ABSTRACT
Reasoning patterns occurring in complex tasks cannot be modeled only by means of a pure classical logic approach. This is due to several factors, for instance, incompleteness of the available information, need of using and representing uncertain or imprecise knowledge, combinatorial explosion of classical theorem proving when knowledge bases become large, or a lack of methodology in building complex and large knowledge bases (KBs).
To deal with these problems an advanced architecture for knowledge base systems (KBSs) must combine control, implicit and explicit, with modularization techniques, together with an approximate reasoning component based on many-valued logics. In this section we survey the main control techniques needed by means of a generic modular architecture. The implicit control is usually based on a subsumption mechanism together with a rule elimination process. The explicit control, declarative in nature, is, in advanced architectures, a metalevel approach, based on reflection techniques and equipped with a declarative backtracking mechanism. Reflection and subsumption techniques can be used to tackle the problem of incompleteness of the available knowledge. The metalevel approach permits assumptions based on the current state of the object deductive process, and the reflection technique makes them effective at the object level. If, later on, these assumptions are proved to be erroneous a declarative backtracking mechanism allows them to be retracted. Other complex reasoning tasks can be implemented using a combination of the overall set of control methods.
The paper starts with an introduction in which the problem of the control of inference is introduced in the context of a simple KBS, that is a system based on pure classical logic. Next, the main control methods for complex KBSs, using a generic KB architecture, are described and finally some conclusions are presented.
