Coarse-grained parallel genetic algorithms: Three implementations and their analysis

Although solutions to many problems can be found using direct analytical methods such as those calculus provides, many problems simply are too large or too difficult to solve using traditional techniques. Genetic algorithms provide an indirect approach to solving those problems. A genetic algorithm applies biological genetic procedures and principles to a randomly generated collection of potential solutions. The result is the evolution of new and better solutions. Coarse-Grained Parallel Genetic Algorithms extend the basic genetic algorithm by introducing genetic isolation and distribution of the problem domain. This thesis compares the capabilities of a serial genetic algorithm and three coarse-grained parallel genetic algorithms (a standard parallel algorithm, a non-uniform parallel algorithm and an adaptive parallel algorithm). The evaluation is done using an instance of the traveling salesman problem. It is shown that while the standard course-grained parallel algorithm provides more consistent results than the serial genetic algorithm, the adaptive distributed algorithm out-performs them both. To facilitate this analysis, an extensible object-oriented library for genetic algorithms, encompassing both serial and coarse-grained parallel genetic algorithms, was developed. The Java programming language was used throughout