TL;DR
This paper introduces a novel unified framework for analyzing queueing systems with customer abandonment, regardless of the specific queueing model or how customer patience is scaled. This approach simplifies the diffusion analysis of such systems, making it as straightforward as analyzing queues without abandonment.
What Problem Does This Paper Address?
Queueing systems are ubiquitous, from call centers to manufacturing lines. A common real-world phenomenon is customer abandonment, where customers leave if they wait too long. Analyzing these systems mathematically, especially with diffusion approximations (which provide continuous, tractable models for complex systems), has been challenging. Existing methods often require different techniques for different queueing models (e.g., single-server vs. many-server) and different ways of modeling customer patience (e.g., fixed patience time vs. scaled patience time). This paper aims to bridge this gap by providing a single, unified approach that handles these variations.
Methods
The authors propose a two-step approach:
- Asymptotic Relationship: They first establish an asymptotic relationship between the customer abandonment process and the queue length process. This is achieved by using the patience-time distribution to connect abandonment to the virtual waiting time, which can then be approximated by the queue length. This step requires the queue length processes to be "tight" (meaning they don't fluctuate too wildly).
- Continuous Mapping: To overcome the difficulty of verifying the "tightness" condition, they introduce a "continuous mapping technique." This involves constructing a mapping that relates the system's state (like queue length) to its underlying random components (arrival and service times). This mapping, if it has certain "nice" properties, helps verify the tightness conditions and allows for the application of the continuous mapping theorem to obtain diffusion approximations.
This unified framework is designed to work across different heavy-traffic regimes (conventional, Halfin-Whitt, and non-degenerate slowdown) and different patience-time scaling methods (no scaling, hazard-rate scaling, and mixtures).
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Unification is Key: The most significant insight is the development of a single, unified framework that simplifies the analysis of queues with abandonment. This eliminates the need for specialized techniques for different scenarios.
- Abandonment Becomes "Easy": The approach makes the diffusion analysis of queues with abandonment essentially the same as the analysis of queues without abandonment, a major simplification.
- Tightness as a Bridge: The paper highlights the importance of "tightness" of stochastic processes in proving diffusion approximations and introduces a method to establish this tightness even when dealing with complex abandonment dynamics.
- General Patience Scaling: The framework is flexible enough to handle various ways customer patience can be modeled, going beyond just fixed or hazard-rate scaling.
Key Results
The paper demonstrates the application of its unified approach to three specific queueing systems:
- A single-server queue with Markov-modulated service speeds in the traditional heavy-traffic regime.
- Many-server queues in the Halfin-Whitt regime with general service times.
- Many-server queues in the non-degenerate slowdown regime with general patience-time distributions.
For these systems, the paper establishes diffusion approximations for the queue length processes, showing that the unified approach yields accurate results and validates previous approximations.
Limitations and Future Work
The authors acknowledge that verifying the "tightness" conditions (specifically the modulus of continuity) can still be challenging in practice, although their framework simplifies the process. They suggest that their approach can be extended to analyze customer abandonment behaviors in a broader range of queueing models and scaling scenarios. The paper also provides a foundation for further theoretical work on the properties of these diffusion approximations.
Summary updated