PointFix: Learning from fixing individual condition violations

TL;DR: An algorithm that can be added to a scheduling system for fixing selected violations and a method for exploiting users' interactions with this algorithm for automatically tuning the importances of conditions in an existing nurse rostering system, leading to better rosters.

Abstract: Building a duty roster for staff at an organization can be a very difficult task due to complex and interdependent conditions that a scheduler has to follow. When an automated scheduling system is used to make this process smoother, the users can face other problems. Most automated scheduling systems are designed for producing rosters almost entirely from scratch. This does not address the needs of users looking to correct a few selected violations individually while rest of the roster changes only minimally. Furthermore, the complex nature of the conditions makes the customization of nurse scheduling systems difficult, often requiring specialists in order to capture fine details before the scheduling system can be personalized for use. This can be very time consuming and costly, while not guaranteeing completely accurate representations of conditions. Modifying the representation later manually may also not be an easy option for the end users because of the complexity of the scheduling systems. In this paper we introduce an algorithm that can be added to a scheduling system for fixing selected violations. We then introduce a method for exploiting users' interactions with this algorithm for automatically tuning the importances of conditions in an existing nurse rostering system, leading to better rosters.