Cars and Congress
- I also took this topology and with the slightest tweaks
made it run for two of the Cal Irvine tasks.
- The first was the congressional voting task. Given
votes on 15 bills, categorise a congressman as Democratic or
Republican.
- I put all 16 binary features in.
- It's a natural category that we'd used before getting 89%.
- Here we got around 83%.
- Similarly, I did a car task. There were 6 input features and
4 categories for the output feature.
- This was a task for C4.5 decision tree learning.
- We got around 85%.
- What was interesting was that I tried to learn and test on
some subsets of the input.
- I could perfectly learn 10 items, but at 20 I went down to the
85% score.
- It didn't matter how large I made the gas.
- I think this may provide a nice entry problem into richer
learning.
