Methodology
1. Select algorithm
2. Select set of input distributions
3. Factor out known sources of hardness
4. Choose features
5. Generate instances
6. Compute running time, features
7. Learn a model of running time