Investigation of a characteristic bimodal convergence-time/mutation-rate feature in evolutionary search

The use of evolutionary algorithms to determine optimum load and data distribution in large distributed databases has been investigated in earlier publications both by the authors and others. This paper reports on some interesting results arising from comprehensive examination of the performance profile of various techniques we have investigated on this problem and others. In particular, we see that when too little mutation is available to the system, the number of evaluations that the algorithm is able to exploit before premature convergence occurs seems near linearly proportional to population size, regardless of evaluation (time) limit, selection strategy, or other features. More interestingly, however, as mutation is increased, there seem to exist characteristic peaks and troughs in the tuned performance landscape indicating an optimal mutation rate independent of population size; this is a trough between the two peaks in a robust bimodal feature in the curve of convergence time against mutation-rate. These features are demonstrated over a range of evaluation limits, algorithm designs, and application landscapes. The continued re-appearance of the bimodal feature leads us to postulate that it may be a relatively problem-independent feature of evolutionary search, with general application to parameter tuning issues.