Académique Documents
Professionnel Documents
Culture Documents
This paper proposes to apply analytical target cascading (ATC) for configuring
assembly supply chains with convergent structures. Individual enterprises in a
supply chain are represented as separate elements in an ATC hierarchy. They are
able to maintain autonomous and heterogeneous decision systems for optimising
their private decision variables and objectives. They collaborate vertically and
laterally along the ATC hierarchy through their common decisions to achieve the
overall consistency and optimality. ATC offers a competitive balance between
computational efficiency and effectiveness while providing an opportunity for
parallel computation to further improve the efficiency. This paper also
investigates the feasibility for individual enterprises to set local targets while
participating in the supply chain configuration (SCC). SCC usually involves
discrete decision variables, causing significant difficulties for existing ATC
techniques to achieve system consistency. Therefore, a new consistency scheme
has to be proposed in this paper, including two techniques: importance weighting
factor (IWF) and dynamic constraints (DC). A case study is used to demonstrate
the application of the ATC method for solving typical SCC problems. A series of
comparative analyses are conducted to identify the strengths of the ATC method
and demonstrate the effectiveness of the proposed consistency scheme.
Keywords: analytical target cascading; supply chain configuration; autonomous
decision systems; distributed problem solving
1. Introduction
A supply chain is composed of a complex sequence of processing stages, ranging from the
raw materials supplies, parts manufacturing, components and end-products assembly, to
the delivery of finished products (Yu et al. 2008, Chan and Chan 2009). This paper
considers a product’s assembly supply chain where each stage refers to a specific
assembling process for those components supplied from upstream stages. Normally, the
manufacturer corresponding to each stage has multiple alternative processing options to
choose from, e.g., a component manufacturer may have alternative outsourcing companies
while a final product manufacturer has alternative OEM partners (Chan et al. 2007,
Chiang et al. 2007). Supply chain configuration (SCC) aims to define a unique network
with certain optimal performance indicators through choosing one from the alternative
options for each stage and assigning the operational parameters between adjacent stages
(Huang et al. 2005a, Zhang et al. 2008).
Researchers have been searching for efficient and effective solution methods for solving
SCC problems. Graves and Willems (2001) developed a two-state dynamic program (DP)
to solve a notebook computer SCC problem that was modelled as spanning trees. For the
same problem, Huang et al. (2005a) adopted a genetic algorithm (GA) as a solution
approach and achieved better optimisation efficiency. Park (2001) built a comprehensive
decision model to simultaneously determine the strategy for developing product families
and configuring the associated global supply chain. In these works, SCC problems are
defined as centralised decision models solved by all-in-one (AIO) approaches. Although
AIO approaches have demonstrated acceptable computational efficiency and effectiveness,
two intrinsic limitations are hard to overcome. First, decision details of all supply chain
stages will be shared in the centralised decision model. Therefore, it is difficult to protect
the business privacy, such as local targets or suppliers, of each autonomous enterprise, not
to say that some enterprises want to use their own decision facilities. Second, an AIO
approach is computationally expensive. It may be acceptable to spend five minutes (for
DP) or one minute (for GA) solving an SCC problem with 17 stages each of which has
only two decision variables (Huang et al. 2005a) but the optimisation time would be
unexpectedly long if the number of stages and/or decision variables is further increased.
Analytic target cascading (ATC) is a model-based, multilevel, hierarchical optimisation
method for system design (Huang et al. 2005b). ATC is potentially suitable for addressing
the two limitations mentioned above. In ATC, a system is hierarchically partitioned into a
set of individual elements, each of which has its respective decision models. As a result, a
large complex system design problem could be converted into several smaller sub-
problems. Sub-problems are solved starting from the top-level element, and design targets
for the lower level are cascaded. If the cascaded targets are not achievable in a lower-level
element, the upper-level is required to iteratively adjust the targets; otherwise targets are
cascaded until the bottom level is reached. When the consistency among all elements is
reached, their sub-solutions form the final system design.
ATC has been applied for solving optimal design problems in various engineering
fields. Since initially introduced to automotive vehicle design (Kim 2001,Kim et al. 2001,
Kim et al. 2003, Nyström et al. 2003), ATC has been successfully used for the optimisation
of building design (Choudhary et al. 2003, 2005), product family design (Kokkolaras et al.
2002), linking marketing and product design (Michalek et al. 2005), manufacturing
applications (Li et al. 2004), and also aircraft design (Allison et al. 2006). When ATC is
applied to SCC, each supply chain stage is modelled to an individual ATC element and the
whole convergent supply chain is converted into a hierarchical ATC model. Enterprises of
different stages enjoy their autonomy to decentralise and formulate their own local
problems and choose solution methods (Qu et al. 2008). Their decision support systems
can be achieved through parallel distributed computation which further improves the
overall optimisation efficiency. Qu et al. (2009) develops a service-oriented ATC platform
called atcPortal for realising such a distributed SCC optimisation.
Several specific research objectives are addressed in this paper: (1) identifying a general
procedure of applying ATC to solve SCC problems; (2) demonstrating the effectiveness of
ATC in solving SCC problems; (3) identifying advantages and limitations of ATC
compared to AIO approaches; and (4) dealing with special issues in applying ATC for SCC
problems, including the presence of discrete variables in the ATC optimisation where
traditional weighting update method (Michalek and Papalambros 2005) is not applicable.
International Journal of Production Research 6885
The rest of the paper is organised as follows. Section 2 presents an overview of the
ATC method and the five key steps. The five steps are then exemplified with an SCC
problem, each in a separate section. Section 8 discusses the two techniques of importance
weighting factor (IWF) and dynamic constraints (DCs).
T
Top level P1,1
R2,1 , R2, 2 , ~
x1,1 R1,1 R21,1 R22,1 R21, 2 R22, 2
r1,1 ( R2,1 , R2, 2 , ~
1 2
x1,1 ) y 2 ,1 y 2 ,1 y12, 2 y22, 2
Intermediate P2 ,1 P2 , 2
level (s)
R32,1 R33,1 R32, 2 R33, 2 R3,1 , R3, 2 , ~
x2,1 , y2,1 R2,1 R2, 2 ~
x2, 2 , y 2, 2
2 3 2 3
y y y y
3,1 3,1 3, 2 3, 2
r2,1 ( R3,1 , R3, 2 , ~
x2,1 , y2,1 ) r2 , 2 ( ~
x2 , 2 , y 2 , 2 )
Bottom level(s) P3,1 P3, 2
~ T
x3,1 , y3,1 R3,1 R3, 2 ~
x3 , 2 , y 3 , 2 P0
r3,1 ( ~
x3,1 , y3,1 ) r3, 2 ( ~
x3, 2 , y3, 2 ) P i : Design problem i
ri (x): Analysis model in P i
r0 ( x)
The third step of an ATC analysis is to formulate the local optimisation problems for
each element in the ATC hierarchy. Deviations of responses and linking variables are
included in the objective function for minimisation of an element (Kim et al. 2003). These
deviations are reduced through each iteration of cascading, eventually becoming
acceptable according to the given tolerances. In this case, the overall system is considered
to reach a consistent state. In Figure 1, P2,1 receives response target, R12,1 , and linking
variable target, y12,1 , from P1,1. At the same time, it also receives lower-level responses, R33,1
and R33,2 , and lower-level linking variables, y33,1 and y33,2 , from P3,1 and P3,2. Therefore, the
corresponding deviation terms, kR2,1 R12,1 k22 , k y2,1 y12,1 k22 , "1 (for constraining
kR3,1 R33,1 k22 ), "2 (for constraining kR3,2 R33,2 k22 ), "3 (for constraining k y3,1 y33,1 k22 ),
"4 (for constraining k y3,2 y33,2 k22 ) will constitute the local objective function. k k22 is the
squared l2-norm. The index used in superscript indicates the level from which the targets
(responses) are cascaded (backtracked).
Once the ATC problem has been formulated, the ATC process is ready to be executed
and then the results are analysed and discussed. These five steps are generally followed in
applying ATC for solving SCC problems.
This supply chain has 13 processing stages. The two most downstream stages, stage1,1
and stage0,1, represent the final product demand and assembly stages respectively. They
will be undertaken by the same enterprise and thus conceptually combined to a single
stage, noted as stage1,10 . Hence, only 12 stages are actually considered in ATC modelling.
Through partitioning this supply chain by stages, a valid ATC model could be easily
obtained. The hierarchical index used in the figure follows the ATC convention, that is,
‘stagei,j’ denotes the jth stage in the ith level of the supply chain.
According to the business scopes, these stages are categorised into three sets. The first
set includes all the most upstream (leaf) stages which procure raw materials and then make
parts, being defined as procurement stages (set R). stage2,2, stage3,2, stage3,3, stage3,4,
stage3,5, stage4,1, stage4,2, stage4,3, stage4,4 belong to this set. The second set includes all the
intermediate stages with both upstream and downstream stages. They perform assembling
processes and are defined as assembly stages (set P). stage2,1 and stage3,1 belong to this set.
The third set includes only the most downstream stage which represents the demand/
delivery of end products, i.e., notebook computer B in this case, and is defined as demand
stage (set E ). stage1,10 is the only stage in this set.
between ATC elements. Since the formulations for the same set of stages are
mathematically alike, only one stage in each set will be selected for illustration.
Where:
AM1, 10 ðTCÞ: TC1,10 ¼ SCCC1,1 = 1,1 H1,1 þ SCCC0,1 = 0,1 H0,1 þ TC2,1 þ TC2,2
AM1, 10 ðST Þ: ST1,10 ¼ ST0,1
X
cð0,1Þ,SO0,1 yð0,1Þ,SO0,1 c0,1 ¼ 0 ð2Þ
SO0,1 2Sð0, 1Þ
X
Tð1,1Þ,SO1,1 yð1,1Þ,SO1,1 T1,1 ¼ 0 ð3Þ
SO1,1 2Sð1, 1Þ
International Journal of Production Research 6889
X
cð1,1Þ,SO1,1 yð1,1Þ,SO1,1 c1,1 ¼ 0 ð4Þ
SO1,1 2Sð1, 1Þ
X
1, if SO0,1 is selected
yð0,1Þ,SO0,1 ¼ and yð0,1Þ,SO0,1 ¼ 1 ð5Þ
0, otherwise SO0,1 2Sð0, 1Þ
X
1, if SO1,1 is selected
yð1,1Þ,SO1,1 ¼ and yð1,1Þ,SO1,1 ¼ 1 ð6Þ
0, otherwise SO1,1 2Sð1, 1Þ
2
TC2,1 TC32,1 "ð1,10 Þ, 1 ð9Þ
2
TC2,2 TC32,2 "ð1,10 Þ, 2 ð10Þ
2
CC2,1 CC32,1 "ð1,10 Þ, 3 ð11Þ
2
CC2,2 CC32,2 "ð1,10 Þ, 4 ð12Þ
n o2
STv1,1 max ST32,1 , ST32,2 "ð1,1Þ,5 ð13Þ
. Objective function: tTC and tST are two demand targets for total cost (TC) and
service time (ST). Therefore TC1,10 and ST1,10 are identified as two stage responses.
To achieve the minimal total cost and customer order fulfil time of the supply
chain, tTC and tST are normally set to zero. Weighting coefficients, wTC ST
1,10 and w1,10 ,
are attached to the deviation terms to indicate customer preferences. The specific
reason and usage of weighting coefficients will be discussed in Section 8.
Corresponding to the lower-level responses, five tolerance variables, "ð1,1Þ,1 , "ð1,1Þ,2 ,
"ð1,1Þ,3 , "ð1,1Þ,4 , and "ð1,1Þ,5 are used to keep upstream linking.
. Analysis model: AM1, 10 ðTCÞ and AM1, 10 ðSTÞ are the analysis models for TC1,10
and ST1,10 . In this problem, TC of a stage is defined as the cumulative unit SCC
cost (noted as SCCC) of the current stage and all its upstream stages. Therefore,
TC1,10 is the summation of the unit SCCC of stage1,10 and TC of stage2,1 and
stage2,2. The analysis model for ST1,10 is simply defined as ST of stage0,1. The
reason for using unit cost in the formulation is to ensure that TC has the same
value magnitude as other decision variables, e.g., ST and SO (supplier option).
6890 T. Qu et al.
. Constraints: Constraints (1) to (6) ensure that only one option will be selected in
each stage. Constraint (7) indicates that the maximal service time from upstream
stages will be used as the actual input service time, STv1,1 , because stage1,1’ cannot
start assembling until components from all the upstream stages become available.
Constraint (8) indicates that all the lower-level responses for TC and ST should
be positive values. The next five tolerance constraints mean that variations of all
the lower-level key links should be limited within corresponding tolerance
variables. Each tolerance variable is in turn constrained by a dynamic constraint
function, gð1,1Þ0 , i ðIÞ, i.e., Constraint (14). Along with the target cascading
iterations, I, the dynamic function will gradually tighten up the tolerable
deviation ranges. Details about the dynamic constraint function will be discussed
in Section 8.
Where:
X5
AM2, 1ðTCÞ: TC2,1 ¼ SCCC2,1 = 2,1 H2,1 þ TC3,k
k¼1
X
5
AM2, 1ðCCÞ: CC2,1 ¼ c2,1 þ CC3,k
k¼1
AM2, 1ðSTÞ: ST2,1 ¼ ST2,1
" ! #
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1 X
5
SCCC2,1 ¼ h2,1 CC2,1 2,1 2,1 STv2,1 þ T2,1 ST2,1 þ þ CC3,k 2,1 T2,1
2c2,1 k¼1
þ H2,1 c2,1 2,1 :
Subject to:
X
Tð2,1Þ,SO2,1 y2,1,SO2,1 T2,1 ¼ 0
SO2,1 2Sð2, 1Þ
International Journal of Production Research 6891
X
cð2,1Þ,SO2,1 yð2,1Þ,SO2,1 c2,1 ¼ 0
SO2,1 2Sð2, 1Þ
X
1, if SO2,1 is selected
yð2,1Þ,SO2,1 ¼ and yð2,1Þ,SO2,1 ¼ 1
0, otherwise SO2,1 2Sð2, 1Þ
TC3,k , CC3,k 0, k ¼ 1, 2, 3, 4, 5
2
4
TC3,k TC3,k "ð2,1Þ,k
2
CC3,k CC43,k "ð2,1Þ,kþ5 , k ¼ 1, 2, 3, 4, 5
n o2
4 4 4 4 4
STv2,1 max ST3,1 , ST3,2 , ST3,3 , ST3,4 , ST3,5 "ð2,1Þ,11
. Objective function: as has been explained, TC2,1 , CC2,1 and ST2,1 are identified as
responses in stage2,1 (ST2,1 is also a linking variable). The targets for these
responses, namely TC12,1 , CC12,1 and ST12,1 , come from the optimisation results
of stage1,10 . Totally 11 lower-lever responses have to be coordinated by the
corresponding tolerance variables.
. Analysis model: three responses in stage2,1 necessitate the existence of three
analysis models in this formulation. They are AM2,1(TC), AM2,1(CC) and
AM2,1(ST), whose outputs are TC2,1 , CC2,1 and ST2,1 , respectively. Note that
AM2,1(ST) is so simple that the input and output are both ST2,1 . This is because
ST2,1 acts as both local design variable and response.
. Constraints: the constraints in this formulation are formed in a similar way as that
in stage1,10 , except that the tolerance constraints increased to 11 to limit the
domain of variation of all the lower-level responses.
Subject to:
X
Tð3,2Þ,SO3,2 yð3,2Þ,SO3,2 T3,2 ¼ 0
SO3,2 2Sð3, 2Þ
X
cð3,2Þ,SO3,2 yð3,2Þ,SO3,2 c3,2 ¼ 0
SO3,2 2Sð3, 2Þ
X
1, if SO3,2 is selected
yð3,2Þ,SO3,2 ¼ and yð3,2Þ,SO3,2 ¼ 1
0, otherwise SO3,2 2Sð3, 2Þ
. Objective function: stage3,2 has no upstream stage, thus the tolerance variables are
omitted from the objective function, leaving only target deviation terms existing
to coordinate with its downstream stage.
. Analysis model: the analysis models in the bottom-level stage are very
straightforward. CC3,2 contains only the production cost incurred in this stage,
namely c3,2 . ST3,2 also serves as both local design variable and response.
. Constraints: as no lower-level response needs to be constrained, this part contains
only those constraints of SCC problem.
6. ATC solution
6.1 ATC solution strategies
The fourth step of applying ATC to SCC problems is to solve the formulated ATC
problems by adopting suitable solution strategies. Solution strategies for ATC could be
discussed from both macro and micro aspects. From the macro aspect, the strategy mainly
refers to the overall solution sequence and coordinating strategy of the ATC model; from
the micro aspect, it refers to the local optimisation method adopted by each ATC element.
The first macro issue is solution sequence. There are five types of convergent
sequence in implementing ATC (Michelena et al. 2003): (1) lower levels converge first;
(2) middle levels converge first then upward levels; (3) middle levels converge first then
downward levels; (4) upper levels converge first; and (5) lower levels and upper levels
International Journal of Production Research 6893
converge concurrently. In this paper, we choose the first type, namely ‘lower levels
converge first’ for the SCC problem. This sequence is most suitable for the problems that
have unattainable top-level targets, because earlier propagating the unattainableness to the
entire hierarchy could avoid unnecessary iterations. The SCC problem to be solved here is
of such a case. The second macro issue is the system coordination strategy, which aims at
coordinating all the elements in the partitioned system. For hierarchical problems under
convex assumption, the hierarchical overlapping coordination (HOC) has been proven for
its convergence, and therefore is selected to coordinate our ATC system of solving the SCC
problem. Interested readers may refer to Michelena et al. (2003) for details about
this issue.
ATC mainly defines a partitioning and coordinating mechanism for the overall system,
but has no restriction to the decision details of each partitioned element. The most suitable
optimisation method for each local problem is totally determined by the application
specifics. Therefore, while the macro issues for most ATC applications are all similar, the
micro issue of local optimisation methods varies significantly. Recently, the genetic
algorithm (GA) has been successfully used to solve SCC related problems (Ding et al.
2005, Huang et al. 2005a), thus it will also be adopted here to solve the individual
optimisation problems of stages.
As can be seen, Pe2,1 is the summation of 12 penalty terms each of which indicates the
extent that constraint is violated. Pe coi is the scaling weight to ensure that each penalty
term is in the comparable magnitude with the objective function value. In this way, the
better a solution is, the smaller Pe2,1 it will produce, and thus the bigger fitness value it
will have.
In implementing our GA program, free software named JGAP (Java Genetics
Algorithms Package) (JGAP Team 2003) was adopted. All the tests were executed on a PC
of 2.8 GHz, and the whole optimisation of ATC process takes 12–14 minutes.
Processing Production
Stage Stage name Option time (T ) cost (c)
TC22,2 ¼ 15 CC22,2 ¼ 15
stage2,1 TC33,1 ¼ 805 CC33,1 ¼ 805 3
STv2,1 ¼0
TC33,2 ¼ 300 CC33,2 ¼ 300
TC33,3 ¼ 200 CC33,3 ¼ 200
TC33,4 ¼ 225 CC33,4 ¼ 225
TC33,5 ¼ 40 CC33,5 ¼ 40
stage3,1 TC44,1 ¼ 30 CC44,1 ¼ 130 4
STv3,1 ¼0
TC44,2 ¼ 200 CC44,2 ¼ 200
TC44,3 ¼ 155 CC44,3 ¼ 155
TC44,4 ¼ 200 CC44,4 ¼ 200
6896 T. Qu et al.
Index Name SO ST SO ST
Stage SO ST TC SO ST TC
Local target setting facilitates the strategic balancing between global supply chain
objectives and suppliers’ local benefits and thus enhances the flexibility of suppliers’
participation.
This section will use a hypothetical scenario to test whether local target setting could be
effectively accommodated by ATC: assume the enterprises in stage3,1 and stage2,1 has an
internal agreement that the component supply adopts ‘zero-service-time’ mode. Therefore,
a local target term ðST3,1 0Þ2 will be added into the objective function of stage3,1. Table 4
contrasts the optimisation results with and without this local target. As can be seen from
the row with ‘*’, when the local target is set for stage3,1, its ST is reduced from 30 days to
zero days. That means, ATC could effectively accommodate local target setting. However,
the total supply chain cost has also been increased from $1866.82 to $1873.08. It is because
the decreased ST of stage3,1 brings extra average-on-hand (AOH) inventory cost to its
upstream stages. This cost increase ($1873.08 $1866.82 ¼ $6.26) seems not to be
significant because the calculation is on the basis of a single piece and the service time
reduction is relatively small (from 10 to zero days). Taking the annual demand rate
( ¼ 125 360) into consideration, the overall effect will be significant.
value of its parent element. Specifically, they are calculated through the following two
steps:
Step 1: After the local optimisation of element Pi,j , giving each lower-level response
variable an increment of one unit from its current optimum while keeping other
variables unchanged. Recording the discrepancy of objective function values by
this increment. Formally:
Dfi,j Riiþ1 ¼ fi,j Riiþ1 þ1 fi,j Riiþ1 , ð17Þ
where fi,j denotes the objective function in Pi,j , Riiþ1 denotes the vector of all its
lower-level responses, and hi denotes the th element of the vector.
Step 2: Based on the ratios of the discrepancies computed in Step 1, figure out a set of
weighting factors and cascade them to the child elements together with the
corresponding response targets. The computing equation is as follows:
R
i
wiþ1 1 : : wRiþ1 : ¼ Dfi,j Riþ1 1 : : Df i,j R i
iþ1 : ð18Þ
For each iteration of target cascading, a set of IWFs could be found for each element
through the above steps. In this way, responses having greater influence to the parent
element will be attached with the bigger IWFs and thus have less target deviations.
Step 4: With the two fitting points A and B, an exponential function could be shaped out:
8
( >
> ln AB
aIA þb
>
< a¼
e ¼A IB IA
) ð19Þ
e aðIA þiÞþb
¼B >
> ln AB
>
:b ¼ IA þ ln A
IB IA
This consistency scheme involves no derivative calculation and thus could be applied to
the problems with discrete variables.
Stage SO ST CC TC SO ST CC TC
(136.90 þ 205.58 136.41 204.75) due to the longer processing time. Knowing from
Table 5, TC response affects the objective function of stage3,1 nearly 10 times over CC
response. Hence, sacrificing $1.32 lower TC to only 5.6 (7.41/1.32) times lower CC is
obviously unreasonable.
For the service time (ST) terms, the bigger coefficients forces all the STs to decrease for
five days. However, these ST decreases will cause the increase of average-on-hand (AOH)
cost in all stages and consequently increase the TC values. As can be seen from the TC
column, the aggregated TC increases of all the upstream stages of stage1,10 is $2.37
(1782.69 þ 15.74 1780.34 15.72). Due to the similar reason as for CC, although five
days lower input service time to stage1,10 could bring lower TC increase in stage1,10 yet
sacrificing $2.37 lower TC to only 2.1 (5/2.37) times lower ST is also unreasonable.
From the above analysis, it could therefore be concluded that in the target cascading
process, ignorance of important target or over-emphasis of unimportant target will both
lead to a non-optimal solution. Therefore, it is very necessary to adopt IWF.
Table 7. Comparison of optimisation results with and without applying dynamic constraints.
targets and actual response values of service times respectively, and similarly for total cost
(TC). As the tolerances for CC are controlled in the same way as for TC, the discussion for
CC will be omitted.
The left part of the table shows the results from applying fixed tolerance constraints:
the tolerances for all lower-level TC and CC responses are set to $5.00, and those for
lower-level ST responses are set to be one day. The inconsistency between stages is
therefore unavoidable due to these fixed tolerances. As can be seen, TCU for all stages are
all nearly $5 lower than their actual response values, TC, and STU for stage3,3 and stage2,1
are one day lower than their actual ST response values (data with ‘z’ in ST and STU
columns). Owing to these deviations, a pseudo lower total supply chain cost of $1813.63 is
rendered, but is at the cost of sacrificing the consistency of the whole supply chain.
It is easy to understand the TC deviation in all stages, but why are there only two
stages that have ST deviations? This is because TC of each stage is required by its
downstream stage as responses, thus the downstream stage is able to squeeze each TC
target to the lower bound of tolerance. For ST, however, only the upstream stage with
maximum ST could be squeezed to the lower bound of tolerance, while other sibling stages
may still have room to satisfy the cascaded targets. Therefore, ST deviations do not exist in
all stages.
After the DC is applied, the deviation of each lower-level response could be reduced to
less than 103. Since all ST values are integer type and TC values are double-precision real
type, the deviations for them are virtually eliminated. As can be seen from the right part of
the table, all the targets and actual response values are consistent and the results are the
same as those obtained by the AIO approach.
9. Concluding remarks
This paper has applied analytical target cascading (ATC) to solve the configuration
problems of convergent assembly supply chains. A case study has been used to
demonstrate the five ATC steps, comprising (1) hierarchical modelling of a supply
chain; (2) identification of key links between supply chain stages; (3) ATC problem
formulation of individual stages; (4) solution of the overall ATC SCC problem; and (5)
analysis of ATC results.
The results verified the effectiveness of the ATC method in solving SCC problems.
Meanwhile, ATC is able to offer other features essential to SCC problems that AIO
methods cannot provide. For example, ATC models do not centralise decision variables of
individual enterprises involved in the same supply chain. Instead, they can retain their
relative privacy, their own decision models and their chosen solution methods, without
affecting each other, as long as linking decisions are properly communicated with each
other. Individual enterprises are able to locally set their own targets for their chosen
objectives, without affecting the effectiveness and efficiency of the optimisation process of
the entire SCC problems. An additional advantage is that decision support systems of
individual enterprises can be implemented and deployed as web services distributed on the
Internet. This configuration not only accommodates parallel computation thus improving
computational efficiency especially when the number of enterprises becomes large, but also
realises autonomous and collaborative problem solving.
This research also makes an important contribution to the ATC method when
unattainable targets are involved under the presence of discrete variables. A consistency
6904 T. Qu et al.
scheme has been proposed to guarantee the achievement of a consistent system design. It is
composed of two techniques. They are importance weighting factor (IWF) and dynamic
constraints (DCs). IWFs are applied to response target deviation terms where more
important targets receive heavier weightings. IWFs are cascaded together with the
corresponding targets to all child elements in the ATC process, enabling the top-level
targets to be fully supported by the whole system. DCs are applied to linking variable
deviation terms and tolerance variables. The system consistency is gradually controlled.
Constraints are relaxed in early iterations so that the whole ATC system reaches a near-
optimal state quickly. Tighter constraints are imposed towards late iterations to fine tune
the final system consistency.
This work is considered as an initial attempt in introducing ATC to solve SCC
problems. While potentials and advantages have been unravelled, further research is
essential for further progress in the field. One of the common decisions in SCC is supplier
selection, that is, to select and determine appropriate supplier enterprises among all the
alternatives for specific stages in the supply chain network. As a result, supplier selection
decisions are represented as alternative ‘OR’ branches in the ATC hierarchy. This has not
yet been addressed in the ATC literature. The second issue worth further investigation is
how to decompose and represent a hierarchical supply chain into an ATC tree.
The number of autonomous stages and the number of links between these stages define
the ATC complexity, and thus affect the computational efficiency. Comparative studies
should be conducted to test the performance of different decomposition strategies for
example, according to available computational resources or scope of autonomies of
enterprise decisions. Best decomposition strategies must be established to minimise the
ATC complexity and therefore maximise the efficiency for a specific SCC problem.
Finally, a computational framework should be established for integrating and coordinat-
ing distributed and autonomous enterprise resources in the supply chain. This framework
should not only take advantage of the latest service-oriented architecture (SOA) for
collaborative web services, but also incorporate ATC mechanisms for coordinating
distributed computational resources.
Acknowledgements
The authors are grateful to HKU Research Committee, Hong Kong SAR RGC and ITF (GHP/042/
07LP) for providing partial financial supports. Cluster de Recherche Gestion et organisation des
systèmes de production et del’innovation de la région Rhône-Alpes, GOSPI.
References
Alexandrov, N.M. and Lewis, R.M., 1999. Comparative properties of collaborative optimization and
other approaches to MDO. NASA/CR-1999–209354, ICASE technical report 99–24. NASA
Langley Research Center, Hampton, Virginia.
Allison, J., et al., 2005. On the use of analytical target cascading and collaborative optimization for
complex system design. In: Proceedings of the 6th world congress on structural and
multidisciplinary optimization, 30 May–3 June, Rio de Janeiro, Brazil, 1–10.
Allison, J., et al., 2006. Analytical target cascading in aircraft design. In: 44th AIAA Aerospace
Sciences Meeting and Exhibit, Reno, Nevada, 9–12 January 2006.
International Journal of Production Research 6905
Chan, F.T.S. and Chan, H.K., 2004. A new model for manufacturing supply chain networks: multi-
agent approach. Proceedings of the Institution of Mechanical Engineers Part B: Journal of
Engineering Manufacture, 218 (4), 443–454.
Chan, F.T.S., Swarnkar, R., and Tiwari, M.K., 2007. Infrastructure for co-ordination of multi-
agents in a network-based manufacturing system. International Journal of Advanced
Manufacturing Technology, 31 (9–10), 1028–1033.
Chan, H.K. and Chan, F.T.S., 2009. Effect of information sharing in supply chains with flexibility.
International Journal of Production Research, 47 (1), 213–232.
Chiang, D., Guo, R.S., and Chen, A., 2007. Optimal supply chain configurations in semiconductor
manufacturing. International Journal of Production Research, 45 (3), 631–651.
Choudhary, R., Malkawi, A., and Papalambros, P.Y., 2003. A hierarchical design optimization
framework for building performance analysis. In: Proceedings of the eight international
Building Performance Simulation Association (IBPSA) conference, 11–14 August, Eindhoven,
Netherlands, 11–14.
Choudhary, R., Malkawi, A., and Papalambros, P.Y., 2005. Analytic target cascading in simulation-
based building design. Automation in Construction, 14 (4), 551–568.
Ding, H.W., Benyoucef, L., and Xie, X.L., 2005. A simulation optimization methodology for
supplier selection. International Journal of Computer Integrated Manufacturing, 18 (2–3),
210–224.
Graves, S.C. and Willems, S.P., 2000. Optimizing strategic safety stock placement in supply chains.
Manufacturing and Service Operations Management, 2 (1), 68–83.
Graves, S.C. and Willems, S.P., 2001. Optimizing the supply chain configuration for new products.
Working paper. Leaders for Management Program and A.P., Sloan School of Management,
MIT.
Holland, J.H., 1975. Adaptation in natural and artificial systems. Ann Arbor, MI: University of
Michigan Press.
Huang, G.Q., Lau, S.K., and Mak, K.L., 2003. The impacts of sharing production information on
supply chain dynamics: a review of the literature. International Journal of Production Research,
41 (7), 1483–1517.
Huang, G.Q., Zhang, X.Y., and Liang, L., 2005a. Towards integrated optimal configuration of
platform products, manufacturing processes, and supply chains. Journal of Operations
Management, 23 (3–4), 267–290.
Huang, G.Q., et al., 2005b. Extensible multi-agent system for optimal design of complex systems
using analytical target cascading. International Journal of Advanced Manufacturing, 30 (9–10),
917–926.
Huang, G.Q. and Qu, T., 2008. Extending analytical target cascading for optimal configuration of
supply chains with alternative autonomous suppliers. International Journal of Production
Economics, 115, 39–54.
JGAP Team, 2003. Java Genetics Algorithms Package 1.0 release candidate (JGAP 1.0-RC1).
Available from: http://jgap.sourceforge.net/ [Accessed 11 May 2007].
Kim, H.M., et al., 2000. Target cascading in optimal system design. In: Proceedings of the 2000
ASME design automation conference, DAC-14265, 10–13 September, Baltimore, Maryland,
USA.
Kim, H.M., 2001. Target cascading in optimal system design. Dissertation (PhD). Department of
Mechanical Engineering, University of Michigan, Ann Arbor, Michigan, USA.
Kim, H.M., et al., 2001. Analytical target cascading in automotive vehicle design. In: Proceedings
of the 2001 ASME design automation conference, 9–12 September, Pittsburgh, Pennsylvania,
USA.
Kim, H.M., et al., 2003. Target cascading in optimal system design. Journal of Mechanical Design
(Transactions of the ASME), 125 (3), 474–480.
Kokkolaras, M., et al., 2002. Extension of the target cascading formulation to the design of product
families. Journal of Structural and Multidisciplinary Optimization, 24 (4), 293–301.
6906 T. Qu et al.
Li, Z.J., et al., 2004. Product tolerance allocation in compliant multistation assembly through
variation propagation and analytical target cascading. In: Proceedings of 2004 ASME
international mechanical engineering congress and exposition, 13–19 November, Anaheim,
California, USA.
Michalek, J.J., Feinberg, F.M., and Papalambros, P.Y., 2005. Linking marketing and engineering
product design decisions via analytical target cascading. Journal of Product Innovation
Management, 22 (1), 42–62.
Michalek, J.J. and Papalambros, P.Y., 2005. An efficient weighting update method to achieve
acceptable consistency deviation in analytical target cascading. Journal of Mechanical Design,
127 (2), 206–214.
Michelena, N.F. and Papalambros, P.Y., 1997. A hypergraph framework for optimal model-based
decomposition of design problems. Computational Optimization and Applications, 8 (2),
173–196.
Michelena, N.F., Kim, H.M., and Papalambros, P.Y., 1999. A system partitioning and optimization
approach to target cascading. In: Proceedings of the international conference on engineering
design (ICED), 24–26 August, Munich, Germany, 1109–1112.
Michelena, N.F., Park, H.A., and Papalambros, P.Y., 2003. Convergence properties of analytical
target cascading. AAIA Journal, 41 (5), 897–905.
Nyström, M., et al., 2003. Linking analytical target cascading to engineering information systems for
simulation-based optimal vehicle design. In: Proceedings of the international conference on
engineering design, 19–21 August, Stockholm, Sweden, 42–62.
Park, B.J., 2001. A framework for integrating product platform development with global supply chain
configuration. Dissertation (PhD). GIT, Gergia.
Qu, T., et al., 2008. Extending analytical target cascading for optimal supply chain network
configuration of a product family. International Journal of Computer Integrated
Manufacturing, doi: 10.1080/09511920802616807. [In press].
Qu, T., et al., 2009. A generic analytical target cascading optimization system for decentralized
supply chain configuration over supply chain grid. International Journal of Production
Economics, doi:10.1016/j.ijpe.2009.08.008. [In press].
Shukla, M., et al., 2009. Integrated model for the batch sequencing problem in a multi-stage supply
chain: an artificial immune system based approach. International Journal of Production
Research, 47 (4), 1015–1037.
Tosserams, S., et al., 2006. An augmented Lagrangian relaxation for analytical target cascading
using the alternating direction method of multipliers. Structural and Multidisciplinary
Optimization, 31 (3), 176–189.
Wong, H., Wikner, J., and Naim, M., 2009. Analysis of form postponement based on optimal
positioning of the differentiation point and stocking decisions. International Journal of
Production Research, 47 (5), 1201–1224.
Yu, Y.G., Chen, H.X., and Chu, F., 2008. A new model and hybrid approach for large scale
inventory routing problem. European Journal of Operational Research, 189 (3), 1022–1040.
Yu, Y.G., Chu, F., and Chen, H.X., 2009a. A Stackelberg game and its improvement in a VMI
system with a manufacturing vendor. European Journal of Operational Research, 192 (3),
929–948.
Yu, Y.G., Huang, G.Q., and Liang, L., 2009b. Stackelberg game theory model for optimizing
advertising, pricing and inventory policies in vendor managed inventory (VMI) supply chains.
Computers and Industrial Engineering, 57 (1), 368–382.
Zhang, X.Y., Huang, G.Q., and Rungtusanatham, M.J., 2008. Simultaneous configuration of
platform products and manufacturing supply chains. International Journal of Production
Research, 46 (21), 6137–6162.
International Journal of Production Research 6907
Appendix: Nomenclature
c production cost;
CCi,j cumulative cost of stagei,j;
f( ) objective function;
Fiti,j( ) fitness function in GA;
gi,j( ) inequality constraint for an optimisation problem;
gi,j(I ) dynamic constraint function;
hi,j( ) equality constraint for an optimisation problem;
H time interval of interest;
I cycle of target cascading;
li variable indicating whether the ith alternative is selected;
P0 original optimal design problem;
Pi,j jth sub-problem at the ith level of an ATC model;
Pei,j penalty factor;
Pe_coi scaling weight for penalty factor;
ri,j set of analysis models in element i,j;
R set of procurement stages;
Ri,j set of responses in element i,j;
Rii,j set of responses backtracked by element i,j;
SCCCi,j supply chain configuration cost generated in stagei,j;
Si,j set of supplier options in stagei,j;
SOi,j a supplier option in stagei,j;
SSi,j selected supplier in stagei,j;
STi,j service time by stagei,j;
stagei,j the jth stage in the ith tier of supply chain;
T production time;
TCi,j total cost by stagei,j;
x~ i,j set of local decision variables in element i,j;
yi,j set of linking variables in element i,j;
yii,j set of linking variables from element i,j;
service factor;
standard deviation of the demand;
"i,j vector of tolerance variables in Pi,j;
common review period for inventory;
mean of the demand rate;
wR i,j importance weighting factors for responses;
wyi,j importance weighting factors for linking variables;
k k22 squared l2-norm;
hi th element of a vector.
Copyright of International Journal of Production Research is the property of Taylor & Francis Ltd and its
content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's
express written permission. However, users may print, download, or email articles for individual use.