What Students Should have Learned

What's so Special About Case-Based Reasoning?

Case-based reasoning is not the first artificial intelligence
method to combine reasoning and learning, but it makes learning
little more than a by-product of reasoning.  That feature makes it
unique among artificial intelligence methods.  In other words, a
case-based reasoner that remembers its experiences also learns as
it reasons; feedback from early experiences gives it insight into
solving problems at a later date.

What the Heck is a Case?

    A case is a contextualized piece of knowledge representing
    an experience that teaches a lesson fundamental to
    achieving the goals of the reasoner.

The basis for case-based reasoning is that situations recur with
regularity; what was done in one situation is likely to be
applicable in another situation of similar context.  In essence, we
begin the case-based reasoning effort by starting with what worked
in a previous situation that is similar to the present.

Artificial Intelligence and Case-Based Reasoning

Some scientists, who have been trying to explain how the human mind
thinks, are putting their efforts into something called artificial
intelligence.  In some areas of artificial intelligence, a great
effort first goes into developing a model of how a thought or
decision-making processes works.  Then, that model is generally
applied to all problem situations involving that particular
process.
General knowledge has some important advantages, one being the
economy of storage that comes through minimizing the amount of
knowledge needed to solve the problem.  General knowledge also
allows people to deal with uncertainty.

General knowledge, however, has certain disadvantages.  One being
something called operationalization, which is the difficulty we
encounter when we try to operate something specific given general
rules.  Sometimes, we need to have specific knowledge when
performing specific problem-solving tasks.  Another disadvantage of
general knowledge is that, although it covers the normal, it
doesn't tell us how to reason about situations that are different
from normal.

So, What do Cases Represent?

Cases represent specific knowledge that is associated with
specific situations.  They represent knowledge that is at the
operational level.  They make that knowledge explicit for
performing a task or using a piece of knowledge or using a strategy
when trying to accomplish a goal.  Cases also represent an
experienced situation.  That situation, when remembered later,
forms a context in which the knowledge embedded in the case is
presumed applicable.

What does a Case Contain?

Cases come in many shapes and sizes.  They may cover a situation
that evolves over time or they may represent a situation that must
be solved at one point in time.  In addition, they may represent
problem-solving episodes or they may associate the description of a
situation with a particular outcome, or they may do both at the
same time.

Which Cases are Worth Remembering?

Normally, tasks that are routinely performed are not considered as
cases worth remembering.  These are the normal, daily, routinely-
performed tasks (i.e., they represent the "norm").  We shall call
these schemata.  Those that are not routinely performed, however,
are usually worth remembering.  These are experiences that are
different from the "norm", and we shall call these "cases".  Stated
in another way, cases that are worth remembering are experiences
that are different in some way from what was expected.  They record
major variations from the norm.

What constitutes a "major variation from the norm"?  Well,
researchers have formulated a general rule which states, "If the
difference is instructive such that it teaches a lesson for the
future that could not have been inferred easily from the cases
already recorded, then record it as a case."  A lesson is learning
the normal way of doing something if it is currently unknown how to
do it.  Therefore, cases represent current differences from the
current norm.

A Summary of the Discussion thus far:

   1) a case represents specific knowledge tied to a context.  It
      records knowledge at an operational level.

   2) cases can come in many different shapes and sizes, covering
      large or small time slices, associating solutions with
      problems, outcomes with situations, or both.

   3) a case records experiences that are different from what is
      expected.  Not all differences are important to record,
      however.  Cases worthy of recording as cases teach a useful
      lesson.

   4) useful lessons are those that have the potential to help a
      reasoner achieve a goal or set of goals more easily in the
      future or that warn about the possibility of a failure or
      point out an unforeseen problem.

   5) Each case serves two purposes when it is recalled:

      a) it provides a suggestion on how to solve a problem, and

      b) it provides a context for understanding or assessing a
         situation.

Case retrieval is the first primary process of case-based
reasoning.  This is not done blindly, as only those cases that
partially to completely match the context of the problem are
retrieved.  Retrieving cases involves two intensive efforts: 1)
searching the cases and 2) matching cases with the current problem.
This is done by a computer program that rapidly scans the library
of cases on hand, selects those that at least partially match the
context of the present problem, and places them into non-permanent
memory.

A ballpark solution is then proposed.  This is done by extracting
the solution from some of the retrieved cases and proposing it as
the solution to the new case.  Note that, at this stage of case-
based reasoning, only a general solution is proposed; hence use of
the term "ballpark" (i.e., within the general area).  Ballpark
solutions are sometimes selected by the case-based reasoner when it
determines that certain portions of the ballpark solution are
appropriate for the current problem.  Alternatively, the case-based
reasoner can select ballpark solutions by determining if one
portion of the ballpark solution is more appropriate for solving
the problem than others.  A third way that a case-based reasoner
can select a ballpark solution is to "check" with the goals of the
decision-maker to determine which portion or portions of a
potential ballpark solution are consistent with those goals.

Next, adaptation or justification of the proposed solution is done.
When using case-based reasoning to solve problems, adaptation is
normally used.  This is accomplished by suggesting small changes in
the ballpark solution to propose a different solution that is
better-adapted to solving the specific problem.  Two steps are
taken in adaptation: 1) identifying what needs to be changed, and
2) making the changes.

When using case-based reasoning to interpret a ballpark solution,
justification is employed.  In case-based reasoning, justification
is the process of creating an argument for the proposed solution by
comparing and contrasting the new situation with prior cases.  The
reasoner then looks both for similarities between the new situation
and the selected cases that justify the desired result and
differences between the new situation and the selected cases that
imply that other factors must be taken into account.

Criticism of the proposed solution is then done as a critique
before the solution is applied to the new problem.  When all the
knowledge necessary to evaluate the proposed solution is known,
this step is called validation.  When all such knowledge is not
available to conduct a validation, alternate cases are called from
memory and their proposed solutions are compared and contrasted
with one another as well as with the currently-considered solution.
This provides the decision-maker potential scenarios through which
the proposed adapted solution may succeed - or fail.

Evaluation of the solution is then conducted after it has been
selected, applied, and carried through to completion.  This step
includes feedback about how the solution was implemented, what went
right, as well as what went wrong.  It also includes brief
explanations about what could have been done to prevent some of the
complications when implementing the solution.  Of all the steps of
the case-based reasoner, this is the most important because it
provides the information necessary for the case-based reasoner to
learn.  It allows the case-based reasoner to "notice" the
consequences of its reasoning.  In essence, evaluation is the
process of judging the goodness of a proposed solution.

Storage of the new case is then accomplished by the computer
program.  Remember, if this new case meets the criteria of
representing an appreciable departure from the current norm, and in
its solution it teaches a lesson, and also satisfies one or more
goals of the reasoner, then it should be saved and indexed for
future use.  Perhaps the most important part of the memory storage
process is the indexing.  Decisions on how to index this case has a
direct bearing on how efficient the searching and matching
algorithms will be.

Decision-makers use case-based reasoning when they remember
previous situations that are similar to the present for helping
them solve a problem.  Case-based reasoning is commonly used for:

  1) adapting an old solution to a new problem,

  2) identifying possible failures when adapting old solutions to
     new problems, and

  3) interpreting the present situation by comparing it to many
     past situations.

Case-based reasoning is really a model of one kind of reasoning
that makes use of problem solving, understanding, and learning.
Further, it connects or integrates these three processes within
something called memory.

Five Premises of the Case-Based Reasoning Model

A premise is something that has been proposed - and proved - as a
basis for argument.

1) Referencing or recalling the circumstances of old situations or
   cases is advantageous when dealing with new situations.  In
   addition, such referencing is usually necessary to understand,
   interpret, and/or solve problems that are involved with novel
   (new) situations.  Therefore, remembering a case to use in a
   later problem-solving endeavor, and integrating that case with
   what is already known, is a necessary learning process.

2) Interpreting - (understanding) - the details of a problem is
   necessary for its solution because descriptions of problems, past
   and present, are rarely complete.  Such interpretation needs to
   be done by the decision-maker before case-based reasoning can
   begin.  As the decision-maker begins to understand the various
   aspects of the problem at hand, then similar cases can be
   recalled to help her or him solve it; as more details are
   understood, more cases can be recalled, evaluated, and either
   kept or discarded.

3) Adapting the solution of an old case to a new situation is
   almost always necessary because old cases are almost never
   identical to new situations.  In essence, adaptation compensates
   for the differences between old cases and new situations.

4) Learning occurs as a natural consequence of reasoning and
   experience.  When a novel solution is derived when solving a
   complex problem, and if that solution is correct, a connection
   is made between that new solution, the set of circumstances
   involved, and the set of cases used to come up with that
   solution.  These circumstances and connections and the new
   solution are stored in memory and indexed so they can be
   retrieved when they are needed.  On the other hand, if that
   particular solution is not a correct solution, that set of
   connections to the circumstances and past cases is indexed so
   the decision-maker can be warned that potential troubles can
   arise if these solutions are used.  Such learning (i.e., the
   processes of making connections, remembering past cases,
   understanding details of the present case, indexing the
   solutions and such and storing and recalling those  solutions)
   occurs in increments.

5) Evaluating the reasoning process is useful to the learning
   process.  This is usually done through something called
   "feedback" and "analysis".  Feedback is the objective evaluation
   of the case-based reasoning process, determining what went
   "right" and what went "wrong".  The analysis of that feedback is
   accomplished by conducting alternate procedures/recalls, running
   them through the mental process, and determining potential
   outcomes.  Such evaluation is a useful part of the
   reasoning/learning cycle.

These premises strongly suggest that the quality of reasoning when
a reasoner uses case-based reasoning depends on the
following:

1) its breadth and depth of experience,

2) its ability to understand new situations and correctly interpret
   them in relation to those old experiences,

3) its ability to adapt old solutions to new problems,

4) its ability to conduct the feedback and evaluation processes and
   make "repairs" when the solution is not working out well, and

5) its ability to make the correct connections between the new
   situation, old experiences, old solutions, and new solutions
   (learning).


Recall Starts the Case-Based Reasoning Process

After the decision-maker understands the details of the new
problem, she or he must begin the reasoning process by selecting a
set of representative cases.  This set of cases must involve the
goals and sub-goals of the decision-maker.
Using these selected
cases, past successful attempts to solve the problem are proposed
as solutions to the new problem.  Also when using these selected
cases, past failed attempts serve to warn the decision-maker of a
potential failure to meet those goals.

Interpretation: Understanding a New Problem in Terms of Old
Experiences

Interpretation is the process of comparing the new situation to the
recalled experiences.  This is done by the case-based reasoner when
it compares and contrasts the old experiences with the situation of
the new problem; this is an interpretation of the new problem.
Interpretation includes inferred knowledge about the new problem
and, sometimes, a classification of the new problem.  (To
infer something is to come to some state of judgment or
conclusion based on facts or information that are known or gathered
about the subject.)  When solutions to the new problems are
compared to solutions of past but similar problems, the case-based
reasoner gains an understanding of the positive and negative
consequences of choosing some solutions over others.
Interpretation is generally used when the problem is not well
understood and there is a need to criticize the solution.  In
contrast, when a problem is well understood (i.e., there is little
to no missing information), there is little need for the
interpretive process.

Adaptation: Fixing Old Solutions to Solve New Problems

Fixing solutions that were used to solve previous but similar
problems is called adaptation.  Adaptation is necessary because no
old situation is exactly the same as a new one.  Adaptation can
happen during the formulation of a solution or after feedback that
is derived from evaluating the results of an old solution to a
similar problem.  Generally, adaptation can be accomplished by
substitution or transformation.

Substitution is accomplished by substituting, for some part of an
old solution, a replacement action or substance.  Often this is
deemed necessary because of ethical, environmental, logistic,
economic, or other constraints.  Currently, there are six different
recognized methods of substitution (see Kolodner 1993 for more
information on these).

Transformation methods are useful for changing an old solution into
one that will work for the new problem.  There are two forms of
transformation methods, one called commonsense heuristics and the
other called model- or principle-guided repair.  The former method
involves a small set of general heuristics, or rules, that use
knowledge about the relative importance of different parts of a
solution to determine whether deletions or substitutions should be
made.  The latter method is used when the decision-maker has
knowledge of the cause and effect "connections" that exist in the
system.

Improving the Performance of the Case-Based Reasoning Process

Artificial intelligence programs that use case-based reasoning to
solve problems must have the ability to learn from their
experiences.  This is done through feedback, evaluation, and
repair.  Without these, the program would get faster at generating
solutions but not better.  In other words, they or they would
repeat their mistakes.  There are two ways that performance is
improved through the use of case-based reasoning: 1) becoming more
efficient, and 2) becoming more competent.  The former occurs when,
in the reasoning process, old solutions are simply remembered and
then adapted to solve new problems rather than going through the
entire case-based reasoning process when solving each new problem.

The latter occurs when the reasoner indexes each problem situation,
indicating whether or not the solution was effective.  Over time,
the depth and breadth of problem situations grows and the
accompanying indexing system is quite useful in sorting through
acceptable and unacceptable solutions for similar problems.  In
light of these performance-enhancing characteristics, it should be
no surprise that performances of case-based reasoners improve when
they accumulate new cases and indexes.  Accumulating new cases will
provide additional familiar contexts for solving problems or
providing evaluations.  Accumulating new indices will allow the
"reasoner" to fine-tune their or its recall of relevant cases.

Advantages:

1)  Case-based reasoning allows the reasoner to propose solutions
    to problems quickly, avoiding the time necessary to derive
    those answers from scratch.

2)  Case-based reasoning allows a reasoner to propose solutions in
    knowledge areas that are not completely understood by the
    reasoner.

    Problem situations in some knowledge areas are impossible to
    fully understand because the solution often depends on as yet
    unpredictable facets, such as the price of winter wheat in the
    middle of the next winter.  In addition, some knowledge areas
    are not yet understood but the decision-maker must make a
    decision anyway.

3)  Remembering previous experiences is particularly useful when
    warning the decision-maker about potential problems that have
    occurred in the past, alerting a reasoner to take actions to
    avoid repeating those past mistakes.

    Remember, cases can also be helpful in pointing out incorrect
    solutions.  How do they do this?  Through the evaluative
    process.  Remember that feedback and analysis of what went
    "wrong" with a solution to a problem situation is also recorded
    and stored in memory.  In future situations, when the case-
    based reasoner calls up that case, it reviews the solution as
    well as the feedback and analysis parts of the case.  All of
    these are taken into account when proposing a new solution to
    the present problem situation.

4)  Cases help a reasoner focus its reasoning on important parts
    of a problem by simply pointing out what features of a problem
    are important.

    This advantage rests on the premise that what was important in
    similar, past cases, will be important in current problem
    situations.  Be aware that important parts of cases include
    those that contributed to successes as well as failures.
Disadvantages:

1)  A case-based reasoner might be tempted to use old cases
    blindly, relying on previous experience without criticizing or
    validating it context of the new situation.

2)  A case-based reasoner (or decision-maker) might allow cases to
    bias them when solving a new problem.

3)  A case-based reasoner may not be able to identify and thus
    consider the most appropriate cases when solving the present
    problem with case-based reasoning.

    These all are potential weaknesses of case-based reasoners.
    They should serve as a warning to decision-makers everywhere
    that they should not shirk their duty of using their own minds
    and experiences to criticize and evaluate potential solutions
    of case-based reasoners.  Remember, case-based reasoners are
    tools that produce information to be considered when we are
    trying to make decisions about difficult problems.  They are
    not in themselves, decision-makers.

Return to: Case-Based Reasoning

Next to: Geographic Information Systems in Agroecosystem Management

• razonamiento basado en casos

• técnicas del pensamiento humano

12.jul.2000

Pulsar tecla de vuelta

Vuelta a Portada

Glosario de Carlos von der Becke.