Datadriven Inference and Optimization for Imputing Non-observable Human Preferences in Staff Scheduling

Abstract

In a staff scheduling problem, while a schedule planner (called an agent in this research) has the knowledge of the criteria and work rules, a typical challenge is that she produces schedules based on her implicit preferences and is uncertain about the relative importance of the multiple objectives. Moreover, due to the limited human capability of achieving optimality, the observed schedules produced manually may not reflect the agent’s actual preferences. Another challenge is that the agent occasionally allows violation of certain work rules. We formalize the above challenges by three main properties of the problem: model uncertainty, bounded rationality and compromised infeasibility. To determine the agent’s implicit preferences in staff scheduling, we propose to develop a data-driven inverse optimization framework. Instead of the common practice which asks the agent to guesstimate the objective weights and degree of allowed infeasibility, the agent is given sample staff schedules to rank according to her realized preferences. The ranked staff schedules will be fed into the inverse optimization model for inferring the problem parameters.