In this subsection, experience and additional thoughts regarding the refinement process are presented.
Using constraints to describe policies supports the idea of policies for management, where policies are used to achieve an objective. Therefore, it is important to specify the desired objective in the beginning. Constraint based-policies are used for this purpose. As in the other approaches, it is necessary to refine (high-level) policies, because using constraint-based policies does not change the fact that an objective is vague at first.
The objects used in the constraint statements must be well known and provided by an appropriate information model of the infrastructure. In discussions2.4 and work on the refinement process of constraint-based policies it became clear that the possibility of adding new objects and relationships between objects to the information model during the refinement of a policy would make this task easier. At least at the moment, an applicable generic information model has not been developed for most management needs. This possibility of adding objects would not be as important, if in the future, customised or branch-specific information models (as for insurances, banks, production plants, warehouses, etc.) were available. These information models could be the basis for applying predefined branch specific policies.
The strong orientation of the policy approaches towards obligating actions as described above results in a refinement process where mostly actions are refined. In the constraint-based approach, the focus is more on the constraints. This includes conditions on states and state transitions and thus on objects and relations between objects. Typical tasks in the refinement of a constraint-based policy are the refinement of objects. This could be supported by the information model with object hierarchies or inheritance combined with specialisation. Otherwise, additional constraints are used to specify a subset of the initial set of objects. This could be accomplished for example with domain scope expressions or constraints on attributes of the objects. Refinement of relationships between the objects must also be done. When the refinements are treated like objects (entities) hierarchical structuring and inheritance with specialisation could again be used for accomplishing this task. In addition to refining the relationships and objects, the constraints themselves are to be refined. When refinement of already specified objects and relationships is not enough, needed objects and relationships are added. Splitting of a policy into a number of policies for the purpose of refinement is also possible. At the end of the refinement process, a policy is a passive policy or an active policy with operations.
The operations as the final implementation of an active policy are added during the final phase of the policy refinement process. The operations added to an active policy are called implementables. An operation is sufficient if it fulfils all of the constraints of the policy. Choosing the operation near the end of the process minimises unwanted side effects as a result of choosing inappropriate operations, because at this level of refinement most of the constraints are known.
Support from an interactive planning system could provide suggestions as to which set of operations is able to transform a system from one state into another. Such a planning system is given constraints, and by satisfying these, a set of possible actions is calculated. Using such a system would help to prune the search space of actions in the refinement of policies [BLMM 97,BLMM 98,BaBa 98,BaMi 98].