Using Genetic Algorithms to Design Bayesian Reliability Experiments

The purpose of a lifetime reliability experiment is to quantify the effect of one or more factors on the lifetime of some device of interest. In such experiments, each tested device either fails or is still functioning when the experiment (under the given factor conditions) is terminated at a given time (right censored lifetimes/time-truncated testing). The purpose of this paper is to introduce genetic algorithms as a practical means for finding near-optimal Bayesian experimental designs for such experiments. The design criterion is the expected Shannon information gain provided by the experiment. The methodology is illustrated using a subset of the eleven-factor thermostat experiment considered by Bullington et al. (1993). A lognormal lifetime distribution is assumed.