Solutions to Parallel and Distributed Computing Problems: Lessons from Biological Sciences

2.6 DISCUSSION OF RESULTS

Based on the implementation of parallel evolutionary algorithms described earlier, a large number of tests have been run to observe the influence of various parameters of the island model and of the neighborhood model. In particular, we tested the migration interval (5, 10, 15, 30 generations), the direction of migration (north; north and east; north, east, south, and west), two different migration rates, the number of islands, the mating scheme and the selection methods within the islands, and, for the neighborhood model, the type of neighborhood. The results of these experiments have been documented and discussed in [15]. Some of our experiences and results have also been reported in [16]. In the following, we focus on the results of our observations with respect to the influence of the number of islands and the type of neighborhood, respectively, since these turned out to be the most important parameters.

In every test run of a particular variant of the evolutionary algorithm we recorded the best and average objective function values for every generation. For every number of islands we observed the behavior of the algorithm for

• Six problem instances (three flow-shop, two RCPSP, one TSP)
• Four different migration intervals (5, 10, 15, 30 generations)
• Three different selection methods (see below)

and used four independent test runs for every combination of parameter values. Thus, we observed 48 different test runs for each of these six problem instances ...