Group boarding for airplanes: benchmarking static policies and optimizing dynamic assignment with deep reinforcement learning

Beyond Fixed Rules: Solving the Airplane Boarding Bottleneck

Instead of using fixed rules for boarding, airlines can use AI to assign passengers to groups one by one as they check in. This method looks at where people are sitting and how much luggage they have to prevent aisle blockages, making boarding faster and more comfortable.

Efficient airplane boarding is a cornerstone of airline logistics. Reducing turnaround time—the period an aircraft spends on the ground between flights—directly impacts airplane utilization. A Boeing study cited in the paper found that cutting turnaround time by 10 minutes can increase utilization by 8 percent. Currently, most airlines manage this process using static, seat-based rules. These rules assign passengers to a handful of sequential boarding groups based on fixed criteria, such as row number or cabin zone, decided long before the plane reaches the gate.

The fundamental flaw in this status quo is its rigidity. As passengers check in, they arrive in a random sequence. Their specific seat choices and luggage requirements create a shifting landscape of potential congestion. Static rules ignore the "state" of the cabin—the reality of which seats are already occupied and which aisles are currently blocked. For example, a rule might send a group to the rear of the plane. However, if several passengers in that group are already seated in the middle of the aisle, they create "seat interference," forcing others to stand up and block the path .

Similarly, "aisle interference" occurs when a passenger stops to stow luggage, creating a temporary dam in the flow of traffic .

Existing optimization methods often solve this as an offline problem. This means they calculate the perfect sequence only after everyone has checked in, which is too late to print group numbers on boarding passes.

A Markov Decision Process for the Cabin

The authors propose treating boarding group assignment as a dynamic, sequential decision problem. They formulate this as a Markov Decision Process (MDP)—a mathematical framework for modeling decision-making where outcomes are partly random and partly under the control of a decision-maker. In this model, the "state" is a snapshot of the cabin at the moment a passenger checks in.

To make sense of this complex environment, the researchers employ a deep reinforcement learning (RL) approach. The core of their architecture is a Convolutional Neural Network (CNN) designed to "read" the airplane cabin like an image. The cabin is represented as a grid. Each cell contains data about a seat: whether it is occupied, which boarding group is assigned to it, its location (window, middle, or aisle), and the amount of luggage the occupant has declared. This spatial encoding allows the model to recognize patterns of congestion, much like a computer vision system identifying objects in a photo.

The training process utilizes Proximal Policy Optimization (PPO), a specific type of RL algorithm that limits how much a policy can change in a single update to ensure stable learning. This uses an actor-critic framework. In this setup, the "actor" is the part of the model that chooses the boarding group, while the "critic" is the part that evaluates those choices by estimating their value. Crucially, the model is trained on a terminal reward that balances two often-conflicting goals: total boarding time (operational efficiency) and average individual boarding time (passenger experience). By sweeping the weight of these two metrics, the authors can generate a "Pareto frontier"—a set of optimal solutions where one metric cannot be improved without degrading the other. This allows an airline to choose a policy that fits its specific business priorities.

Outperforming the Back-to-Front Standard

The study benchmarks the RL policy against several established static strategies. These include back-to-front (BF), modified-Steffen, and alternating-block policies. While the authors find that the back-to-front policy is the strongest of the static competitors, the dynamic RL policy demonstrates superior performance across all tested layouts.

The researchers report that on a representative narrow-body airplane, the RL policy outperforms the optimal four-group back-to-front benchmark by up to 9.8% in total boarding time. It also improves average individual boarding time by up to 22.8% . This gap is driven by two main factors. First, the RL policy utilizes seat-column information (knowing if a passenger is in a window versus an aisle seat) while still ensuring that travel companions remain in the same group. Second, the policy is trained over thousands of stochastic episodes. This allows it to implicitly learn how to navigate the randomness of passenger arrivals and luggage loads.

The impact of luggage information is particularly notable. The authors find that while seat-column data primarily drives reductions in total boarding time, declaring luggage quantities is the key lever for improving the passenger experience. For a four-group policy, adding luggage observation to the state can reduce the average individual boarding time by 18.1% (from 618.8 to 506.6 seconds) . Furthermore, the RL policy proves remarkably robust. The authors demonstrate that a single trained policy maintains its lead even when faced with "out-of-distribution" conditions, such as varying passenger load factors or different distributions of luggage and companion sizes .

Limitations in the Virtual Sky

Despite the impressive metrics, several gaps remain between this simulation and a real-world airport terminal. The authors acknowledge that their simulation simplifies the boarding process by omitting several chaotic variables. For instance, the model does not account for late-arriving passengers who disrupt the planned sequence. It also does not model the specific delays that occur during boarding-pass checks at the gate.

Additionally, the simulation treats luggage stowage as a somewhat generic temporal delay. In reality, the time it takes to stow a bag is highly dependent on the available space in overhead bins and the current density of the aisle. The authors also note that while their model is calibrated using empirical data, the results have not yet been validated against real-world flight data. For a practitioner, this means the reported 22.8% improvement in passenger experience should be viewed as a theoretical upper bound rather than a guaranteed field result. Finally, the reliance on "declared" luggage quantities introduces a human-factor variable. The policy's effectiveness is tied to how accurately passengers report their belongings during check-in.

The Verdict: Deployable via the App

The evidence suggests that the era of "one-size-fits-all" boarding rules is nearing its end. If an airline wants to stick to simple, manual rules, the authors recommend using back-to-front boarding with highly unequal group sizes to maximize efficiency. However, for carriers looking to maximize both turnaround speed and passenger satisfaction, the path forward is digital.

The RL policy is ready for prototyping, provided it is integrated into existing digital touchpoints. Because the assignment happens at check-in, the logic can be embedded directly into mobile check-in apps or kiosk software. The results are clear: moving from a static rule to a dynamic, state-aware assignment provides a measurable edge in both operational throughput and customer experience. The technology is not just a marginal gain; it is a way to turn a logistical bottleneck into a managed, optimized flow.

Figures from the paper