Neural network ensembles have been used increasingly in recent years to improve classifier’s generalization ability. There are several methods of designing neural network ensembles, such as weighted average linear ensemble. The output of the weighted average linear ensemble is a weighted average of the output of each component neural network, with the weights determined by a function. Based on the characteristic of the classification issue, a function is defined, which is the ratio of the pattern separability within class to the pattern scatter between classes. The minimum of the function corresponds to the optimal weights, so an optimal linear ensemble is obtained. The optimal weights are searched by a genetic algorithm. The rationale behind the function is also analyzed, showing that it accords with the Bayesian decision rule. Finally, to estimate the performance of this linear ensemble, two handwritten Chinese character feature sets and four data sets from UCI machine learning depository are used. Empirical study shows that the optimal linear ensemble method can produce ensemble of the neural network with a stronger generalization ability.