Learning Dynamic Time Preferences in Multi-Agent Meeting Scheduling

In many organizations, people are faced with the task of scheduling meetings subject to conflicting time constraints and preferences. We are working towards multi-agent scheduling systems in which each person has an agent that negotiates with other agents to schedule meetings. Such agents need to model the scheduling preferences of their users in order to make effective scheduling decisions. We consider that a user’s preferences over meeting times are of two kinds: static time-of-day preferences, e.g., morning versus afternoon times; and dynamic preferences which change as meetings are added to a calendar, e.g., preferences to schedule meetings back-to-back (i.e. in succession). The dynamic nature of preferences has been understudied in previous work. In this paper, we present an algorithm that effectively learns static time-of-day preferences, as well as two important classes of dynamic preferences: back-to-back preferences and spread-out preferences (i.e. preferences for having gaps between meetings).