The number and importance of AI-based systems in all domains is growing. With the pervasive use and the dependence on AI-based systems, the quality of these systems becomes essential for their practical usage. However, quality assurance for AI-based systems is an emerging area that has not been well explored and requires collaboration between the SE and AI research communities. This paper discusses terminology and challenges on quality assurance for AI-based systems to set a baseline for that purpose. Therefore, we define basic concepts and characterize AI-based systems along the three dimensions of artifact type, process, and quality characteristics. Furthermore, we elaborate on the key challenges of (1) understandability and interpretability of AI models, (2) lack of specifications and defined requirements, (3) need for validation data and test input generation, (4) defining expected outcomes as test oracles, (5) accuracy and correctness measures, (6) non-functional properties of AI-based systems, (7) self-adaptive and self-learning characteristics, and (8) dynamic and frequently changing environments.
翻译:在所有领域,基于AI的系统的数目和重要性都在增加,随着对基于AI的系统的普遍使用和依赖,这些系统的质量对其实际使用至关重要,然而,基于AI的系统的质量保证是一个新兴领域,尚未进行充分探讨,需要SE和AI研究界之间开展协作。本文件讨论了基于AI的系统为确定这一目的的基线而在质量保证方面的术语和挑战。因此,我们按照人工制品类型、过程和质量特征的三个层面界定了基本概念和基于AI的系统的特点。此外,我们阐述了以下关键挑战:(1) AI模型的可理解性和可解释性,(2) 缺乏规格和界定的要求,(3) 验证数据和测试投入生成的需要,(4) 将预期结果界定为测试或触手,(5) 准确和正确性措施,(6) 基于AI的系统不起作用的特性,(7) 自我适应和自学特点,以及(8) 动态和经常变化的环境。