Three Stages of Intelligent Testing: Computational, Perceptual, and Cognitive AI

Software testing is often perceived as a low‑value activity, but the rapid development of AI and big data technologies offers new opportunities. Since 2018, Baidu's MEG Quality Efficiency Platform has explored AI applications in testing, defining intelligent testing as the combination of data and algorithms to empower quality activities.

Stage 1 – Computational Intelligence : This stage embeds behavior data generated by software processes, simple algorithms, and computing power into quality activities to assist or predict testing actions. The focus is on proving that data and algorithms can significantly improve testing, without demanding cutting‑edge models. Examples include using genetic algorithms for task queuing, DTW for memory‑leak detection, Pearson correlation and bucket algorithms for precise replay of billions of traffic records, and JC distance for test case selection, all of which have reduced execution time and improved efficiency. However, limitations such as coverage‑based case recommendation that cannot detect faults highlight the need for the next stage.

Stage 2 – Perceptual Intelligence : Building on richer algorithms and computing resources, this stage aims to perceive risk like a human and make decisions. It addresses two realities: not every code change introduces risk, and not every test uncovers risk, leading to resource waste. Applications include visual techniques for front‑end automated test case generation, pop‑up removal, and UI diff; Bayesian‑CatBoost models for risk‑aware test case recommendation (improving recommendation rates from 50% to 10%); logistic‑regression‑based project risk prediction enabling 70% low‑risk, unattended releases; and deep‑learning‑based white‑box defect detection. These successes demonstrate AI can replace humans in risk perception, identification, and decision‑making.

Stage 3 – Cognitive Intelligence : Extending perceptual intelligence, this stage seeks to let machines not only perceive risk but also react autonomously. Research includes AST‑driven smart abnormal unit test generation for C++ (catching exceptions, infinite loops), UCB‑based priority traversal for higher page coverage, intelligent failure localization to reduce manual debugging, and self‑healing CI pipelines. Although still nascent, the authors anticipate a major breakthrough as AI technologies mature.

The article concludes with a recruitment notice for test developers, Java, C++, mobile, and machine‑learning engineers to join Baidu MEG’s intelligent testing team.