Summary of Results

      All the tests run produced the same accuracy and the exact same set of rules. From the long traces it can be seen that the same steps were taken to arrive at these rules. The rules produced can be seen in the results for the first test.

To recap, three approaches were tried.

1. all non-continuous attributes were left as possibly ordered types
   • This lets Foil decide which attributes are ordered and which are not
2. The Education attribute was specified as an ordered type
3. Education specified as ordered type and
   Race and Sex were specified as unordered type

The accuracy over all test data was 81.3341%.

The best results over the learning methods so far

Foil produced the third best accuracy, and arguably, the easiest to comprehend classifier to date.

Things to try to improve accuracy include

• Trying more combinations or ordered and unordered attributes.
• Specifying theory constants
• Everything else that worked for other people in the class