International Conference February 10 - 13, 2010 CYBERNETICS AND INFORMATICS VYŠNÁ BOCA, Slovak Republic COMPARISON OF USING SIMPLE GENETIC ALGORITHM AND PARALLEL GENETIC ALGORITM IN HEAT TRANSFER MODEL OPTIMIZATION Michal Oravec, Stanislav Števo, Ivan Sekaj Institute of Control and Industrial Informatics Slovak University of Technology, Faculty of Electrical Engineering and Information Technology Ilkovičova 3, 812 19 Bratislava, Slovak Republic Tel.: +421 2 60291 621 , +421 2 60291 403 e-mail: michal.oravec@stuba.sk , stanislav.stevo@stuba.sk , ivan.sekaj@stuba.sk T Abstract: In this paper the comparison of using simple genetic algorithm and parallel genetic algorithm is presented. As the optimization problems the parameter setting of the heat transfer model of a building and the building’s model calibration were chosen. The model simulation requires huge computing capacity and it is time consuming. Therefore the pressure of simulation evaluations number is concerned and the use of parallelism is desirable. Genetic algorithms and parallelization were implemented in Matlab and the simulation of heat transfer model, which is the part of the fitness function, is performed in Comsol Multiphysics. Keywords: genetic algorithm, parallel genetic algorithm, building model optimization, heat transfer, Matlab, Comsol Multiphysics 1 INTRODUCTION Genetic algorithms (GA) are effective stochastic optimization approaches imitating natural evolution process (Sekaj, 2005). Despite the fact, that there has been progress in the area of GA, the premature convergence sometimes occurred and large computing capacity is needed. Especially when more complicated system is to be optimized or a model simulation takes a lot of time. In such cases it’s necessary to reduce the number of the cost function (fitness) evaluations (simulations). There are many options to improve GA’s. Most common is to tune GA setting to reach the best algorithm performance. However, it is sometimes not possible to tune the algorithm to be able to achieve a sufficient convergence rate to the global optimum. Therefore another option is to use parallelism. Parallel genetic algorithms (PGA) are able to improve the performance of simple genetic algorithms with a single population (Cantú-Paz, 1995). This paper presents practical comparison of using simple genetic algorithm (SGA) with a single population and parallel genetic algorithm with population distributed into several interconnected subpopulations. 2 PARALLEL GENETIC ALGORITHM In parallel genetic algorithms (PGA) the evolution is distributed into many more or less isolated subpopulations, where the transfer of genetic information among these subpopulations has an important influence on the evolution process. In this case we don’t consider parallelisation into more processors or more computers respectively, which can extend the computational power of the computer system. Let us consider such parallelisation, wich is realised on a single processor or PC. 1