Professional Documents
Culture Documents
REFERENCES
Linked references are available on JSTOR for this article:
http://www.jstor.org/stable/2631911?seq=1&cid=pdf-reference#references_tab_contents
You may need to log in to JSTOR to access the linked references.
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted
digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about
JSTOR, please contact support@jstor.org.
INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Management Science
This content downloaded from 137.189.171.235 on Sat, 16 Jul 2016 01:35:15 UTC
All use subject to http://about.jstor.org/terms
MANAGEMENT SCIENCE
Vol. 35, No. 2, February 1989
Printed in U.S.A.
JOHN MILTENBURG
Faculty of Business, McMaster University, Hamilton, Ontario, Canada L8S 4M4
Mixed-model assembly lines are used to produce many different products without carrying
large inventories. The effective utilization of these lines requires that a schedule for assembling
the different products be determined. For Just-In-Time (JIT) production systems, which require
producing only the necessary products in the necessary quantities at the necessary times, the
objective is to keep a constant rate of usage of all parts used by the line. This is called levelling
or balancing the schedule. This paper develops a theoretical basis for scheduling these systems,
and presents new scheduling algorithms and heuristics.
(PRODUCTION/SCHEDULING-LINE BALANCING; INVENTORY/PRODUCTION-
JUST-IN-TIME; FACILITIES/ EQUIPMENT PLANNING)
1. Introduction
Many companies use mixed-model assembly lines for diversified small-lot production.
This helps them satisfy their customers' demands, for a variety of products, without
holding large inventories. Just-in-time (JIT) production systems, which require producing
only the necessary products in the necessary quantities at the necessary time, often use
mixed-model assembly lines. The effective utilization of these lines requires that the
following problems be solved (Okamura and Yamashina 1979);
1. Determination of line cycle times,
2. Determination of the number and sequence of stations on the line,
3. Line balancing,
4. Determination of the sequence schedule for producing different products on the line.
This paper deals with the fourth problem-the problem of scheduling the sequence
of products to be assembled by the line. The sequence will vary depending upon the
goals of the company. There are two possible goals;
1. Levelling the load (total assembly time) on each station on the line,
2. Keeping a constant rate of usage of every part used by the line.
Goal 1 recognizes that all products do not have the same operation time at each station
on the line. Some may even have operation times at certain stations that exceed the pre-
determined cycle time. The assembly line may be able to adjust to this without slowing
down or stopping. However, if products with relatively longer operation times are suc-
cessively scheduled, delays and line stoppages will result. Okamura and Yamashina (1979)
present a heuristic procedure for sequencing products (with different operation times)
on a mixed-model assembly line. Their objective is to minimize line stoppages. Other
work on this problem includes Kilbridge and Wester (1963), Thomopoulos (1967, 1970),
Macaskill (1972, 1973), and Dar-El and Cother (1975).
In this paper we consider the second goal. (In Miltenburg and Sinnamon 1986 we
show that both goals can be considered together.) Although both goals are important and
need to be considered for all mixed-model assembly lines, the second goal is considered
to be the more important goal for JIT production systems (Mondem 1983, p. 182). That
is, the quantity of each part used by the mixed-model assembly line per unit time should
* Accepted by David G. Dannenbring, former Departmental Editor; received August 12, 1986. This paper
has been with the author 2! months for 1 revision.
192
0025- 1909/89/3502/0 192$O0 1.25
Copyright ?i 1989, The Institute of Management Sciences
This content downloaded from 137.189.171.235 on Sat, 16 Jul 2016 01:35:15 UTC
All use subject to http://about.jstor.org/terms
MIXED-MODEL ASSEMBLY LINES IN JIT PRODUCTION SYSTEMS 193
be kept as constant as possible. There should be very little variability in the usage of each
part from one time period to the next. This is called levelling, or balancing the schedule
(Hall 1983, p. 123, Japanese Management Association 1985). Mondem (1983, pp. 182-
192) discusses this and shows how it is handled at Toyota. The purpose of this paper is
to develop a theoretical basis for scheduling mixed-model assembly lines in JIT production
systems, and to present a number of new algorithms and heuristics.
While very little has been published on this topic, there is a considerable literature on
scheduling multi-stage production systems. A good categorization of the literature is
included in Billington, McClain and Thomas (1983). Our paper is an extension of this
work, in that we develop a schedule for the final stage of a multi-stage production system.
The objectives of JIT require that variability in usage be minimized, rather than the
traditional objective of minimizing setup and holding costs. It is also a more detailed
schedule than is typical. For example suppose di,, is the demand for product i, i = 1, 2,
... , n, in period t. In traditional models the decision variables would be xi,, the amount
of product i to be produced in period t. Typical time periods are one week or one month.
Under JIT the decision variables are Si,k, k = 1, 2, . . , P, where Si,k = 1 if product i is
produced in stage k and 0 otherwise, and P is the number of stages per time period. In
theory, once this schedule is determined the schedules at all other levels in the JIT pro-
duction system will be fixed. In practice lot-sizing, capacity planning, and lead time
considerations are important and many of the techniques used in the literature can be
used. (See, for example, Zahorik, Thomas and Trigerio 1984.)
The paper is organized as follows. The scheduling problem is described in ?2. A math-
ematical model is formulated, and its properties are analyzed in ?3. New solution algo-
rithms and heuristics are presented in ?4. A property which is important for scheduling
large problems is discussed in ?5. ?6 describes other work that is being done in this area,
while ?7 gives a summary and suggests extensions.
Each product assembled on the mixed-model assembly line requires a variety of parts.
Often these parts vary from product to product. Scheduling large lots of each product
requires large lots of parts. When a part is only needed for certain products, its usage
will be high when those products are being assembled and will be low otherwise. This is
what we wish to avoid. JIT systems only work when there is a constant rate of usage for
all parts. (See, for example, Hall 1983 or Japanese Management Association 1985, for
a review of this and other JIT concepts.) To minimize the variation in the usage of each
part, products will be sequenced in very small lots. In this paper we assume that products
require approximately the same number and mix of parts. In this case we can achieve a
constant rate of part usage by considering only the demand rates for the products (and
ignoring the resulting part demand rates). The objective then is to schedule a constant
rate of production for each product. The results developed for this problem can be used
for the general case of products with differing parts requirements. (See ?6.)
3. Mathematical Model
n products with demands di, d2, . . ., d, totalling DT I 7=1 di units are to be pro-
duced. If ri = di/DT is the proportion of product i demand to the total demand, then
the objective is to schedule the assembly line so that the proportion of product i produced
(over a time period) to the total production is as close to ri as possible (for all time
periods). This is accomplished as follows.
Let Si,k, i = 1, 2, .. ., n, k = 1, 2,.. ., DT, where Si,k is either 0 or 1, be a production
schedule. If 5i,k = 1 then product i will be produced during stage k. z in=1 Si,k = 1,k for all
This content downloaded from 137.189.171.235 on Sat, 16 Jul 2016 01:35:15 UTC
All use subject to http://about.jstor.org/terms
194 JOHN MILTENBURG
k, because only one product can be produced during each stage. Let Xi,k = sj,j be
the total production of product i over stages 1 to k. Clearly Xi,k is a nonnegative integer
and I Xi,k = k, for all k. Let Si*,k (or xi*k) be the optimal schedule. The objective might
be one of the following.
DT n Xik/
minimize k- X= ( k r (1)
k=1I=
DT n
DT n Xi,k
minimize kz | _ ri- (3)
k=1 Ii= k(3
DT n
These objective functions seek to minimize the variation (either squared deviations or
absolute deviations) of the actual production from the desired production. Neither type
of variation is more tractable mathematically. Objective functions (1) and (3) try to keep
the actual proportions of the production mix (Xi,k/k) close to the desired proportions
(ri) at all times (k). Objective functions (2) and (4) try to keep the actual number of
units produced (xik) close to the desired number of units (kri) at all times.
Both of these objectives are reasonable. In fact they give similar schedules. For example
consider (1),
DT n \2 DT n
To schedule a product for stage k = j, we could decide to consider only the terms
7=l (1 /j2)(x1, - Jr1)2. The product i selected would minimize this quantity; which is
equivalent to minimizing I i (xi,j - jri)2, since j is a constant. Doing this for all stages,
would minimize kD=Tl _ =I (Xi,k - kri)2 which is objective function (2).
Let us now investigate a number of solution techniques for objective function (2).
Similar solution techniques can be developed for (1), (3) and (4).
DT n
If we consider only the first constraint, the problem reduces to minimizing a convex
function subject to a set of linear constraints. The solution is found by inspection:
Xi,k = kri. The objective function is equal to zero and the constraints are satisfied:
i= 1 Xi,k = i=1 kri = k in,= 1 ri = k. It is a relatively simple task to adjust this solution to
accommodate the last constraint. Define the point Xk = (X1,k, X2,k, ..., xn,k) E Rn
where Xi,k = kri, z i Xi,k = k, and R is the set of real numbers. Our problem is to find
the "nearest" integer point Mk = (MIl,k, M2,k, . .. Mn,k) E Zn to the point Xk
where zi7=1 Mi,k = k, Z is the set of nonnegative integers and "nearest" means
minimize - 7=1 (?nl,k - XI,k)2*
This content downloaded from 137.189.171.235 on Sat, 16 Jul 2016 01:35:15 UTC
All use subject to http://about.jstor.org/terms
MIXED-MODEL ASSEMBLY LINES IN JIT PRODUCTION SYSTEMS 195
Algorithm 1
The following algorithm finds the nearest integer point M = (mI, M2, , Mn) E Zn
to a point X = (xI, x2, . . ., xn) E Rn where m7= i = z =l xi = k.
1. Calculate k = ~1= I xi
2. Find the nearest nonnegative integer mi to each coordinate xi. That is, find mi so
that mi-xi < , i = 1,2, . . ., n.
3. Calculate km = 1=I mi.
4. (a) If k- km = 0 stop. The nearest integer point is M = (mI, M2,..., Mn).
(b) If k-km > O go to step 5.
(c) If k-km<Ogoto step 6.
5. Find the coordinate xi with the smallest mi - xi. Increment the value of this mi;
mi -- mi + 1. Go to step 3.
6. Find the coordinate xi with the largest mi - xi. Decrement the value of this mi;
mi -- i - 1. Go to step 3.
The optimal production over 5 stages is (2, 2, 1) while the optimal production over 6
stages is (3, 3, 0). During the sixth stage one unit of product 1 and one unit of product
2 must be produced while one unit of product 3 must be destroyed. Of course this is
impossible. Only one product can be assembled during a stage and products assembled
earlier cannot be destroyed. However, based on these results, we can develop a number
of "feasible" schedules.
This content downloaded from 137.189.171.235 on Sat, 16 Jul 2016 01:35:15 UTC
All use subject to http://about.jstor.org/terms
196 JOHN MILTENBURG
Of these feasible schedules, 3 or 5 is the best (with 6.2% more variation than the initial
infeasible schedule). We now present a number of algorithms for generating good feasible
schedules.
The solution to P 1 is a level or balanced schedule for the mixed-model JIT production
system problem. However the schedule may not be feasible. The following algorithm
ensures that a feasible schedule is found.
Algorithm 2
1. Solve problem P1 (using Algorithm 1), and determine whether the schedule is
feasible. (It is feasible ifMi,k - mi,k1I 2 0 for all i, k.) If the schedule is feasible, stop.
This is the optimal schedule. Otherwise go to step 2.
2. For the infeasible schedule determined in step 1, find the first (or next) stage 1
where mi - mij,-I < 0. Set a = number of products i, for which mi, - mij,-I < 0.
Reschedule stages 1 - a, 1 - a + 1, . . ., 1 + 1 by considering all possible sequences that
begin with the schedule for stage 1 - a - 1 and end with the schedule for stage 1 + 1.
3. Repeat step 2 for other stages where mi,k - mi,kl <0. Then stop.
Algorithm 3
1. Solve problem P1 (using Algorithm 1), and determine whether the schedule is
feasible. (It is feasible if mi,k - mi,k 1> 0 for all i, k.) If the schedule is feasible, stop.
This is the optimal schedule. Otherwise go to step 2.
2. For the infeasible schedule determined in step 1, find the first (or next) stage 1
where min,1- mi,1_ < 0. Set a = number of products i, for which mi,1 - mi,1_ < 0, and
This content downloaded from 137.189.171.235 on Sat, 16 Jul 2016 01:35:15 UTC
All use subject to http://about.jstor.org/terms
MIXED-MODEL ASSEMBLY LINES IN JIT PRODUCTION SYSTEMS 197
TABLE 1
Scheduling Example 2 Using Algorithm 2
Main Procedure
Algorithm I Rescheduling
1 1 0 0 1
2 1 10 2
3 2 1 0 1
4 220 2
5 2 2 1 3 2 2 1 3
6 330 1 2-3** 32 1 1
7 331 3 331 2
Stage 4 Schedule 2 2 0
Stage 7 Schedule 3 3 1
Possible
Sequences for Variation for
stages 5, 6, 7 stages 5, 6, 7
1 - 2 - 3 1.361
1 - 3 - 2 1.976
2- 1-3 1.361
2 - 3 - 1 1.976
3 - 1 - 2 1.822
3 - 2 - I 1.822
beginning at stage I-,a use Heuristic 1 or Heuristic 2 to schedule stages 1-a, I-'a + 1,
. . .1 I + co; where co > 0. 1 + c is the first stage where the schedule determined by the
heuristic matches the schedule determined in step 1.
3. Repeat step 2 for other stages where mik - mi,k1 < 0. Then stop.
Heuristic 1. For a stage k, schedule the product i with the lowest Xi,kI - kri.
The justification for Heuristic 1 is as follows. Consider a stage k. If product h is sched-
uled, the affected terms in the objective function of P1 for stage k are:
This content downloaded from 137.189.171.235 on Sat, 16 Jul 2016 01:35:15 UTC
All use subject to http://about.jstor.org/terms
198 JOHN MILTENBURG
TABLE 2
Nearest
Mk = (mI, M2 ..M7)
Stage Product Total
k Xk = (XI,.X2,= X7) Initial Final Scheduled Variation Variation
1 0.223 0.223 0.223 0.223 0.036 0.036 0.036 0 0 0 0 0 0 0 1 0 0 0 000 1 0.757 0.757
2 0.446 0.446 0.446 0.446 0.072 0.072 0.072 0 0 00 000 11 0 0 0 0 0 2 1.027 1.784
3 0.670 0.670 0.670 0.670 0.107 0.107 0.107 1 1 1 1 000 1 1 1 0000 3 0.810 2.594
4 0.893 0.893 0.893 0.893 0.143 0.143 0.143 1 1 1 1 000 1 1 1 1 000 4 0.107 2.701
5 1.116 1.116 1.116 1.116 0.179 0.179 0.179 1 1 1 1000 1 1 1 1 1 00 5 0.792 3.493
6 1.339 1.339 1.339 1.339 0.214 0.214 0.214 1 111 0 I 0 2 2 1 1 000 1 2 -5** 1.241 4.734
7 1.563 1.563 1.563 1.563 0.250 0.250 0.250 2222000 2 2 2 1 0 0 0 3 1.078 5.812
8 1.786 1.786 1.786 1.786 0.286 0.286 0.286 2 2 22000 2 22 2000 4 0.429 6.241
9 2.009 2.009 2.009 2.009 0.321 0.321 0.321 2222000 2 2 2 2 1 0 0 5 0.667 6.908
10 2.232 2.232 2.232 2.232 0.357 0.357 0.357 2 2 2 2 000 2 2 2 2 1 1 0 6 1.170 8.078
** Infeasible.
This is a myopic heuristic, in that it does not consider the effect of its current decision
on the variation in future stages.
EXAMPLE 3. There are n = 7 products with demands d, = d2= d3= d4= 25 and d5
= d6 = d7 = 4. Hence r1 = r2= r3= r4= 25/112 and r5 = = r7= 4/112.
The schedule in Table 2 was obtained by using Algorithmn 1. For each stage k, a
point Xk is calculated (where Xi,k = kri). The nearest integer point is calculated (Mk-
initial). If I 1=I mi # k the point is adjusted to give Mk-final. The schedule for stages 1
to 10 is 1-2-3-4-5-(1, 2, -5)-3-4-5-6 giving a total variation of 8.078. Unfortunately this
is not a feasible schedule. In stage 6 one unit of product 5 is destroyed (M5,6 - M5,5 = 0
- 1 < 0) while both product 1 and product 2 are scheduled.
Suppose we use Heuristic 1 to calculate an entire schedule. (See Table 3.) For each
stage k, Xi,kI - kri, i = 1, 2, . .. , n, is calculated and the product i with the smallest
value is scheduled. Ties are arbitrarily broken. (Note that xi,0 = 0, for all i.) The schedule
is 1-2-3-4-5-1-2-3-4-6 for 10 stages. The total variation is 9.953 which is considerably
more than Algorithm 1's schedule. This is not the optimal schedule.
In Table 4, Algorithm 3 with Heuristic 1 is used to determine a schedule. It begins by
using Algorithm 1. Stage 6 is found to be infeasible and so Heuristic 1 is used to reschedule
stages 5, 6, .. ., until a stage is reached where Mk matches the original schedule. As we
see in the second part of Table 4, stages 5 to 9 are scheduled with Heuristic 1. At stage
9, Mg from Heuristic 1 matches the original schedule-namely 2 2 2 2 1 0 0. The final
TABLE 3
This content downloaded from 137.189.171.235 on Sat, 16 Jul 2016 01:35:15 UTC
All use subject to http://about.jstor.org/terms
MIXED-MODEL ASSEMBLY LINES IN JIT PRODUCTION SYSTEMS 199
TABLE 4
Main Procedure-Algorithm 1
Algorithm 1 (from Table 2) Heuristic 1
Product
Stage Sched- Total
k X,k_-k l -2,k - kr2 I3,k - kr3 -X4,k - kr4 - ,k- kr5 -X6,k - kr6 -X7,k - kr7 uled Alk -Final Variation Variation
4 1111000 2.701
5 -0.116 -0.116 -0.116 -0.116 -0.176 -0.176 -0.176 5 1 1 1 11 0 0 0.792 3.493
6 -0.339 -0.339 -0.339 -0.339 0.786 -0.214 -0.214 1 2 1 111 0 0 1.491 4.984
7 0.437 -0.562 -0.562 -0.562 0.750 -0.250 -0.250 2 2 2 111 0 0 1.703 6.687
8 0.214 0.214 -0.786 -0.786 0.7 14 -0.286 -0.286 3 2 2 2 11 0 0 1.429 8.116
9 -0.009 -0.009 -0.009 -1.009 0.676 -0.321 -0.321 4 2 2 2 2 1 0 0 0.667 8.783
Schedule
Stage, k 1 2 3 4 5 6 7 8 9 10
Algorithm I 1 2 3 4 6
Heuristic 1 5 1 2 3 4
schedule is 1-2-3-4-5-1-2-3-4-6. Notice that this schedule is the same as the schedule
produced by Heuristic 1 alone (Table 3).
Unfortunately, this is not the optimal solution. The schedule 1-2-3-4-1-2-3-4-5-6, with
a total variation of 8.203, is much better.
The problem with Heuristic 1 is illustrated in Figure 1. The top frame corresponds to
stage k = 4, from Example 3. There are seven bars, one for each of the seven products,
of length Xi,k-1 - kri. If Xi,k-1 - kri is positive, the bar is above the line. Usually this
quantity is negative and the bar is below the line. The other frames depict stages 5 and
6. Heuristic 1 is a greedy heuristic, in that at each stage k it selects the product with the
lowest Xi,kI - kri (the largest bar below the line). It will schedule products 4, 5 and 1
for stages 4, 5 and 6 respectively. At k = 5 product 5 was scheduled because its bar was
lower than all other bars. (Actually the bars for products 5, 6 and 7 are same and we
arbitrarily selected product 5.) This gives the lowest variation for stage 5. However, there
is a dramatic effect on the next stage, k = 6. Although the bars are used for scheduling
purposes, they are also a rough gauge for the variation of a product's actual production
(Xi,k) from its ideal production (kri). Scheduling product 5 in stage 5 results in high
variability in stage 6 (as indicated by the large positive bar for product 5 in stage 6).
This content downloaded from 137.189.171.235 on Sat, 16 Jul 2016 01:35:15 UTC
All use subject to http://about.jstor.org/terms
200 JOHN MILTENBURG
k=4 F- FL L L
SCHEDULE PRODUCT 4
k=5J LiLL
SCHEDULE PRODUCT 5
k=6 }
SCHEDULE PRODUCT 1
Heuristic 1 is myopic-it does not consider the effect on future stages of its current
decision. Its great advantage is that it is a one-pass algorithm. It does one calculation for
each product and then makes a selection. The complexity is 0(n) for each stage. While
many people find this a satisfactory algorithm-Toyota uses a version of it (see ?6) -it
can result in much larger than necessary variations. Because of the myopic nature of
Heuristic 1 the following two-pass heuristic, of complexity 0(n2) for each stage, was
developed.
EXAMPLE 3 (continued). Table 5 shows how Heuristic 2 would be used with Algorithm
3. In the first section Algorithm 1 is used to calculate the k nearest integer points Mk.
Stage 6 is found to be infeasible. The algorithm backtracks to stage 5, and uses Heuristic
2 to schedule stages 5, 6, . . ., until a stage is reached where Mk matches the original
schedule. For this problem stages 5 and 6 are scheduled with Heuristic 2. At stage 6 the
schedule produced by Heuristic 2 matches the original schedule-namely 2 2 1 1 000O.
This content downloaded from 137.189.171.235 on Sat, 16 Jul 2016 01:35:15 UTC
All use subject to http://about.jstor.org/terms
MIXED-MODEL ASSEMBLY LINES IN JIT PRODUCTION SYSTEMS 201
TABLE 5
Main Procedure
Algorithm 1 (from Table 2) Heuristic 2
Stage k Stage k + 1
n n
Schedule
Stage k 1 2 3 4 5 6 7 8 9 10
Algorithm 1 12 3 4 3 4 5 6
Heuristic 1 1 2
The next part of the table shows the detailed calculations for Heuristic 2. First stage
k = 5 is considered. The first row shows that the variation in stage 5 when product 1 is
scheduled is V 1 = 0.917. Given this schedule for the current stage, product 2 should be
scheduled for the next stage because product 2 has the lowest Xi,k - (k + 1) ri. If product
2 is scheduled for stage 6, the variation in stage 6 is V2 = 1.241. Hence the variation
over stages 5 and 6, when product 1 is scheduled for stage 5, is approximately 0.917
This content downloaded from 137.189.171.235 on Sat, 16 Jul 2016 01:35:15 UTC
All use subject to http://about.jstor.org/terms
202 JOHN MILTENBURG
+ 1.241 = 2.158. When this procedure is repeated for each product we see that it is best
to schedule product 1 for stage 5 because V 1 + V2 is the lowest. (Actually any of
products 1 through 4 could be scheduled.)
Notice that if we are myopic when scheduling a product for stage 5, we will schedule
product 5 because the variability, 0.792 during stage 5, is lowest for product 5. However
scheduling product 1 now results in a better schedule in the next stage. Heuristic 2
suspects this and schedules product 1 because V1 + V2 = 2.158 for product 1 versus
2.283 for product 5. After stage 6 we stop using Heuristic 2 because the schedule produced
by the heuristic at stage 6 matches the initial schedule.
The schedule for the first 10 stages is 1-2-3-4-1-2-3-4-5-6 with a total variation of 8.203.
This is slightly higher than Algorithm I's infeasible schedule (8.078-a good lower bound),
and is much better than when Heuristic 1 was used-9.953. Two stages were rescheduled
with Heuristic 2 (approximately 2n2 calculations), compared to the five stages that were
rescheduled with Heuristic 1 (approximately 5n calculations).
We are now ready to schedule larger problems. Suppose there are n products, each
with a demand di, i = 1, 2, . . ., n. A total of DT = I= di units are to be produced,
during stages 1, 2, . .. , DT (since one unit is produced during each stage). That is, the
schedule is a sequence of length DT. For real problems, n > 5 and DT > 1000. Were it
not for an important property this would be quite a computational task. The property
is that the schedule is usually made up of a short sequence which is repeated again and
again. This is illustrated in the following example. A proof is given in Appendix 2.
EXAMPLE 4. There are n = 3 products with demands of 600, 600 and 100, for a total
of 1300 units. Algorithm 3, using either Heuristic 1 or Heuristic 2, produces the following
schedule for 1300 stages:
1-2-1-2-1-2-3-1-2-1-2-1-2-1-2-1-2-1-2-3-1-2-1-2-1-2-
1-2-1-2-1-2-3-1-2-1-2-1-2- * * * -2-3-1-2-1-2-1-2.
This schedule is simply the sequence 1-2-1-2-1-2-3-1-2-1-2-1-2 repeated 100 times. Notice
1 600 6 6/13
2 600 6 6/13
3 100 1 1/13
1300 13
Most good algorithms will produce a sequence of 13 stages (the sum of the "relative
demands rates") which will then be repeated 100 times.
The algorithms discussed in this paper schedule a product to a stage so that the total
variation is minimized. We have seen that the total variation increases as each successive
stage is scheduled. However our algorithms will always encounter a stage, say 1, where
no additional variation is produced. Usually / << DT. In Appendix 2 we prove that the
schedule for stages 1, 2, . . . , / will be repeated in stages / + 1, 1 + 2, . . ., 21 and again
for stages 21 + 1, 21 + 2, ..., 31 and so forth. In Example 4, 1 = 13, the sum of the
relative demand rates. This property significantly reduces the amount of computation
required to produce a schedule.
This content downloaded from 137.189.171.235 on Sat, 16 Jul 2016 01:35:15 UTC
All use subject to http://about.jstor.org/terms
MIXED-MODEL ASSEMBLY LINES IN JIT PRODUCTION SYSTEMS 203
In this paper we assumed that each product requires approximately the same number
and mix of parts. (This is often a reasonable assumption since these products are all
assembled on the same mixed-model assembly line.) Hence we achieve a constant rate
of usage of all parts used by the line (an objective of JIT production system) by considering
only the demand rates for the products, and ignoring the resulting part demand rates.
However, in many cases not all products have approximately the same number and
mix of parts. For these problems, the resulting part demand rates must be considered
when scheduling the mixed-model line. The results developed in the previous sections
can be used to determine schedules for these problems. In the general problem there are
n, different products (or final assemblies), p levels of subassemblies, with nj different
subassemblies at each level (j = 2, 3, . . . , p + 1). The objective is to keep a constant
rate of usage for products and subassemblies at all levels. Mondem (1983, pp. 182-192)
discusses a particular version of this problem-n products and one level of subassembly.
An appropriate objective function for his problem would be:
DT nI DT n2
7. Summary
Many companies use mixed-model assembly lines. The sequence of products to be
assembled will vary depending upon the objectives of the company. For Just-In-Time
production systems the most important objective is to keep a constant rate of usage of
every part used by the line, and the schedule selected is called a balanced or level schedule.
The problem to be solved is
DT n
We call P2 the "JIT usage problem" to differentiate it from other JIT problems. The
usage problem was studied in detail, and a number of good solution algorithms and
heuristics were developed. Similar results can be derived for the other objective functions
This content downloaded from 137.189.171.235 on Sat, 16 Jul 2016 01:35:15 UTC
All use subject to http://about.jstor.org/terms
204 JOHN MILTENBURG
discussed in ?2. When products have differing parts requirements are to be scheduled
the derived part demand rates must also be considered. The results developed in this
paper are used to develop schedules for these problems.'
' This research was supported in part by Grant A5474 from the Natural Sciences and Engineering Research
Council of Canada. I am grateful for the research assistance provided by Mr. Gordon Sinnamon, a Ph.D. student
at McMaster University.
Appendix I
THEOREM. ForpointsX = (x, X2...x) E Rn andM = (m,, n2,..,M)C- E Z, whereR is the set of
real numbers and Z is the set of nonnegative integers, such that k = zI" xi and k = miL, i the following
statements are equivalent.
1. M is the nearest integer point to X. (That is, I M - X I = X ,=, (i - x1)2 is minimized.)
2. If Pe Zn such thatJ L1=, pi = k then IM- XI c IP - X].
3. For all i, j, (mi-xi)-(mji-xj) 1.
4. For alli,j, I (mi-xi)-(mj-xj)I < 1.
5. There exists an a e R such that Jbr all i we have a - 1 < mi - xi < a.
6. There exists an a E [0, 1] such that for all i we have a - 1 ' mi - xi c a.
TI
2.
T1
6. ?t5.* 3. -2?4.
1. ? 2. 2 is just a restatement of 1.
4. 3. (mi - xi) - (mj - xj) < I (min- xi) - (mj - xj) I < 1.
3. 4. Either I (mi-xi)-(mji-xj) = (mi-xi)-(nj-xj) < I
or I(mi - xi) - (inj - xj) = (mj - xj) - (mi - xi) .
6. 5. Trivial since t0, 1] is a subset of R.
5. 6. Note that 2.., (mi - xi) = 0,
n
a - I < min (mi - xi) < - z (mi - xi) = 0,
ni=1
In
aE 2 max (mi - xi) 2-z (mi- xi) O .
ni=,
n n
iEF iEG
This content downloaded from 137.189.171.235 on Sat, 16 Jul 2016 01:35:15 UTC
All use subject to http://about.jstor.org/terms
MIXED-MODEL ASSEMBLY LINES IN JIT PRODUCTION SYSTEMS 205
COROLLARY. Let r (r1, r2 . rn) be the demand ratio vector, Xk = r and Xk.+ = (k + 1)r, where k
is an integer and k > 1. Suppose M is the nearest integer point to Xk and Ni the nearest integer point to XE,+l .
Then for each i, I mi - ni I < 1.
Hence for some] # i we have mj - nj < 0 and since they are integers mj - <-1. Now for this j,
=3
= [(ni - (ki + l)ri) - (nj - (k + 1)rj)] + [(imj - krj) - (mi -- kr)] + ri -r,
? I + I +r- rj
REMARK. We have proved that mi - ni I ? 1; and so, unfortunately the. case n, = mi - t can occur. Since
the assembly line does not destroy units it is not possible for any algorithm to produce a schedule that is always
at ideal production levels.
To give a simple example, let r = (6/ 13, 6/ 13, 1 / 13) and k = 5. The integer point M = (2. 2, 1) is nearest
to 5r, but the point (3, 3, 0) is nearest to 6r. To achieve the optimal sequence at stage k = 6, one unit of product
I and one unit of product 2 must be assembled and one unit of product 3 must be destroyed. This is impossible-
because only one product can be assembled during a stage k, and products previously assembled cannot be
destroyed.
For the time being we will ignore the above remark and use the theorem to justify Algorithm 1. The theorem
(specifically 1 3) and the following observations will be used:
This content downloaded from 137.189.171.235 on Sat, 16 Jul 2016 01:35:15 UTC
All use subject to http://about.jstor.org/terms
206 JOHN MILTENBURG
Appendix 2
The total number of units of product i produced over stages 1, 2, . , kI (under sequence Uk ) and stages kI
+ 1, ki + 2, . . ., kl + j (under sequence Tk2) is just the total number of units produced at stage ki under
sequence Ukl, plus the number of units produced during the first j stages under sequence Tk2.
Claim 1. If Dkl(k,) = 0 then Dkl+I(crkl + Tk2) = Dj(Tk2) forj < k2, for any sequence Tk2.
Conclusion
Suppose an algorithm produces a schedule for which Dkl (Ukl) = 0. According to Claim 1, all subsequent
variations are independent of Ukl, giving the same result as if Ukl had never been scheduled. Since the algorithm
depends on the variations alone for its choice of product, it will compose a schedule after Ukl just as it did at
the beginning. Hence the sequence akl will be repeated again and again.
References
DAR-EL, E. M. AND R. F. COTHER, "Assembly Line Sequencing For Model Mixing," Internat. J. Production
Res., 13 (1975), 463.
BILLINGTON, P. J., J. 0. MCCLAIN AND J. L. THOMAS, "Mathematical Programming Approaches To Capacity-
Constrained MRP Systems: Review, Formulation and Problem Reduction," Management Sci., 29, 10
(1983), 1126.
HALL, R. W., Zero Inventories, Dow Jones-Irwin, Homewood, IL, 1983.
Japanese Management Association, Kanban: Just-In-Time At Toyota, Productivity Press, Stamford, CT, 1985.
This content downloaded from 137.189.171.235 on Sat, 16 Jul 2016 01:35:15 UTC
All use subject to http://about.jstor.org/terms
MIXED-MODEL ASSEMBLY LINES IN JIT PRODUCTION SYSTEMS 207
KILBRIDGE, M. AND L. WESTER, "The Assembly Line Model-Mix Sequencing Problems," Proc. Third Internat.
Conf. Operations Research, Oslo, English Universities Press, Paris; 1963, 247.
MACASKILL, J. L., "Production-line Balances for Mixed-Model Lines," Management Sci., 19 (1972), 423.
, "Computer Simulation for Mixed-Model Production Lines," Management Sci., 20 (1973), 341.
MILTENBURG, G. J. AND G. SINNAMON, "Scheduling Mixed-Model Multi-Level Just-In-Time Production Sys-
tems," Internat. J. Production Res., to appear.
MONDEM, Y., Toyota Production System, Institute of Industrial Engineers Press, Norcross, GA, 1983.
OKAMURA, K. AND H. YAMASHINA, "A Heuristic Algorithm for the Assembly Line Model-Mix Sequencing
Problem to Minimize the Risk of Stopping the Conveyor," Internat. J. Production Res., 17 (1979),
233.
THOMOPOULOS, N. T., "Line Balancing-Sequencing for Mixed Model Assembly," Management Sci., 14, 2
(1967), 159.
, "Mixed-Model Line Balancing with Smoothed Station Assignments," Management Sci., 16 (1970),
593.
ZAHORIK, A., L. J. THOMAS AND W. W. TRIGEIRO, "Network Programming Models for Production Scheduling
in Multi-Stage, Multi-Item Capacitated Systems," Management Sci., 30, 3 (1984), 308.
This content downloaded from 137.189.171.235 on Sat, 16 Jul 2016 01:35:15 UTC
All use subject to http://about.jstor.org/terms