Learning

   Learning may be defined as a change in an internal
representation that leads to a (positive) change in behavior.


   Induction may be defined as that sort of learning based on
observations.

   Induction may be distinguished, for example, from learning by
reading or hearing.
   
   
   General reinforcement learning involves the modification of
behavior based on rewards or punishments for particular attempts.
   
What may be learned?
   
   
Concepts (as relations or predicates)

   Learning relations is typically a form of supervised learning.

   Training examples are typically presented to the learner as
positive examples of concepts or negative examples of concepts.

   Experiment generation: The examples may be generated and
labeled by the teacher.

Or they may be generated by the learner and presented to the
teacher for classification. Classification

   Classification refers to learning patterns in the input.

   Classification involves partitioning a space of items.

   It may be unsupervised.

   It differs from the learning of relations or functions because
the number and names of the classes to be learned need not be
provided to the learner.

Learning a function

   Sometimes it is enough to learn a good approximation to a
function.  Issues in representation


   Learning may be considered as a search for a hypothesis in a
hypothesis space.


   Sometimes an appropriate representation is part of what is to
be learned.

   Often a specific or generic representation is provided to the
learner.

   For example, certain features of a domain can be identified to
the learner as being worth considering.

General issues in learning

Training data (set) vs. test data (set)

   training data is data used to construct the new internal
representation.

   test data is data used to test whether the representation is an
improvement.

   the test data should be disjoint from the training data.

Incremental learning
   is all of the test data available at once?

Experiment generation
   if the learner can generate training data, how should this be
done?

Noise
   what happens if there is noise in the training data?

Bias
   which portions of a large hypothesis space should be searched
first?
   do some hypotheses remain unconsidered?

Expressiveness vs. efficiency
   allowing more hypotheses to be formulated can mean more time to
search through them.

Credit assignment
   if a sequence of actions leads to a reward, should some actions
get more credit than others for the result?

Overfitting
   tailoring a solution too closely to the training data can
prevent generalization.


Simplicity
  should simpler solutions be rewarded?
   e.g., when learning a function from R to R, for any n training
pairs there’s an (n-1)-degree polynomial that fits the data
exactly.


Is prior knowledge available?
   about the domain?
   about the world?
   linguistic knowledge?


Understandability
   Will the resulting representation be understandable to humans?

   Issues of legal liability can arise.

Nearest neighbor strategies

Idea:
    give test input the same value as the nearest training input

    or a combination of values of the kth nearest neighbors

Issues:
    can k values be combined?

    how can distance be defined?

    how can distance be computed efficiently, for a large set of
training inputs?

    are nearest neighbors near enough
      (in high-dimensional spaces)?