Information and Knowledge Representation in Fusion Problems - Information Fusion in Signal and Image Processing - page 59

The different types of knowledge represented in these systems are divided into the
following categories:
- declarative (how things are);
- procedural (how things are done);
- episodic (related to the previous experience);
- and meta-knowledge (knowledge about knowledge).
Generally, it is extremely difficult to change from one type of knowledge to an-
other (and often even impossible) and therefore these categories of knowledge are, by
nature, quite different and may all be necessary.
Control consists of searching for paths between initial knowledge and goals, us-
ing techniques called forward chaining (applying inference rules when new data is
declared, with the consequences possibly triggering new inference rules), or backward
chaining (applying inference rules when new requests are stated, in which case the
premises of these rules that have not been verified generate new rules).
The most common examples of KBS include:
- production rules;
- frames;
- semantic networks;
- systems with uncertainty: Mycin [SHO 76], etc.
Production rule systems (of the type if. . . and/or. . . then. . . ) are systems that are
easy to adapt or extend, and the way they operate and their results can easily be ex-
plained. They have the drawback of having a fragmented representation of knowledge
which causes a lack of efficiency. Their power of expression depends mostly on the
type of logic used. For example, first-order or predicate logic make it possible to han-
dle variables and quantifiers, whereas in propositional logic, everything is constant.
Frames constitute a declarative form of KBS in which a list of attributes or proper-
ties is supplied with characteristics and values for these characteristics. They are useful
in describing general concepts, classes of objects. Classes with different granularities
can be handled using links of hierarchy, inheritance, specialization and instantiation.
Most of the time, these systems are static, but some dynamics can be introduced by
assigning procedures to attributes.
Semantic networks rely on a graphic representation of the knowledge base, in
which the nodes represent concepts and objects, and arcs represent relations. The infer-
ence rules are based on inheritance properties when taking an arc from one class to a
more specific class. These networks are often used in natural language processing, for
example.