next up previous
Next: Dissertation Contributions Up: Introduction Previous: Software Uncertainty Modeling

Premises and Hypothesis


To arrive at the approach taken in this dissertation, one must follow the statements in this section. Most statements, such as ``software complexity hinders its understanding'' or ``users get disoriented in large hypertext spaces,'' are either widely accepted or substantiated later. These statements therefore constitute the premises for this dissertation. One statement, however, is less accepted and requires further investigation: ``Explicit modeling of software uncertainties improves human judgment and decision making during software development.'' This statement therefore constitutes the hypothesis for this dissertation and will be further elaborated and validated.

Our initial observation is that software is extremely complex, i.e., that ``software systems are perhaps the most intricate of man's handiworks'' [Bro95]. Software complexity is not only ubiquitous, it is also intrinsic: ``The complexity of software is an essential property, not an accidental one'' [Bro87]. ``By essential we mean that we may master this complexity, but we can never make it go away'' [Boo94]. Complexity hinders software understanding, yet understanding must be afforded if development is to be successful.

To manage complexity, software development concerns are often separated in time and space. Thus, software is typically developed in distinct phases by different teams of developers. Large-scale software is often long-lived and delivered in multiple version, thereby comprising a family of related programs [Par79].

The product of software development is typically a large and complex collection of software elements of diverse types. Software artifact collections often contain not only millions of source code lines, but also versioned releases, volumes of user manuals and documentation, requirements and design specifications, test cases and test results, and so forth. Moreover, these software objects are typically integrated and correlated within themselves and with each other in surprisingly intricate and subtle ways [SZ92]. This poses substantial impediments to software understanding, particularly with respect to traceability, visualization, and navigation issues, thereby leading to the following research question.

Research Question:

Given the intricacy and complexity of software artifact relationships, what can be done to improve traceability, visualization, and navigation of large artifact collections?

As a first step, it is proposed that software systems be viewed as large hypertexts, since hypertext offers a coherent and consistent metaphor for viewing both inter- as well as intra-artifact relationships (cf. [ZO95, SZ92]). Here, it suffices to define hypertext informally as providing node-and-link views and navigation of software artifact collections. Hypertext is defined more formally and completely in Chapter gif. We therefore contend that a hypertext metaphor of software artifacts and relations should improve developer's ability to trace or track related elements in large and complex software spaces.

Despite the above, hypertext is no ``silver bullet'' for relieving software artifact complexities and uncertainties. Specifically, hypertext introduces several concerns regarding its efficacy for software engineering [ZO95]. Here, we are particularly concerned with the well-known navigation problem of user confusion and disorientation in large hypertext spaces as well as the fact that hypertext systems generally are not designed with software engineering in mind. This means that hypertext does not necessarily address the unique needs of software developers and users nor does it model or make use of the unique characteristics of software systems. Additional means should, therefore, be provided to further facilitate understanding of software artifact collections. The following statements concern this facilitation.

We claim that software engineering is fraught with uncertainties. Software uncertainties contribute significantly to the overall complexity and unpredictability of software development. Like complexity, uncertainty is inherent to the engineering of software systems. This observation is summarized succinctly in [ZRK96] as the Maxim of Uncertainty in Software Engineering:

Uncertainty is inherent and inevitable in software processes and products.

Examples of software uncertainties abound - many are provided in this dissertation. Starting in Chapter Four, we focus on uncertainties associated with developers' confidence levels in software artifacts, including, among others, requirements specifications, design elements, code modules, and testing information. These confidence levels fluctuate frequently during development. Their modeling should, therefore, include a scheme for confidence revision and updating.

We now arrive at the high-level statement of the hypothesis for this dissertation.

High-level Research Hypothesis:

Explicitly modeling uncertainty improves human judgment and decision making during software development.

This hypothesis is too vague to be investigated and validated directly. Instead, we require a more specific formulation, including a specific technique for software uncertainty modeling. To model software uncertainties, we turn to artificial intelligence research in modeling and management of uncertainty. We select a specific uncertainty modeling technique called Bayesian belief networks [Pea88]. Detailed reasons for this choice are provided in Chapter Four.

Here, it suffices to say that Bayesian networks offer a clear conceptual model of causality among related elements and include algorithms for belief revision and updating. This leads to our specific hypothesis below. The hypothesis speaks in terms of software artifacts, which include, among others, requirements elements, design nodes, code modules, and test information, relationships among those artifacts, established as part of the development process, and confidence levels, which reflect developer confidences in certain qualities and properties of software artifacts.

Specific Research Hypothesis:

Bayesian-network models of software artifact uncertainties improves understanding of associated confidence levels, compared to simply following linked information, ultimately leading to better human decision making.

In the long run, we believe that proper adoption and use of uncertainty modeling techniques will improve human judgment and decision making during software development. For the purposes of this dissertation, we focused on the specific hypothesis above and took steps toward its validation, as follows:

  1. An existing software system was selected as case study. The system of choice, called CEquencer, is currently under development at Beckman Instruments in Fullerton, CA. CEquencer is designed to control and communicate with various hardware devices that in turn are used by biologists, chemists, and other scientists to separate laboratory specimens into molecular constituents in order to determine their DNA sequences. CEquencer software artifacts are described in Chapter Five.
  2. CEquencer requirements were investigated throughly, and a hypertext model as well as a Bayesian-network model of those requirements were constructed.
  3. A key subsystem of CEquencer, called FSM Run, was investigated thoroughly, resulting in a hypertext model of software artifacts, including all FSM Run code modules as well as related requirements.
  4. Software uncertainty information was collected for FSM Run artifacts. The information was elicited from Beckman experts and developers and organized in a Bayesian network.
  5. Our main hypothesis was evaluated against FSM Run code modules and related requirements, as identified in (2) above. In a questionnaire session with the Beckman team, developers were asked to trace artifacts and associated confidence levels with and without the Bayesian-network representation of confidence values.
  6. Results were gathered for the case study and analyzed for statistical significance. Results of the questionnaire session are summarized in Chapter Six.

next up previous
Next: Dissertation Contributions Up: Introduction Previous: Software Uncertainty Modeling

Hadar Ziv
Fri Jun 20 16:22:31 PDT 1997