Performance of Parallel Genetic Algorithms on Distributed Memory Architectures

The Genetic Algorithms draw a similarity from the Genetic mutation and Cross Over within populations from biology. The genetic algorithms are highly parallel in nature. These can be used to solve many important problems like Graph Partitioning, Travelling salesman problems, 0-1 Integer linear programming problem etc. When these are implemented, there exists a trade-off between Genetic search qualities and execution performance. In order to improve the execution performance of algorithms, those implementations with lesser communications between populations are considered best. In this direction, we try present an algorithm by discrete small subpopulations. However this implementation reduces the quality of search of the algorithm. Therefore we can improve the quality of search by having a centralized population. In this paper, we review some of the alternatives of implementation of these algorithms on distributed memory architectures in which centralized data can be implemented. We also present an example in which we implement these alternatives of parallel algorithms for predicting the tertiary protein structure. In the final section, we try to provide a performance analysis of the various proposed architectures. Genetic Algorithms is a heuristic search technique which draws its inspiration from the popular " survival of the fittest " principle of natural evolution. First pioneered by John Holland in the 60s, Genetic Algorithms has been widely studied. In genetic algorithms, a collection of possible solutions, known as population is maintained. The search algorithm proceeds in steps called generations. Each generation involves transformation of individual solutions based on their fitness function and results in a new population of solutions. There are two operators of transformation: crossover and mutation. In cross over the pieces of solution are swapped to give new solutions and their fitness is tested. After crossover, mutation is applied which is randomly done to select new characteristics for the solution. The termination is based on either pre specified number of steps or a pre specified level of optimality. GA's provide alternative methods to solving problem and consistently outperform other traditional methods. Many of the real world problems involved finding optimal parameter may be difficult for traditional methods but ideal for GA's. Although these are highly deft for solving hard problems, they might take longer execution times. Driven by the need to reduce the execution time, scientists resorted to exploit the parallel nature of GA's. These came to known as parallel genetic algorithms. The implementation of parallel genetic algorithms involves similar issues …