Software reliability prediction based on support vector regression with binary particle swarm optimization for model mining

Data-Driven Software Reliability Modeling (DDSRM) is an approach in software reliability prediction problem which only relies on software failure data. There are two kinds of model architecture in this modeling, which are Single-Input Single-Output (SISO) and Multiple-Delayed-Input Single-Output (MDISO). In MDISO architecture, the prediction process involves having multiple inputs from the failure data to predict single output in the future. Most MDISO literatures use underlying assumption that a failure is correlated with a number of most recent failures. In more “generic” model of MDISO, a failure can be correlated with some of the previous failures. The process of searching which time lags to use as inputs in this model is sometimes referred to as a model mining process. This paper proposes to apply Binary Particle Swarm Optimization (BPSO) algorithm as model mining in software reliability prediction problem in terms of failure count number with Support Vector Regression (SVR) as predictor. Initial experiment shows that the proposed SVR-BPSO method yields more accurate prediction result than a prediction without model mining.