I.5.2.2: ACE

The ACE methods, developed by Breiman and Friedman \cite{brfr}, minimize over all smooth functions.

A problem with ACE is that smoothers, at least most smoothers, often do not minimize a loss function, or the same loss function as is used for the remaining parameters.

In any case, ACE is less general than ALS, because not all least squares problems can be interpreted as computing conditional expectations.

Another obviously related area in statistics is the Generalized Additive Models discussed extensively by Hastie and Tibshirani \cite{hastib}.