Program Behavior Representation Learning for Software Testing

点击翻译按钮获取中文摘要

Software has been incorporated into an ever-increasing number of mission and safety-critical systems such as self-driving cars, air traffic control systems, communication platforms, e-banking infrastructures and e-commerce systems. Such systems provide central and crucial services to our society and thus require high quality software. However, history is full of software failures that have caused numerous problems: aircraft have crashed, patients have died from incorrect medication, and key financial systems have broken down. My research is directed towards improving the quality of software. My program focuses on software testing as the most commonly used method of quality assurance in the software industry. Recent research has shown that systematic test automation potentially increases the effectiveness of testing and reduces its cost. The ultimate goal in software testing is to make sure the actual behaviour of the software at run-time matches the expected behavior of the software. However, manually defining the behavior is quite costly. Thus, most existing test automation techniques are source code-driven and miss many challenging behavioral defects. The core idea of this research is to change the current tradition of source code-driven testing of software programs to behavior-driven testing, without the extra cost of manual modeling. I propose an approach where using machine learning, the behavioral model of the program can be inferred and optimized. Such inferred models can then be used to further test the software. The models will be optimized toward test generation. In other words, they distinctly identify the part of behavior space that requires more testing. In addition, they provide mechanisms to actually generate such tests. The proposed techniques will be empirically evaluated on large software systems and compared against existing software test automation techniques. Since software testing is one of the most commonly used techniques in almost every software development process, Canadian software communities in many sectors will be well positioned to benefit from the outcomes of this research, through decreased costs and increased software quality leading to enhanced competitiveness, internationally. Though this proposal is only about functional testing, the long-term goal is to generalize its core idea of learning a specialized behavioral model to run-time analysis in other domains such as debugging, program comprehension, and security. Thus, this research has the potential to affect many sub-domains within computer science and software engineering. The proposed research program will train 11 highly qualified personnel (HQP): 2 PhD and 4 Master and 5 undergrad summer students. All HQPs will work closely with Canadian companies, gaining an appreciation of the challenges facing practitioners on a daily basis and helping them to improve their products and processes.