CSE255  TUESDAY FEBRUARY 10, 1998



ADMINISTRATION

Remember the midterm on Thursday.



COMMENTS ON THE SECOND PROJECT REPORTS

MINUS TWO POINTS

(1) Must explain clearly and precisely exactly what the project
objectives are: what is to be predicted, and how will this be 
financially beneficial?  How are objectives prioritized?

(2) Must discuss whether it is plausible that the available data 
contain enough information (in particular, enough features) to
achieve the prediction objectives.

(3) Mis-interpreted error bars on accuracy estimates.



MINUS ONE POINT

(1) Bad organization, e.g. talking about data columns before
describing the data, or not having separate sections and paragraphs
for separate topics.

(2) Not discussing the significance of a decision tree with just one node.

(3) Missing general conclusions about (e.g.) improved data collection
and operator training.

(4) Missing error bars on accuracy estimates.

(5) Incorrect explanation of decision trees.

(6) Not discussing the direction of causation in a relationship.

(7) Missing labels on figures.



MINUS ZERO POINTS

(1) Titles that are not informative about the specific project.

(2) Table of contents with some or all page numbers missing.

(3) Not following the rules explained by Strunk and White.

(4) Using a MineSet feature without explaining it,
e.g. "column importance".

(5) Giving raw numerical values without explaining them.

(6) A bibliography rather than citations and references.

(7) Don't use predictors that are computed in part from what is to be
predicted.  (Cf. the January effect in the stock market.)

(8) Missing learning curves.



SYSTEM VERSUS USER INITIATIVE

A simple forward-chaining system requires all needed immediate
conditions to be known, i.e. asserted, before reasoning begins.

This is reasonable is the information comes from automated sensors.

If the information comes from a human, s/he must essentially
fill out a form that asks all possible relevant questions in advance.

Often humans don't want to do this.  They want
	(1) the system to ask them only for information that is
	    important for the problem at hand, or
	(2) they want to give partial information in any order they want,
	(3) or both.

(1) is called "system initiative", and (2) is called "user initiative".
(3) is called "mixed initiative" and is usually best.

In constructive forward chaining, having separate stages is one easy
way to achieve restricted mixed-initiative behavior.

The user does not need to supply all relevant information
at the start of problem-solving.  Procedural rules can
request information for each stage at the start of that stage.

Another way is to ask the user whether the unknown premise of a rule
is true or false, whenever all the premises of a rule except one are
known to be true.



DECISION TREES ALLOW SYSTEM-INITIATIVE

A tree is a graph with internal nodes and leaves.
Each leaf is labeled with a final condition, i.e. a root cause or prediction.
Each internal node is labeled with a test.
Each edge out of an internal node is labeled with one
of the possible answers to the test.

The simplest tests just look at what the value of one condition is.
Each possible answer is one of the alternative values.

Here is an example decision tree for predicting what sport to do,
depending on the weather.	

                            sky
                      /      |     \
                     /       |      \
                 sunny   overcast   rainy
                 /           |          \
                /            |           \
        humidity          cycling         wind
         /    \                           /   \
      high    normal                  strong   weak
       |         |                       |      |
       |         |                       |      |
      golf    cycling                  home   rowing



A decision tree is used to solve a problem with system-initiative
as follows:
	(1) start at the root
	(2) ask the user what the answer is to the test at this node
	(3) move along the edge labeled with the given answer to a child node
	(4) if the child node is a leaf, return this diagnosis/prediction
	(5) if not, repeat from (2)



WHAT TO NOTICE ABOUT DECISION TREES

The system does not always ask the user the same questions.
The second question is different depending on the answer to the
first one, and so on.

Each subtree corresponds to a set of problems.  When all the problems
in a set have the same answer, then the subtree is just a leaf giving
the answer.

Decision trees allow the representation of generalized knowledge.  
For example, the middle leaf above represents the knowledge that
"regardless of humidity and regardless of wind, cycling is recommended
if the sky is overcast."

Logically, a decision tree is equivalent to a set of rules.

Procedurally, decision trees can be implemented using forward-chaining rules.

A decision tree is a hybrid of declarative and procedural knowledge:
factual knowledge about the domain (e.g. "if it's overcast go cycling")
is mixed with control knowledge (e.g. "find out the sky condition first").