SmartCalendar: Improving Scheduling Through Overcoming Temporal Inconsistencies
Jiahua Tang, Du Zhang · 2018
Conventional machine learning systems are trained-once-used-everywhere. The Achilles heel of such systems is that in its highly specialized scope of tasks, the system's knowledge learnt with one single and static training dataset can hardly keep up with the demand from a changing and dynamic environment. Using a calendar system as an example, it may be equipped with initial knowledge of a user's preferences. But as a user encounters different types of conflicting events when scheduling her daily activities, the calendar system must continuously learn to refine its scheduling knowledge to adapt to the user's dynamic and changing circumstances and preferences. It turns out that none of the existing calendar systems is capable of detecting any type of conflicts, say a temporal inconsistency between two events, not to mention how to handle those conflicts. Inspired by inconsistency-induced learning, in this paper, we propose a calendar system that can continuously improve its scheduling performance through overcoming temporal inconsistencies. With Interval Temporal Logic, we can identity and detect different types of temporal inconsistencies. We propose distance-based heuristics for solving temporal inconsistencies between events. Knowledge will continuously be refined so as to incrementally improve the performance of event scheduling. We also compare our approach with related work.