Asynchronous Strategy of Parallel Hybrid Approach of GA and EDA for Function Optimization Said Mohamed Said Information Engineering Department University of the Ryukyus 1 Senbaru, Nishihara, Okinawa 903-0213 JAPAN saidy87@hotmail.com Morikazu Nakamura Information Engineering Department University of the Ryukyus 1 Senbaru, Nishihara, Okinawa 903-0213 JAPAN morikazu@ie.u-ryukyu.ac.jp Abstract—This paper adapts parallel master-slave estima- tion of distribution and genetic algorithms (GAs and EDAs) hybridization. The master selects portions of the search space, and slaves perform, in parallel and independently, a GA that solves the problem on the assigned portion of the search space. The master’s work is to progressively narrow the areas explored by the slave’s GAs, using parallel dynamic K-means clustering to determine the basins of attraction of the search space. Coordination of activities between master and slaves is done in an asynchronous way (i.e. no waiting is entertained among the processes). The proposed asynchronous model has managed to reduce computation time while maintaining the quality of solutions. Keywords-Hybrid, Estimation of Distribution Algorithm, Genetic Algorithms, Synchronous, Asynchronous, Parallel processing, Master-Slave, K-means clustering; I. I NTRODUCTION Researches in [1][3][4] and many others hint high capa- bility of Evolutionary Computation (EC) to solve various problems in computer science domain. However, with rapid growth of real world problem size and complexity, higher computational cost is needed to solve these prob- lems. While hybridization boosts the performance of EC [10], parallel processing helps to speed up searching pro- cess as shown in [11] and [12]. This paper is a continuation of proposed algorithms in [6] and [18] to achieve both higher quality solutions and good computational speed based on GAs and EDAs hybrid approach. The latter outperformed the former by improving solution quality but in a cost of additional computation time. The increased complexity was due to additional load in dynamic K means clustering algorithm even when parallelly executed. In [18] strategic synchronous master-slave formulation of EDA and GA was used similar to that in [6] except that master part was emphasized by parallel dynamic K means clustering for more reasonable estimation. Both GAs and EDAs have shown promising achieve- ments and have been used in variety of problem domains [5],[14],[13],[2]. The aim of this research is to reduce computation time by introducing asynchronous strategy to the present master-slave hybrid scheme. The proposed change has been well analyzed and proved to be fruitful in [9] and [20]. Both of them assure significant reduction of computation time in an asynchronous mode. Our approach suits well in an asynchronous mode due to the fact that it uses shared memory multiprocessor unit with several parallel threads working over one common population. From a parallel processing point of view, reducing un- necessary communication among processors is essential to avoid performance degradation [20]. In asynchronous master-slave scheme, slaves perform independent evolu- tionary computation using GA with un-identical number of generations(i.e slave terminates searching when predefined target fitness value has been reached) and master controls the searching using EDA, whenever pre-defined number of solutions in the Database(DB) returned by slaves is reached. Furthermore the master EDA has to follow four phase strategy adopted in [6], with every phase initiated by parallel dynamic K means clustering except in the first and last phases. The phase defines the manner in which EDA probabilistic estimation vectors are obtained by the master. The experiment was done using Real-Parameter Black Box Optimization Benchmarking system on noiseless testbed and compared with performance of [18] and GA. The results suggest maintained solution quality with notable reduction in computation time. II. HYBRID ASYNCHRONOUS MASTER-SLAVE SCHEME Among the merits of master-slave formulation are; i) it is a simple transposition of the single processor evolution- ary algorithm onto multiple processor architectures that allows reproducibility of results, ii) there is no permanent loss of information when a slave fails or is unreachable by the master, iii) it is appropriate for networks of computers where availability is sometimes limited (e.g. available only during night time or when screen saver is on) as nodes can be added or removed dynamically with no loss of information, and iv) it is made of a centralized repository of the population which simplifies data col- lection and analysis as elaborated in [17]. In changing our approach to be asynchronous, we maintained its basic master-slave architecture to retain the above mentioned benefits. The whole algorithm runs in a fixed number of repetitive iterations, with all slaves executing GA and re-initialization taking place at the beginning of every iteration. The manner in which population members are initialized in each iteration is controlled by master using probabilistic estimations of EDA with the help of K means clustering algorithm. Using EDA on local optimal solutions returned by slaves, master can guess the areas 2012 Third International Conference on Networking and Computing 978-0-7695-4893-7/12 $26.00 © 2012 IEEE DOI 10.1109/ICNC.2012.80 420