Intelligent sectioning for searching of unimodal data
Milan Milatovic, Adedeji B. Badiru · 2002
This paper proposes an improvement to the original Cantor trisectioning search technique that was specialized only for search domains where the distribution was approximately bell-shaped, but performed poorly when searching through skewed data. In this study, a new formula has been derived, which, in terms of only five specific percentile values and regardless of the database size, estimates the position of the mode in unimodal curves with an accuracy of more than 95%. This enhanced the search by being able to start approximately at the mode instead at the middle of the search space as previously proposed. In addition, a relation between the choice of 1/n sectioning and the distribution peakedness has been proposed, such that the sectioning interval, n, equals 2 when searching uniform distributions, and approaches infinity when searching very "spiky" distributions.