Professional Documents
Culture Documents
www.elsevier.com/locate/dsw
a
Department of Industrial Engineering, Chonbuk National University, Dukjin-gu, Chonju 561-756, South Korea
b
Department of Industrial Engineering, Korea Advanced Institute of Science and Technology, Yusong-Gu, Daejon 305-701, South Korea
Received 1 September 1998; accepted 1 February 1999
Abstract
This paper focuses on a production scheduling problem in a tree-structured assembly system operating on a make-to-
order basis. Due dates are considered as constraints in the problem, that is, tardiness is not allowed. The objective of the
problem is to minimize holding costs of ®nal product inventory as well as work-in-process inventory. A mixed integer
linear programming model is given and a branch and bound (B&B) algorithm based on a Lagrangian relaxation
method is developed. In the algorithm, a subgradient method is used for obtaining good lower bounds of subproblems
(or partial solutions) while a Lagrangian heuristic is used for upper bounds. Results of computational experiments on
randomly generated test problems showed that the B&B algorithm found optimal solutions of problems of moderate
sizes in a reasonable amount of computation time. Ó 2000 Elsevier Science B.V. All rights reserved.
0377-2217/00/$ - see front matter Ó 2000 Elsevier Science B.V. All rights reserved.
PII: S 0 3 7 7 - 2 2 1 7 ( 9 9 ) 0 0 1 0 8 - 3
M.-W. Park, Y.-D. Kim / European Journal of Operational Research 123 (2000) 504±518 505
This paper is organized as follows. The sched- Mk set of items to be processed on machine k
uling problem considered in this study is described F set of ®nal products
in more detail and a mathematical formulation for pi processing time (or workload) required to
the problem is given in the next section. A La- produce item i
grangian relaxation method and a Lagrangian di due date of item i (de®ned only if the item is
heuristic for the problem are given in Sections 3 a ®nal product)
and 4, respectively. Section 5 presents a B&B al- hi inventory holding cost of item i per unit per
gorithm, and performance of the algorithm is unit time
tested on a number of randomly generated test L a large positive number
problems and the test results are given in Sec-
tion 6. Decision variables
si starting time of item i
yij 0±1 variable that is equal to 1 if item i
2. Problem description and formulation precedes item j, and 0, otherwise
The assembly system considered in this study Now, a mathematical formulation for the
consists of machines in which assembly operations scheduling problem is presented.
are performed. For each machine, there may be one X
or more immediate upstream machines, but there
P minimize z hi
s/
i ÿ si
i62F
exists at most one immediate downstream machine X
in the system. Machines are indexed in such a way hi
di ÿ si
1
that a machine does not have a larger index than its i2F
sults of recent applications of the Lagrangian re- sj pj ÿ si 6 L yij for all i; j 2 Mk
i < j;
3
laxation method to various scheduling problems 0
can be found in Chen and Hsia (1997), Czerwinski yij f0; 1g for all i; j 2 Mk
i < j;
7
and Luh (1994), Hoogeveen and van de Velde
(1995, 1998), and Luh (1993) among others. ki P 0 for all i 2 Mk ;
11
In this study, problem (P) is relaxed by dual-
izing constraints (4) and (5) with Lagrangian si P lk for all i 2 Mk ;
12
multipliers ki P 0, i 1; . . . ; I. The resulting re-
laxed problem is si pi 6 uk for all i 2 Mk :
13
508 M.-W. Park, Y.-D. Kim / European Journal of Operational Research 123 (2000) 504±518
The second and third terms of the objective Since (DPk ) is solved for given ki values, wi s can be
function in (LRk ) are omitted in the decomposed given and these three sets are ®xed as well in (DPk ).
problems, since these are constants when k is Recall that it is assumed that (DPk ) is feasible in
given. If the optimal solution value of (DPk ) is the following proposition.
denoted
PK by LP k (k), then L(k) is P de®ned as L(k)
L
k1 k
k
e d
i i u
i k p
i i ÿ i2F ki di . In (DPk ), Proposition 1. Item i precedes item j (yij 1 and
constraints (12) and (13) are newly introduced to yji 0) in an optimal solution of (DPk ), if
consider the lower and upper bounds on starting wi =pi P wj =pj .
times of items to be produced on machine k. Here,
lk and uk denote a lower bound on the starting time Proof. The scheduling problem, (DPk ), can be
and an upper bound on the completion time of the transformed to a single machine scheduling prob-
items, respectively. Since each item should be lem with the objective of minimizing the weighted
completed before its item due date that is obtained sum of completion times of the jobs, if si is
by subtracting the sum of the processing times of replaced with ci ÿ pi + lk for all i. Here, ci
its successor items from the due date of its ®nal corresponds to the completion time of job i in
product, uk is set as uk maxi2Mk fdu
i ÿ
P the single machine
P problem with the objective of
j2U
i pj g in this study, where U
i is the set of all minimizing
P {wi c i + wi (lk ÿ pi )}. Here, the term
successors of item i including itself. Similarly, since wi (lk ÿ pi ) can be removed from the objective
the processing of each item can be started after all function, since it is a constant. We consider the
its predecessors are following four cases: (i) i; j 2 Mk ; (ii) i; j 2 Mkÿ ;
P completed, lk is set as lk
mini2Mk f minj2W
i
l2H
i;j pl ÿ pi g, where H
i; j (iii) i; j 2 Mk0 ; and (iv) i 2 Mk [ Mk0 and j 2 Mkÿ , or
is the set of all items on the path connecting items i i 2 Mk and j 2 Mkÿ [ Mk0 .
and j in the product structure, and W
i is the set of Case I (i; j 2 Mk ): Since the weighted sum of
predecessor items of item i that do not have any completion times is minimized by a weighted
predecessor. Instead of Eqs. (12) and P (13), tighter shortest processing time (WSPT) sequence (Smith,
constraints can be used, such as si P l2H
i;j P pl ÿ 1956), in which job i precedes job j if pi =wi 6 pj =wj
pi for all j 2 W
i and si pi 6 du
i ÿ j2U
i pj . (or equivalently wi =pi P wj =pj ), the proposition
Although tighter constraints may make solutions follows.
of (DPk ) be feasible for (P) more often, it is much Case II (i; j 2 Mkÿ ): Since wi is negative, the
more dicult to obtain an optimal solution of objective
P 0 function can be replaced with ``maximize
(DPk ) with such tighter constraints. Therefore, in w i ci '', where w0i ÿ wi . Since the weighted sum
this study, loose but easier-to-deal-with con- of completion times is maximized by a weighted
straints are used to save computation time for the longest processing time (WLPT) sequence, in
overall solution procedure. which job i precedes job j if pi =w0i P pj =w0j or
Note that (DPk ) can be considered as a single equivalently ÿwi =pi 6 ÿ wj =pj . Therefore, in an
machine scheduling problem in which items to be optimal sequence, i precedes j if wi =pi P wj =pj .
produced on machine k should be scheduled in the Case III (i; j 2 Mk0 ): The solution value is not
time interval lk ; uk . The following proposition can aected by the sequence of the items if wi wj 0.
be used to obtain Lk
k for given k, ifP(DPk ) is Case IV (i 2 Mk [ Mk0 , j 2 Mkÿ or i 2 Mk ,
feasible. Note that (DPk ) is feasible if i2Mk pi 6 j 2 Mk0 [ Mkÿ ): In this case, either wi is positive and
uk ÿ lk . Let wi denote the cost P coecient of item i wj is not positive, or wi is nonnegative and wj is
in (DPk ), i.e., wi ki ÿ ei ÿ j2K
i kj , and let negative. In both cases, the objective function
value is smaller when item i precedes item j.
Mk fi: wi > 0 and i 2 Mk g;
In this study, a sequence in which item i pre-
Mk0 fi: wi 0 and i 2 Mk g; and cedes item j if wi =pi P wj =pj for all pairs of two
items i and j is referred to as a generalized
Mkÿ fi: wi < 0 and i 2 Mk g: weighted shortest processing time (GWSPT) se-
M.-W. Park, Y.-D. Kim / European Journal of Operational Research 123 (2000) 504±518 509
quence, since it is similar to WSPT sequence but Initially, the value for ln is set to be equal to 2
can be applied to a general case with a non-posi- (i.e., l0 2), and ln is halved when the solution of
tive weights. The above proposition leads us to the (LRk ) has not increased for a given number of
following property that can be used to ®nd an iterations, f. It is reported by Fisher (1981) that
optimal solution of (DPk ), or an optimal sequence this rule performs well empirically, even though
of items in the scheduling problem (DPk ). Once an convergence to the optimal solution is not guar-
optimal solution of (DPk ) is obtained by the fol- anteed. The initial multipliers, k0i s are set to zero
lowing corollary, the optimal solution value of in this study. The subgradient method is termi-
(LRk ) for given k can be computed easily. nated if the iteration count reaches a predeter-
mined maximum count, x, or if a relative error,
Corollary 1. The scheduling problem (DPk ) is
z ÿ LB=LB, becomes smaller than a predeter-
optimized Pif items in Mk are scheduled between lk mined error limit, e, where LB is the best lower
and lk i2M Mk0 are scheduled
Pk pi , items in P bound obtained so far. In this study, the three
between lk i2M pi and uk ÿ i2M ÿ pi , and
P items parameters, e, f and x were set to 0.0001, 10 and
k k
in Mkÿ are scheduled between uk and uk ÿ i2M ÿ pi , 1000, respectively, after a series of preliminary
k
all according to a GWSPT sequence. tests.
Once yij values are obtained from the nth iter-
Proof. It is obvious that it is better to schedule ation of the subgradient method, starting times of
items in Mk earlier, since their cost coecients are the items can be computed easily from constraints
positive. Similarly, items in Mkÿ should be sched- (2)±(6) of the original problem, (P). However,
uled as late as possible, since their cost coecients since {yij }n is a solution of the relaxed problem
are negative. On the other hand, since cost (LRk ) for given kn , it may be infeasible to (P). In
coecients of items in Mk0 are zero, the solution other words, for certain Lagrangian vectors gen-
value is not aected by the sequence of these items. erated during the subgradient optimization rou-
This, with Proposition 1, proves the optimality of tine, resulting {yij }n values do not satisfy
the schedule constructed as speci®ed. constraints of (P). In the suggested B&B algo-
rithm, feasible solutions are not sought for all
As mentioned earlier, to solve (PL) or to ®nd Lagrangian vectors kn . Such infeasible Lagrangian
best k, a subgradient method is used in this study. vectors are disregarded since it is possible to ®nd
In the subgradient method, a sequence of La- feasible solutions from various other Lagrangian
grangian multiplier vectors {kn } is generated as vectors during the subgradient optimization
follows: starting from given initial values k0i , let routine.
In the suggested algorithm, the starting times of
kin1 the items for given {yij }n values are determined as
(
maxf0; kni tn
sni pi ÿ di g for all i 2 F ; follows:
8
maxf0; kni tn
sni pi ÿ sn/
i g for all i 62 F ;
>
>
> min s/
i ; min sj ÿ pi if i 62 F ;
<
14 j2C
i
si
16
where
sn1 ; sn2 ; . . . ; snI is an optimal solution to >
>
>
: min di ; min sj ÿ pi if i 2 F ;
LRkn and tn is a positive scalar step size de®ned as j2C
i
ln
z ÿ L
kn where C
i fj: yij 1g is the set of items that
tn P 2 P 2
:
n
i2F
si pi ÿ di
n n
i62F
si pi ÿ s/
i
are scheduled after item i on the (same) machine
for the item. Since the objective function, (8), of
15
problem (P) is a non-increasing function of start-
Here, ln is a scalar satisfying 0 < ln 6 2 and z is ing times of items, larger values for si give a better
an upper bound on the optimal solution value of or at least an equal solution value. Note that the
(PL). starting times of items satisfying Eq. (16) are the
510 M.-W. Park, Y.-D. Kim / European Journal of Operational Research 123 (2000) 504±518
largest values that the items can have without a new sequence by moving item l to the last po-
violating due dates and precedence constraints, sition in qk , modify values of yij s according to the
although si < 0 for some i. new qk , and obtain C(i) and si using yij . Let
l l ÿ 1 and go to step 2.
Step 4. For h |Mk | ÿ 1 down to 1 do:
4. Lagrangian heuristic If sh ÿ
shÿ1 phÿ1 P pl and
shÿ1
phÿ1 pl 6 dl , where l is the index of the
It is well known that solutions obtained from lth item in qk , obtain a new sequence by moving
Lagrangian problems, such as (PL), are rarely item l to the hth position in qk . Modify values
feasible for the original problem, i.e., (P). Al- of yij s according to the new qk , and obtain C(i)
though these solutions can be used as lower and si using new yij values.
bounds in the B&B algorithm suggested in this Let l l ÿ 1. Go to step 2.
paper, the algorithm needs a good upper bound
(feasible solution) as well. To make solutions of In step 1, an initial schedule is generated ac-
such relaxed problems feasible, a heuristic called cording to Eq. (16). This initial schedule is de-
the Lagrangian heuristic is often used (Fisher, composed into K schedules for K machines, and
1981). they are improved sequentially from the schedule
The Lagrangian heuristic suggested in this re- for machine 1 to the schedule for machine K in
search tries to make infeasible solutions obtained steps 3 and 4. In step 3, an item is moved to the last
by the subgradient method feasible by reschedul- position of the current sequence qk if such a move
ing items in the solutions. That is, using yij values does not make the item tardy. On the other hand,
obtained by the subgradient method, the Lagran- in step 4, an item is moved to a later position (not
gian heuristic ®rst generates an initial schedule, to the last position) if it can be moved into an idle-
i.e., starting times of items, and then improves the time slot that is inserted between items to be pro-
schedule if possible. As discussed earlier, yij values cessed later in qk .
obtained from (DPk ) may be infeasible. Therefore,
in the Lagrangian heuristic, the initial schedule is
improved by changing the assembly sequence. 5. B&B algorithm
Also, the starting time of each item is delayed as
much as possible without making any item tardy In this section, the B&B algorithm suggested in
or violating precedence constraints since the solu- this study is described. In the scheduling problem
tion value decreases as starting times are increased. under consideration, a sequence of items should
The procedure for the heuristic is summarized in be determined for each machine. Let qk be a full
the following. sequence on machine k and rk be a partial se-
quence on machine k (if the current sequence is
Procedure (The Lagrangian heuristic) not a full sequence), which is a sequence of a
Step 0. Find a sequence qk of items on machine subset of items included in Mk that are to be
k from given yij values for all k. Let k 1 placed at the end in qk . Each node in the B&B tree
Step 1. If k > K, stop. Otherwise, let C
i be the corresponds to a subproblem, which is de®ned by
set of items to be scheduled after item i, i.e.,
q1 ; . . . ; qkÿ1 ; rk , i.e., a set of full sequences on
C
i fj: yij 1 or yji 0g, for all i 2 Mk . Let machines 1; 2; . . . ; k ÿ 1, and a partial sequence on
di s/
i for all i 2 Mk n F , si minfdi ; machine k.
minj2C
i sj g ÿ pi for all i 2 Mk , and let l |Mk | Branching: When a node branches, one or more
ÿ 1, where || is the number of elements in set . nodes are de®ned by adding one more item in front
Step 2. If l 6 0, let k k + 1, and go to step 1. of the partial sequence on machine k associated
Otherwise, go to step 3. with the node branching now. If the current node
Step 3. If dl < djMkj pl , where l is the index corresponds to a full sequence, that is, if all items
of the lth item in qk , go to step 4. Otherwise, obtain to be processed on machine k are scheduled, a new
M.-W. Park, Y.-D. Kim / European Journal of Operational Research 123 (2000) 504±518 511
partial sequence is formed for machine k 1 in the F 0
F [ fij/
i 2 Sg \ U . For items i 2 F 0 , due
new node. To select a node to branch, the depth- dates (di 0 ) in (P0 ) are de®ned as
®rst (newest active node) rule is employed in the
minfsmin ; s/
i ; di g if i 2 Mk ;
algorithm. In this rule, a node with the most items di0
17
minfs/
i ; di g if i 62 Mk ;
in the corresponding partial and full sequences is
selected for branching. In case of ties, a node with where smin denotes the minimum starting time of
the minimum lower bound is selected. items included in rk . Therefore, u(i), the index of
Computing bounds and fathoming: For each of the ®nal product of item i is accordingly changed
the nodes generated in the algorithm, a new sub- to u0 (i) for all i 2 U in (P0 ), i.e.,
problem, denoted by (P0 ), is constructed. Set S of
u
i if u
i 2 U ;
items included in q1 ; . . . ; qkÿ1 , and rk , are excluded u0
i
18
j if u
i 62 U ;
from (P0 ), since the sequence of these items are ®xed
and their optimal starting times can be calculated where j denotes the index of the item in F 0 which is
with Eq. (16). Lower and upper bounds corre- on the path from i to u(i) in the product structure
sponding to a node are obtained by solving (P0 ) and of which the immediate successor is included
using the Lagrangian relaxation method and the in S. Also, in (P0 ), the set Mk0 of items to be pro-
Lagrangian heuristic. The Lagrangian heuristic cessed on machine k0 is set to Mk0 \S, for
given in the previous section is used to ®nd a good k 0 k; k 1; . . . ; K.
initial upper bound at the root node. At the other Therefore, the lower bound (LB) and upper
nodes, the heuristic is applied once for each node bound (UB) for the node can be calculated as
using the ®nal yij values resulting from the subgra- LB LB0 + Q and UB UB0 + Q, where LB0 and
dient method in order to save the computation time. UB0 are the lower bound obtained from the sub-
Let the set of items considered in (P0 ) be U, i.e., gradient method and the upper bound from the
U f1; . . . ; Ig n S. Then, the set F 0 of ®nal prod- Lagrangian to (P0 ), respectively,
P heuristic appliedP
ucts considered in (P0 ) is de®ned as a collection of and Q i2A ei
du
i ÿ si i2U ei
du
i ÿ du
i :
items in U which are elements of F or of which the A dangling node is deleted from further con-
immediate successors are elements of S, i.e., sideration (fathomed) if the lower bound for the
Fig. 1. System con®gurations used in the test problems (squares denote machines).
512 M.-W. Park, Y.-D. Kim / European Journal of Operational Research 123 (2000) 504±518
Fig. 2. Product structures used in the test problems (numbers denote machines in which the corresponding items are processed, not
item indices).
M.-W. Park, Y.-D. Kim / European Journal of Operational Research 123 (2000) 504±518 513
P
node is greater than or equal to the incumbent times of the items, i.e., i2Mk pi > uk ÿ lk for
solution, i.e., the smallest upper bound of all some k.
nodes generated so far, or if the subproblem (P0 )
associated with the node is found to be infeasi-
ble. Note that a subproblem cannot be feasible if 6. Computational experiments
the sum of processing times of items to be pro-
cessed on any one machine is greater than the To test the performance of the B&B algorithm,
dierence of upper and lower bounds on starting computational experiments were done on ran-
Table 1
Test results for three-machine problems
System Number of r Number Lagrangian heuristic B&B algorithm
orders of items PE CPU nN CPU
3 0.0 13 0.00 0.66 10 1.81
0.1 13 0.00 0.72 7 1.32
0.2 13 0.00 0.71 8 1.48
0.3 13 0.00 0.60 4 0.60
0.4 13 2.57 0.71 11 2.08
0.5 13 0.00 0.71 7 1.26
domly generated problem instances. For the test, the test problems are illustrated in Fig. 1. Product
144 problems were generated, one for each of all structure of each product was randomly selected
combinations of three levels for the number of from the product structures illustrated in Fig. 2.
orders or products (3, 5 and 7), eight cases of In the problems, processing time for each item
system con®gurations, and six levels for the rela- was randomly generated from DU1; 100, the
tive range of due dates (0.0, 0.1, 0.2, 0.3, 0.4 and discrete uniform distribution with the range of
0.5). Con®gurations of assembly systems used in 1±100. Due dates of orders were generated using a
Table 2
Test results for four-machine problems
System Number of r Number Lagrangian heuristic B&B algorithm
orders of items PE CPU nN CPU
3 0.0 19 0.00 1.32 16 5.28
0.1 18 0.36 1.27 10 2.97
0.2 19 0.00 1.37 8 2.69
0.3 16 0.00 1.05 7 1.76
0.4 19 1.40 1.38 7 2.31
0.5 17 0.00 1.21 8 2.36
Table 3
Test results for ®ve-machine problems
System Number of r Number Lagrangian heuristic B&B algorithm
orders of items PE CPU nN CPU
3 0.0 24 0.00 1.93 8 3.75
0.1 25 2.40 2.09 8 4.18
0.2 26 0.00 2.20 7 3.57
0.3 25 0.37 2.03 8 3.95
0.4 26 0.00 2.20 7 3.57
0.5 26 2.71 2.20 8 4.18
5 5 0.0 46 0.00 5.11 79 68.77
0.1 41 0.00 4.45 119 89.53
0.2 43 3.70 4.72 148 117.48
0.3 43 8.13 4.78 74 50.37
0.4 45 0.71 5.11 31 20.82
0.5 47 3.57 5.44 28 23.02
7 0.0 64 2.97 7.20 5472 7713.46
0.1 64 3.03 7.36 6532 9192.77
0.2 64 0.00 7.36 283 352.02
0.3 61 6.14 6.86 1477 1824.24
0.4 64 4.41 7.36 196 240.73
0.5 64 4.39 7.19 556 665.36
parameter called the relative range of due dates, r. immediate predecessors of item i and u a random
The due dates were randomly generated from number generated from the uniform distribution
DUP
1 ÿ r=2; P
1 r=2, where P is the sum of with the range of 1±2.
processing times of all items. The inventory hold- Tables 1±4 show results of the test. In the ta-
ing cost of an item was set in such a way that it was bles, PE denotes the percentage error which is
greater than the sum of the inventory holding costs calculated as 100 ´ (UB0 ÿ z )/z , where z is the
of its immediate predecessors. Here, hi was set to
P optimal solution value and UB0 is the upper
j2K
i hj u, for all i, where K
i is the set of bound obtained by the Lagrangian heuristic at
516 M.-W. Park, Y.-D. Kim / European Journal of Operational Research 123 (2000) 504±518
Table 4
Test results for six-machine problems
System Number of r Number Lagrangian heuristic B&B algorithm
orders of items PE CPU nN CPU
3 0.0 26 0.00 1.64 8 3.67
0.1 26 2.44 1.76 8 3.80
0.2 26 0.00 1.65 7 3.03
0.3 26 0.00 1.75 7 3.17
0.4 26 0.00 1.70 8 3.73
0.5 26 0.00 1.76 7 3.08
7 5 0.0 44 1.81 3.79 150 120.19
0.1 43 2.08 3.79 68 54.16
0.2 43 2.73 3.73 70 56.18
0.3 44 0.46 3.84 39 24.11
0.4 43 1.66 3.73 78 51.79
0.5 43 2.49 3.79 19 11.98
7 0.0 60 2.75 6.42 2582 3266.90
0.1 61 2.58 8.35 1602 2082.51
0.2 61 4.37 8.40 246 296.59
0.3 60 3.55 8.13 187 217.12
0.4 61 3.72 8.41 224 205.87
0.5 61 2.12 8.35 273 234.70
the root node, i.e., the solution of Lagrangian It can be seen from the tables that the number
heuristic applied to (P). Also, nN denotes the of orders has more impact on the diculty (in
number of nodes generated in the B&B algorithm, terms of nN and CPU time) of the problem than
and CPU denotes computation time (in seconds) the number of machines. This may be because the
required for each test problem. The suggested diculty is aected by the number of items to be
algorithms were coded in Pascal and run on a scheduled. Note that the number of items to be
personal computer with a Pentium processor (150 produced on each machine increases quickly as the
MHz). number of orders increases. Although it was ob-
M.-W. Park, Y.-D. Kim / European Journal of Operational Research 123 (2000) 504±518 517
served that the Lagrangian relaxation method gave an objective, such as a problem to minimize the
better lower bounds when larger values for f and x sum of tardiness penalties of ®nal products and
were selected for the subgradient method, the inventory holding costs.
values are set as they were (f 10 and x 1000) to
reduce computation time for the whole branch and
bound algorithm. With larger value for f and x, References
less nodes were generated but it took much longer
time to compute a lower bound for each node. Chen, T.R., Hsia, T.C., 1997. Scheduling for IC sort and test
Also, it can be seen from the tables that CPU facilities with precedence constraints via Lagrangian
relaxation. Journal of Manufacturing Systems 16,
times for the B&B algorithm increase as the range
117±128.
of due dates (r) decreases. This may be because it is Clark, A.J., Scarf, H., 1960. Optimal policies for a multi-
harder to solve a problem with a smaller r since echelon inventory problem. Management Science 6, 475±
dierences among lower bounds on nodes in the 490.
B&B tree are smaller, and hence the B&B algo- Czerwinski, C.S., Luh, P.B., 1994. Scheduling products with bill
of materials using an improved Lagrangian relaxation
rithm cannot fathom many nodes at an earlier
technique. IEEE Transactions on Robotics and Automation
stage of the algorithm. The most dicult problem 10, 99±111.
among the tested was solved within 2.6 hours. Fisher, M.L., 1981. The Lagrangian relaxation method for
Therefore, it can be argued that the suggested solving integer programming problem. Management Science
B&B algorithm can be used to ®nd optimal solu- 27, 1±18.
Fisher, M.L., 1985. An applications oriented guide to Lagran-
tions for problems of moderate or practical sizes.
gian relaxation. Interfaces 15, 10±21.
The Lagrangian heuristic also seems to work well Georion, A.M., 1974. Lagrangian relaxation for integer
since heuristic solutions obtained at the root nodes programming. Mathematical Programming Study 2, 82±
were not far from optima. The overall average 114.
percentage error was 2.9%, although performance Hastings, N.A.J., Yeh, C.-H., 1990. Job oriented production
scheduling. European Journal of Operational Research 47,
of the heuristic seems to deteriorate as the number
35±48.
of orders increases. Hariri, A.M.A., Potts, C.N., 1997. A branch and bound
algorithm for the two-stage assembly scheduling problem.
European Journal of Operational Research 103, 547±556.
7. Concluding remarks Hoogeveen, J.A., van de Velde, S.L., 1995. Stronger Lagran-
gian bounds by use of slack variables: Applications to
machine scheduling problems. Mathematical Programming
In this paper, we considered a production 70, 173±190.
scheduling problem in a tree-structured assembly Hoogeveen, H., van de Velde, S., 1998. Scheduling by
system operating on a make-to-order basis. The positional completion times: Analysis of a two-stage ¯ow
objective of the problem is to minimize holding shop problem with batching machine. Mathematical Pro-
costs of items under the due date and precedence gramming 82, 273±289.
Kim, Y.-D., 1987. On the superiority of a backward approach
constraints. A B&B algorithm was developed to in list scheduling algorithms for multi-machine makespan
®nd an optimal solution of the problem. In the problems. International Journal of Production Research 25,
B&B algorithm, Lagrangian relaxation methods 1751±1759.
are used to obtain lower and upper bounds. Re- Kim, Y.-D., 1995. A backward approach in list scheduling
sults of tests on the performance show that the algorithms for multi-machine tardiness problems. Comput-
ers and Operations Research 22, 307±319.
suggested algorithm can ®nd optimal solutions of Lalsare, P., Sen, S., 1995. Evaluating backward scheduling and
problems of moderate sizes in a reasonable sequencing rules for an assembly shop environment. Pro-
amount of time. In this study, it is assumed that duction and Inventory Management Journal 36, 71±78.
tardiness is not allowed. In many real circum- Lee, C.Y., Cheng, T.C.E., Lin, B.M.T., 1993. Minimizing the
stances, however, it is impossible to meet given due makespan in the 3-machine assembly type ¯owshop sched-
uling problem. Management Science 39, 616±625.
dates. Therefore, it may be necessary to study a Luh, P.B., 1993. Scheduling of manufacturing systems using the
production scheduling problem in which meeting Lagrangian relaxation technique. IEEE Transactions on
due dates is not considered as a constraint but as Automatic Control 38, 1066±1079.
518 M.-W. Park, Y.-D. Kim / European Journal of Operational Research 123 (2000) 504±518
Park, M.-W., Kim, Y.-D., 1999. A heuristic algorithm for a assembly scheduling problem: Complexity and approxima-
production scheduling problem in an assembly system, tion. Operations Research 43, 346±355.
Journal of the Operational Research Society 50, 138±147. Smith, W.L., 1956. Various optimizers for single stage produc-
Potts, C.N., Sevast'janov, S.V., Strusevich, V.A., Van Was- tion. Naval Research Logistics Quarterly 3, 59±66.
senhove, L.N., Zwaneveld, C.M., 1995. The two-stage