We consider a parallel distributed decision fusion system consisting of a bank of local sensors and a fusion center. Each local sensor makes binary decisions based on its own observations. The decision is to accept one of the hypotheses, H0 or H1. Each local sensor transmits its decisions to the fusion center over an error free channel. The fusion center combines all the local decisions to obtain a global decision. For observations that are statistically independent conditioned on the hypothesis and fixed local decisions, the Chair-Varshney fusion rule minimizes the global Bayesian risk. However, this fusion rule requires knowledge of local sensor performance parameters and the prior probabilities of the hypothesis set. In most applications, these are unavailable. Moreover, local sensor performance may be time varying. Several studies attempted on-line estimation of the unknown local performance metrics and prior probabilities. We develop a fusion rule that applies a genetic algorithm to fuse the local sensors' binary decisions. The rule adapts to time varying local sensor error characteristics and provides near-optimal performance at the expense of a larger number of observations and higher computational overhead.