
cells t of a capacity seller, C
0
t
, is set to 1250, 1975, 775,
1350, 2700, 1325, 2350, 2450, 2375, 11,200, respectively,
in this case. The other data are presented in Appendix A.
Table 1
Setting parameters
Parameters Values
a 1
q, c 0.01
N
ants
75
C 1000
RF 25
300000
350000
400000
0 1000 2000 3000 4000 5000
CPU time (in seconds)
V
a
l
u
e
o
f
r
e
s
u
l
t
a
n
t
r
e
m
a
i
n
i
n
g
c
a
p
a
c
i
t
y
o
m
a
c
h
i
n
e
s
Fig. 1. Fitness evolutions of MAA for case I.
0
1
2
3
4
5
0 1 2 3 4 5 6 7 8 9 10
Periods
#
M
a
c
h
i
n
e
s
M1
M2
M3
M4
M5
M6
M7
M8
M9
M10
Fig. 2. Resource portfolios of case I.
2836 K.-J. Wang, M.-J. Chen / Expert Systems with Applications 36 (2009) 28312842
The problem is highly complex. ILOG OPL optimiza-
tion software (2006) was applied but failed to solve the case
in 32 h.
The best solutions obtained by MAA (in 5000 CPU sec-
onds) are 3.920 10
5
. Fig. 1 depicts the evolution of objec-
tive values computed by MAA. Fig. 2 is the resulting
resource portfolio of machines.
5.2. Case study of model II
The initial portfolio of manufacturing cells is
(1, 1, 1, 1, 1, 1, 1, 1, 1, 1). The unit prots of orders are
(20, 30, 40, 20, 30, 40, 20, 30, 40, 40), respectively. The avail-
able operational time, target utilization, interest rate, bud-
get are the same as those in case I. The detail data are
presented in Appendix B.
The best solutions obtained by MAA (in 2000 s) are
7.379 10
7
. Fig. 3 displays the evolution of the objective
values of the algorithm. Fig. 4 shows the resulting resource
portfolio of the manufacturing cells.
5.3. Case study of model III
The remaining capacity of the manufacturing cells
obtained in Section 5.1 and the remaining orders obtained
in Section 5.2 are used in this case. It is assumed that the
capacity unit price (C
tj
) at the capacity seller is propor-
tional to l
tj
by multiplying a ratio R. Note that R = 0
imply that resources are free. The other parameters are
the same as the ones in case I.
The algorithms nd promising solutions in only a few
CPU seconds. Fig. 5 depicts the cost-prot structure for
0
1000
2000
3000
4000
5000
6000
7000
8000
0 200 400 600 800 1000 1200 1400 1600 1800 2000
CPU time (in seconds)
P
r
o
f
i
t
Fig. 3. Fitness evolutions of MAA for case II.
0
1
2
3
4
0 1 2 3 4 5 6 7 8 9 10
Periods
#
M
a
c
h
i
n
e
s
M11
M12
M13
M14
M15
M16
M17
M18
M19
M20
Fig. 4. Resource portfolios of case II.
K.-J. Wang, M.-J. Chen / Expert Systems with Applications 36 (2009) 28312842 2837
dierent resource outsourcing cost (in a ratio of R). As
shown in Fig. 5, the gross prot of the capacity buyer
decreases as the price of the required capacity price
increases. When R = 1.4 the gross prot converges to zero
which means no deal is reached. Note that when R = 0.6,
the system will reach balance that the two factories will
have equal prot.
5.4. Comparison to other algorithms
The author has applied a GA algorithm and compared
the performances of the MAA, GA and ILOG OPL with
small, executable cases. As shown in Fig. 6, the proposed
MAA and GA reach almost the same performance at the
end.
6. Comparison of the proposed cooperative capacity planning
and resource allocation model to a centralized model
In this section, an integrated planning for all the orders
and resources of the two factories is performed assuming
that there is a centralized decision-maker. The result of
such a centralized planning, in terms of prot-sharing
structures (eectiveness), serves as a benchmark (i.e., the
0
500
1000
1500
2000
2500
3000
3500
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4
R
P
r
o
f
i
t
The net profitof
capacity demander
(a)
The gross profitof
capacity demander
(a+b)
Production cost (b)
Fig. 5. Prot structure of dierent outsourcing cost in R.
0
50.000
100.000
150.000
200.000
250.000
300.000
350.000
400.000
0 1.100 2.200 3.300 4.400 5.500
CPU time (in seconds)
P
r
o
f
i
t
s
MAA CP GA ILO GOPL
Fig. 6. Objective function evolutions of CPGA, MAA and ILOG OPL.
2838 K.-J. Wang, M.-J. Chen / Expert Systems with Applications 36 (2009) 28312842
global optimization assuming the existence of a centralized
decision-maker) for the comparison with the negotiated
results of the two independent parties.
Assume that two factories, I and II, coordinate for
capacity planning. There are four types of parts committed
to customers by each of the two factories. The two factories
play as a capacity seller and capacity buyer for its partner.
The proposed MILP mathematical model III is applied to
both of the two factories and their mediator as well, to
search for prot maximization at each level.
There are ten periods of orders committed to customers
by each factory. There are two types of resources for each
factory. A factory has a choice to be processed either by
alternative machines within the factory or outsourcing to
another factory, whereas the other requires to selling
resources.
The other input data include (i) capacity unit price, C
kj
which is proportional to machine throughput (with a mul-
tiplier R due to machines technological level of capacity).
It is assumed that the cost of outsourcing is cheaper than
that of make-in-factory. (ii) The initial resource set,
N
0
= (2, 1, 1, 2). The available operation time of each
resource, W, is 1800 h for all periods. The target utilization
u
tk
of each kind of machine is 100%. The interest rate I is
6% in all periods.
The price of capacity (of both factories) is increased by
multiplying R throughout resource unit prices. The result-
ing local, negotiated solutions of the two factories, and the
globally near optimal solution are calculated. The lines of
prot in Fig. 7 are numbered in L1 (the global solution),
L2 (Factory I+ Factory II+ Mediator), L3 (Factory I+
Factory II), and L4 (Factory I). The solutions of Factories
I and II are computed by MAA separately, and lines L4
and L3 are then drawn for dierent R values, or equiva-
lently resource unit prices. The remaining orders and
capacity of the two factories are then matched using
MAA by the mediator for a cooperative capacity plan of
resources. Line L2 is depicted by adding the solution values
of Factories I, II, and Mediator. Finally, based on the
assumption that a centralized decision-maker exists, the
same mathematical model is used to compute the global
optimum by considering all of orders and capacity of
resources. L1 is thus drawn by running the MAA.
Fig. 7 indicates that (i) the dierences between L1 and
L2 are small and the proposed cooperative capacity plan-
ning and resource allocation model can approach the glob-
ally near-optimal model for a broad range of the
outsourcing cost, and (ii) the dierence between L2 and
L3 is the prot gain resulting from cooperation between
the cooperative parties. Experimental results revealed that
the proposed model and solving method improves total
prots for both the isolated (a single factory) and negotia-
tion-based (between factories) environments.
7. Conclusions
This study has addressed the inter-factories capacity
planning and resource allocation problem. We have
developed three mixed integer linear programming mod-
els with the objective either to minimize the total cost
or to maximize the company prot of planning a
resource portfolio. Issues that pertain to decisions to pur-
chase, sell machines and allocate machines economically
to produce orders were resolved. The other contribution
of this study is to solve this complex resource portfolio
problem by proposing a modied ant algorithm with a
repair mechanism, which enables articial ants always
walk in feasible solution regions, reducing computational
eort.
Furthermore, this study has developed a cooperation-
based supply demand orchestration framework in which
the issues of coordinated resource allocation between
0
2000
4000
6000
8000
10000
12000
0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90
R
P
r
o
f
i
t
L4: Factory I
L3: Factory I + Factory II
L2: L3+Mediator
L1: Centralized Model
Fig. 7. Prot share structure for dierent R.
K.-J. Wang, M.-J. Chen / Expert Systems with Applications 36 (2009) 28312842 2839
factories tasks can be resolved economically. Experiments
reveal that the proposed ant algorithm found good solu-
tions in both of isolated (a single factory level) and cooper-
ation-based (at a mediator level) environments.
Some interesting future work in this eld is suggested. In
the area of the resource portfolio decision-making in capi-
tal-intensive industries such as semiconductor manufactur-
ing and testing, alternative means of acquiring resources,
such as purchasing new facilities, renting from competitors,
transferring from other plants and selling equipment, can
be incorporated into the model. Beside, the proposed ant
algorithm can incorporate local search methods to improve
further the solution it nds and the eciency with which it
nds them. The proposed model limits two negotiating par-
ties. With some extension, our math model and solution
can handle capacity planning of multiple-negotiating par-
ties. In order to apply it to real supply chain scenarios, a
further study for multiple-negotiating parties is
encouraged.
Appendix A. Data of for case I
See Tables A1A4.
Table A1
Order demands (d
kj
: ten thousands)
Periods O1 O2 O3 O4 O5 O6 O7 O8 O9 O10
1 6 7 3 4 5 5 2 0 7 6
2 6 7 3 4 5 5 2 0 7 6
3 6 6 3 4 8 5 3 0 6 8
4 6 6 3 4 8 5 3 0 6 8
5 5 5 5 4 8 5 3 0 5 5
6 5 5 5 4 8 5 3 30 0 5
7 5 4 5 4 6 6 2 30 0 2
8 5 4 7 4 6 6 2 30 0 2
9 5 4 7 4 5 6 2 30 0 1
10 5 4 7 4 5 6 2 30 0 1
Table A2
Procurement cost of machines
*
(c
kt
: ten thousand)
Periods M1 M2 M3 M4 M5 M6 M7 M8 M9 M10
1 350 550 300 450 700 450 600 500 650 400
2 350 550 300 450 700 450 600 500 650 400
3 350 550 300 450 700 450 600 500 650 400
4 350 550 300 450 700 450 600 500 650 400
5 350 550 300 450 700 450 600 500 650 400
6 350 550 300 450 700 450 600 500 650 400
7 350 550 300 450 700 450 600 500 650 400
8 350 550 300 450 700 450 600 500 650 400
9 350 550 300 450 700 450 600 500 650 400
10 350 550 300 450 700 450 600 500 650 400
*
The unit salvage of each machine is set as half of procurement cost.
Table A3
Machine throughput corresponding to orders (l
kt
: units/h)
Periods O1 O2 O3 O4 O5 O6 O7 O8 O9 O10
M1 10 10 35 20
M2 20 25 30 45 25
M3 25 15 25
M4 30 30 30
M5 40 40 30 30 30
M6 25 45 20
M7 30 40 50 35
M8 30 50 55
M9 25 35 40 10 50
M10 25 35 10 10
2840 K.-J. Wang, M.-J. Chen / Expert Systems with Applications 36 (2009) 28312842
Table A4
Resultant remaining capacity of machines (N
0
kt
Periods M1 M2 M3 M4 M5 M6 M7 M8 M9 M10
1 3.444 1.79 2.148 3.814 3.814 3.061 3 1.828 2.777 2.079
2 1.341 2.172 3.555 2.703 3.675 3.876 2.841 1.939 2.777 2.015
3 3.722 2.462 0.592 2.888 3.814 2.827 2.365 2.363 3 2.619
4 1.738 2.851 3.333 3.444 4 3.382 3 1.525 2.5 1.952
5 2.174 2.166 1.777 3.259 3.814 4 2.861 2.05 2.722 2.365
6 0.309 3.753 1.925 3.074 3.675 3.876 2.702 2.595 2.277 2.841
7 1.46 2.753 1.851 2.703 4 4 2.841 2.898 2.777 2.301
8 2.293 3.308 2.888 2.888 3.861 3.5 2.841 2.898 2.277 0.365
9 2.333 3.197 3.333 2.333 4 3.629 2.841 3 2.638 1.888
10 2.293 1.993 4 2.703 4 3.722 3 2.713 2.777 1.253
Table B1
Order demands (d
kj
: ten thousands)
Periods O1 O2 O3 O4 O5 O6 O7 O8 O9 O10
1 15 5 0 2 8 10 13 3 7 6
2 15 7 0 2 8 10 13 3 7 6
3 8 7 0 3 8 10 15 4 6 5
4 8 10 5 3 8 10 15 4 6 5
5 5 10 5 10 8 10 15 5 5 4
6 5 10 5 10 8 0 13 5 5 4
7 5 3 8 10 8 0 13 6 4 3
8 0 3 8 7 8 0 13 6 4 3
9 0 2 15 7 8 0 10 7 3 3
10 0 2 15 5 8 0 10 7 3 3
Table B2
Procurement cost of machines (c
kt
: ten thousand)
Periods M11 M12 M13 M14 M15 M16 M17 M18 M19 M20
1 150 350 300 200 250 500 450 200 450 400
2 150 350 300 200 250 500 450 200 450 400
3 150 350 300 200 250 500 450 200 450 400
4 150 350 300 200 250 500 450 200 450 400
5 150 350 300 200 250 500 450 200 450 400
6 150 350 300 200 250 500 450 200 450 400
7 150 350 300 200 250 500 450 200 450 400
8 150 350 300 200 250 500 450 200 450 400
9 150 350 300 200 250 500 450 200 450 400
10 150 350 300 200 250 500 450 200 450 400
Table B3
Machine throughput corresponding to orders (l
kt
: units/h)
Periods O1 O2 O3 O4 O5 O6 O7 O8 O9 O10
M11 20 10 10
M12 10 30 20 30
M13 35 15 15 10
M14 30 20 15
M15 20 25 25
M16 15 25 20 50
M17 25 25 55
M18 10 30 25
M19 25 35 45
M20 25 50 25
Appendix B. Data of case II
See Tables B1B4.
K.-J. Wang, M.-J. Chen / Expert Systems with Applications 36 (2009) 28312842 2841
References
Bauer, A., Bullnheimer, B., Hartl, R. F., & Strauss, C. (2000). Minimizing
total tardiness on a single machine using ant colony optimization.
Central European Journal of Operations Research, 8(2), 125141.
Cachon, G. P., & Zipkin, P. H. (1999). Competitive and cooperative
inventory policies in a two-stage supply chain. Management Science,
45(7), 936953.
Chang, T. S. (2001). An auction system for resource capacity. Taiwan
Patent No. 503361, 29 (27).
den Besten, M., Stutzle, T., & Dorigo, M. (2000). Ant colony optimization
for the total weighted tardiness problem. Parallel problem solving from
nature PPSN VI 6th international conference, Paris, France,
September.
Dorigo, M., & Di Caro, G. (1999). The ant colony optimization meta-
heuristic. In D. Corne, M. Dorigo, & F. Glover (Eds.), New ideas in
optimization (pp. 1132). McGraw-Hill.
Dorigo, M., & Stu tzle, T. (2000). The ant colony optimization metaheu-
ristic: Algorithms applications and advances. Technical Report IRI-
DIA/2000-32, IRIDIA, Universite Libre de Bruxelles, Belgium.
Gravel, M., Price, W. L., & Gagn, C. (2002). Scheduling continuous
casting of aluminum using a multiple objective ant colony optimization
metaheuristic. European Journal of Operational Research, 143(1),
218229.
Hsu, J. S. (1998). Equipment replacement policy A survey. Journal of
Production and Inventory Management, 29(4), 2327.
Huang, S. Z. (2002). An agent-based order exchange model of wafer fab.
Master thesis, National Chiao Tung University.
ILOG (2006). ILOG Company Web homepage. <http://www.ilog.com/>
Accessed 28.10.06.
Jiang, I. S. (2000). A framework of capacity trading in semiconductor
manufacturing industry. Master thesis, Department of Industrial
Engineering, National Taiwan University.
Mayer, B. C. (1993). Market obsolescence and strategic replacement
models. The Engineering Economist, 38, 209222.
Merkle, D., & Middendorf, M. (2000). An ant algorithm with a new
pheromone evaluation rule for total tardiness problems. In EvoWork-
shops 2000 (pp. 287296).
Meyer, B. (2005). Constraint handling and stochastic ranking in ACO.
IEEE CEC congress on evolutionary computation, Edinburgh,
September.
Meyer, B., & Ernst, A. (2004). Integrating ACO and constraint propa-
gation. Ant colony optimization and swarm intelligence (ANTS 2004),
Brussels, September.
Rajagopalan, S. (1994). Capacity expansion with alternative technology
choices. European Journal of Operational Research, 392402.
Wang, S.-M., Chen, J., & Wang, K.-J. (2007). Resource portfolio planning
of make-to-stock products using a constraint programming based
genetic algorithm. Omega International Journal of Management
Science, 35(2), 237246.
Wang, K.-J., & Lin, S. H. (2002). Capacity expansion and allocation for a
semiconductor testing facility with a constrained budget. Production
Planning and Control, 13(5), 429437.
Ying, K.-C., & Liao, C.-J. (2003). An ant colony system approach
for scheduling problems. Production Planning and Control, 14(1),
6875.
Table B4
Remaining orders (d
0
kj
: ten thousands)
Periods O1 O2 O3 O4 O5 O6 O7 O8 O9 O10
1 0 5 0 0 0 5 13 3 7 6
2 0 7 0 0 0 5 13 3 7 6
3 0 7 0 0 0 5 15 4 6 5
4 0 10 0 0 0 5 15 4 6 5
5 0 10 0 0 0 5 15 5 5 4
6 0 9 0 0 0 0 13 5 5 4
7 0 2 0 0 0 0 13 6 4 3
8 0 2 0 0 0 0 13 6 4 3
9 0 1 0 0 0 0 10 7 3 3
10 0 1 0 0 0 0 10 7 3 3
Total 0 54 0 0 0 25 130 50 50 42
2842 K.-J. Wang, M.-J. Chen / Expert Systems with Applications 36 (2009) 28312842