During the years, softgoals have enjoyed much attention in the requirement engineering (RE) community for handling non-functional requirements (NFRs). As pointed out by RE researchers in recent years, softgoal may not be the right concept for non-functional requirements since it also captures early functional requirements (FRs) that are vaguely defined, e.g., “increase profit”. Alternatively, quality goals have been proposed. However, existing reasoning techniques on softgoals also have some deficiencies and can not be directly adopted: qualitative approaches lack sufficient accuracy and expressiveness for decision making, while the numbers used in existing quantitative procedures are subjective. This paper proposes a two-staged approach to tackle the problem of reasoning with quality goals: in the first stage we measure the satisfaction of leaflevel quality goals based on prototype values and graded membership, and in the second stage we propagate the satisfaction degree using similarity measurement. Our approach aims to address the aforementioned issues in the current reasoning procedures of quality goals.
Non-functional requirements (NFRs) have been commonly acknowledged as a
critical class of requirements in RE as it significantly influences user satisfaction
and the success of software/services. In the RE community, many techniques
were proposed for dealing with NFRs, and goal-oriented approaches have been
argued to be the major effort in dealing with NFRs in depth [
In goal-oriented requirements engineering (GORE), requirements are
captured as goals that represent state-of-affairs in the world and need to be brought
about by a system-to-be, and non-functional requirements are captured as
softgoals that are vague for success. As one of the key concepts in the NFR
Framework [
That is, softgoals constitute a useful abstraction for early requirements, both
functional and non-functional, rather than just non-functional ones [
Copyright © 2016 for this paper by its authors. Copying permitted for private and academic purposes.
With goal-oriented techniques, the AND/OR structure of goal models
naturally lends itself to an analysis of goal satisfaction [
In this work, we focus on goal satisfaction analysis, more specifically, the forward satisfaction analysis of quality goals. We propose a two-staged reasoning procedure: first, we measure the satisfaction of leaf-level QGs based on prototype values and graded membership; then, we propagate the satisfaction of lower-level QGs to higher-level QGs based on similarity measurement.
The rest of this paper is organized as follows. Section 2 briefly reviews related work on goal satisfaction analysis, Section 3 describes our treatment of NFRs as requirements referring to qualities, Section 4 presents the two-staged reasoning procedure, Section 5 concludes the paper and offers suggestions for future work. 2
In general, goal satisfaction analysis can be categorized into two classes:
qualitative and quantitative [
Both qualitative and quantitative approaches use some propagation rules to
propagate initial satisfaction labels of lower-level goals to higher-level ones. For
example, the NFR Framework [
In our observation, the aggregation from the satisfaction of lower-level goals to that of higher-level ones is often non-linear. For example, supposing that we have a softgoal “enjoyable trip” being AND-decomposed into “low cost”, “enjoyable activities” and “delicious food”, how about if we finally have “enjoyable activities” and “delicious food” but a “high cost”? It is probably not the minimum satisfaction value in this case, at least for some classes of users.
A second issue here is how to determine the satisfaction degree of leaf-level
goals, which are often manually assigned in existing qualitative and quantitative
approaches. Letier et al. [
Our treatment of NFRs [
Vague NFRs are captured as quality goals (QGs) and denoted in the form of
“Q (SubjT ) :: QRG”, which is an abbreviation for “ 8q 2 Q; x 2 SubjT; inheres_
in(q; x) ! has_value_in(q; QRG)”. Using this syntax, the trip example can be
captured as a QG “ Cost (T rip) :: Low”, which is vague and will be further
operationalized as a quality constraint (QC) “ Cost (T rip) :: 600 (Euros)”,
meaning that “for each trip, its cost shall be low (less than 600 Euros)”. Note
that QCs share the same syntax with QGs, but have measurable regions [
QGs can be refined via its quality, subject, desired region, or a combination
of them [
We show some example snippets of QG refinement in Fig. 1. There are two
points to be noted. First, we allow alternative operationalizations of a QG. This
is particularly useful when dealing with context, which is captured as part of
the subject of a QG, e.g., a low cost trip means differently to professors versus
students. Second, we do not allow refinements from QCs, which are specification
elements, to QGs, which represent requirements, because our main purpose is
to turn informal stakeholder requirements into an eligible specification; we do
support refinements of QCs to QCs after operationalization (e.g., weakening [
This section presents a two-staged reasoning procedure for the satisfaction
analysis of quality goals. The first stage is previous work on measuring the satisfaction
of leaf-level QGs [
Measuring the satisfaction of leaf-level quality goals In general, complex quality goals will be reduced to sub-QGs and finally operationalized as measurable quality constraints (QCs) during software design. And, a QC is satisfied if the perceived quality values fall into the expected region, or unsatisfied otherwise. That is, such a QC is akin to a binary function that returns 1 or 0. This may be too strong a condition since in many cases “good enough” is sufficient. For example, if one has operationalized “Low” as “ 600 (Euros)”, how about a cost of 620 Euros? As we can see, the satisfaction degree, not the make-or-break status, is what actually matters. So, when operationalizing a QG, we need to derive a smooth function, not a binary one.
To address this issue, we have proposed to measure the satisfaction degree of
leaf-level QGs based on gradable membership [
The interesting point here is that we can use prototype values to represent a concept, and then use Voronoi diagrams to reason about the graded membership without the need of inventing made-up numbers as in existing quantitative goal satisfaction analysis procedures. 4.2
Propagating the satisfaction value of quality goals upstream In practice, a QG may refer to a composite quality that has a set of sub-qualities as its offsprings, e.g., “Security” has sub-qualities “Confidentiality”, “Integrity” and “Availability” (see the ISO/IEC 25010:2011 standard for a reference quality hierarchy), or a complex subject that has many different parts as it components, e.g, a system. In such cases, we need to first refine the complex QG to simple ones via its quality/subject, then use the technique discussed in Section 4.1 to measure the satisfaction of leaf-level QGs, and finally propagate up the satisfaction of lower-level QGs along the refinement hierarchy.
Recall that a QG can be reduced to a set of sub-QGs using our reduce operator. For such a refinement, we represent the QG as a vector of its subQGs. For example, we denote the refinement of a quality goal QG to QG1 and QG2 as QG =: [QG1; QG2]. We set the expected satisfaction degree of each subQG to be 1.0, and express the expected satisfaction of QG as a 1 2 vector: SatEQG = [SatEQG1 ; SatEQG2 ] = [1:0; 1:0]. Supposing that QG1 and QG2 are leaf-level goals associated with single-dimensional spaces, we can derive their satisfaction degrees by using the technique discussed in Section 4.1. On obtaining the satisfaction degree of QG1 and QG2, say sat1 and sat2, we have the vector SatRQG = [SatRQG1 ; SatRQG2 ] = [sat1; sat2]. Finally, we use Eq. 1 to compute the the satisfaction degree of QG, where n is the dimension of SatEQG (here, n = 2), d is the Euclidean distance between SatEQG and SatRQG, dmax is their maximum distance when SatRQG = [0:0; 0:0] and is in fact a constant pn.
SatQG = 1 d = 1 s Pin=1 (SatEQGi SatRQGi )2 (1) dmax n
The rationale of this approach is to reduce a composite quality in a highdimensional space to low-dimensional spaces (i.e., sub-qualities) and measure the similarity between reality and expectation. The geometric meaning of this approach is shown in Fig. 2. As shown in the figure, “Security” is represented as a point E in the three-dimensional space and can be refined to “Confidentiality”, “Integrity” and “Availability”. We denote the expected satisfaction of “Security” as [1.0, 1.0, 1.0], and suppose that the real satisfaction R to be [0.8, 0.4, 0.75]. The distance d between the expectation and the reality is represented by ER, the maximal distance dmax is OE, and the similarity between the reality R and the expectation E is (1 ER=OE). This approach can also be applied to subjects, which can be refined to constituting parts.
Integrity E 1.0 0.0
O vectors, with each of its sub-goals being a feature, and calculate their similarities using the metric defined in Eq. 1. Moreover, unlike Machine Learning techniques, our approach does not need to learn the satisfaction function of a leaf-level QG from big data; instead, we able to achieve the purpose by using prototype values, which are elicited from some stakeholders and are affordable in practice. 5
In this paper, we have proposed a two-staged procedure for the satisfaction analysis of quality goals. We first measure the satisfaction of leaf-level QGs based on graded membership and then propagate the satisfaction value upstream based on similarity measurement. Our approach is promising in addressing some of the deficiencies in existing reasoning procedures, e.g., the lack of sufficient expressiveness for qualitative approaches, and the meaningless of subjective numbers used in quantitative approaches.
We are going to implement this procedure in Desiree 4, a prototype tool that
has been developed to support our requirement engineering framework [
Acknowledgments. This work is supported by the European Research Council advanced grant (no. 267856), the Key Project of National Natural Science Foundation of China (no. 61432020) and the Intelligence Innovation Center, Alibaba. 4 The tool is available at https://goo.gl/oeJ9Fi.