Monday, April 25, 2005

Goal-Oriented Requirements Engineering and its relevance to SOA

The i* approach

Last week I attended a seminar organized by the BCS RESG. Modelling your System Goals: The i* approach. The main speaker was Eric Yu (University of Toronto), the original developer of i*. The remaining speakers reported their experiences using and extending i* in a range of projects.

The i* approach is a methodology for goal-oriented requirements engineering (GORE). Professor Yu characterized the distinct nature of Goal-Oriented RE as follows.

Conventional RE Goal Oriented RE
Describes how things work (AS-IS) and how they should work (TO-BE). Describes why things work (or should work) that way.
Models of behaviour. Models of intention and rationale.

i* is particularly focused on so-called "early stage" requirements engineering - before you know what "the system" is going to be. It involves two key models: strategic dependency models and strategic rationale models.

The strategic dependency model identifies goal dependencies between actors. For example, a CarOwner actor depends on "CarBeRepaired", and this can be fulfilled by the BodyShop actor. Put crudely, a goal dependency consists of a pair: {"I want", "I can"}.

i* distinguishes between two different kinds of goal dependencies: hard goals and soft goals. Although this distinction was explained in terms of precision, this seems unsatisfactory to me, since surely precision is itself a matter of degree and timing and negotiation. However, the soft goals that were used as examples seemed to me to possess some other interesting features, which may be relevant to making this distinction clearer. Firstly, the soft goals were generally not amenable to precise specification by the consumer of the goal. Secondly, there is a sense that what counts as satisfaction of the goal is context-dependent and shifts over time. Thirdly, the perceived outcome would generally affect the relationship of trust between the consumer and the provider, and thus the future behaviour of the consumer towards the provider.

This led me to an alternative way of formulating the distinction - according to who owns the goal. In some cases, the consumer determines and decomposes the problem, and then delegates the solution. I think this produces a hard goal that is (or should be) relatively easy to specify. In other cases, the consumer delegates the problem (formulation of) as well as the solution (provision of). I think this would produce what i* calls a soft goal.

So for a soft goal such as FairSettlement, what counts as fairness is delegated either to the service provider, or to a trusted third party such as a regulator. (In some cases, notions of fairness are implicitly delegated to virtual third parties, such as TheCommunity or PublicOpinion, but these are notoriously unstable and unreliable mechanisms.) The significance of this distinction for goal modelling is that identifying something as a soft goal introduces an additional set of concerns for the requirements analyst.

The strategic rationale model identifies causal connections between goals and subgoals for a given actor. Note that the behaviour of an actor may depend on perceived causal connections and associations, rather than actual ones. So if we are going to model rationale properly, we need to include some representation of the actor's beliefs. However, last week's i* presentation did not introduce any notation for belief.

SOA interpretation

Much of the material presented was pre-SOA in outlook, and there were some limiting assumptions about the nature of the systems being built. However, I was looking to push i* beyond these assumptions. What I wanted to explore was the possibility of using i* to design services and service-oriented solutions.

There appears to be a very straightforward mapping from goal dependencies on the strategic dependency model to business services. And these requirements are supported by an understanding of the rationale of each actor. The reason for this is that we want to design business services in terms of added value, and this seems to rely on our having some notion of value from the perspective of the service consumer. SOA is geared towards flexibility, and an appreciation of the possible rationale of each actor also helps build solutions that support a good range of different scenarios.

Another SOA-related motive for modelling dependency and rationale is to analyse compliance (including so-called non-functional requirements). For example, we may identify some system vulnerability, and then identify a reciprocal dependency that provides some enforcement mechanism. We can then look at the system dynamics within a collaboration, to determine the adequacy of this mechanism, as well as any possible side-effects.

In general terms, I am convinced that some modelling of intentions and outcomes is an important aspect of modelling for SOA. In the Boxer-Cohen Triple-Articulation approach to Asymmetric Design, this is known as the Deontic Articulation. i* is much simpler, and widely practised - but is it powerful enough?

Complexities, difficulties and future opportunities

Modelling intentions. There are difficulties involved in modelling the intentions of both organizations and machines, and these difficulties cause some people to say that it doesn't make sense to attribute intentions to either organizations or machines - only to people. I don't agree, for three reasons. The first is that many of these difficulties also apply when we are trying to understand the intentions of people, since people often have confused or concealed intentions. The second is that there are established techniques for tackling these difficulties, which help us to understand the intentions of all kinds of system, at different levels. The third is that we simply cannot make sense of business strategy and organization at a reasonable level of abstraction without somehow talking about organization goals.

Goal mismatch
. In practice, there is unlikely to be a neat and exact match between "I want" and "I can". There may be both pragmatic mismatch (crudely: what the provider does is not exactly what the consumer wants) and semantic mismatch (crudely: what the consumer understands is not what the provider understands). This is a form of impedance or asymmetry.

For example, a car driver's real goal may be to have a reliable car with low total maintenance cost. But no service provider offers exactly this service. So the car driver has to compose something that approximates to his real goal, using a combination of car maintenance services and car warranty (insurance) services. (There may be significant commercial opportunities for a value-adding service or service platform in this kind of space. Hence the relevance of goal modelling to business strategy.)

Within i*, the strategic dependency assumes an exact match between the consumer's view of the goal and the provider's view of the goal. To the extent that there is some unfulfilled remnant, this is analysed within the strategic rationale of that actor. But it is not clear to me how this approach would support architectural reasoning about the unfulfilled remnants and the strategic opportunities for developing new added-value services.

moreDeconfliction. Given that a complex situation contains considerable goal conflict, we need a systematic way of resolving goal conflict. Deconfliction means organizing operations in a way that minimizes the potential risk of interference and internal conflict. Conversely, we need ways of resolving goal synergy.

Note that from a security perspective, goal synergy between (supposedly) independent actors may not be a good thing - especially where it indicates opportunities for collusion and fraud.

Variation. A robust service economy typically needs to accommodate considerable variation in the goals and rationale of the actors. So instead of modelling a single standard strategic rationale, we need to understand the nature of the variation between different rationales, and the implications of this for producing robust and agile systems.

For example, a healthcare system makes some assumptions about the rationale of a patient. But a patient that happens to be a professional athlete (with particular concerns about stamina and performance) will have a very different strategic rationale in relation to healthcare, as compared to a patient that happens to be a healthcare professional (with above average exposure to infection and other risk).

Dynamic and recursive rationale modelling. The causal relationships between goals and subgoals are subject to dynamic effects. For example, many processes operate in terms of surrogate goals - outcomes that are valued not for their own sake but because they are correlated with some real goal. The surrogate goals are available for measurement before the real goals, and may therefore provide useful predictive metrics. (See discussion of Surrogate EndPoints in relation to SOA Pharma.) This means that cause-effect modelling may need to include system dynamics such as delay and oscillation.

There is also a problem that the rationale of one actor may depend on that actor's beliefs about the rationale of another actor. For example, Professor Yu presented an example from insurance, where reorganizing the strategic dependency model could align the insurance broker with the interests of the customer (instead of the broker being aligned with the interests of the insurance provider). In this situation, what the insurance customer demands from the insurance broker depends crucially on the customer's belief as to whose side the broker is on. But that in turn may depend on the insurance broker's beliefs about the customer's acting in "good faith", and about the future commercial tactics of the insurance company. So the whole thing becomes recursive, rather like the Knots identified by R.D. Laing.


I think the overall approach of i* is extremely interesting for service-oriented architecture, and is broadly consistent with the general approach to business modelling for SOA we have been developing.

A version of this post was published in the RESG Newsletter Requirements Quarterly (June 2005).

No comments: