Computers and Structures 79 (2001) 1605±1616

www.elsevier.com/locate/compstruc

An application of genetic algorithms to precast

production scheduling
Weng-Tat Chan *, Hao Hu
Department of Civil Engineering, National University of Singapore, 10 Kent Ridge Crescent, Singapore S119260, Singapore
Received 11 October 1999; accepted 22 February 2001

Abstract
There are two alternatives for production organisation in precast factories, namely the comprehensive method and
the specialised method. Production scheduling under the specialised alternative has been found to be a dicult opti-
misation problem if heterogeneous elements are involved. A ¯ow shop sequencing model is developed for this kind of
production scheduling that considers the constraints encountered in actual practice. The model is optimised using a
genetic algorithm (GA) approach. The results are compared with those obtained using classical heuristic rules in two
examples that involve the objective of minimising the makespan or the total tardiness penalty. The comparison shows
that the GA can obtain good schedules for the model, giving a family of solutions that are at least as good as those
produced by the heuristic rules. Ó 2001 Civil-Comp Ltd. and Elsevier Science Ltd. All rights reserved.

Keywords: Planning and scheduling; Precast production; Genetic algorithms; Optimisation

1. Introduction solutions are required. The two principal methods used

to solve scheduling problems are based on operations
Precast technology has been pursued to increase research (OR) and arti®cial intelligence (AI). OR algo-
construction productivity as well as the buildability and rithms are theoretically rigorous and guarantee optimal
quality of the buildings in the past decades. In the ap- solutions, but are dicult to use in real life applications
plication of prefabrication in the construction industry, due to the fact that their performance does not scale well
it is necessary to look into material selection, element on larger problems. As a consequence, they have been
design, transportation and site assembly methods as well enhanced with the addition of heuristics. Unfortu-
as pay close attention to factory layout, inventory con- nately, the heuristic enhanced methods, based on rules
trol and production scheduling. In this paper, we focus of thumb, do not guarantee an optimal solution even
on production scheduling, noting that it impacts both though they are easy to use. In AI methods all the
site construction and plant operations. constraints of a scheduling problem are represented ex-
Scheduling involves the allocation of resources over a plicitly, even if some of the available information is
period of time to perform a collection of jobs subject to uncertain. Among the various AI methods, a new ap-
known system constraints. Scheduling is known to be proach, genetic algorithm (GA) has been widely used in
NP-complete and has proven to be a dicult task for recent years by many researchers and engineers to over-
human planners and schedulers, particularly if optimal come the drawbacks of the mathematical models. The
GA was proposed by Holland [12] and has been suc-
cessfully applied to a great variety of problems, includ-
*
Corresponding author. Tel.: +65-874-2576; fax: +65-874- ing process scheduling and resource allocation [15] in
2248. many disciplines. In the area of construction manage-
E-mail address: [email protected] (W.-T. Chan). ment, Chan et al. [3] proposed a GA approach for
URL: http://itcive.nus.edu.sg/frontdesk/. construction scheduling and resource allocation. Li and

0045-7949/01/$ - see front matter Ó 2001 Civil-Comp Ltd. and Elsevier Science Ltd. All rights reserved.
PII: S 0 0 4 5 - 7 9 4 9 ( 0 1 ) 0 0 0 3 6 - 0
1606 W.-T. Chan, H. Hu / Computers and Structures 79 (2001) 1605±1616

Love [16] presented an application of GAs for time-cost

optimisation problems in construction planning.
A lot of e€ort has been devoted to improving precast
production planning and control activities. Warszawski
and Ishai [22], an early paper in this area, describes the
long range planning that is necessary for the prefab-
rication industry in a national economy and which
involves consideration of the number, location and ca-
pacity of plants for the production of precast elements
in the various stages of development. Warszawski [23]
discussed the ecient planning of short (speci®c orders)
and long (continuous demand for standard elements)
production series on one or several moulds. Dawood and
Neale [8] developed a computer-based capacity planning
model for precast concrete building products in order to
help production managers make better planning deci-
sions and explore options open to them. The present
study focused on another aspect of planning, that of
scheduling the production in the precast plants. There is,
as yet, no formal method for production scheduling that
takes into consideration the practical constraints pre-
vailing in the industry. The standard objective encoun-
tered in most production planning models is meeting
customer demand at minimum cost. Therefore, precast-
ers are more concerned with issues like which, when and
how many elements should be produced to meet their
delivery dates. Moreover, production scheduling issues Fig. 1. Production system: (a) stationary moulds, (b) movable
like the order of component production, the organisa- moulds.
tion of men and equipment have not been studied in
depth. These details are usually left to the supervisory The major types of moulds are table moulds, long
personnel on the production ¯oor. The result may be casting beds, stacks and battery moulds. Two basic types
unnecessary idle waiting time and factory production of component production systems may be employed in a
below the optimal plant capability. Chan and Hu [4] ®rst prefabrication plant, namely the stationary system and
applied formal scheduling models to precast production the travelling system [24]. With the stationary system, as
planning. However, only generic scheduling situations shown in Fig. 1a, all basic production activities are per-
were considered and only simplistic production con- formed in one place and horizontal or battery moulds
straints were included. As a result, the model could not are usually used. In the travelling system, as shown in
readily be used in actual production scheduling. In this Fig. 1b, moulds are moved on a rolling line or a con-
paper, we report on the work that has been done to veyor from one workstation to another, with di€erent
re®ne the ¯ow shop sequencing model to remedy this. activities being performed at each one of them. Almost
The re®nements were the result of careful investiga- always, a movable mould line is connected to a curing
tions of actual precast plant operations in Singapore. chamber where the concrete mix undergoes an acceler-
A variant of this model suitable for plants employing the ated but controlled curing process.
specialised mode of production organisation is presented Typically, the production process in a precast factory
here. The model is solved using a GA-based approach to consists of the following sequence:
give a family of high-quality schedules that can be used
in production planning. 1. Mixing of concrete (including aggregates and cement
storage and handling, and moving of concrete from
mixing centre to moulds).
2. Setting of moulds: cleaning and oiling of mould sur-
2. Precast production and scheduling face, and fastening of side frames.
3. Placing of reinforcements, ®xtures, electrical con-
2.1. Precast production duits, and inserts etc. to be contained in the compo-
nents.
The main pieces of equipment in a prefabrication 4. Casting: pouring, compacting and levelling of con-
plant are the casting moulds made of steel or timber. crete.
 W.-T. Chan, H. Hu / Computers and Structures 79 (2001) 1605±1616 1607

5. Curing: through an arti®cial (usually heating) or natu- move from one mould to another. This system will work
ral process. smoothly if all components are identical so that a bal-
6. Demoulding: stripping the side frame and taking out anced rhythm of production can be maintained. How-
the components. ever, when heterogeneous elements are involved with
7. Finishing, patching and repairing of components. di€erent production times, there is a loss of eciency
8. Placing the completed components in the stockyard due to the diculty of balancing the work at all the
to achieve the delivery strength. workstations (crews). Some components have to wait for
9. Transporting the components to the construction the next operation and some workstations (crews) have
site. to be waiting idly for the arrival of new components.
Ecient scheduling is simple under the specialised
In studying the process found in the precast plant, we method of organisation when identical components are
concluded that the production process is a form of ¯ow involved. However, it is rather dicult to determine an
shop process, where every component has to be pro- ecient schedule if the components are di€erent in the
cessed in the sequence described above. specialised alternative. Unfortunately, this is the pre-
vailing situation in the industry where it is very common
2.2. Modelling the two methods of prefabrication organi- to produce many di€erent elements to satisfy several
sation contract orders at the same time. We focus on this latter
problem and propose a scheduling model for this dicult
The purpose of the production scheduling function but very practical situation encountered in the industry.
in a precast plant is to ensure timely prefabrication of
components with the most ecient utilisation of pro-
duction resources. According to Warszawski [24], there 3. A ¯ow shop sequencing model for precast production
are two alternatives of production organisation in plants.
Under the ®rst alternative, the same crew performs all Based on investigation and observation of precast
operations outlined in the section above. After casting plants, precast production under the specialised method
and completing the processing of one component, the possesses many of the characteristics of the traditional
crew moves to the next mould and starts to work on the ¯ow shop sequencing problem.
second component, repeating the same sequence of ac-
tions. Under the second alternative, the total process is 3.1. Traditional ¯ow shop sequencing problem
broken into several activities which are performed by
di€erent crews with specialised tools and work methods. The ¯ow shop sequencing problem is generally de-
Although the ®rst alternative, which may be called scribed as follows: There are m machines and n jobs;
the comprehensive method, is easy to employ, it is usu- each job consists of m operations, and each operation
ally less ecient than the other in terms of labour, tools requires a di€erent machine. The n jobs have to be pro-
and workspace utilisation. It also creates problems of cessed in the same sequence on all the machines. The
co-ordinating access to common resources by the dif- processing time of job i on machine k is given by tik
ferent work teams. For example, under a casting cycle of i ˆ 1; . . . ; n; k ˆ 1; . . . ; m†. The objective is to ®nd the
1 day, all teams will want to use crane for demoulding sequence of jobs minimising the maximum ¯ow time,
hardened elements in the morning when they start their which is technically called the makespan of the pro-
shifts. The co-ordination of concrete supply may pose duction schedule. The main assumptions for this prob-
another problem, again for the same reason, when all lem are as follows:
crews progress at approximately the same rate.
The second method of production organisation, · every job has to be processed on all machines in the
which may be called the specialised method, requires order 1; 2; . . . ; m;
division of the total work into several tasks, such as · every machine processes only one job at a time;
demoulding, mould preparation, casting, and so on, with · every job is processed on one machine at a time;
each task being performed by a di€erent crew. This · the machines are not pre-emptable;
method is very ecient in terms of crew and equipment · the set-up times for the operations are sequence-inde-
usage since there is little idle waiting time when all the pendent and are included in the processing times;
crews work at a balanced rate. The system is particularly · the operating sequences of the jobs are the same on
ecient with a movable production line, that is, moulds every machine, and it is this common sequence that
moving between workstations with tasks performed at has to be determined.
each one of them, eliminating the need for crews to walk
between and set up at di€erent stations. Although this Let C ji ; k† denote the completion time of job ji on
method is well suited to a movable production line, it machine k and let fj1 ; j2 ; . . . ; jn g denote a job permuta-
can also be employed by static moulds where crews tion. Supposing a machine begins its processing when it is
1608 W.-T. Chan, H. Hu / Computers and Structures 79 (2001) 1605±1616

possible to do so. Then the completion time for an n-job ¯ow shop. It may be that the modelling assumptions
m-machine ¯ow shop sequencing problem can be calcu- made in the standard ¯ow shop sequencing model are
lated as: overly restrictive and idealistic; practical constraints are
not considered in the model and the criterion of mini-
C j1 ; 1† ˆ tj1 1 ; mising makespan is not the primary consideration in the
C j1 ; k† ˆ C j1 ; k 1† ‡ tj1 k ; k ˆ 2; 3; . . . ; m: industry. In the next section, we describe a modi®ed ¯ow
C ji ; 1† ˆ C ji 1 ; 1† ‡ tji 1 ; i ˆ 2; 3; . . . ; n: 1† shop sequencing model (FSSM) which addresses these
C ji ; k† ˆ maxfC ji 1 ; k†; C ji ; k 1†g ‡ tji k : two points.
i ˆ 2; 3; . . . ; n; k ˆ 2; 3; . . . ; m:
3.2. A ¯ow shop sequencing model for precast operation
The makespan is calculated as follows:
In the application of the FSSM for specialised pre-
makespan ˆ Cmax ˆ C jn ; m†: 2† cast production, the following issues should be carefully
considered:
Johnson [14] published the ®rst paper on the ¯ow (1) The de®nition of jobs and machines: In our ¯ow
shop sequencing problem and proposed an ``easy'' al- shop sequencing model, jobs correspond to the compo-
gorithm for the two-machine problem with makespan as nents to be produced (in the various moulds), and ma-
the objective criterion. Since then, the m-machine (m > chines correspond to the distinct operations (performed
2) problem has held the attention of many researchers. by specialised crews at the di€erent workstations). The
Over the past four decades, extensive research has been precast production process in this study is divided into
done on the pure ¯ow-shop problem, but there are no six operations, namely mould setting, reinforcement
corresponding easy algorithms like that proposed by setting, casting, curing, demoulding, with ®nishing/re-
Johnson that can provide an optimal solution for these pairing as the last one. This means that it is a six-
latter problems. Analytic methods, employing integer machine problem. Each of these operations is done by a
programming and branch-and-bound techniques etc., crew. Although each crew is strongly identi®ed with a
have been proposed to ®nd the optimal solution [13]. particular operation, in practice, a crew may help out
However, they are not very e€ective on large problems with the work of another crew if it is not otherwise
or even medium-sized ones. The ¯ow shop sequencing engaged. This complicates the scheduling problem since
problem has been proved to be NP-complete by Co€- there is the added dimension of crew allocation. In the
man [5]. As a result, most researchers turned away from traditional ¯ow sequencing model, when a machine ®n-
optimisation and focussed instead on approximate al- ishes one job, it waits idly for the arrival of the next job.
gorithms by developing heuristics to provide good and We follow the traditional setting and do not take into
quick solutions. Gen and Cheng [10] included some of consideration this opportunistic behaviour in our cur-
the well-known heuristics in their survey, including those rent model. We assume that each crew works only on its
proposed by Palmer [19], Gupta [11], the CDS heuristic own task.
[2], the RA heuristic [6] as well as the NEH heuristic [18]. (2) Distinction between normal-working time and o€-
However, these heuristics are developed for the tradi- normal-working time: Time in the traditional FSSM is
tional problem of minimising the schedule makespan; continuous, there being no distinction between working
few algorithms have been published for optimising on and non-working hours. In precast plants, workers
the Earliness and Tardiness (E & T) criteria. In industry, typically work normal 8-h shifts that are punctuated by
the commonly used heuristic for the tardiness problem o€-normal-time hours before a new 8-h shift begins.
is the earliest due date (EDD) rule. Sule [20] applied the They can be asked to work overtime if necessary but
EDD rule for the single machine sequencing problem, overtime charges are incurred. But the curing activity,
which can be brie¯y described as follows: A schedule is once begun, continues irrespective of whether it is nor-
developed considering the due dates of the jobs. The mal or o€-normal hours. Usually overtime hours are
resulting sequence arranges and processes the jobs in an limited for one shift, no more than 4 h in the examples in
ascending order of their due dates. EDD rule can also be Section 5.
applied to the ¯ow shop sequencing problem. Some jobs (3) Distinction between pre-emptive and non-pre-emp-
may have identical due dates, in which case the tardiness tive operations: In the traditional FSSM, all operations
penalty also should be considered. It is apparent that the are non-pre-emptive, which means that a task, once
job with a larger unit penalty should be processed ®rst. started, cannot be interrupted until it is completed. An
If there is still a tie, the job with the smallest total pro- operation is pre-emptive if it can be interrupted and
cessing time would be the one selected. resumed later on the same machine. According to this
However, Dudek et al. [9] pointed out that there was de®nition, casting and curing are non-pre-emptive op-
lack of papers in the literature describing applications of erations while the others are pre-emptive ones. This
the algorithms minimising the makespan for the static complicates the scheduling of these two types of tasks.
 W.-T. Chan, H. Hu / Computers and Structures 79 (2001) 1605±1616 1609

For the pre-emptive operations, workers can stop work- distinction between normal and o€-normal hours, and
ing if they have to go beyond normal working hours the possibility of parallel curing, it is sometimes ad-
(since this minimises overtime costs); this is not the case vantageous to switch the processing order of jobs for the
for casting and curing, which may require extension into di€erent tasks. For example, since the curing times be-
o€-normal working hours. Let TD and TN denote the tween di€erent components may di€er, some compo-
daily normal-working time and o€-normal-working time nents with shorter curing times may ``overtake'' other
respectively. TN ˆ 24 TD . And let OT represent work- components with longer curing times but may end up
ing overtime. The completion time of pre-emptive op- waiting needlessly if a rigid job scheduling order is fol-
erations can be computed as: lowed.
 (6) Criteria used: For many years, scheduling re-
t if t 6 24d ‡ TD searchers focused on minimising schedule makespan.
C ji ; k† ˆ 3†
t ‡ TN if t > 24d ‡ TD However, the most common consideration encountered
in industrial production is to meet due dates. Moreover,
where t ˆ maxfC ji 1 ; k†, C ji ; k 1†g ‡ tji k ; d ˆ int t=
with the current interest in just-in-time production,
24† i ˆ 2; 3; . . . ; n; k ˆ 2; . . . ; m.
earliness as well as tardiness should be discouraged.
The completion time of the non-pre-emptive opera-
Let j1 ; j2 ; . . . ; jn † denote the job processing order, di
tions can be calculated as Eq. (4).
the due date, and ci the completion time for job ji . As-
C ji ; k† ˆ sociated with each job is a unit earliness penalty ai > 0
8 and a unit tardiness penalty bi > 0. Tardiness Ti is de-
<t
> if t 6 24d ‡ TD or if t > 24d ‡ TD
®ned as Ti ˆ maxf0; ci di g which represents the late
and OT > 0 completion of job ji after it fails to meet its due date.
>
:
24 d ‡ 1† ‡ tji k if t > 24d ‡ TD and OT ˆ 0 Likewise, earliness Ei is de®ned as Ei ˆ maxf0; di ci g;
4† it represents the early completion of job ji if it does not
occur on its due date. The job can be given a completion
where t ˆ maxfC ji 1 ; k†; C ji ; k 1†g ‡ tji k ; d ˆ int t= time window ai within which no penalties are incurred if
24† i ˆ 2; 3; . . . ; n; k ˆ 2; . . . ; m. the completion time of job i is within the time interval
(4) Curing: Curing is identi®ed as a special operation di ai , di ‡ ai †. Let P be the set of all possible job se-
in our model and is di€erent from the traditional ¯ow quences and r be an arbitrary sequence. Assuming that
shop operation in several respects. Firstly, curing occu- the penalty functions are linear, the generic E & T model
pies almost no labour resources. But processing time, can be formulated as Eq. (5).
moulds and production space are required. In addition,
this curing time is rather long compared to the other X
n

operations. For example, 12±16 h under natural curing minf r† ˆ ai Ei ai †U Ei ai † ‡ bi Ti ai †U Ti ai ††;

r2P
iˆ1
with an ambient temperature of 25°C to 30°C and about 5†
6 h under steam curing. Therefore, it is better to try to
schedule curing during the o€-normal working time.
Secondly, curing is a non-pre-emptive operation that where U x c† is the unit step function de®ned as
cannot be interrupted. Thirdly, but most importantly, 
curing is a parallel process. In the traditional FSSM, a 0 if x 6 ai ;
U x ai † ˆ
new operation ji ; k† of job ji on a machine k cannot be 1 if x > ai :
started until ji ¯ows from machine k 1 and job ji 1 has
been ®nished on machine k. However, di€erent compo- Following the discussion above, our FSSM for pre-
nents can be cured simultaneously. The beginning of a cast production scheduling can be described as follows:
component's curing is only dependent on the ®nish of its There are n jobs (components) to be processed in the
casting (i.e. the processing of job ji on machine k 1). A same task sequence by m machines (workstations or
component can be cured as soon as it is cast and it need crews, m ˆ 6). A distinction is made between the nor-
not wait for the completion of curing of another com- mal-working time and the o€-normal-working time. All
ponent (curing space in the chamber permitting). How- operations are pre-emptive with the exception of casting
ever, if the normal-working time is TD and the ®nish time and curing. Every job is processed on one machine at a
of the curing activity falls during the o€-normal period, time, and every machine processes one job at a time
i.e. between 24d ‡ TD † and 24 d ‡ 1†, the demoulding except for curing, which is a parallel process. Reschedu-
operation which follows, cannot be processed until the ling is possible after the curing operation if it is advan-
next working day, i.e. the 24 d ‡ 1†th h. tageous. The objective is to ®nd the processing sequence
(5) Rescheduling: In the traditional FSSM, once we of the precast elements that gives the minimum make-
decide the processing order of the jobs, this order re- span, meets delivery dates or satis®es other quanti®able
mains unchanged till the end. However, because of the criteria.
1610 W.-T. Chan, H. Hu / Computers and Structures 79 (2001) 1605±1616

4. Genetic algorithms for precast production scheduling tion operation: selection. The genetic operations mimic
the process of heredity of genes to create new o€spring
4.1. Basic genetic algorithms in each generation. The evolution operation mimics the
process of Darwinian evolution to create populations
GAs are inspired by the process of natural evolution from generation to generation.
and the principle of ``survival of the ®ttest''. GAs iter- Crossover: Crossover is the principal mechanism by
atively generate new solutions from currently available which GA arranges for good schemes present on dif-
solutions and replace some or all of the existing mem- ferent chromosomes to aggregate on a single individual.
bers of the current solution pool with the newly created Single-point crossover involves exchange of a part of
members. The quality of the solution pool should im- each chromosome in a pair across a randomly chosen
prove with the passage of time as a consequence of the point (as depicted in Fig 3a). Although crossover is
selection procedure. According to Davis [7], a GA op- principally thought of as a mechanism that improves the
erates on a population (of solutions) of ®xed size (P) as quality of solutions, it is also possible that crossover will
follows: disrupt a good schema already present, especially those
of higher order.
· initialise a pool, known as the parent pool, of P ran- Mutation: Mutation operates on a single chromo-
domly created individuals; some and produces a new genotype by making a random
· measures the goodness of each individual in the par- change to values of one or more of the genes. Mutation
ent pool with respect to the problem evaluation func- is performed at the gene level but not every gene needs to
tion and converts the absolute objective values to be (or should be) mutated. The frequency of mutation is
relative ®tness values; often kept very low to avoid disruption of good solu-
· select individuals in parent pool for the creation of tions; the principal use of mutation is to reintroduce
the next generation with a probability relative to their genetic diversity to avoid getting trapped in local op-
®tness; tima. Fig. 3b is an example of constant mutation. If gene
· create new individuals, known as the o€spring, by i is to be mutated, a constant value will be added (sub-
means of GA operators (crossover and/or mutation) tracted) to (from) the existing allele value.
on the selected parent genotypes; Selection: Selection is the procedure by which better-
· use a selection scheme to select the next parent pool than-average solutions are determined for recombination
for the new generation; to generate new o€spring. Above-average individuals,
· if pre-speci®ed stopping criteria is met, then stop the which contain good schemata, have an above-average
GA cycle; otherwise return to step 2. chance of passing on their schemata to the next genera-
tion. In the so-called simple selection scheme, the parent
Each individual solution is represented by a single genotypes would be assigned a number of o€spring based
string-like entity called a chromosome. A chromosome on their ratio of its ®tness values with the aggregate ®t-
typically consists of a number of genes, which may be ness of the parent pool. One form of simple selection
visualised as boxes arranged in a linear fashion (Fig. 2).
Two attributes are associated with each gene: its posi-
tion and its contents. When applied to scheduling, each
individual in the population corresponds to one possible
solution of the scheduling problem. A chromosome is
the internal representation of the solution and consists
of a linear string of genes which can be transformed
directly or decoded indirectly to produce an actual
schedule.
There are two kinds of operations in GA: (1) Genetic
operations: crossover and mutation, and (2) the evolu-

Fig. 2. Chromosomal representation. Fig. 3. (a) Single-point crossover, (b) constant mutation.
 W.-T. Chan, H. Hu / Computers and Structures 79 (2001) 1605±1616 1611

scheme is the roulette selection where the ratios of the ‰0:11 0:98 0:03 0:23 0:87 0:91 0:36 0:51 0:72 0:84Š
parent genotypes can be ®tted onto a biased roulette
wheel. At the time of o€spring creation a simple spin of where position i in the list represents job i. The random
the roulette wheel yields the ``lucky'' genotype and highly number in position i determines the processing order of
®t genotypes get more chances to transfer their properties job i in a schedule. Random keys solve the problem of
to the next generation [7]. illegal o€spring that may be generated during genetic
There are several parameters that can determine the operations on chromosomes at the expense of an en-
performance of GAs but their optimal values cannot be larged search space.
ascertained by applying ®xed rules: these are the popu- Crossover and mutation: The traditional crossover
lation size, the number of iterations performed, the can be used without modi®cation in the random key
number of new individuals created in each generation representation. We used a two-point crossover and a
and the probability of crossover and mutation. constant mutation respectively. In constant mutation, a
suitable constant value will be added (or subtracted) to
(or from) the existing gene value of gene i (Fig. 3b).
4.2. Genetic algorithm model for precast production Decoding: The decoding of the random key repre-
scheduling sentation is a simple process. The random values are
used as sort keys, usually in ascending order, to decode
Chan et al. [3] proposed a GA model for the schedu- the solution and to get a legal schedule. For the above
ling and resource allocation problem. We use this same chromosome, the random keys are sorted in ascending
GA process for the scheduling of precast production (as order:
depicted in Fig. 4).
0:03 0:11 0:23 0:36 0:51 0:72 0:84 0:87 0:91 0:98

4.3. A implementation for precast production By comparing the new list with the initial permuta-
tion, we get a processing schedule: 3 ! 1 ! 4 ! 7 !
Representation: Two di€erent kinds of representation 8 ! 9 ! 10 ! 5 ! 6 ! 2. We use this to evaluate the
can be adopted in the process scheduling problem: one is above chromosome to acquire the ®tness of the schedule
based on random keys, the other is based on permuta- produced from it.
tion ordering. The permutation representation is a sim- Objective/evaluation functions: Makespan and total
ple and natural representation that encodes the identity tardiness penalties are selected as the evaluation criteria
of the element to be scheduled in each gene; the order in in this paper. For the ¯ow time problem, makespan is
which the genes appear in the chromosome string gives calculated as in Eqs. (1) and (2), whilst for the E & T
the schedule sequence. Unfortunately, this kind of rep- problem, the total E & T penalty is calculated as in
resentation is prone to the generation of illegal o€spring Eq. (5).
if the traditional one-point or two-point crossover op-
erator is used. We used random key representation in-
stead in our adaptation of the ¯ow shop sequencing
5. Results and discussions
model for the precast production. The random key
representation was ®rst proposed by Bean [1] and en-
Stationary moulds and the specialised production
codes a solution with random numbers from (0,1). These
method are adopted for the examples discussed in this
random values are used as sort keys to represent the
section. As stated before, curing is an operation which
scheduling priority to be used to decide the schedule. A
does not require a special crew, thus the workers were
chromosome based on random keys may look like this:
divided into ®ve crews, namely mould, reinforcement,
casting, demoulding and ®nishing/repairing, moving
from one mould to another.
Production times: In precast factories, the production
times of di€erent components by di€erent crews depend
on the component size and complexity, and they may
vary even for the same kind of component. The pro-
duction times used in the examples of this section were
the average values observed in practice. Natural curing
was used and the duration is 12 h.
Overtime: In our examples, the maximum overtime
per working day is 4 h. The following policy regarding
overtime was used: the casting of an element will be
Fig. 4. GA model for the scheduling of precast production. processed using overtime if it can ®nish within the
1612 W.-T. Chan, H. Hu / Computers and Structures 79 (2001) 1605±1616

maximum overtime period for the day (4 h); or if its due inevitable that there would be some settling time where
date falls on the next day and it can be ®nished within all the task crews (except the ®rst) had to wait for the
the maximum overtime period. arrival of the ®rst component. This initialisation lowers
Due dates: Due dates are taken as givens/constraints plant productivity compared to what it would be when
imposed by the planning department which takes into production is on-going.
consideration factors like the contracted delivery dates, Resource availability: We assume there is no resource
the inventory level and the time for concrete to develop limitation for the FSSM, particularly, the number and
sucient strength for transportation to the site. the type of the molds can always meet production de-
Earliness penalty: The components are delivered to mand.
the site in batches and stays in storage from the time the Our ®rst example considers a simple precast situation
last task on it is completed. Storage costs are incurred comprising six heterogeneous components. Components
which are modelled indirectly by including earliness 1 and 2 are utility room and balcony elements to be
penalties. The quantum of the penalty value remains a added to an existing apartment, components 3 and 4 are
dicult question with no hard and fast rules. The pen- elements for a proposed covered walkway and compo-
alty values used here were determined after some ex- nents 5 and 6 are lift shaft and lobby elements to be
perimentation to ensure that the values used did not trap constructed at one of the gable ends of an apartment
the GA in a sub-optimal part of the search space. block. Production times, due dates and E & T penalties
Adding new components to a schedule: Production is a of these components are listed in Table 1. The compu-
continuous process in the precast plants. Components tation results are summarised in Table 2. The second
are scheduled in batches over ®nite planning horizons, example expands the problem considerable by taking a
with each batch being independent of each other. New 10-type 44-component situation. Table 3 compares the
scheduled components are added to the end of the pre- schedules obtained by our GA scheduling model against
vious schedule. those obtained by di€erent heuristic algorithms. The
Initialisation of the examples: As explained above, results indicated that the classical heuristic algorithms
new components are added to the end of the previous (Palmer, Gupta, CDS and RA) were good at shortening
schedule in practice. However, in our examples, we as- the makespan while the EDD rule was good at reducing
sumed that there were no prior schedules. Thus, it was the total tardiness penalty and/or total E & T penalty

Table 1
The quantity, production times and unit E & T penalties of the components
Components Crews Due dates E&T penalties
(h)
No. Name Moulds Rebars Casting Demoulding Finishing ai (earliness) bi (tardiness)
(2 men) (3 men) (4 men) (2 men) (2 men) (units/h) (units/h)
1 Balcony (6 M2 ) 1 0.8 1.2 1.5 0.5 28 2 10
2 Utility room 1.7 2 2 1.5 2.5 28 2 10
(12 M2 )
3 Footing (2.5 T) 0.4 0.5 0.6 0.5 0 28 1 10
4 Stump (1.5 T) 0.3 0.4 0.5 0.4 1 28 1 10
5 Lobby (12.0 T) 1.5 1.8 1.2 1.5 1.5 52 2 10
6 Shaft (13.1 T) 1.5 1.6 1.5 1.8 0.8 32 2 10

Table 2
Results for the 6-component example
Heuristic used Schedule Makespan Tardiness penalty E & T penalty (units) Overtime (h)
(h) (units)
Palmer's heuristic 4-2-1-5-3-6 50.6 210.0 249.8 1.0
Gupta's heuristic 4-2-5-6-1-3 50.0 407.0 447.8 0.9
CDS heuristic 4-2-5-6-1-3 50.0 407.0 447.8 0.9
RA heuristic 5-2-6-1-4-3 49.4 399.0 445.0 2.0
EDD rule 3-4-1-2-6-5 51.0 16.0 17.9 0.9
GAs 4-5-6-2-1-3 48.5 384.0 429.8 2.4
GAs 4-2-6-1-3-5 51.0 0.0 0.6 1.3
 W.-T. Chan, H. Hu / Computers and Structures 79 (2001) 1605±1616 1613

Table 3
Results for the 44-component example
Heuristic used Schedule Makespan (h) Tardiness Total E&T Overtime (h)
penalty penalty
(units) (units)
Palmer's heu- 33-34-35-36-37-38-39-40-1-2-3-4-5-6-7-8-41-42- 175.1 3776.0 5177.0 7.1
ristic 43-44-17-18-25-26-27-28-29-30-31-32-23-24-
19-20-9-10-11-12-13-14-15-16-21-22
Gupta's heu- 33-34-35-36-37-38-39-40-41-42-43-44-5-6-7-8-1-2- 174.4 4781.5 6154.9 7.2
ristic 3-4-17-18-23-24-9-10-11-12-13-14-15-16-21-
22-19-20-25-26-27-28-29-30-31-32
CDS heuristic 33-34-35-36-37-38-39-40-41-42-43-44-5-6-7-8-1-2- 174.4 5980.0 7472.5 7.2
3-4-18-17-24-23-16-15-14-13-12-11-10-9-22-
21-20-19-32-31-30-29-28-27-26-25
RA heuristic 1-2-3-4-16-15-14-13-12-11-10-9-24-23-18-17-20- 192.3 7368.5 8319.4 10.0
19-8-7-6-5-22-21-40-39-38-37-36-35-34-33-44-
43-42-41-32-31-30-29-28-27-26-25

EDD rule 9-10-25-26-27-28-33-34-35-36-41-42-1-2-11-12- 193.5 155.0 654.2 9.1

17-19-21-23-5-6-3-4-13-14-29-30-31-32-43-44-
37-38-39-40-18-20-22-24-7-8-15-16
GAs 39-35-25-36-33-28-27-29-32-10-2-9-34-26-12-30- 174.1 0.0 1006.4 9.9
1-44-11-38-6-3-40-13-21-23-8-5-4-43-19-17-37-
31-7-41-14-20-42-15-18-22-24-16
GAs 25-28-36-35-2-27-34-40-9-33-26-10-44-11-1-39-5- 171.5 0.0 919.4 9.2
12-7-37-21-4-43-32-29-3-41-18-8-31-42-23-7-
38-30-22-13-6-19-14-24-16-15-20

GAs 34-36-26-35-28-38-30-9-1-10-27-25-33-42-12-11- 175.3 0.0 865.3 8.2

2-32-13-31-44-3-7-37-21-19-29-6-39-14-43-4-
22-20-40-5-17-23-41-15-18-24-16-8
GAs 36-26-9-33-34-25-10-28-27-35-1-2-11-12-42-19-5- 192.5 0.0 372.4 7.3
32-3-13-23-38-39-17-4-44-40-14-30-6-31-21-
43-37-29-15-24-18-16-8-7-20-22-41
GAs 34-33-41-18-40-17-11-31-20-39-29-36-26-37-15- 170.8 4940.5 6242.4 10.9
42-35-7-3-13-28-8-21-24-30-9-6-43-10-22-25-2-
14-1-5-4-23-44-38-19-16-12-27-32

(Tables 2 and 3). This is expected since these heuristic schedules with the minimum makespan of 48.5 h and the
algorithms were developed speci®cally for these pur- minimum total tardiness penalty of 0.0 units were both
poses. We cannot conclude which heuristic is better acquired by GAs. Similarly, the schedules in Table 3
based on the results of these two examples because it is with the minimum makespan of 170.8 h and the mini-
well accepted that heuristics do not work well for all mum total tardiness penalty of 0.0 units were obtained
problems and their performance is sensitive to the input by the GA. The results provide evidence that the GA is a
data. Taillard [21] compared the solutions to the make- very e€ective algorithm for tardiness or makespan
span problem obtained by classical heuristics. NEH problems. Another point to note is that the GA deter-
appeared to be the best heuristic among them, but it was mined not just one solution with the minimum total
the most complicated and it required a long computa- tardiness penalty, but four. The GA also returned many
tional process if n was large. The RA or Palmer's heu- other good solutions which are not listed here. This il-
ristic also returned good solutions and only required lustrates another important advantage of GA which is
modest computational times. that it can produce a family of high quality (near opti-
Tables 2 and 3 compare the makespan, tardiness mal) solutions in a short time.
penalty, E & T penalty and overtime obtained by the A good schedule with a shorter makespan might
Palmer, Gupta, CDS, RA heuristics and the EDD rule incur a large tardiness penalty, and vice versa. In Table
against those obtained by the GA. In Table 2, the 3, it can be seen that the schedule with the minimum
1614 W.-T. Chan, H. Hu / Computers and Structures 79 (2001) 1605±1616

makespan of 170.8 h has a rather high total tardiness makespan or the tardiness penalty and the job sequence
penalty of 4940.5 units. This suggests that a good itself decides the performance.
schedule under a speci®c criterion may have a poor
performance under another one. On the other hand, the
schedule obtained by the GA has a short makespan as 6. Conclusions
well as a small total tardiness penalty, for example, 171.5
h/0.0 units and 175.3 h/0.0 units in Table 3. This is We categorised the production organisation alterna-
probably a stroke of luck but the existence of many tives in precast factories as the comprehensive method
good solutions increases the possibility of ®nding bal- and the specialised method. A modi®ed FSSM for the
anced schedules. GAs have already been demonstrated specialised precast production was proposed which was
to be good for handling multi-objective problems [17] solved using a GA. A distinction is made between nor-
although that has not been considered here. mal-working time and o€-normal-working time, and
Reliability and stability are two important charac- operations are divided into pre-emptive and non-pre-
teristics for an algorithm. We not only care that an al- emptive tasks. Curing is a non-pre-emptive task which
gorithm is able to ®nd the best solution, but also the also happens to be a parallel process. This enables re-
spread between the worst, average and current solutions scheduling to occur which is found to be bene®cial. With
and the repeatability of the results are also important these modi®cations, the proposed FSSM is better able to
considerations in actual production scheduling. Fig. 5 model the situation occurring in industry.
shows a typical run of the total tardiness penalty mini- We also conclude from our experiments that the GA
misation described in Example 2. It shows how the best, is a useful and e€ective method for solving the FSSM.
worst, average and online solutions evolve during a GA The main advantages of the GAs over the heuristics
run; even the worst solution is still acceptable compared compared against are that GAs produce equal or better
with the results determined by the heuristics. The results eciently and are able to give a family of optimal
makespan for thirty consecutive runs of Example 2 is and/or near optimal results in a short time. GAs have
depicted in Fig. 6. The results indicate that the GA also been proven to be reliable and stable, and are able
performance is quite good and stable. Fig. 6 also shows to produce good results under a wide variety of oper-
the e€ect of di€erent population sizes on the best solu- ating conditions.
tion obtained. The result shows that larger population The FSSM needs to be tested on larger data sets
size is more likely to produce better solutions. under a wider variety of operating conditions. Also, the
Finally, we also found that, contrary to what we requirement that work crews can only do a ®xed task
might expect, the schedules which employed more needs to be relaxed to better match actual industry
overtime work did not always have a smaller makespan practice. This will make the solution of the FSSM more
and total tardiness penalty (Tables 2 and 3). This shows dicult since the both scheduling and resource assign-
that the use of overtime value is not directly related to ment have to be considered simultaneously. Further

Fig. 5. The best, worst, average and online solutions for Example 2 in a typical run.
 W.-T. Chan, H. Hu / Computers and Structures 79 (2001) 1605±1616 1615

Fig. 6. The results of 30 consecutive runs of Example 2.

work to address both points is currently being planned [8] Dawood NN, Neale RH. A capacity planning model for
for the next phase of development. precast concrete building products. Build Environ 1993;
28(1):81±95.
[9] Dudek RA, Panwalkar SS, Smith ML. The lessons of
¯ow-shop scheduling research. Oper Res 1992;40(1):7±
Acknowledgements 13.
[10] Gen M, Cheng RW. Genetic algorithms and engineering
The authors are indebted to the sta€ and ocers of design. New York: Wiley; 1997.
the Precast Technology Centre of the Housing and De- [11] Gupta J. A functional heuristic algorithm for the ¯ow
velopment board, Singapore for their support of the shop scheduling problem. Oper Res Quart 1971;22:39±
research that resulted in the ®ndings reported in this 47.
[12] Holland J. Adaptation in natural and arti®cial systems.
paper.
Ann Arbor: University of Michigan Press; 1975.
[13] Ignall E, Schrage L. Application of the branch and bound
technique to some ¯ow shop scheduling problem. Oper Res
References 1965;13:400±12.
[14] Johnson SM. Optimal two and three-stage production
[1] Bean J. Genetic algorithms and random keys for sequenc- schedules with set up time included. Naval Res Logist
ing and optimisation. ORSA J Comput 1994;6(2):154± Quart 1954;1:61±8.
60. [15] Kopfer H. Evolutionary search and the job shop: investi-
[2] Campbell H, Dudek R, Smith M. A heuristic algorithm for gations on genetic algorithms for production scheduling.
the n-job m-machine sequencing problem. Mgmt Sci Heidelberg: Physica-Verlag; 1996.
1970;16B:630±7. [16] Li H, Love P. Using improved genetic algorithms to
[3] Chan WT, Chua DKH, Kannan G. Construction resource facilitate time-cost optimisation. J Comput Engng Mgmt
scheduling with genetic algorithms. J Construct Engng ASCE 1997;123(3):223±37.
Mgmt ASCE 1996;122(2):125±32. [17] Liong SY, Khu ST, Chan WT. Derivation of pareto front
[4] Chan WT, Hu H. Process scheduling using genetic with accelerated convergence genetic algorithm ACGA.
algorithms for construction industry. In: Chen SW, editor. Hydroinformatics '98, 24±26 August 1998. Copenhagen,
Proceedings of the 3rd International Conference on Man- Denmark.
agement, 25±28 July 1998. Shanghai: CHEP and Springer. [18] Nawaz M, Enscore E, Ham I. A heuristic algorithm for the
[5] Co€man E. Computer and job-shop scheduling theory. m-machine n-job ¯ow shop sequencing problem. Omega
New York: Wiley; 1976. 1983;11:91±5.
[6] Dannenbring D. An evaluation of ¯ow shop sequencing [19] Palmer D. Sequencing jobs through a multi-stage process
heuristics. Mgmt Sci 1977;23:1174±82. in the minimum total time ± a quick method of obtaining a
[7] Davis L. Handbook of Genetic Algorithms. New York: near optimum. Oper Res Quart 1965;16:101±7.
Van Nostrand Reinhold; 1991. [20] Sule DR. Industrial scheduling. Boston: PWS; 1997.
1616 W.-T. Chan, H. Hu / Computers and Structures 79 (2001) 1605±1616

[21] Taillard E. Some ecient heuristic methods for the ¯ow [23] Warszawski A. Production planning in prefabrication
shop sequencing problem. Eur J Oper Res 1990;47:65±74. plant. Build Environ 1984;19(2):139±47.
[22] Warszawski A, Ishai E. Long range planning of prefabri- [24] Warszawski A. Industrialisation and robotics in building: a
cation industry in a national economy. Build Environ managerial approach. New York: Harper and Row;
1982;17(1):47±54. 1990.