Oct 16, 2006
-------------

- New topic: machine learning
        - using past experience to improve
          behavior

- Examples
	- bank clients, characteristics, and credit risks 
        - students, question papers, and grading

- Different ways to slice learning
        - based on nature of feedback

- Classification based on type of feedback
        - instructive (supervised learning)
        - evaluative (reinforcement learning)
                - need to do credit assignment
        - no feedback (unsupervised learning)

- Example of supervised learning
        - passing 4804 midterm, given
          solution sketches for "similar problems"
	- finding what makes a person a good credit risk,
	  given past history of the bank loans

- Example of unsupervised learning
        - just observing and noticing patterns
        - no goal in mind, so cannot effectively
          use what is learnt for any purpose

- Example of reinforcement learning
	- no "right" or "wrong" answers given
	- only "smiles" or "slaps"
		- must balance exploration/exploitation

- More ideas for learning
        - curve fitting example
                - how many possible answers are there?

- Lesson 1: Learning requires bias
        - e.g., polynomials

- Lesson 2: there might still be many answers
        - choose the one that is "simplest"
          (called Occam's razor)

- Lesson 3: bias can be too limiting

- Lesson 4: bias-variance dilemma holds

- Learning methodology
        - use training set for learning
        - evaluate learning on test set

- Why is bias important?
        - motivation from #examples needed for
          learning

- Example: learning boolean functions of n boolean variables
        - possibilities: |H| = 2^(2^n)
        - for two variables, leads to 16

- With every additional data point (example)
        - can rule out half of this

- Profile of narrowing down |H| with every example
        - but this is simple memorization

- Need generalization for good learning
        - what does the profile look like in this case?

- Lets adopt bias as conjunctions (n=2 case)
        - 10 possible hypotheses
                - 1
                - 0
                - a
                - not a
                - b
                - not b
                - not a and b
                - not a and not b
                - a and b
                - a and not b

- Consider the example data given by f=b
        - will need all four examples
          to learn!

- Trick
        - reorder last two examples
        - what does this show?

- PlayTennis handout
	- let us learn a decision tree from this data