Can “ Very Noisy ” Information Go a Long Way ? An Exploratory Analysis of Personalized Scheduling in Service Systems

Problem definition. In this work, we focus on the implementation of personalized scheduling policies which exploit noisy individual customer information. In particular, we consider the shortest job first policy with noisy service time predictions. Academic/Practical Relevance. The extant queueing-theoretic literature on personalized scheduling typically assumes perfect information about customers, while in practice such information tends to be “noisy”. Therefore, it remains unclear to what extent the promise demonstrated in theoretical work can translate into practice. In this paper, we investigate the tradeoff between information availability and operational performance by studying the performance of noisy service-time-based scheduling. Methodology. We use a combination of three methodologies. Through our empirical study, we investigate how personalized scheduling using service-time predictions from a real call-center performs in practice. Motivated by our empirical results, we use both simulation modeling and queueing theory to deepen our understanding into that operational performance. Results. We present empirical and theoretical evidence that “very noisy” information can go a long way in personalized service-time-based scheduling. First, we quantify the improvement in predicting service times when exploiting the individual histories of callers, which enables us to implement customer-specific scheduling. Second, we perform detailed simulation studies which quantify the performance of noisy service-timebased scheduling in a variety of queueing contexts. Third, we derive sufficient conditions under which the expected waiting time with noisy service-time-based-scheduling is smaller than for any finite priority-based scheduling policy. Managerial Implications. Our analysis shows that call center managers can benefit substantially from implementing personalized scheduling disciplines, even when personalized customer information is considerably noisy. We derive insights on the performance of noisy service-time-based scheduling in general service systems. For example, we show that such scheduling performs better in large and congested systems.