A Discrete Artificial Bee Colony Algorithm for the Permutation Flow Shop Scheduling Problem with Total Flowtime

Criterion
M. Fatih Tasgetiren, Quan-Ke Pan, P. Nagaratnam Suganthan, Angela H-L Chen
the completion time for the n-job and m-machine problem is calculated as follows:
C ( 1 ,1) = p( 1 ,1)

AbstractVery recently, Jarboui et al. [1] (Computers & Operations Research 36 (2009) 2638-2646) and Tseng and Lin [2] (European Journal of Operational Research 198 (2009) 8492) presented a novel estimation distribution algorithm (EDA) and a hybrid genetic local search (hGLS) algorithm for the permutation flowshop scheduling (PFSP) with the total flowtime (TFT) criterion, respectively. Both algorithms generated excellent results, thus improving all the best known solutions reported in the literature so far. However, in this paper, we present a discrete artificial bee colony (DABC) algorithm hybridized with an iterated greedy (IG) and iterated local search (ILS) algorithms embedded in a variable neighborhood search (VNS) procedure based on swap and insertion neighborhood structures. We also present a hybrid version of our previous discrete differential evolution (hDDE) algorithm employing the IG and VNS structure too. The performance of the DABC and hDDE is highly competitive to the EDA and hGLS algorithms in terms of both solution quality and CPU times. Ultimately, 43 out of 60 best known solutions provided very recently by the EDA and hGLS algorithms are further improved by the DABC and hDDE algorithms with short-term search.

C ( 1 , k ) = C ( 1 , k 1) + p ( 1 , k )

C j ,1 = C j 1 ,1 + p j ,1

(
(

) (
)

) (

j = 2,..., n
k = 2,..., m

(1) (2)

C j , k = max C j 1 , k , C j , k 1 + p j , k

{(

) (

) (

)}

(3)

j = 2,.., n; k = 2,.., m

completion time C ( j , m ) of job j on the last machine m

flowtime of job j . Clearly F ( j ) is equivalent to the

(4) As for the total flowtime criterion, let F ( j ) represent the

since the release times of all jobs are zero. The total flowtime TFT ( ) of a permutation can be computed by summing flowtimes or completion times of all jobs. Then, the total flowtime of a permutation is defined as
TFT ( ) = F j = C j , m . Therefore, the PFSP with the
j =1 j =1 n

( )

the PFSP, solutions are represented by the permutation of n jobs, i.e., = { 1 , 2 ,..., n } . Each job is composed of m operations and every operation is performed by a different machine. Jobs, once initiated, cannot be interrupted (preempted) by another job on the same machine and the release times of all jobs are zero. Thus, given the processing time p( j , k ) for job j on machine k, the PFSP
* * * was to find the best permutation of jobs * = 1 to , 2 ,.., n be processed on each machine subject to the minimization of total flowtime criterion. Let C ( j , m ) denote the completion

In

I. INTRODUCTION

time of job j on machine m. Given the job permutation ,

This research is partially supported by National Science Foundation of China under Grants 60874075, 70871065, 60905039, and Science Research and Development of Provincial Department of Public Edu cation of Shandong under Grant J09LG29. M. Fatih Tasgetiren is with the Department of Industrial Engineering at Yasar University, Izmir, Turkey; Phone: +90 232 411 5301; e-mail: fatih.tasgetiren@yasar.edu.tr). Q-K Pan is with the School of Computer Science, Liaocheng University, China. (e-mail: panquanke@gmail.com). P. N. Suganthan is with the School of Electrical and Electronics Engineering at Nanyang Technological University 639798, Singapore (e-mail: epnsugan@ntu.edu.sg). Angela H-L Chen is with the Department of Finance, Nanya Institute of Technology, Taiwan 320, R.O.C email:achen@nanya.edu.tw)

total flowtime criterion is to find the optimal permutation * in the set of all permutations of such that TFT * TFT ( ) for each permutation belonging to . In order to solve the PFSP with respect to TFT criterion, exact algorithms such as branch and bound algorithm were presented in [3, 4, 5, 6]; constructive heuristics were developed in [7, 8, 9,10, 11, 12, 13, 14]; genetic algorithms (GAs) were presented in [15, 16]; ant colony optimizations (ACOs) were developed in [17, 18]; particle swarm optimizations (PSOs) were presented in [19, 20, 21]; discrete differential evolution (DDE) was developed in [22]; and an iterated local search (ILS) was developed in [23]. Very recently, an estimation of distribution algorithm (EDA) is hybridized with a variable neighborhood search (VNS) has been introduced by Jarboui et al. [1] whereas a hybrid genetic local search (hGLS) has been proposed by Tseng and Lin [2]. Both algorithms have further improved almost all the best known solutions so far in the literature. In general, swarm intelligence is based on collective behavior of self-organized systems. As a typical example of swarm intelligence, the bee swarming around her hive has received significant interest from researchers. Recently, by modeling the specific intelligent behaviors of honey bee swarms, an artificial bee colony (ABC) algorithm is developed by Karaboga in [24-28] to optimize multi-variable and multi-modal continuous functions. Numerical comparisons demonstrated that the performance of the ABC algorithm is competitive to other population-based algorithms with an advantage of employing fewer control

( )

978-1-4244-8126-2/10/$26.00 2010 IEEE

parameters [24-28]. Recently, a discrete version of ABC algorithm is applied to the lot-streaming flowshop scheduling problem in [36]. Since there is no published work to deal with the PFSP with the TFT criterion by using the ABC algorithm, we present a novel discrete ABC (DABC) algorithm and the hybrid version of our previous IG based discrete differential evolution (hDDE) algorithm in [22] for solving the PFSP with the TFT criterion in this paper. The proposed algorithms are hybridized with IG algorithm embedded in the VNS procedure based on swap and insertion neighborhood structures. The main purpose of the hybridization stems from the fact that DABC and DDE carry out the global search by the exploitation of the search space whereas IG fused into VNS is responsible for intensifying the search on the local minima. Therefore, the balance in both global and local search has been effectively achieved. The performance of the proposed algorithms is very competitive to the very recent two best performing algorithms in terms of solution quality and CPU times. Ultimately, 43 out of 60 best known solutions provided very recently by both algorithms are further improved by the proposed algorithms with the short-term search. The long-term search results are not provided for the sake of space limitation. The remaining paper is organized as follows. Section II introduces the DABC whereas Section III presents the hDDE algorithm. The details of the local search algorithms developed for the PFSP with TFT criterion are provided in Section IV. Section V discusses the computational results over benchmark problems. Finally, Section VI summarizes the concluding remarks. II. DISCRETE ARTIFICIAL BEE COLONY ALGORITHM The artificial bee colony (ABC) algorithm is a new swarm intelligence based optimizer proposed by Karaboga [24-28] for multi-variable and multi-modal continuous function optimization. Inspired by the intelligent foraging behavior of honeybee swarm, the ABC algorithm classifies the foraging artificial bees into three groups; namely, employed bees, onlookers and scouts. A bee that is currently exploiting a food source is called an employed bee. A bee waiting in the hive for making decision to choose a food source is named as an onlooker. A bee carrying out a random search for a new food source is called a scout. In the ABC algorithm, each solution to the problem under consideration is called a food source and represented by an n-dimensional real-valued vector, whereas the fitness of the solution corresponds to the nectar amount of the associated food resource. Similar to the other swarm intelligence based approaches, the ABC algorithm is an iterative process. It starts with a population of randomly generated solutions or food sources. Then the following steps are repeated until a termination criterion is met [24-28]: Step 1: Initialization. Step 2: Place the employed bees on their food sources. Step 3: Place the onlooker bees on the food sources depending on their nectar amounts.

Step 4: Send the scouts to the search area for discovering new food sources. Step 5: Memorize the best food source found so far. Step 6: If a termination is not satisfied, go to step 2; otherwise stop the procedure and output the best food source found so far. A. Initialization The permutation based representation is used in both algorithms and a solution is represented by a permutation of jobs = { 1 , 2 ,..., n } . For the initial population in the DABC and hDDE algorithms, the first solution is constructed with the NEH heuristic of [29]. The rest of the solutions in the population are constructed randomly.

B. Employed bee phase

According to the basic ABC algorithm, the employed bees generate food sources in the neighborhood of their current positions. As to the permutation based neighborhood structure, insert and swap operators are commonly used to yield neighboring solutions in the literature [30]. The insert operator of a permutation is defined by removing a job from from its original position j and inserts it into another position k such that (k { j , j 1}) , whereas the swap operator produces a neighbor of by interchanging two jobs of . To enrich the neighborhood structure and diversify the population, six neighboring strategies, denoted as S i , based on the insert, swap operators and destruction and construction procedure (DC) of IG algorithm [30] are separately utilized to generate neighboring food sources for the employed bees as follows:
S1 : Performing one insert operator to a permutation . S 2 : Performing one swap operator to a permutation . S 3 : Performing two insert operators to a permutation . S 4 : Performing two swap operators to a permutation . S 5 : Performing a DC with the destruction size of d = 8 S 6 : Performing a DC with the destruction size of d = 12

Each method for the generation of neighboring food sources may have different performance during the evolution process. Therefore, each food source (individual) in the population is assigned to one of the six strategies to generate a neighboring food source. After generating a neighboring food source, a local search is applied to further improve the solution quality (nectar amount). As for the selection, new good source is always accepted if it is better than the current food source the same as in the basic ABC algorithm which carries out a greedy selection procedure. The local search procedure will be explained in detail in Section IV.

C. Onlooker bee phase

In the basic ABC algorithm, an onlooker bee selects a food source k depending on its winning probability value which is similar to the wheel selection in GAs. However, the tournament selection is widely used in GA applications due to its simplicity and ability to escape from local optima. For this

reason, we propose a tournament selection with the size of 2 in the DABC algorithm. In the tournament selection, an onlooker bee selects a food source k in such a way that two food sources are randomly picked up from the population, and compared to each other, then the better one is chosen. In addition, an onlooker bee utilizes the same strategy as used by the employed bee to produce a new neighboring solution. Then a local search is employed to further improve the nectar amount of the onlooker bee. If the new food source obtained is better than or equal to the current one, the new food source will replace the current one and become a new member in the population. The size of the onlooker bees is 2*NP individuals. The local search procedure will be explained in detail in Section IV.

32]. To obtain the mutant individual, the following equation can be used:
t 1 insert i vit = t 1 swap i

( ) ( )

if

r < Pm

otherwise

(5)

where it 1 is the individual in the target population; Pm is the perturbation probability; and insert is the simple random insertion move whereas swap is the simple interchange of two randomly chosen jobs. A uniform random number r is generated between [0,1]. If r is less than Pm then the perturbation operator is applied to generate the mutant individual as vit = insert it 1 ; otherwise, the individual is

( )

D. Scout bee phase

In the basic ABC algorithm, a scout bee produces a food source randomly in the predefined search space. This will decrease the search efficacy, since the best food source in the population often carries better information than others during the evolution process, and the search space around it could be the most promising region. Therefore, in the DABC algorithm, a tournament selection with the size of two is again used to discard a solution in such a way that two random food sources are picked up and the worst one is selected. Then the scout generates a food source by performing a DC with a destruction size of four (d = 4 ) to the best solution in the population and replaced with the food source determined by tournament selection. The size of the scout bees is 0.2*NP individuals. III. DISCRETE DIFFERENTIAL EVOLUTION ALGORITHM Differential evolution (DE) is one of the latest evolutionary optimization methods proposed by Storn and Price in [31]. Similar to the other evolutionary-type algorithms, DE is a population-based stochastic global optimizer. DE is originally a continuous algorithm where, individuals are represented by chromosomes based on floating-point numbers. A DE algorithm mutates individuals in such a way that the weighted difference between two randomly selected population members is added to a third member to generate a mutated solution. Then, a trial solution is generated in such a way that the mutated solution is recombined with the target solution. Thereafter, a selection operator is applied to compare the fitness function value of both competing solutions, namely, target and trial solutions to determine who can survive for the next generation. Since DE has a continuous nature, it cannot tackle discrete/combinatorial optimization problems. To remedy this drawback, Tasgetiren et al. [32] and Pan et al. [22] proposed a simple and novel discrete DE (DDE) algorithm whose solutions are based on discrete job permutations. In the DDE algorithm, the target individual is represented by a permutation of jobs = { 1 , 2 ,..., n } . The mutant individual is obtained by perturbing each individual in the target population, which is different than the one presented in [22,

perturbed by the swap operator as vit = swap it 1 . Following the perturbation phase, the trial individual is obtained such that: t t 1 CR ( vi , i ) if r < Pc uit = (6) vit otherwise

( )

where CR is any type of crossover operators; and Pc is the crossover probability. In other words, if a uniform random number r is less than the crossover probability Pc, then the crossover operator is applied to generate the trial individual u it = CR vit , it 1 . Otherwise the trial individual is chosen as

. By doing so, the trial individual is made up from the outcome of the perturbation operator or from the crossover operator. Finally, the selection is based on the survival of the fitness among the trial and target individuals such that:

uit

vit

it = ti1 i

t u

if f ( uit ) < f ( it 1 ) otherwise

(7)

IV. HYBRIDIZATION WITH LOCAL SEARCH In order to intensify the search on the local minima and improve the solution quality of DABC and DDE, a common approach is to hybridize it with some local search methods as mentioned in [33]. For this reason, the local search algorithm is, simply a VNS, fused into the DABC and hDDE algorithms. VNS is a metaheuristic proposed in [34] systematically exploiting the idea of neighborhood change, both in descent to local minima and in escape from the valleys containing them. For the PFSP with the TFT criterion, we propose the following VNS local search in Fig. 1 which is embedded in the DABC and DDE algorithms: The key procedure in the DABC and hDDE algorithms is the destruction and construction heuristic. It should be noted that they are used in the both algorithms in different ways. In the hDDE algorithm, the best solution in the population is chosen and IG with VNS is applied to the best solution in the population. However, in the DABC algorithm, they are used as different strategies which are assigned to each individual in the population randomly.

procedure VNS ( ) flag = false

0 =
do

1 = InsertLS ( 0 ) 2 = SwapLS ( 1 ) if ( f ( 2 ) < f ( 0 ))

flag = true 0 =2 else flag = false

swaps and insertions are systematically carried out in such a way that if any improvement is made, the search starts from scratch on the improved solution. In other words, the first improvement pivoting rule has not been employed in both of them.
procedure SwapLS ( )

0 =
i =1 do j = i +1 do

while( flag == true )

=0
return endprocedure
Fig. 1. VNS Algorithm

1 = 0 swap ( 1 , i, j ) if ( f ( 1 ) < f ( 0 )) 0 = 1
i=0 j = i +1 else

procedure DC ( , d )

D = Destruct , R
D

( = Construct (
return

j = j +1

, R

endif while( j < n ) while(i < n ) i = i +1

endprocedure
Fig. 2. Destruction and Construction Procedure

=0
endprocedure
Fig. 3. Swap Local Search Procedure

In the destruction step, a given number d of jobs, randomly chosen and without repetition, are removed from the solution, thus resulting in two partial solutions. The first one with the size d of jobs is denoted as x R and includes the removed jobs in the order in which they are removed. The second one with the size n d of jobs is the original solution without the removed jobs, which is denoted as x D . Finally, the construction phase requires a constructive heuristic procedure. We employ the NEH insertion heuristic of [29]. In order to reinsert jobs in x R into the destructed solution x D , the first jobs x1R is inserted into all possible n d positions in the destructed solution x D generating n d partial solutions. Among these n d partial solutions including job x1R , the best partial solution with the minimum total flowtime is chosen and kept for the next iteration. Then the R is considered and so on until x R is empty or a second job x 2 final solution is obtained. Hence x D is again of size. It is obvious that the speed-up method as in the case of makespan criterion is not applicable for the TFT criterion. For this reason, the NEH heuristic is employed without any speed-up method, i.e., partial solutions are evaluated using the TFT criterion. The VNS local search has two neighborhood structures, namely InsertLS() and SwapLS(), which are similar to those in [1]. InsertLS() evaluates all possible insert moves of pairs of job position (i, j ) as shown in Fig. 3, whereas SwapLS considers possible interchange of pairs of job positions (i, j ) as shown in Fig. 4. Note that in both local search procedures,

procedure InsertLS ( )

0 =
i =1 do j = i +1 do

1 = 0 insert ( 1 , i, j ) if ( f ( 1 ) < f ( 0 )) 0 = 1
i=0 j = i +1 else j = j +1 while( j < n ) i = i +1 while(i < n ) endif

=0
return endprocedure
Fig. 4. Insert Local Search Procedure

procedure InsertFPV ( )

0 =
k =1 counter = 1;

while(counter < n ) k = (k + 1)%n

1 = remove job k from 0 1 = best permutation obtained by inserting k in all possible positions of 1 if ( f ( 2 ) < f ( 0 )) 0 =2
counter = 0 else counter = counter + 1 endif endwhile

However, we also propose an insertion procedure, which is called InsertFPV() and based on the first improvement pivoting rule. In InsertFPV(), a job is removed from a permutation and inserted into n 1 positions, then the permutation with the best out of n 1 insertions is retained for the next iteration. The pseudo code of the InsertFPV() is given in Fig. 5. It is important to note that VNS is applied to the best solution in the target population at each generation. On the other hand, SwapLS() and InsertFPV() are sequentially applied to each trial individual generated by the hDDE algorithm with a small probability p LS of 0.01. Finally, the pseudo code of the hDDE algorithm is given in Fig. 6. As for the DABC algorithm, the following computational procedure is used and the pseudo code is also given in Fig. 7. Step 1: Set the population size NP, p Ls S i for each food source and S max . Note that in the DABC algorithm, the number of employed bees is equal to the number of solutions in the population. However, or the onlookers is 2*NP. Step 2: Initialize the population: = { 1 , 2 ,..., NP } and evaluate each solution in the population. Step 3: Employed bee phase: For i = 1,2,..., NP , repeat the following sub-steps: 1. Produce a new food source u i for the ith employed bee who is associated with the strategy S i and evaluate the new solution. 2. If r < p LS , perform local search to u i . 3. If u i is better than or equal to i , let i = ui and update best so far solution too. Step 4: Onlooker bee phase. For 2 * NP , repeat the following sub-steps: 1. Select a food source in the population for the onlooker bee k by using the tournament selection with size of 2 (Better TFT is chosen) 2. Generate a new solution for the onlooker by using the S i and apply VNS and evaluate it. Then update food source k and best so far solution in the population 3. If the generated solution is better than or equal to the selected one, update the population. Step 5: Scout phase. A tournament selection with the size of 2 is again used to discard a solution in such a way that two random food sources are picked up and the worst one is selected. Then the scout generates a food source by performing a DC with a destruction size of 4 to the best solution in the population and replaced with the food source determined by tournament selection. The size of the scout bees is 0.2*NP individuals. Step 6: Memorize the best solution achieved so far. Step 7: If the termination criterion is reached, return the best solution found so far; otherwise go to Step 3.

=0
return endprocedure
Fig. 5. Insert Local Search Procedure with First Pivoting Rule

i = [ 1 , 2 ,.., NP ] 1 = NEH ( 1 ) b = arg min ( i )

i =1, 2,.., NP

procedure hDDE

do insert ( i ) if r < Pm vi = swap ( i ) else u i = CR (vi , i ) if

i =1, 2,.., NP i =1, 2 ,.., NP

= InsertFPR(u i ) = SwapLS ( i )
i =1, 2 ,.., NP i =1, 2 ,.., NP

(r < PLS )

endif if f ( ) < f (u i ) i =1, 2,.., NP ui =

i =1, 2,.., NP

if

( f ( ) < f ( b ))

b =
endif endif

b = DestructConstruct ( b ) b = VNS ( b )
while(NotTer min ation ) return b endprocedure
Fig. 6. hDDE Algorithm

 = [ 1 , 2 ,.., NP ] 1 = NEH ( 1 ) B = arg min ( i )

i =1, 2,.., NP

procedure DABC
TABLE I COMPARISON TO THE BEST PERFORMING ALGORITHMS:

S i = rand ()% S max

nm 50 5

Tmax = 0.4 n m
hGLS 64853 68173 63367 68281 69551 67013 66294 64560 63029 69037 87599 83001 80224 86787 86646 86826 88996 86860 85841 88293 126073 119300 116856 121028 118736 121066 123580 122770 121872 124354 254619 243817 239075 228291 241255 233583 241458 232283 249269 243879 300634 277209 290198 303669 287136 273172 281306 293628 304276 293465 370603 375982 373554 376236 373524 374705 376998 388058 378474 383283 EDA 64817 68066 63240 68287 69478 66882 66274 64418 62981 68843 87238 83116 80132 86725 86626 86735 89014 87025 85688 88149 125831 119247 116696 120834 118457 120820 123271 122820 121872 124486 254250 243227 238580 228520 241397 233161 241213 231865 249038 243647 301001 275601 288943 303443 286646 271956 281090 293067 303893 293492 368641 374838 372423 374832 371268 375348 376353 387189 377729 381623

SECONDS

do // Employed ui = i u i = S i (u i )
i =1, 2,.., NP i =1, 2 ,.., NP

Bee Phase

if

u i = InsertFPR(u i ) u i = SwapLS (u i ) if
i =1, 2,.., NP i =1, 2,.., NP

(r < p LS )

50 10

DABC 64803 68086 63162 68242 69448 66878 66271 64381 63081 68989 87340 83068 80139 86525 86453 86687 88946 86883 85637 87998 125842 119270 116712 120897 118457 120850 123043 122529 121872 124079 254738 243834 239242 228925 241959 234017 241727 232238 249884 244335 301204 276470 289400 303062 286742 272282 281716 293071 304457 293775 369297 374321 373210 374205 371334 373689 375188 387582 377113 380725

hDDE 64803 68051 63226 68345 69360 66841 66271 64365 63015 68906 87143 82949 80105 86547 86511 86730 89024 86886 85646 88139 125877 119270 116628 120983 118767 120703 123084 122672 122018 124327 254319 243410 238772 228518 241243 233534 240918 231716 249180 243838 300201 275920 289366 303403 285950 271601 280921 292664 303742 293147 368702 374894 372057 375540 370646 373826 376807 386803 377730 380773

( f (u i ) < f ( i ))
i =1, 2,.., NP i =1, 2,.., NP

i = ui

if

( f (u i ) < f ( B ))
i =1, 2,.., NP

50 20

B = ui
endif endif endif

i =1, 2,.., NP

// Onloo ker Bee Phase k = TournamentSelect ( NP )

u k = S k ( k ) if
i =1, 2,.., 2* NP

100 5

u k = VNS (u k )

( f (u k ) < f ( i ))
i =1, 2,.., NP

i =1, 2,.., 2* NP

i = uk

if

( f (u k ) < f ( B ))
i =1, 2,.., 2* NP

i =1, 2,.., 2* NP

B = uk

100 10

else S i = rand ()%S max endif

//

k = TournamentSelect ( NP )
u k = DestructionConstruction( B )
i =1, 2,.., 0.2* NP

Scout

Bee Phase

100 20

while(NotTer min ation ) return B endprocedure

Fig. 7. DABC Algorithm

k = uk

V.

COMPUTATIONAL RESULTS

increased CPU times.

The DABC and hDDE algorithms were coded in Visual C++ and run on an Intel Pentium IV 3.0 GHz PC with 512MB memory. They were applied to the 90 benchmark instances of Taillard in [35] ranging from 20 jobs with 5 machines to 100 jobs with 20 machines. Regarding the parameters of the hDDE algorithm, a small population size of NP = 10 is employed. the destruction and construction procedure with destruction size of 8 (d = 8) is used. The crossover and mutation probabilities are taken as 0.9 and 0.2, respectively. The probability that a local search applied to each trial individual is taken as p LS = 0.01 . To be fair especially with Jarboui et al. [1], the same termination criterion is used as Tmax = 0.4 n m seconds. Note that similar machine speed is also used so that the comparisons will be fair enough. As to the parameters of the DABC algorithm, again the population size is fixed at 10. The number of employed bees are NP , onlooker bees are 2 NP and scout bees are 0.2 NP . Strategies are determined as explained in Section II. R=10 runs were carried out for each problem instance. The best out of 10 replications is reported to be compared to those yielded by the best performing algorithms in the literature. The computational results are given in Table I where the results for hGLS and EDA were taken from [1] and [2], respectively. We do not report the results for 20x5, 20x10 and 20x20 instances since both DABC and hDDE were able to find the best known solution for each instance for each of 10 runs. As seen in Table I, DABC and hDDE algorithms were able to further improve 43 out 60 best known solutions where most of them were reported by Jarboui et al. [1]. In Tseng and Lin [2], extremely long runs are carried out for 50x5, 50x10 and 50x20 instances and the current results are further improved. However, some of our current results are even better than those long runs reported in Tseng and Lin [2]. For the sake of space limitation, we do not report our long runs which are either very competitive or even better than Tseng and Lin [2].
TABLE II COMPUTATIONAL TIME OF ALGORITHMS COMPARED

VI.

CONCLUSIONS

In this paper, we considered the application of DABC and hDDE algorithms to the PFSP with the total flowtime criterion. DABC algorithm is a new one and takes advantage of looking for food sources in the solution space by using destruction and construction procedures and some enhanced local search algorithms. In fact, IG was not used with VNS local search so far in the literature. The paper also indicates the power of IG employing VNS algorithm in the framework of DDE and DABC algorithms. Ultimately, 43 out of 60 best known solutions are further improved by hDDE and DABC algorithms, respectively. REFERENCES
[1] Jarboui B., Eddaly M., Siarry P., An estimation of distribution algorithm forminimizing the total flowtime in permutation flowshop scheduling problems. Computers and Operations Research 36 (2009) 2638 2646 [2] Tseng L.-Y., Lin Y.-T., A hybrid genetic local search algorithm for the permutation flowshop scheduling problem. European Journal of Operational Research 198 (2009) 8492 [3] Chung C.S., Flynn J., Kirca O., A branch and bound algorithm to minimize the total flow time for m-machine permutation flowshop problems. International Journal of Production Economics 79 (2002) 185196. [4] Van de Velde S.L., Minimizing the sum of the job completion times in the two- machine flow shop by Lagrangian relaxation. Annals of Operations Research 26 (1990) 257268. [5] Bansal S.P., Minimizing the sum of completion times of n-jobs over m-machines in a flowshop a branch and bound approach. AIIE Transactions 9 (1977) 306-311. [6] Ignall E., Schrage L., Application of the branch and bound technique to some flow-shop scheduling problems. Operations Research 13 (1965) 400412. [7] Woo H.S., Yim D.S., A heuristic algorithm for mean flowtime objective in flowshop scheduling. Computers and Operations Research 25 (1998) 175182. [8] Framinan J.M., Leisten R., An efficient constructive heuristic for flowtime minimisation in permutation flow shops. Omega 31 (2003) 311317. [9] Liu J., Reeves C.R., Constructive and composite heuristic solutions to the P | Ci scheduling problem. European Journal of Operational Research 132 (2001) 439452. [10] Ho C.J., Gupta J.N.D., Flowshop scheduling with dominant machines. Computers and Operations Research 22 (1995) 237246. [11] Framinan J.M., Leisten R., Ruiz-Uzano R., Comparison of heuristics for flowtime minimisation in permutation flowshops. Computers and Operations Research 32 (2005) 12371254. [12] Rajendran C., Heuristic algorithm for scheduling in a flowshop to minimize total flowtime. International Journal of Production Economics 29 (1993) 6573. [13] Allahverdi A., Aldowaisan T., New heuristics to minimize total completion time in m-machine flowshops. International Journal of Production Economics 77 (2002) 7183.

nm 50 5 50 10 50 20 100 5 100 10 100 20

Avg Machine

hGLS 16.55 40.33 82.23 94.17 251.62 588.86 119.39 AMD 1.83

VNS 31.98 56.18 142.08 174.26 324.37 644.98 152.78 PIV 3.2

EDA 57.22 105.45 240.96 124.55 266.02 570.27 151.95 PIV 3.2

hDDE 43.46 84.94 175.00 174.81 340.21 659.66 164.40 PIV 3.0

DBAC 66.11 111.74 201.99 185.59 364.18 734.03 184.85 PIV 3.0

Regarding the computational times of the algorithms compared, Table II summarizes the results. It seems from Table II that the DABC was computationally the most expensive one. hDDE is competitive to EDA and VNS of Jarboui et al. [1]. The less expensive one was the hGLS of Tseng and Lin [2]. However, hDDE and DABC algorithms yielded very good solution qualities at the expense of slightly

[14] Li X., Wang Q., Wu C., Efficient composite heuristics for total flowtime minimization in permutation flowshops. Omega 37 (2009) 155164. [15] Vempati V.S., Chen C.L., Bulllngton S.F., An effective heuristic for flow shop problems with total flow time as criterion. Computers and Industrial Engineering 25 (1993) 219222. [16] Zhang Y., Li X., Wang Q., Hybrid genetic algorithm for permutation flowshop scheduling problems with total flowtime minimization. European Journal of Operational Research 196 (2009) 869-876. [17] Gajpal Y., Rajendran C., An ant-colony optimization algorithm for minimizing the completion-time variance of jobs in flowshops. International Journal of Production Economics 101 (2006) 259272. [18] Rajendran C., Ziegler H., Ant-colony algorithms for permutation flowshop scheduling to minimize makespan/total flowtime of jobs. European Journal of Operational Research 155 (2004) 426438. [19] Jarboui B., Ibrahim S., Siarry P., Rebai A., A combinatorial particle swarm optimisation for solving permutation flowshop problems. Computers and Industrial Engineering 54 (2008) 526538. [20] Tasgetiren M.F., Liang Y.-C., Sevkli M., Gencyilmaz G., A particle swarm optimization algorithm for makespan and total flowtime minimization in the permutation flowshop sequencing problem. European Journal of Operational Research 177 (2007) 1930-1947. [21] Tasgetiren M.F., Sevkli M., Liang Y.-C., Gencyilmaz G., Particle swarm optimization algorithm for permutation flowshop sequencing problem. In: Proceedings of Ant Colony Optimization and Swarm Intelligence (ANTS2004), LNCS 3172, 2004, Springer-Verlag, pp. 381-389. [22] Pan Q.K., Tasgetiren M.F., Liang Y.-C., A discrete differential evolution algorithm for the permutation flowshop scheduling problem. Computers and Industrial Engineering 55 (2008) 795-816. [23] Dong X., Huang H., Chen P., An iterated local search algorithm for the permutation flowshop problem with total flowtime criterion. Computers and Operations Research 36 (2009) 1664 1669 [24] Karaboga D., A new design method based on artificial bee colony algorithm for digital IIR filters. Journal of the Franklin Institute 346 (2009) 328-348. [25] Karaboga D., Basturk B., On the performance of artificial bee colony (ABC) algorithm. Applied Soft Computing 8 (2008) 687-697. [26] Karaboga D., Akay B., A comparative study of artificial bee colony algorithm. Applied Mathematics and Computation 214 (2009) 108-132. [27] Karaboga D., An idea based on honey bee swarm for numerical optimization, Technical Report TR06, Computer Engineering Department, Erciyes University, Turkey, 2005. [28] Karaboga D., Basturk B., A powerful and efficient algorithm for numerical function optimization: artificial bee colony (ABC) algorithm. Journal of Global Optimization 39 (2007) 459-471. [29] Nawaz M., Enscore Jr. E. E., Ham I. A., Heuristic algorithm for the m-machine, n-job flow shop sequencing problem. OMEGA 11(1) (1983) 91-95. [30] Ruiz R., Sttzle T., A simple and effective iterated greedy algorithm for the permutation flowshop scheduling problem. European Journal of Operational Research 177(3) (2007) 2033-2049.

[31] Storn R., Price K., Differential evolution a simple and efficient adaptive scheme for global optimization over continuous spaces. ICSI, Technical Report TR-95-012, 1995. [32] Tasgetiren M.F., Pan Q.-K., Liang Y.-C., A discrete differential evlution algorithm for the single machine total weighted tardiness problem with sequence dependent setup times. Computers and Operations Research 36 (2009) 1900-1915. [33] Lozano J.A., Larranaga P., Inza I., Bengoetxea E., Towards a New Evolutionary Computation Advances on Estimation of Distribution Algorithms. Berlin: Springer; 2006. [34] Mladenovic N., Hansen P., Variable neighborhood search. Computers and Operations Research 24 (1997) 1097-1100. [35] Taillard E., Benchmarks for basic scheduling problems. European Journal of Operational Research 64 (1993) 278-285. [36] Pan Q-K, Tasgetiren MF, Suganthan P. N, Chua TJ, A discrete artificial bee colony algorithm for the lot-streaming flow shop scheduling problem, Inform. Sci. (2010), doi:10.1016/j.ins.2009.12.025