EN 
DE Can't the computer write the requirements itself?
Author: Peter Schedl, IBM Germany GmbH
Contribution – Embedded Software Engineering Congress 2018
Data drives the world today. Algorithms, artificial intelligence (AI), and machine learning are crucial in the transformation to a digital society. System and software development also involve a vast amount of data. Products can consist of thousands or even millions of requirements. How can these larger datasets be better utilized to combine the core elements of successful development with adherence to high quality standards and rapid responses to market developments?
The presentation will showcase applications of AI, demonstrating how it can assist engineers in defining requirements by identifying duplicate or conflicting requirements in specifications. Furthermore, using IBM Watson services as an example, it will illustrate how artificial intelligence programs can provide users with a quality assessment of their requirements, thereby increasing development efficiency.
Good requirements management is strenuous.
Product development typically involves multiple levels of requirements and thus requirements documents. At a minimum, there is always a requirements specification, which expresses the client's wishes, and a functional specification, which represents the supplier's response. Even at this highest level, there are thousands of requirements, and as products become increasingly complex, this number continues to grow.
In addition, requirements are often not clearly defined. In the past, there have been approaches to improve requirements by using rules, such as which words to avoid, or by using fixed sentence structures. However, all these efforts lead, on the one hand, to more – albeit better – requirements, and on the other hand, make requirements management significantly more complex and therefore more strenuous.
A requirements expert as an assistant?
Artificial intelligence is on everyone's lips, so why not use it and let AI generate good requirements? What could that look like? We would need an assistant that analyzes our spoken or typed requirements and derives good requirements from them.
What skills does our assistant need? First, it must understand our natural language. Such systems are called NLP (Natural Language Processing) and are already used in speech recognition and translation. Secondly, the assistant must know what constitutes good user behavior. It must have been trained for this through machine learning (ML).
How does machine learning work?
First, a large amount of representative data is needed. Since this involves semantic analysis, the data must be annotated using a type system and divided into a training and a test set. A trained machine learning model is then created from this data. The type system is of particular importance, as it defines the various entities and relationships that, in turn, form the basis for evaluating the requirements.
As an example, consider the requirement "The flight control system should use Bernoulli's equation as its control algorithm." This requirement has three entities:
- Flight control subsystem type
- Bernoulli's equation of the type Physical principle
- Control algorithm as a control method
Furthermore, the subsystem (flight control) has a relationship "has control method" to the control method.
This knowledge is added to the training data via annotation, thus training the ML algorithm. Subsequently, test data is used to check whether the AI is already "good enough." Three basic metrics are used for this. See Figure 1 (PDF).
The assistant brought to life
A prototype assistant was built using standard IBM services. Essentially, the cloud services IBM Watson Natural Language Understanding and IBM Watson Knowledge Studio [1] were used to qualitatively evaluate the requirements from the DOORS Next Generation requirements management tool [2].
Each requirement is scored from 0 to 100 points, with 100 representing a good requirement. The AI can also provide a justification for each score, see Figure 2 (PDF).
Result
It has been demonstrated that higher-quality requirements can be met with today's available AI tools. These techniques can also be applied to industry-specific domains, thus enabling user-specific assessments.
It is foreseeable that such applications will be incorporated into requirements management tools like DOORS Next Generation, and then the age of cognitive requirements engineering will dawn.
Bibliography and list of sources
[1] IBM Cloud Services
[2] DOORS Next Generation
author
Peter Schedl has been working in the field of mechatronic systems development for 25 years. Early on, he recognized the added value of a model-based approach to creating system, hardware, and software architectures. Through various embedded projects, he first gained experience implementing methodological approaches and later shared this knowledge as a consultant. He is currently focusing on the emerging challenges of the Internet of Things.
Contact
E-mail: peter.schedl@de.ibm.com
Requirements – MicroConsult Training & Coaching
Do you want to bring yourself up to date with the latest technology?
Then find out more here MircoConsult offers training courses/seminars/workshops and individual coaching on the topic of requirements/embedded and real-time software development.
Training & coaching on the other topics in our portfolio can be found here. here.
Requirements – Expertise
Valuable expertise in the field of requirements/embedded and real-time software development is available. here Available for you to download free of charge.
You can find expertise on other topics in our portfolio here. here.
