Environmental Engineering Reference
In-Depth Information
to achieve the objectives. Model-based systems can be extremely effective at
creating plans in the face of one or more damaged subsystems.
One can argue that many planners are model-based since their database
of potential actions implicitly defines the system and its capabilities. While
there is some merit to this argument, it misses the fact that this implicit
knowledge is usually generated by human programmers and may not be com-
plete. These systems are, therefore, artificially limited in what they can plan.
A model-based planner, using its deep understanding of the system it con-
trols, can come up with novel approaches to achieve the objectives. This is
most important when the system is damaged. Model-based systems can use
their model to work around the damage. In other approaches, the human pro-
gramming must explicitly define the strategy to use when failure occurs, and if
they did not make a plan for the failure, the system will respond suboptimally.
The advantages of model-based approaches are that they only need the
model description of the system being controlled and the rest is automatically
determined. Their disadvantage is that reasoning on models can be very slow
and the necessary models are often hard to construct and possibly incomplete.
Model-based approaches are often used in support of other forms of planners.
5.1.5 Case-Based Planners
Case-based planners are systems that represent their planning knowledge as
a database of previously solved planning cases. Case-based planners exploit
the idea that the best way to solve a new problem is probably very similar
to an old strategy that worked. Their cases can be a mixture of symbolic and
numeric information. Case-based planners are a class of case-based reasoning
(CBR) technology [
1
,
116
,
138
].
Figure
5.6
shows the structure of a case-based planner. When a case-
based planner is given a new goal, it attempts to find a similar case in its
case database. The cases are indexed on the goals being solved and the envi-
ronmental conditions when they were solved. If similar cases are found, they
are extracted, adapted to the current situation, and analyzed to see whether
they will achieve the goal being pursued. Often the case will be simulated
in the current environment to determine its actual behavior. This simulation
can discover defects in the plan so that repair strategies can be applied to
customize the plan for the current situation. The new case is simulated and
the cycle repeated. Once a plan has met all the necessary requirements, it is
executed. The results of execution are examined, and if it did not achieve the
objective, new cases are retrieved and the cycle is repeated. Cases that are
successful may be placed in the database for future planning activity. In this
way the case-based planner learns and becomes more proficient.
Two diculties that arise in case-based planners pertain to the methods
used to index the cases and the reasoning component of the system. How cases
are indexed determines whether appropriate cases will be found when they are
needed to solve a problem. If the indexing is too restrictive, an appropriate
