In order to improve accuracy and dependability of using neural network for software reliability prediction

a multi-objective optimization-based improved Elman recurrent network method（Mop-IElman） was proposed.First

on the basis of the Elman network

a self-delay feedback of the output layer as another context layer was designed.Second

the network architecture and the initial outputs of these two context layers were taken as variables of network configuration setting

and NSGA-II was employed to simultaneously optimize prediction performance and robustness

then the Pareto solution was obtained.After that

by maximizing the sum of prediction performance and robustness

the final network configuration setting was determined.Finally

the proposed method was compared with the feed-forward neural network

the Elman network

both the single-objective and the multi-objective optimization Elman networks with respect to two real software failure data.It demonstrated that the proposed Mop-IElman achieves higher prediction accuracy and de-pendability.