Low-Ievel Teamwork Hybridization (L TH) for P-metaheuristics : A Review and Comparative S.Masroml, Siti Z.Z. Abidin', P.N.Hashimahl, A.S.Abd.

Rahman2

1 Faculty of Computing and Mathematical Sciences Universiti Teknologi MARA, Malaysia {suray078, phashi655}@perak.uitm.edu.my sitizaleha533@salam.uitm.edu.my

2Computer and Information Science Department Universiti Teknologi PETRONAS, Malaysia sanirahman@petronas.com.my

Abstract. Inspired by nature, many types of Population based metaheuristics or P-metaheuristics is cropping out of research labs to help solve real life problems. Since every metaheuristics has its own strength and weaknesses, hybridizing the algorithms can sometimes produce better results. Among hybridization techniques available we are particularly interested in the Low-level Teamwork Hybridization (L TH). L TH operates at the internal structure of algorithms, thus allowing researchers to finely select the components from the algorithms to be hybridized. However, L TH technique is quite complicated since it requires the restructuring of the original algorithms. This paper identifies and describes the internal structures of P-metaheuristics that are suitable candidates for hybridization using L TH. The review and comparative study of several implementations of this technique is also presented.

Keywords: P-metaheuristics;intemal structures;hybridization; LTH.

1 Introduction

Metaheuristics have been considered among the best method for many types of application especially in the optimization problems such as scheduling, routing, resource allocating and time tabling. To date, a variety of metaheuristics have been introduced with different and specific characteristics. Depends on the characteristics, a metaheuristic can be grouped either as single based(S-metaheuristics) or population based (P-metaheuristics). Genetic Algorithm (GA) [1], Ant Colony Optimization (ACO) [2], Particle Swarm Optimization (PSO) [3], Bee Colony (Be) [4] and Differential Evolution (DE) [5] are a few examples of P-metaheuristics. As for the single based, the common metaheuristics are Simulated Annealing (SA) [6] ,Tabu Search (TS) [7] and Variable Neighborhood Search (VNS)[8].

Over the last years, many researchers have agreed that relying on sole metaheuristics was quite restricted in achieving best solution for optimization problem. Moreover, the "no free lunch theorem" [9] has explained that no single optimization strategy will be always better than any other. Therefore, hybrid metaheuristics have been widely accepted as an effective approach compared to the single implementation.

The numerous implementations of hybrid metaheuristics have attracted many researchers to cluster hybrid metaheuristics techniques into several classifications and introduced different taxonomy [10]. The classification schemes however, are not specifically grouped with regards on the internal structures of P-metaheuristics. Identification of the internal structures that are suitable for hybridization has been given in this paper.

The remaining content of this paper is organized as follows. Section 2 gives a brief background about different classification schemes and taxonomies that have been proposed for hybrid metaheuristics. Then the different hybrid classification is presented in section 3. Section 4 and 5 provide review and comparison study on several implementations of L TH respectively. Finally, in section 6, we conclude the paper with a short summary.

2 Related Works

In this section, the focus is directed towards other research works that tackles relevant issues in classification, methodology and taxonomy of metaheuristics hybridization. As introduced in [10], the hybrid metaheurisitics scheme can be divided into three general forms. There are component exchange among metaheuristics, cooperative search from different metaheuristics and integration with others methods. The book however, does not provide detailed explanation on the first category, which mostly related to metaheuristics internal structure. Many of the implementation examples in the rest of chapter are rather focused on the third category only.

A more detailed classification can be seen in [11]. The authors primarily distinguish hybrid metaheuristics according to four criteria namely the kinds of algorithms, the level of hybridization, the order of execution, and the control strategy. Although level of hybridization was considered in the classification scheme, it is however did not involved the elements of internal structure.

Mohammed El-Abd and Mohamed Kamel have introduced new taxomomy for hybrid metaheuristics [12]. The taxonomy is created according to the algorithms involved in hybrid and space decomposition. Nevertheless, it is just focus on parallel implementation of multiple metaheuristics that is not required for internal structure modifications. On the contrary, a classification scheme that was concentrated with hybrid evoloutionary algorithms (EAs) has been described in [13]. In this work, different methodologies and architectures of hybrid EAs has been illustrated with variation of implementations but all of them were mostly associated with general combination of algorithms.

In a more detailed view, E. G. Talbi in the Journal of Heuristics [14] have introduced another taxonomy for hybrid metaheuristics techniques. The classification

has been grouped into hierarchical and flat scheme. Low-level Relay Hybridization (LRH), Low-level Teamwork Hybridization (LTH), High-level Relay Hybridization (HRH) and High-level Teamwork hybridization (HTH) are the four classifications regarding to the hierarchical scheme. Each class from the hierarchy scheme can be associated to the certain criteria from the flat scheme such homogeneous versus heterogeneous and global versus partial.

We have gained a lot of inspirations from the Talbi classification scheme especially to the L TH technique. L TH is a hybrid technique that embedded other metaheuristics in P-metahaeuristics. It has been shown from a comparative study that L TH method has sparked the most interest among researchers compared to other schemes [14]. In term of the hybrid effectiveness, LTH has been proved to outperform LRH and HRH[15].

The implementation of L TH can be considered as quite complicated. It is involved directly to the modification of internal structure of the hybridized algorithms. In fact, choosing a suitable metaheuristics combination at the right internal structures will contribute to some significant impacts to the hybridization effectiveness[ 11].

As to concern with the difficulties, a study should be done to identify and describes the internal structures of P-metaheuristics that are suitable candidates for hybridization using L TH. This study can provide some guidance for the L TH implementation thus make new classification is possible.

3 Internal Structure of P-metaheuristics

In P-metaheuristics, the internal structure that is suitable for LTH can be classified as population initialization (PI) and population generation (PG).

PI method has some significant influence in the effectiveness of the metaheuristics. Nevertheless, the determination of PI is often overlooked during the design phase [16]. In P-metaheuristics, the initial populations are naturally diversified which benefits in reducing premature convergence problem of P-metaheuristics. Unfortunately, in some cases, a sole P-metaheuristic are not enough diversified. This is why hybridization is essential at the population initialization.

In P-metaheuristics, there are two categories of PG strategy have been formed.

One is related to evolution process, involved selection and reproduction using variation operators, namely as crossover and mutation. Evolutionary algorithms (EAs) [1] are the most well-known P-metaheuristics in this type. The second category is associated on utilizing shared memory mechanism which is used as input information for the next population generation. ACO for example, uses pheromone matrix information in each population generation while PSO updates particles velocity, location, personal and global best solutions.

4 A Review of L TH

A review on several L TH has been done based on the internal structure of Pmetaheuristics.

4.1 Population Initialization (PI)

A remarkable result has been achieved from PI hybridization technique that combines single-parent evolution into GA population initialization [17]. The hybridization has shown an improvement when applied to the route optimization process for wireless sensor network. Similarly, an insertion of differential evolution (DE) for PI in genetic algorithm (GA) has been proved to improve the solution of job shop scheduling problem [18]. Hybridization between GA, Multiple Phase Neighbourhood Search GRASP (MPNS-GRASP) and PSO [19] has also produced significant results when applied to vehicle routing problem (VRP). In this research, the MPNS-GRASP has been used for the population initialization. Alternatively, ACO has also been applied with EAs based population initialization like GA [20] as to get high population quality. Efficiency and effectiveness of the hybridization of tabu search into ACO has also been achieved when applied to the open vehicle routing problem [21].

4.2 Population Generation (PG)

Hybridization ofPG is the most popular method for P-metaheuristics. The variants of EA for example, have different crossover and mutation operators. Any of these operators are possible to be used in any single p-metaheuristics among the EAs. The hybridization technique has been demonstrated with some empirical experiments. All the results show that hybrid EAs outperform the individual ones in all of the benchmark problems[22].

A similar work that hybrid EA with estimation distribution algorithm (EDA) also proved to give better results than the individual GA and EDA[23]. However, in this work, it has introduced several participation function (PF) which describes a ratio of how many individual has been selected from each EA and EDA. All the PF factors in the hybridization produced successful results when tested on the satisfiability (SAT) problem [24].

In some techniques, S-metaheuristics strategy has been used as a selection method in the EAs for example GA with local search [25], GA with simulated annealing[26], EAs with taboo search [27] and GA with Variable Neighbourhood Search (VNS) [18].

On the other hand, there is a way for researcher to hybrid memory based Pmetaheuristics before the stage of selection process in the evolution based Pmetaheuristics. For example, PSO strategy can be used to evolve each individual of GA population after the crossover and mutation operations[19, 28]. This is to ensure that only the fitness individuals who are successful in adapting to their environment will have a better chance of surviving and reproducing.

Experiment test on some optimization benchmark problems have shown that the performance of particle swarm optimization algorithm could be improved with the integration of Velocity Propelled Averaged Crossover (VPAC) operator [29]. VPAC operator has been created based on the ideas of GA operators but this new parameter has been used to determine the population proportion before proceeding to the reproduction processes.

With respect to increase the particle swarm synergy, a mechanism of mutation from the EA paradigm could be adapted before the process of population update in PSO[30]. This technique has been tested on some benchmark problems and the output was better than the standard PSO. Combining PSO and DE operator [31] is an another successful work in which the results outperformed the two hybridization methods of [29] and [32].

As to improve the pheromone updating rules of ant colony optimization (ACO) algorithm, an employment of PSO mechanism into ACO population generation proved to improve the performance of the algorithm [33]. Through the convergence analysis, the algorithm with the hybridization method has achieved better convergence than the single PSO and ACO. In other way, an introduction to GA operators to ACO population generation has produced an improved result to the single ACO[21].

Hybridization between ACO with local search procedure is one of the examples that combine S-metaheuristics and P-metaheuristics. This technique demonstrates excellent convergence and robustness in uncovering low risk paths or routes in a sparse graph [34]. Hybridization between ACO with simulated annealing and ACO with tabu search with related to the course timetabling problem has produced good solutions compared to others methodologies[35].

Bee colony (BC) is another type of swarm intelligence technique like ACO and PSO. As a P-metaheuristics type, BC population needs for hybridization method for performance improvement. Combination of BC with greedy heuristic and local search strategy for population generation can produce significant and satisfactory results for the optimization problem [36]. Similarly, after testing to the Knapsack Problem, a better result has been achieved from the hybridization of BC with PSO[37].

5 A Comparative Study

Table 1 illustrates the comparison of L TH according to three different aspects: the metaheuristics, internal structure and problem applied in the experiment.

It is shown from comparison table that majority of P-metaheuristics hybridization techniques have been implemented at the population generation (PG). In most cases, researchers prefer to use benchmark optimization problems such as Ackley, Rastrigin, Ellipse, Rosenbrock, Sphere, Schaffer, Griewank, Ellipse and Ellipsoida. However, it has been revealed from various studies that L TH implementation provides effective and improved solution to the real optimization problems like network design and course time-tabling.

Table 1. A comparison of L TH
Research works Metaheuristics Internal Problem
structure
L. Guo, et all 17] GA with single population PI Network routing
G. Zobolas, et al. [IS] GAwithDE PI Scheduling
GA with VNS PG
Y. Marinakis and M. Marinaki GA with MPNS-GRASP, PI VRP
[19] GAwithPSO PG
H. Guangdong, et al [20] ACOwithGA PG Scheduling
X.-Y. Li, et al. [21] ACO with tabu search PI VRP
PG
Antonio LaTorre et al. [23] EA variants(GA,ES) and PG Benchmark
EDA
J. M. Perra, et al. [24] GAwithDE PG SAT
V. Robels et al. [3S] GAwithEDA PG Bench-mark
M. Gen, et al. [25] GA with local search PG VRP
Y. K. Ahn, et al. [26] GAwith SA PG Nonlinear
squeeze film
damper
J. P. Caldeira, et al. [27] EA with tabu search PG Scheduling
F. Juang [2S] GAwithPSO PG Network design
S. Matthew and S. Terence PSO with GA PG Benchmark
[29]
Z. You and T. Ying [30] PSO with GA PG Benchmark
X. Wei and G. Xing sheng [31] PSO with EAs PG Benchmark
B. Shuang, et al. [33] ACO with PSO PG TSP
K. K. Lim, et al. [34] ACO with local search PG Path planning
M. Ayob and G. Jaradat[35] ACOwith SA PG Course time-
ACO with tabu search tabling
S. Pulikanti and A. Singh [36] BC with greedy heuristic PG Knapsack
M. Aurelio, et al. [37] BC with PSO PG Knapsack
6 Conclusion Low-level Teamwork Hybridization is a technique that is used to combine Pmetaheuristics with other metaheuristics. It involves restructuring original algorithms through modification of P-metaheuristics internal structure, i.e. the essential elements that determine the traits and characteristics of original algorithms. The main internal structures of P-metaheuristics that can be embedded with other metaheuristics are population initialization and population generation. Based on our review and comparative study, hybridizations at the population generation is preferred by many

researchers. We have also found that the use of GA IS very common when implementing L TH based hybridization.

Acknowledgement

The authors would like to thank Universiti Teknologi MARA for its financial support to this project.

References

1. Affenzeller, M., et al., Genetic Algorithms and Genetic Programming - Modern Concepts

and Practical Applications. 2009: CRC Press.

2. Dorigo, M. and T. Stutzle, Ant Colony Optimization. 2004: MIT Press Ltd.

3. Clerc, M., Particle Swarm Optimization. 2006: ISTE.

4. Snodgrass, R.E., Anatomy of Honey Bee. 1953, Ithaca, NY: Comstock Publishing Associates.

5. Storn, R. and K. Price, Differential evolution: A simple evolution strategy for fast optimization. Dr. Dobb's Journal, 1997.22(4): p. 18-24.

6. Laarhoven, P.J.M.V. and E.H.L. Aarts, Simulated Annealing: Theory and Applications (Mathematics and Its Applications). 1988: Kluwer Academic Publishers Group.

7. Glover, F. and M. Laguna, Tabu Search. 1998, Boston.: Kluwer Academic.

8. Hansen, P., N. Mladenovic, and J.A.M. Perez, Variable neighbourhood search: methods and applications. Annals of Operations Research, 2009.175(1): p. 367-407.

9. Wolpert, D. and W. Macready, No free lunch theorems for optimization. EEE Transaction on Evolutionary Computation, 1997. 1(1): p. 67-82.

1O.Blum, C. and A. Roli, Hybrid Metaheuristics: An Introduction, in Hybrid Metaheuristics, C.

Blum, et al., Editors. 2008, Springer. p. 1-30.

1 1. Raidl, G.R., J. Puchinger, and C. Blum, Metaheuristic Hybrids, in Handbook of Metaheuristics, Pardalos, et al., Editors. 2010, Springer New York. p. 305-335.

l2.El-Abd, M. and M. Kamel, A Taxonomy of Coorperative Search Algorithm, in Hybrid Metaheuristic 2005, MJ.B.e. al., Editor. 2005, Springer-Verlag Berlin Heidelberg. p. 32-41. 1 3. Grosan, C. and A. Abraham, Hybrid Evolutionary Algorithms: Methodologies, Architectures, and Reviews. Studies in Computation Intelligence (SCI), 2007.75: p. 1-17. l4.Talbi, E.G., A Taxonomy of Hybrid Metaheuristics. Jounal of Heuristics, 2002. 8: p. 541- 564.

1 5. Lau, H.C., et al., A software framework for fast prototyping of meta-heuristiccs hybridization. International Transactions in Operational Research, 2007. 14(2): p. 123-141. l6.Talbi, E.-G., Metaheuristics: From Design to Implementation. 2009: Wiley. 586.

l7.Guo, L., B. Wang, and Q. Tang, A Hybrid Genetic Routing Algorithm in Wireless Sensor Networks, in Advances in Wireless Networks and Information Systems, Q. Luo, Editor. 2010, Springer-Verlag: Berlin Heidelberg. p. 87-92.

l8.Zobolas, G., C. Tarantilis, and G. Ioannou, A hybrid evolutionary algorithm for the job shop scheduling problem. Journal of the Operational Research Society, 2009. 60: p. 221-235.

1 9. Marinakis, Y. and M. Marinaki, A hybrid genetic - Particle Swarm Optimization Algorithm for the vehicle routing problem. Expert System with Applications, 2010. 37: p. 1446-1455. 20.Guangdong, H., L. Ping, and W. Qu. A Hybrid Metaheuristic ACO-GA with an Application in Sports Competition Scheduling. in Eighth ACIS International Conference on Software

Engineering, Artificial Intelligence, Networking, and Parallel/Distributed Computing (SNPD 2007). 2007. Haier International Training Center, Qingdao, China IEEE.

21.Li, X.-Y., P.Tian, and S. Leung, An ant colony optimization metaheuristic hybridized with tabu search for open vehicle routing problems. Journal of the Operational Research Society, 2009.60: p. 101O-102S.

22.LaTorre, A., et al. Hybrid Evolutionary Algorithms for Large Scale Continuous Problems. in GECCO'09, July 8-12,2009.2009. Montreal Quebec, Canada.

23.Pena, J.M., et al., GA-EDA: Hybrid Evolutionary Algorithm Using Genetic and Estimation of Distribution Algorithms, in lEA/AlE 2004, R.O.e. al., Editor. 2004, Springer-Verlag Berlin Heidelberg. p. 361-371.

24.Rodriguez-Tello, E. and J.Torres-Jimenez. ERA: An algorithm for reducing the epistasis of SAT problems. in Genetic and Evolutinary Computation Conference. 2003: Springer Verlag. 2S.Gen, M., L. Lin, and J.-B. Jo, Hybrid Genetic Algorithm for Designing Logistics Network, VRP and AGV Problems, in Intelligent and Evolutionary Systems. 2009, Springer-Verlag Berlin Heidelberg p. 123-139.

26.Ahn, Y.K., et al., Optimal Design of Nonlinear Squeeze Film Damper Using Hybrid Global Optimization Technique. Journal of Mechanical Science and Technology, 2006. 20(8): p. 112S-1138.

27.Caldeira, J.P., F. Melicio, and A. Rosa. Using a Hybrid Evolutionary-Taboo Algorithm to solve Job Shop Problem. in ACM Symposium on Applied Computing. 2004. Nicosia, Cyprus:

ACM.

28. Juang, F., Hybrid of Genetic Algorithm and Particle Swarm Optimization for Recurrent Network Design. IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART B: CYBERNETICS, 2004.34(2).

29. Matthew, S. and S. Terence. Breeding swarms: a GA/PSO hybrid. in Proceedings of the 2005 conference on Genetic and evolutionary computation. 200S. Washington DC, USA:

ACM.

30.You, Z. and T. Ying. Particle swarm optimization with triggered mutation and its implementation based on GPU. in Proceedings of the 12th annual conference on Genetic and evolutionary computation. 2010. Portland, Oregon, USA: ACM.

31.Wei, X. and G. Xing sheng. A hybrid particle swarm optimization approach with prior crossover differential evolution. in Proceedings of the first ACM/SIGEVO Summit on Genetic and Evolutionary Computation. 2009. Shanghai, China: ACM.

32.Yen, G.G. and L. Wen Fung, Dynamic Multiple Swarms in Multiobjective Particle Swarm Optimization. Systems, Man and Cybernetics, Part A: Systems and Humans, IEEE Transactions on, 2009.39(4): p. 890-911.

33.Shuang, B., J. Chen, and Z. Li, Study on hybrid PS-ACO algorithm. Applied Intelligence, 2009.

34.Lim, K.K., et al., Hybrid ant colony algorithms for path planning in sparse graphs. Soft Computing, .2007. 12: p. 981-994.

3S.Ayob, M. and G. Jaradat. Hybrid Ant Colony systems for course timetabling problems. in The Second Conference on Data Mining and Optimization, 2009.2009. Kajang, Malaysia:

IEEE.

36.Pulikanti, S. and A. Singh, An Artificial Bee Colony Algorithm for the Quadratic Knapsack Problem, in ICONIP 2009, C.S.Leung, M. Lee, and J.H. Chan, Editors. 2009, SpringerVerlag Berlin Heidelberg. p. 196-20S.

37.Aurelio, M., et al., Application of the Bee Swarm Optimization BSO to the Knapsack Problem, in Soft Compo for Recogn. Based on Biometrics" P.M.e. al., Editor. 2010, Springer-Verlag Berlin Heidelberg p. 191-206.

38.Robles, V., et al., Extending the GA-EDA Hybrid Algorithm to Study Diversification and Intensification in GAs and EDAs, in IDA 2005, A.F.F.e. al, Editor. 200S, Springer-Verlag Berlin Heidelberg p. 339-3S0.

Sign up to vote on this title
UsefulNot useful