Support for Heterogeneity

There is a strong underlying tension in the HPKB endeavor. This is the tension between the real need to provide strong paradigm oriented guidance in a tool and the equally real need to free the developer from implementational handcuffs that prevent integration and heterogeneity.

Paradigm oriented guidance comes from tailoring a tool to the viewpoint of a problem solving paradigm which dictates the choice of representation and control structure. The paradigm assumes the role of a pair of glasses which helps to see the problem in a clearer light. However, when mismatched to the problem the paradigm blurs rather than focusses, leading to unclear implementation and hackery.

Most applications are heterogeneous in that they incorporate subproblems with different relevant paradigms. A tool which provides strong paradigm oriented guidance might well prevent the flexiblity needed to employ the relevant vehicle in the right place. Rather than a set of focussing glasses it becomes a set of handcuffs.

This tension was addressed in the Shrobe et al.'s JOSHUA system which developed the idea of a Protocol of Inference.

The distinguishing characteristics of this approach are:

• There is a Uniformly Accessible but Heterogeneous Data Base of Statements. The two functions ASK and TELL provide the interface to this database. ASK and TELL may implement their behavior in any manner desired and the statements may be represented using a variety of different data structures.

• There is a Fine-Grained Protocol of Inference in which each distinct step of processing a statement is identified and made accessible. This protocol is hierarchical as well as fine-grained. ASK, TELL, rule compilation, rule triggering and truth maintenance are all parts of this protocol as are their component steps.

• Each step of the protocol is a Generic Function, i.e., an abstract procedure whose concrete implementation is found by dispatching on the data types of its arguments. Statements are regarded as instances of the class associated with the predicate. The protocol steps are implemented as methods.

• Class-based inheritance allows facilities to be reused and overridden. Methods on more abstract classes constitute the default behavior which may be overriden by more specialized classes. To use a novel set of data structures, or to call out to an external problem solver, one needs to change only a few, small steps of the protocol.

We have used this approach in a number of projects at the MIT AI Laboratory.

One of the most important ways in which START's capabilities can be increased is by incorporating the capabilities of other HPKB systems. The Protocol of Inference is a recipe for how to build a system which flexibily incorporates a variety of other capabilities. We plan to redesign START's inference layer along the lines of an improved Protocol of Inference. Since we think that this extended Protocol of Inference will be valuable to other HPKB contractors as well we will also solicit other HPKB contractors to participate in its design and freely share the results.