September 30, 2002

Constraint Programming 2002, Cornell

22

Learning

nClassification:
misleading error measure

nStatistical
regression: learn a continuous function of features that predicts **log of running time
**

nSupervised
learning: data broken into 80% training
set, 20% test set

nStarted
with simplest technique: linear regression

nfind a hyperplane that minimizes root mean squared error (RMSE) on training data

nLinear
regression is useful:

nas
a (surprisingly good) baseline

nyields a very interpretable model with understandable variables