Learning to Predict User Replies in Interactive Job Scheduling

Abstract

We consider a learning task that arises within an interactive job scheduling setting, in which a scheduler plans the execution of jobs that require the presence of human users. Availabilities of these users shall be considered but are only known partially, and thus the scheduler presents queries to the users to receive more information about the users’ availabilities. The replies of the users help creating a better schedule. Having a precise understanding of typical user behavior is crucial for the efficacy of the scheduler. As the scheduling problem must be solved repeatedly over time, e.g., weekly, the knowledge gained from previous instances can be used to learn a user model. In this work we employ Bayesian Learning and investigate three different models to predict user replies to queries and train them by means of the framework of Probabilistic Programming. Two models learn time-independent, respectively time-dependent, probabilities for a user to either become available, stay available, become unavailable, or stay unavailable from one timestep to the next. The third model learns time intervals in which the user is available with normally distributed endpoints. These models are experimentally evaluated and compared on two datasets, one based on artificially generated user availabilities, the other on real-world data. Results show that especially the time-dependent model performs well and near-optimal for the artificial dataset while the time interval model works best on the other dataset.