Showing changes from revision #1 to #2:
Added | Removed | Changed

Contents

Idea

Co-ordinate descent is a technique for optimizing a multivariate function by simply repeatedly optimizing along each co-ordinate axis in turn. Although this is a less “intelligent” choice than utilising the gradient direction, it sometimes turns out allow closed form optimisations in that direction, resulting in being overall faster. A case where co-ordinate descent is often used is in sparse regression, wherepart of the function being optimized is oriented along the co-ordinate axes (eg, $L_1$ regression has $\sum_i |x_i|$, the sum of the absolute values of each of the variables.)

Revised on August 18, 2014 23:44:50
by davidtweed
(0.0.0.0)