Locally Weighted Regression

• construct an approximation from the training examples in the neighborhood of , then calculate , then can be deleted.
• assume that our function is a linear function,
• Possible error methods for gradient descent

• 

• 

• 

• criterion 2 is nice because allows every training example to have an impact, but computation grows linearly with the # of training instances
• criterion 3 is a nice compromise
• the gradient descent training rule becomes

  

• but many more efficient ways are available to directly solve for the coefficients
• Why won't these work for ANNs?