248
Accessing Geoenvironmental Data: From On-line to Expert Systems
An Expert System must be integrated in a GIS kernel, thus requiring a knowledge
base (Bouille, 1984, 1988). While a database is supposed to store "data", a
knowledge base stores the knowledge made of facts and rules; it does not imply an
opposition between both terms; it might be considered as an extension, a
knowledge base simply using a database for ensuring its storage; in an 0.-0.
context, a database stores abstract data types, whatever they semantically represent;
if a class is a class of facts or a class of rules, it is stored likewise; but the main
difference does not come from the knowledge base itself; taken independently, it
is very easily implemented; the difference comes from the way the rules are used,
launched, chained, interpreted and first of all compiled; it is the task of the
inference engine, making this knowledge base "intelligent". The term deductive
database (or deductive knowledge base) is more adequate. As the reader may see,
the "intelligence", is not in the "base", itself but in the engine connected to it. As a
typical example, the toponymy must be stored with each geographical object; but
the way this toponymy is drawn on an output map follows some complex rules
which must be also stored in the same system. At run-time, the inference engine
will generate the most well-suited coordinates for each toponymical element.
The Expert System of a GIS must not be based on the stupid tree-structure, must
not use the obsolete pattern matching and back-tracking, and must work on fuzzy
facts and rules. In the near future, most GIS will include an 0.-0. expert system,
working on a partitioned knowledge base, with classes of facts and rules, the classes
of rules being nothing but particular facts, and the expert system using several
cooperating inference engines. For more than ten years, we can build such expert
systems whose speed does not depend on the number of facts and rules stored,
because they no longer use pattern matching. They use the pre-existing
relationships recognized at the structuring step. Such methods have been
successfully applied since 1987 for toponymy repositioning on maps (Tite 89).
A GIS must be able to allow decision-makers to test some hypotheses; a
geographic dynamic model requires thousands of processes, considered as
persistent objects implicitly stored in the knowledge base of the GIS. This tool is
one of the best ways for testing phenomena in which components move in
interaction: for instance, car flows in an urban area, schedules of transportation
networks, pollution propagation in pipes, plant disease expansion in a forest, etc ....
But sometimes, this discrete simulation must be connected to continuous processes,
for instance when the trajectories of some mobile components must be displayed
on a digital/topological terrain model; or when you build a very large geological
model... Simulation models will be one of the main tools in environmental studies
during the next decade.
Complementary tools are provided by the connection approach. Classical neural
models are essentially based upon matrices, on which various filters are applied.
On the other hand, 0.-0. neurons are objects of classes of neurons, and their
behavior is implicitly launched when an external action is performed on one of
their attributes or links previously specified as a launcher. The behavior is able to
perform any task on any other persistent type(s). Of course if these other types are
launchers of other neurons, the reader may imagine the propagation which can be
built very easily. While the matrices of the classical connection approach are not
extensible, the persistent 0.-0. neurons are unlimited because of the hidden
knowledge base, and new ones can be generated at any time. The behavior is
Accessing Geoenvironmental Data: From On-line to Expert Systems
An Expert System must be integrated in a GIS kernel, thus requiring a knowledge
base (Bouille, 1984, 1988). While a database is supposed to store "data", a
knowledge base stores the knowledge made of facts and rules; it does not imply an
opposition between both terms; it might be considered as an extension, a
knowledge base simply using a database for ensuring its storage; in an 0.-0.
context, a database stores abstract data types, whatever they semantically represent;
if a class is a class of facts or a class of rules, it is stored likewise; but the main
difference does not come from the knowledge base itself; taken independently, it
is very easily implemented; the difference comes from the way the rules are used,
launched, chained, interpreted and first of all compiled; it is the task of the
inference engine, making this knowledge base "intelligent". The term deductive
database (or deductive knowledge base) is more adequate. As the reader may see,
the "intelligence", is not in the "base", itself but in the engine connected to it. As a
typical example, the toponymy must be stored with each geographical object; but
the way this toponymy is drawn on an output map follows some complex rules
which must be also stored in the same system. At run-time, the inference engine
will generate the most well-suited coordinates for each toponymical element.
The Expert System of a GIS must not be based on the stupid tree-structure, must
not use the obsolete pattern matching and back-tracking, and must work on fuzzy
facts and rules. In the near future, most GIS will include an 0.-0. expert system,
working on a partitioned knowledge base, with classes of facts and rules, the classes
of rules being nothing but particular facts, and the expert system using several
cooperating inference engines. For more than ten years, we can build such expert
systems whose speed does not depend on the number of facts and rules stored,
because they no longer use pattern matching. They use the pre-existing
relationships recognized at the structuring step. Such methods have been
successfully applied since 1987 for toponymy repositioning on maps (Tite 89).
A GIS must be able to allow decision-makers to test some hypotheses; a
geographic dynamic model requires thousands of processes, considered as
persistent objects implicitly stored in the knowledge base of the GIS. This tool is
one of the best ways for testing phenomena in which components move in
interaction: for instance, car flows in an urban area, schedules of transportation
networks, pollution propagation in pipes, plant disease expansion in a forest, etc ....
But sometimes, this discrete simulation must be connected to continuous processes,
for instance when the trajectories of some mobile components must be displayed
on a digital/topological terrain model; or when you build a very large geological
model... Simulation models will be one of the main tools in environmental studies
during the next decade.
Complementary tools are provided by the connection approach. Classical neural
models are essentially based upon matrices, on which various filters are applied.
On the other hand, 0.-0. neurons are objects of classes of neurons, and their
behavior is implicitly launched when an external action is performed on one of
their attributes or links previously specified as a launcher. The behavior is able to
perform any task on any other persistent type(s). Of course if these other types are
launchers of other neurons, the reader may imagine the propagation which can be
built very easily. While the matrices of the classical connection approach are not
extensible, the persistent 0.-0. neurons are unlimited because of the hidden
knowledge base, and new ones can be generated at any time. The behavior is