LEARNING

A learning agent
	performance element - takes actions based on 
		knowledge and percepts
	learning element - modifies knowledge based on 
		experience
	critic - evaluates success of knowledge
	problem generator - explores to learn more

Any type of knowledge can be learned:
	direct mappings from percepts to actions
	percept to properties of world mappings
	knowledge about world evolution
	possible results of actions
	utility information about desirability of world states
	action-value information about desirability of action in 
		a particular state
	goal descriptions to identify success of agent



			SOME METHODS OF MACHINE LEARNING

Supervised learning
	told problem and solution
	learning tries to identify mapping between the two

Reinforcement learning
	told problem and whether generated solution is right or 
		wrong
	learning is based on examples of success and failure

Unsupervised learning
	not told anything about outcome other than what it 
		collects through existing percepts
	can learn effect of actions on percepts
	must already have a utility function to learn

Learning system may or may not have prior knowledge.

Technically, we can say learning is generating a function 
mapping inputs to outputs

	all of the types of knowledge in performance element 
		can be viewed as output



			INDUCTIVE LEARNING

Inductive inference -- given a set of examples of function f(x),
return a function h(x) that approximates f(x).

Learning procedures may include a bias towards particular results
(simple solution, exact match, one that does not require specific
types of data, etc.)

Examples:
	curve/line fitting -- given points on plane, find function 
		to match them
	number patterns -- what is next element in sequence

How do we learn?

Consider sequence:
	3, 5, 7, ?             (what do you think is next element?)
	3, 5, 7, 11, ?         (now what is the next element?)
	3, 5, 7, 11, 13, ?     (and now?)
	3, 5, 7, 11, 15, ?     (any guess now?)
	3, 5, 7, 11, 15, 23, 31, 47, 63, 95, ...



			LEARNING DECISION TREES

Decision trees describe how percepts are used in making a decision.

Given a set of examples of percepts and outputs, determine if-then
rules for mapping percepts to correct (or mostly correct) outputs.

Rule learning:
	identify percepts consistently mapping to particular 
		result and create a rule
	use combinations of percepts in form of multiple levels 
		of decision tree.

Probability & Statistics
	Principal component analysis
	Factor analysis
	Minimize information gaps ...

Learning process:
	Collect large set of examples.
	Divide set in two, training and evaluating sets.
	Use the learning method on training group.
	Measure learning by success rate on evaluation set.
	Repeat above steps to compare learning algorithms.