Information Technology Reference
In-Depth Information
PRODIGY is a general problem solver combined with learning modules. The
architecture is depicted in Figure 9.7. Based on problem solver of operator, a
unified control structure for inference rule and operator search is exhibited. The
problem solver comprises of a simple reason maintenance sub-system and
stipulates the influential operators. The search of problem solver can be
navigated by control rules.
Control Knowledge
Problem
Domain theory
Abstraction Generator
User Problem
EBL
Problem sovler
Deriv-
ation
Multi-
level
Plan database
Abstraction level
PS track
Experiment
Derivation
Abstraction
External disposal
Problem solving
Figure 9.7. Architecture of PRODIGY(Picture from Minton et al.,1989)
The EBL module obtains control rule by tracking problem solver. The
explanation is constructed from domain theory and problem solver, and then the
result expression is converted into control rule for searching. The derived module
is the derived analogy, which can recap the whole solution of previous similar
problems. From figure 9.7, it is obvious that analogy and EBL are two
independent mechanisms to get the control information for a concrete domain.
The experimental learning module is to refine the incomplete or incorrect domain
theory. When the plan execution controller detects that there is some difference
between inner expectation and outer expectation, the experimental module is
triggered. Abstraction generator and abstraction level model provides multi-level
plan capability. Based upon the depth first analysis of certain domain knowledge,
the domain theory is divided into several abstraction levels. The PRODIGY
builds abstract results and makes refinement when a problem begins to solve.
PRODIGY can learn from four goal concepts: success, failure, the only
choice and restriction of goals. Every time a goal and an example is given to a
user, the system first decomposes the goal to the leaf node backwardly and gains
