RE4AI: Correctly capture requirements for AI-heavy products (Part 1)

Everyone is talking about AI: but what does it mean for requirements management to make artificial intelligence part of the product? This is called RE4AI, Requirements Engineering for Artificial Intelligence. Ahmad et al. have addressed this highly topical question.

Specifically, the researchers investigated which approaches for RE4AI can be found in the literature. These include methods and tools, as well as their possibilities and limitations. Below are the results of the analysis of 27 studies from the research literature. The result: there is still a lot to do, proven approaches are reaching their limits in AI systems.

This is the first part on AI, which addresses the limitations and challenges of RE4AI, as well as the possible domains of the application and the notations that help to capture requirements.

The research work

The paper with the title What's up with Requirements Engineering for Artificial Intelligence Systems? was published in 2021 by Khlood Ahmad, Muneera Bano, Mohamed Abdelrazek and Chetan Arora (Deakin University, Geelong, Australia) and John Grundy (Monash University).

AI systems are different

Requirements management follows an established process that runs through the following phases: Elicitation, Analysis, Specification and Documentation, Validation and Management. This process ensures that the requirements are correctly collected, documented and managed and that they match the needs of the users.

The software development process is different for systems with AI/ML components. Such systems are not always specified by specifications, but learn a behavior based on data. Therefore, RE techniques need to be adapted to the changes introduced by this new paradigm of AI systems.

The lack of requirements specifications in current ML systems has a significant impact on the quality of the ML model.

What's up with Requirements Engineering for Artificial Intelligence Systems?

In particular, such systems contain comparatively little software code, at least in relation to the training data. Accordingly, the management of training data (MLOps) is very important. The authors have found evidence in the literature that the lack of specification of ML systems has a negative impact on the quality of ML models.

Non-functional requirements such as transparency, trust, privacy, security and reliability are becoming increasingly important.

RE4AI research

This paper is a systematic literature review (SLR). It turned out that most of the papers deal with the use of AI in requirements management. Comparatively less has been published on the use of AI in products, although a lot is happening at the moment. The authors assume that we will see many new, innovative papers on RE4AI in the coming years.

The authors used the literature research to answer the following questions:

Which modeling languages and notations for requirements are used for RE4AI?

The authors first examined how the requirements were recorded in the development process. This does not refer to the ML model. The authors found a total of five modeling languages for capturing requirements.

More than three quarters of the papers use either GORE or UML/SysML for modeling the requirements. In the Goal Oriented Requirements Engineering (GORE) hard and soft targets are defined, which often helped in communication with stakeholders. As the goals are solution-neutral, the approach fits well with machine learning, which is about the "what", not the "how".

Although UML (and SysML) has been used by many researchers, it has not always been as effective as GORE. Several researchers had difficulties with capturing non-functional requirements.

In which domains has RE4AI been used so far?

The authors identified a total of eight domains, with some of the papers examined not mentioning any particular domain. The list is not really surprising. Autonomous driving, data processing and ethics dominate.

If you want to see the whole list, you can consult the research paper linked above. But basically it shows that RE4AI is conceivable in all domains.

What are the limitations and challenges of RE4AI?

This is certainly the most interesting question for those considering incorporating AI into products. Here is a brief summary:

• Overestimation of AI - As we are currently at the peak of the Hype cycle of AI, it is no wonder that many see AI as a panacea.
• Define requirements - As mentioned above, RE has many challenges for AI systems. Specifically, the authors mentioned examples of the difficulties of definitions: What is a "pedestrian"? What does "fair" mean?
• The nature of ML systems - I have already mentioned this too: machine learning systems are fundamentally different, which leads to challenges in RE4AI. Traditional approaches can quickly lead to dead ends.
• Evaluate compromises - Weighing up is already a challenge in "normal" requirements management. With RE4AI, this is even more difficult, as many trade-offs depend on an evaluation, which must then be calculated in a comprehensible manner. For example, do we trade privacy for transparency or fairness for accuracy?
• Responsibility of the requirements engineer - Traditionally, the data scientist is responsible for the data of AI systems. This role has certain overlaps with the tasks of the requirements engineer. Gaps can quickly appear here. Without a data scientist, we have the problem that requirements engineers are not trained in dealing with large amounts of data.
• New requirements and technologies emerge - New requirements regarding data, ethics, trust and transparency are creating new challenges.
• Data requirements - We also have requirements for the data in terms of quality, availability, tests, etc.
• Non-functional requirements - Even if non-functional requirements (NFR) have always been a part of RE, a clear shift has taken place here with RE4AI. Traditionally important NFRs are receding into the background (e.g. compatibility, modularity), while others are becoming much more important (e.g. fairness, transparency).

Outlook for next week

Part 2 on RE4AI will follow next week, which will deal with specific recommendations, in particular a reference to Google PAIR (People + AI Research).

Photo by Emmanuel Ikwuegbu on Unsplash