Multistart Strategy Using Delta Test for Variable Selection

摘要

Proper selection of variables is necessary when dealing with large number of input dimensions in regression problems. In the paper, we investigate the behaviour of landscape that is formed when using Delta test as the optimization criterion. We show that simple and greedy Forward-backward selection procedure with multiple restarts gives optimal results for data sets with large number of samples. An improvement to multistart Forward-backward selection is presented that uses information from previous iterations in the form of long-term memory.