Professional Documents
Culture Documents
School of Mechanical and Production Engineering School of Electrical and Electronics Engineering
Nanyang Technological University Nanyang Technological University
Nanyang Avenue Nanyang Avenue
SINGAPORE 639798 SINGAPORE 639798
This work aims to explore an evolutionary computa- organized model enables proportionate feedback on good-
tion approach, which is based on nectar collection in honey ness of food sources.
bee colonies, to job shop scheduling problems. This re-
search is inspired by the work done by Nakrani and Tovey 3 JOB SHOP SCHEDULING
(2004), on using a new honey bee algorithm for dynamic
allocation of internet servers. In their algorithm, servers Job shop scheduling is concerned with finding a sequential
and HTTP request queues in an Internet server colony are allocation of competing resources that optimizes a particu-
modeled as foraging bees and flower patches respectively. lar objective function. The deterministic job shop schedul-
The experimental results show that the algorithm performs ing problem is the most general of the classical scheduling
reasonably well in the dynamic allocation problem. Based problems (Jain and Meeran 1999). The problem consists of
on similar idea of honey bee colonies and the behavior of a finite set J of n jobs to be processed on a finite set M of m
forager bees, which is characterized by decentralized and machines. Each job Ji must be processed on every machine
elementary interactions, we adapt the algorithm to job shop and consists of a chain of mi operations Oi1, Oi2,…,Oim,
scheduling problem. which have to be scheduled in a pre-determined given or-
This paper first describes how honey bee colonies de- der. Oij is the jth operation of job Ji which has to be proc-
ploy forager bees to collect nectar amongst diverse flower essed on a machine Mx for a processing time period of τij
patches in section 2. Subsequently in section 3, job shop without interruption and preemption. Each machine can
scheduling problem is discussed. In section 4, the mapping process only one job and each job can be processed by
of job shop scheduling meta-heuristic to honey bees forager only one machine at a time. The longest duration in which
deployment is given. Subsequently, the implementation de- all operations of all jobs are completed is referred to as the
tails are discussed in section 5. This is then followed by a makespan Cmax.
comparative study on the performance of the honey bee ap- More specifically, let Ai be the set of ordered pairs of
proach on benchmark problems in section 6. The paper fi- operations constrained by the precedence relations for each
nally ends with conclusions and future works in Section 7. job Ji. For each machine Mx, the set Ex describes the set of
all pairs of operations to be performed on the machine. For
2 HONEY BEE COLONY each operation Oij, let its earliest possible process start time
be tij. Hence, the job shop scheduling problem can be mod-
Colonies of social insects such as ants and bees have instinct eled as:
ability known as swarm intelligence (Nakrani and Tovey
2004, Teodorovic and Dell'orco, 2005). This highly organ- ti(j+1) – tij ≥ τij ∀(Oij, Oi(j+1)) ∈ Ai,
ized behavior enables the colonies of insects to solve prob- tij – tkl ≥ τkl or tkl – tij ≥ τij ∀(Oij, Okl) ∈ Ex,
lems beyond capability of individual members by function-
ing collectively and interacting primitively amongst Although many objective functions can be considered
members of the group. In a honey bee colony for example, in job shop scheduling problems, makespan is the principal
this behavior allows honey bees to explore the environment criterion for research community and is able to capture the
in search of flower patches (food sources) and then indicate fundamental computational difficulty which exists implic-
the food source to the other bees of the colony when they itly in determining an optimal schedule. Further, makespan
return to the hive. Such a colony is characterized by self- minimization problem is well defined and is simple to un-
organization, adaptiveness and robustness. derstand. It is used in our work as a proving ground for the
Seeley (1995) proposed a behavioral model of self- proposed honey bee colony algorithm for job shop schedul-
organization for a colony of honey bees. In the model, for- ing problems.
aging bees visiting flower patches return to the hive with A common representation for job shop scheduling
nectar as well as a profitability rating of respective patches. problem is the disjunctive graph. In the graph, there is a
The collected nectar provides feedback on the current node for each operation. There are also two additional
status of nectar flow into the hive. The profitability rating nodes, known as source and sink to represent the initial and
is a function of nectar quality, nectar bounty and distance final operations respectively. The positive weight of each
from the hive. The feedback sets a response threshold for node is equivalent to the processing time of the corre-
an enlisting signal which is known as waggle dance, the sponding operation. The starting and completion times of
length of which is dependent on both the response thresh- the source and sink represent the starting and finishing
old and the profitability rating. The waggle dance is per- times of the job shop problem respectively. The source is
formed on the dance floor where individual foragers can connected to the initial operation of each job whereas the
observe. The foragers can randomly select a dance to ob- final operation of each job is connected to the sink.
serve and follow from which they can learn the location of A set of directed conjunctive arcs is used to represent
the flower patch and leave the hive to forage. This self- the precedence constraints for each job. A set of disjunc-
tive arcs is used to represent capacity constraints to ensure
1955
Chong, Low, Sivakumar, and Gay
that no two operations processed by the same machine can m1 O11 O22 O32
be executed simultaneously. An example of a 3 x 3 dis-
junctive graph is shown in Figure 1. The conjunctives arcs m2 O12 O23 O33
are shown by solid lines while the dotted lines represent
disjunctive arcs. The set of operations on each machine is m3 O21 O31 O13
given by: M1 = {O11, O22, O32}, M2 = {O12, O23, O33} and Time
M3 = {O13, O21, O31}
Solutions to a job shop scheduling problem can be ob- Figure 3: Gantt Chart of the Feasible Solution
tained by directing the disjunctive arcs of the disjunctive
graph according to the machine permutations (i.e. by mak- 4 HONEY BEE COLONY ALGORITHMS
ing bidirectional arcs to become unidirectional arcs). The
makespan of a solution is the length of the longest directed This section details algorithms to perform job shop sched-
path in the graph. The length of the path is given by the uling inspired by the behavior of honey bee colony. The
sum of the processing times of the operations on that path. challenge is to adapt the self-organization behavior of the
A feasible solution to the earlier job shop scheduling prob- colony for solving job shop scheduling problems. There
lem is presented in Figure 2. are two major characteristics of the bee colony in searching
for food sources: waggle dance and forage (or nectar ex-
3 5 2 ploration). We will discuss in separate sub-sections on how
O11 O12 O13 we map these characteristics of a bee colony to job shop
scheduling.
0 2 3 2 0
0 O21 O22 O23 * 4.1 Waggle Dance
1956
Chong, Low, Sivakumar, and Gay
Table 1: Lookup Table for Adjusting Probability of Fol- Based on the expression, it should be noted that for the
lowing a Waggle Dance first nectar exploration expedition by the foragers, ρ ij will
Profitability Rating ri be assigned the same value for all allowed nodes (since m
Pfi < 0.9Pfcolony 0.60 = 0).
0.9Pfcolony ≤ Pfi < 0.95Pfcolony 0.20 The parameters α and β tune the relative importance in
0.95Pfcolony ≤ Pfi < 1.15Pfcolony 0.02 probability of the ‘weight’ in edges found in the preferred
1.15Pfcolony ≤ Pfi 0.00 path versus the heuristic distance. According to this rule,
edges that are found in the preferred path and that are
4.2 Forage (Nectar Exploration) shorter will have a higher probability to be chosen for the
solution. The heuristic distance is the processing time of
For foraging algorithm, a population of l foragers is de- the operation associated with nodej. When a forager com-
fined in the colony. These foragers cyclically construct so- pletes a full path, the edges it has traveled and the
lutions to the job shop scheduling problems. The foragers makespan of the resulting solution will be kept for the
move along branches from one node to another node in the waggle dance when it returns to the hive.
disjunctive graph and so construct paths representing solu-
tions. A forager must visit every node once and only once 4.3 Algorithmic Framework
in the graph, starting from initial node (i.e. source) and fin-
ishing at final node (i.e. sink), so as to construct a complete A combination of forage and waggle dance algorithms
solution. When a forager is at a specific node, it can only constitutes one cycle (or iteration) in this evolutionary
move to next node that is defined in a list of presently al- computation approach. This computation will run for a
lowed nodes, imposed by precedence constraints of opera- specific number of iterations Nmax. The best solution during
tions. A forager chooses the next node from the list accord- the iteration process will be presented as final schedule at
ing to the state transition rule: the end of run. The algorithmic framework of the schedul-
ing algorithm is presented in Algorithm 1.
β
⎡1⎤
[ρ (t )]
ij
α
⋅⎢ ⎥ ALGORITHM 1: Algorithmic framework for job shop
scheduling problem
Pij (t ) = ⎣⎢ d ij ⎦⎥
β
⎡1⎤ for i = 1 to Nmax
∑ [ρ (t )]
ij
α
⋅⎢ ⎥
for j = 1 to l
Forage
j∈allowed _ nodes ⎣⎢ d ij ⎦⎥ Save best solution
Waggle dance
end for
where, end for
1957
Chong, Low, Sivakumar, and Gay
1958
Chong, Low, Sivakumar, and Gay
concentration of highly volatile chemicals called phero- Table 2: Parameter Settings for the Algorithms
mones in their immediate environment. Since both ant al- Bee Ant Tabu
Parameter
gorithms and bee algorithms are based on self-organization colony colony search
of colonies to forage food sources, we reason it is orderly Maximum number of 1000 1000 500
to compare the performance of both algorithms. iterations, Nmax
Tabu search is a meta-heuristic based on a local search Population size, l Number Number
technique which attempts to exploit the solution space be- of jobs of jobs
yond local optimality. It is a simple technique that attempts Alpha, α 1.0 1.0
to guide a search process away from solutions that appear Beta, β 1.0 1.0
to duplicate or resemble previously achieved solutions.
Rating, ρij 0.9
tabu search is an iterative approach improving on existing
Scaling factor, A 100
solutions and thus requires an initial solution to be con-
Evaporation coeffi- 0.01
structed by other scheduling techniques. A good neighbor-
hood structure plays an important role in tabu search, and it cient, ρ
is one of the reasons why tabu search algorithm developed Maximum number of 20 20
by Nowicki and Smutnicki (1996) is so successful. Their elite solution
algorithm is one of the most efficient tabu search imple- Maximum size of tabu 8
mentations for the performance criteria of makespan list
(Blazewicz et. al. 1996). By considering tabu search heu-
ristic in our study, we can thus benchmark the proposed Table 3: Relative Performance of Bee Colony, Ant Colony
bee algorithm to one of the best in class. and Tabu Search Heuristics
Relative Bee Ant Tabu
6.3 Experiments and Results Improvement colony colony search
Mean (%) 11.08 11.45 5.17
Since both the bee and ant algorithms rely on random dis- Minimum (%) 0 0 0
tribution function to construct solutions, a total of 5 repli- Maximum (%) 38.09 38.70 24.23
cation runs have been performed for each job shop prob- Number of best 14 13 27
lem to obtain average results. Further, we have performed solutions
fine tuning on the parameters for the algorithms, and the Relative execu- 1.09x 1.19x 1x
final settings of major parameters are presented in Table 2. tion time
Our initial experimentation indicates that the machine
centric approach to operation scheduling always lead to Comparing the performance of peers, bee colony and
better solutions comparing to operation centric approach. ant colony heuristics, bee algorithm performs slightly bet-
We therefore adopt machine centric approach in our ex- ter than ant algorithm. Bee algorithm achieves better mean
periments. and maximum percentages as well as higher number of
Table 3 summarizes the results on the relative per- best solutions in comparison to ant algorithm. The time
formance of makespan in terms of percentage for bee col- taken to solve the 82 job shop problems for both heuristics
ony, ant colony and tabu search algorithms. The results is approximately equal with bee colony being slightly
show the average, minimum and maximum percentage dif- faster. Evidently, both heuristics under perform tabu search
ferences from the best known makespan for the 82 job primarily due to: 1) solutions are always constructed from
shop problems. The second last row records the number of scratch, instead of the more efficient operation swaps in
best solutions achieved among the three heuristics. The last tabu search; 2) no clear scheme to escape from being
row exhibits the relative execution time for the three heu- locked into local minimums.
ristics.
From the results, it is obvious that tabu search outper- 7 CONCLUSIONS AND FUTURE WORKS
forms other two heuristics. Tabu search records the closest
results to the best known solutions and has the most num- We have implemented a job shop scheduling algorithm
ber of best solutions. Further, it also manages to achieve based on self-organization of honey bee colony for solving
best results in the shortest execution time. These spectacu- job shop scheduling problems. It is found that the perform-
lar results are attributed to the efficient critical block ance of the algorithm is comparable to ant colony algo-
neighborhoods. Moreover, a tabu list that keeps track of rithms, but gaps behind the efficient tabu search heuristics.
the most recent tabu moves prevents the search algorithm Since the bee algorithm is our first implementation, we be-
to be locked in local minimums. lieve there is much room for improvement.
We intend to test the bee algorithms on semiconductor
manufacturing scheduling problems. One of the works we
1959
Chong, Low, Sivakumar, and Gay
intend to pursue is to deploy the algorithms in a distributed Dorigo, M., Maniezzo, Vittorio, Colorni, Alberto, Ant sys-
computing environment using software agents. Prior work tem: optimization by a colony of cooperating agents.
has already been carried out using agents in symbiotic IEEE Transactions on Systems, Man, and Cybernetics,
simulation of semiconductor assembly and test operation Part B: Cybernetics, 1996. 26(1): p. 29-41.
(Low et. al. 2005). In comparison to ant colony algorithms, Fisher, H. and Thompson G. L., “Probabilistic learning
it will be relatively easier to treat each bee forager as an combination of local job shop scheduling rules,” In-
agent in the bee colony algorithms since the issue of share dustrial Scheduling, (1963), pp. 225-251.
state in maintaining a global pheromone table in ant algo- Ganesan, V. K., Sivakumar A. I., and Srinivasan G, “Hier-
rithms does not occur. archical minimization of completion time variance and
Further works can obviously be done to enhance the makespan in jobshops,” Computers & Operations Re-
bee algorithms. A research area to focus on is to incorpo- search, (In press: date of accept: 03-Aug-2004).
rate local search heuristics based on an effective critical Gere, W. S., Jr., “Heuristics in jobshop scheduling,” Man-
block neighborhood structure on every schedule generated agement Science, Vol. 13, No. 1 (1966), pp. 167-175.
by forage heuristic. Another area of research is to consider Jain. A. S. and Meeran. S., “Deterministic job shop sched-
penalizing edges (lower rating ρ ij ) that are in opposite di- uling: past, present and future,” European Journal of
Operational Research, Vol. 113, No. 2 (1999), pp.
rections to the edges in the preferred path. As operation 390-434.
centric and machine centric scheduling methods have their Lawrence, S. “Resource constrained project scheduling: an
respective strengths and weaknesses, it is proposed that experimental investigation of heuristic scheduling
both methods can be used jointly in the honey bee colony techniques (supplement),” Graduate School of Indus-
algorithm. trial Administration, Carnegie Mellon University,
Pittsburg, 1984.
REFERENCES Low, Y. H., Lye, K. W., Lendermann, P., Turner, S. J.,
Leo, S. and Chin, R., “An agent-based approach for
Abbas, H.A. and J. Teo. A true annealing approach to the managing symbiotic simulation of semiconductor as-
marriage in honey–bees optimization algorithm. in sembly and test operations,” in Proceedings of the
The Inaugural Workshop on Artificial Life (AL'01). 2005 International Conference on Autonomous Agent
2001. Adelaide, Australia. and Multiagent Systems (AAMAS), July 25-29, 2005,
Adams, J., Balas, E. and Zawack, D., “The shifting bottle- Utrecht, The Netherlands.
neck procedure for job shop scheduling,” Management Nakrani, S. and C. Tovey, On honey bees and dynamic al-
Science, Vol. 34, No. 1 (1988), pp. 391-401. location in an internet server colony. Adaptive Behav-
Applegate, D. and Cook, W., “A computational study of ior, 2004. 12(3-4): p. 223-240.
the job shop scheduling problem,” ORSA Journal on Nanot, Y. R. “An experimental investigation and compara-
Computing, Vol. 3, No. 2 (1991), pp. 149-156. tive evaluation of priority disciplines in jobshop like
Balas, E. and Vazacopoulos, A., “Guided local search with queuing networks,” Management Science Research
shifting bottleneck for job shop scheduling,” Man- Project, UCLA, Research report, No. 87 (1963).
agement Science, Vol. 44, No. 2 (1998), pp. 262-275. Nowicki, E. and Smutnicki, C., “A fast taboo search algo-
Blazewicz, J., Domschke, W. and Pesch, E., “The job shop rithm for the job shop problem,” Management Science,
scheduling problem: conventional and new solution Vol. 42, No. 6 (1996), pp. 797-813.
techniques,” European Journal of Operational Re- Rajendran, C. and Holthaus, O., “A comparative study of
search, Vol. 93, No. 1 (1996), pp. 1-33. dispatching rules in dynamic flowshops and job-
Blum, C. and M. Sampels, An ant colony optimization al- shops,” European Journal of Operational Research
gorithm for shop scheduling problems. Journal of Vol. 116, No. 1 (1999), pp. 156-170.
Mathematical Modelling and Algorithms, 2004. 3(3): Seeley, T. D., The Wisdom of the Hive, Publication Har-
p. 285 - 308. ward University Press.
Brucker, P., Jurisch, B. and Sievers, B., “A branch and Storer, R. H., Wu, S. D. and Vaccari, R., “New search
bound algorithm for the job shop scheduling prob- spaces for sequencing problems with application to job
lem,” Discrete Applied Mathematics, Vol 49, No. 1 shop scheduling,” Management Science, Vol. 38, No.
(1994) pp. 109-127. 1 (1992), pp 1495-1509.
Buffa, E. S., Operations Management: Problems and Mod- Sumpter, D.J.T. and S.C. Pratt, A modelling framework for
els, John Wiley & Sons (New York, 1968) pp. 454- understanding social insect foraging. Behavioral Ecol-
460. ogy and Sociobiology, 2003. 53: p. 131-144.
Chong, C.S., Using simulation based approach to improve Teodorovic, D and Dell'orco, M, Bee colony optimization -
on the mean cycle time performance of dispatching A cooperative learning approach to complex transpor-
rules. in Proceedings of the 2005 Winter Simulation
Conference. 2005.
1960
Chong, Low, Sivakumar, and Gay
tation problems, Advanced OR and AI Methods in Professor ROBERT GAY obtained his B. Eng, M. Eng
Transportation, 2005, pp. 51-60 and PhD degrees at the University of Sheffield, England, in
Yamada, T. and Nakano, R. A genetic algorithm applicable 1965, 1967 and 1970 respectively. He was awarded the
to large-scale job-shop problems. In Manner, R., Man- Grouped Scholarship in Engineering and Metallurgy by the
derick, B. (Eds.), Proceedings of the Second Interna- University of Sheffield from 1967 to 1970. Since obtaining
tional Workshop on Parallel Problem Solving from his PhD, he has been involved in Education and R&D
Nature (PPSN'2), Brussels, Belgium, 1992. pp. 281- working in institutions such as Singapore University
290 (1972-1979), Rutherford and Appleton Laboratory (Eng-
land, 1979-1982), NTU (1982-1995 and 1999-present) and
AUTHOR BIOGRAPHIES Singapore Institute of Manufacturing Technology (1989-
1999). He has also been actively involved in promoting in-
CHIN SOON CHONG obtained his degree in Electrical novation in Singapore through work in various commit-
and Electronics Engineering from the City University of tees: Science Quiz (MOE), Science Centre Board, National
London, UK. He joined Singapore Institute of Manufactur- CAD/CAM (NCB), Tan Kah Kee Young Inventors Award
ing Technology (SIMTech), and is currently in the Plan- (TKK Foundation & NSTB), National IT Plan (NCB),
ning Operation Management Group. He obtained his Mas- Technopreneurship incubation center (ITE), Commercenet
ter of Engineering in Computer Integrated Manufacturing Singapore, Singapore Computer Society. Currently Profes-
from Nanyang Technological University, Singapore. He sor Gay is in the School of EEE at NTU and also Director
has been involved in simulation, scheduling and optimiza- and CEO of the ASP Centre. He has more than a hundred
tion related projects in logistic and manufacturing IT do- publications in journals, conference proceedings and
mains. The projects include cargo container operation si- books. His email and web addresses are
mulation, printing process simulation, manufacturing <eklgay@ntu.edu.sg> and <www.ntu.edu.sg/
cycle-time modeling, scheduling and optimization for eee/icis/cv/robertgay.html>.
MNCs. His current research interest includes simulation,
planning, scheduling, optimization in the area of manufac- Associate Professor APPA IYER SIVAKUMAR (Sen-
turing, logistic, and supply chain. He can be reached via ior member IIE) is an Associate Professor in the School of
email at <cschong@simtech.a-star.edu.sg>. Mechanical and Production Engineering at Nanyang Tech-
nological University (NTU), Singapore and a Fellow of
Dr. MALCOLM YOKE HEAN LOW is an Assistant Singapore Massachusetts Institute of Technology (MIT)
Professor in the School of Computer Engineering at the Alliance (SMA). Prior to this he was at Singapore Institute
Nanyang Technological University (NTU), Singapore. of Manufacturing Technology, Singapore (SIMTech). His
Prior to this he was with the Singapore Institute of Manu- research interests are in the area of simulation-based opti-
facturing Technology, Singapore (SIMTech). He received mization of manufacturing performance, supply chain, and
his Bachelor and Master of Applied Science in Computer dynamic schedule optimization. Prior to joining SIMTech
Engineering from NTU in 1997 and 1999 respectively. He in 1993, he held various management positions including
was awarded a Gintic (now SIMTech) Postgraduate Scho- technical manager and project manager for nine years at
larship in 1999. In 2002, he received his D.Phil. degree in Lucas Systems and Engineering and Lucas Automotive,
Computer Science from Oxford University. His current re- UK. He received a Bachelors of Engineering from Univer-
search interest is in the application of parallel/distributed sity of Bradford, UK and a PhD in Manufacturing Systems
simulation, grid computing and agent technology for the Engineering from University of Bradford, UK. He has been
modeling, simulation, analysis and optimization of com- the technical chair and co-edited the proceedings of the 3rd
plex systems. His e-mail and web address are and 4th International Conference on Computer Integrated
<yhlow@ntu.edu.sg> and <www.ntu.edu.sg/ Manufacturing (ICCIM ’95 and ICCIM’97), Singapore.
home/yhlow/>. His email and web addresses are msiva@ntu.edu.sg
and <www.ntu.edu.sg/home/MSiva/>.
1961