Apply Composed Statistical Model In Program Fault Localization

As the increased workload and difficulty in software maintains, research in automatic debug and program fault localization is more and more significant. There are various aspects of the research in program fault localization. Statistical based method is a new and important research topic, which is motivated to inference the location of bugs in programs by statistic theories.Since the common behavior characteristic for programmers is hard to abstract and the logic for different programs is hard to clarify, it may not make sense to follow state of the art methods, which abstracts unitary feature from runtime states and build unitary model on it. Those methods have problems such as weak generalization, poor coverage on bug patterns and strong dependence on test suites.There are two proposals to deal with the issues that happened in unique statistical model. The first is that abstract multi-dimension features from runtime states and build a complex model. The second is that build unique statistical for each dimension of feature and compose them by some mechanism. We take the second proposal to research. We propose composed statistical model to solve the problems referred. Composed statistical model build effective unique model for each feature at the very beginning. These models have fine performance in fault localization under certain environment. To guarantee the basic system performance, performances of different models are complementary. Then we compose these models into one integrated system.Then there may be two questions for this research:1. How to build unique modelsWe abstract three features from runtime states, which are coverage on control flow graph, execution counts in runtime, transfer probability between previous and succeed nodes in control flow graph. Then we build Occupancy model, Relativity model, Execution counts distribution model and Previous/Succeed distribution model for relevant feature.2. How to composeWe compose those four models by boosting algorithm. The effect that each model takes is adjusted by model weights.The experiment results illustrate that our unique statistical model covers some error patterns and find the bugs to this patterns excellently. But those models are strong dependent on bug pattern and hard to generalized to other bugs. Composed statistical model fix this problem perfectly. It can directly localize bugs to various patterns in each experiment.Therefore, composed statistical model, which has stronger capability on fault localization and noise tolerance, can find the bugs in various patterns effectively. The improvement on analysis performance and generalization can be verified by comparative experiments between composed model and unique model.