Abstract
This paper presents a novel solution approach for semi-open queueing networks (SOQNs) with backordering, where customers wait in an external queue if resources are unavailable. We introduce a "lost-customers" approximation that adjusts arrival rates to achieve pairwise identical throughputs and idle probabilities compared to the original system. By combining this with Norton’s theorem, we reduce complex inner networks to a single node, providing a computationally efficient closed-form estimation for performance metrics. We demonstrate the utility of this approach by determining the minimal number of robots required for stable operation in robotic mobile fulfilment systems (RMFS).
The Challenge of Semi-Open Queueing Networks
In modern logistics, robotic mobile fulfilment systems (RMFS) have revolutionized warehouse operations. In these systems, robots transport shelves (pods) to human pickers, rather than pickers walking to the shelves. Mathematically, these operations are modeled as Semi-Open Queueing Networks (SOQNs).
As shown in the figure above, an SOQN is "open" because it accepts external customer arrivals and "closed" because it operates with a fixed pool of resources (e.g., robots). A critical feature of these systems is backordering: if no resource is available when a customer arrives, the customer must wait in an external queue until a resource returns to the pool.
Solving these systems exactly is computationally prohibitive. As the number of resources increases, the state space—the number of possible system configurations—explodes. For a system with even a modest number of robots, exact matrix-geometric methods may require handling billions of entries, making real-time optimization or large-scale design nearly impossible for engineers.
A Three-Step Approximation Strategy
To overcome these computational hurdles, we propose a strategy that transforms the complex backordering system into a series of simpler, solvable models.
The approximation process, illustrated in the figure above, follows three primary steps:
- The Lost-Customer Transformation (Section 3): We first consider a modified system where arriving customers are "lost" if the resource pool is empty. We prove a fundamental equivalence: by "adjusting" the arrival rate of this lost-customer system, we can make its internal throughput and node idling probabilities identical to those of the original backordering system.
- Complexity Reduction via Norton’s Theorem (Section 4.1): The "inner network" where robots move and pick items can be highly complex. We apply Norton’s theorem to collapse this multi-node network into a single "composite" node with state-dependent service rates.
- Closed-Form Estimation (Section 4.2): Finally, we return to the backordering model but with the simplified one-node inner network. This allows us to derive closed-form expressions for average waiting times and external queue lengths.
Key Results and Application to RMFS
We applied this methodology to a simulated RMFS environment with 550 pods, using parameters sourced from established literature. Our findings highlight several critical insights for warehouse management:
- Stability Thresholds: The model identified a clear "maximal arrival rate" ($\lambda_{BO,max}$). For the tested configuration, a minimum of 18 robots was required to prevent the external queue from growing to infinity.
- Throughput Equivalence: We mathematically proved that the local throughputs at each node in the adjusted lost-customer model are pairwise identical to the original backordering model, validating the core of our approximation.
- Diminishing Returns: The results quantify the "knee" of the performance curve. While adding the 19th robot significantly reduces waiting times (sometimes by hours), adding the 40th robot provides negligible benefits, allowing managers to optimize capital investment.
- Computational Efficiency: Unlike exact methods that struggle with 10 robots, our approximation provides accurate results for systems with hundreds of robots in a matter of seconds.
Limitations and Future Outlook
While highly effective, the approximation is less precise in "heavy traffic" regimes—when the system operates very close to its stability limit (e.g., having exactly 18 or 19 robots in the test case). In these cases, the model tends to slightly underestimate waiting times due to high variance. Additionally, the current model assumes robots do not interfere with each other's movement. Future research will focus on incorporating robot congestion and exploring non-exponential service time distributions to further refine the model's real-world applicability.
Summary updated