Measuring the software complexity is an important taskin the management of software projects. In the recent years, many researchershave paid much attention to this challenging task due to the commercialimportance of software projects. In the literature, there are some softwaremetrics and estimation models to measure the complexity of software. However, westill need to introduce novel models of software metrics to obtain moreaccurate results regarding software complexity. In this paper, we will show that neural networkscan be used as an alternative method for estimation of software complexitymetrics. We use a neural network of three layers with a single hidden layer andtrain this network by using distinct training algorithms to determine theaccuracy of software complexity. We compare our results of software complexity obtained byusing neural networks with those calculated by Halstead model. Thiscomparison shows that the difference between our estimated results obtained by Bayesian RegularizationAlgorithm with 10 hidden neurons and Halstead calculated results of softwarecomplexity is less than 2%, implying the effectiveness of our proposed methodof neural networks in estimating software complexity.