8803B 10-23-02

 

Version Spaces

Instance based learning is where the algorithm creates a most-specific model and a most general model and then uses continued instances to merge multiple possible models or versions  into one model.

 

 

 

An Example

Our room mate Bob likes to go out to eat on a regular basis. He often gets sick from this food. We want to learn what causes Bob to be sick.

Bob goes to 3 restaurants, has 1 of 3 meals (breakfast, lunch, and dinner), eats on 7 days a week (Mon, Tues, Weds, Thurs, Fri, Sat, Sun) and has cheap and expensive meals. This means to fully explore every possible meal combination we would have to record Bob's eating habits for a 126 meals. We'd like to learn the rule without having to explore the full space. Version Space for instance based learning lets us do this.

Data

Sample # Name of Restaurant Meal Type Day of Week Cost Bob's Reaction
1 Sam's Breakfast Friday Cheap Sick
2 Jason's Lunch Friday Expensive Not Sick
3 Sam's Lunch Saturday Cheap Sick
4 Sarah's Breakfast Sunday Cheap Not Sick
5 Sam's Breakfast Sunday Expensive Not Sick

Now we want to derive the rule for Bob getting sick from this limited data.

 

 

Goal of the algorithm is to find a rule that divides negative examples from positive examples

 

Algorithm Converges

1. Specializes as needed

2. Specialize to include positive examples and exclude negatives

3. Specialize in a way that the new model is a generalization of the current most specific model

 

Notes

- The initial example needs to be positive or creating the initial  most general and most specific is problematic

- This algorithm assumes there is no noise in the data.

 

Explanation Based Learning

Algorithm to learn from just one example. The algorithm is based on the agents ability to create an explanation for the reason the example works.

 

The algorithm works by giving the agent a functional definition, some domain knowledge, and an example that demonstrates the functional definition.

 

Example

We want to teach a robot what makes a coffee cup work as a coffee cup, so if no coffee cup is available it can find a suitable replacement such as a beer mug.

Functional Description of a coffee cup:

    1. Holds Liquid -         must be concave - Is_Concave

    2. Can Drink From -     Open Cavity     -Has_Open_Cavity

    3. Must be liftable  -                              - Is_Liftable   

Example of a cup:

    Ashok's Cup:

                        Has_Handle

                        Made_of_Porcelain

                        Color_White

                        Is_Light

                        Is_Round

                        Is_Small

                        Has_Depth

                        Has_No_Top

Explanation Based Generilzation

 

From this will be able to look for all objects that satisfy the definition is of Is_Cup.

This does have the flaw that the system must be given Domain Knowledge. In current research this domain knowledge is built from examples. However, this means no learning from just one example.

The example is used instead of just the functional definition, to focus the search of what properties fulfill the functional definition.