Agent-based simulation · passenger boarding
Boarding is a queueing problem with one aisle, no overtaking, and two ways to stop the line: someone stowing a bag, and someone climbing over the people already sitting between them and the aisle. All fourteen methods run here at once, on the identical passenger manifest.
Boarding is the longest and most variable task in a narrow-body turnaround, and the only one whose duration is set by a queueing rule the airline chooses for free.
Two costs, and only one is linear. The APU runs throughout boarding, so minutes saved are minutes of an auxiliary turbine not burning 150–200 kg/h — that converts directly, every flight. Delay cost does not: boarding sits inside the scheduled turn, so it only becomes money when it stops a turn overrunning its slot. Multiplying minutes by €/min overstates the case badly, and this tool reports the two separately.
The industry made it worse on purpose. Checked-bag fees pushed luggage into the cabin — $7.3 bn of US ancillary revenue in 2024 — and created a permanent shortage of bin space. That produced bin anxiety, and bin anxiety produced gate lice, a term the Cambridge Dictionary added in 2025. Cabin crew in most US contracts may not lift passengers' bags.
Which is why the optimum is not used. Calling a scattered group of high-value passengers first destroys any careful ordering, and airlines accept that because the ancillary revenue outweighs the minutes. Push the priority boarding slider up and watch the good methods collapse — that is the trade being made, made visible.
It is all moving right now. United reinstated window-middle-aisle in 2023 and measured up to two minutes a flight. Southwest ends 53 years of open seating on 27 January 2026. Boeing's Space Bins add 50% more bin volume. Biometric gates clear a passenger in 6–10 seconds — but above about 25 pax/min the aisle, not the gate, is the constraint.
Each one sets every parameter below at once. Change anything afterwards and the situation is no longer selected — the numbers are yours.
Type a value. Every one of these changes the ranking, and the two that change it most are how long a bag takes to stow and how accurately people queue.
The stage above is a single game, and a single game cannot rank these methods. Replaying across manifests is what separates the merit of a method from the luck of one planeload: the same method varies by about 27% between a kind manifest and a cruel one, from roughly 10 to 13 minutes. That spread is wider than the gap between most neighbouring methods, so on one run the ranking of any close pair is very nearly a coin toss — measured over sixty replications, adjacent pairs come out in the wrong order 45 to 55% of the time. Press Replay all 14 and the means, the p90 and the spread below are computed over that many fresh planeloads, which is the only version of this table worth quoting.
Each card opens with the seat map of the aircraft currently selected, laid out the way you would see it looking down on the cabin: rows run left to right, nose to tail, seat letters down the side, and the aisle is the gap through the middle. Every seat is shaded by when its passenger was called — pale for the first through the door, dark for the last.
So the shading is the method itself. Window-middle-aisle shows three pale-to-dark bands running the length of the cabin; zone methods show vertical blocks; Steffen shows a checkerboard, because consecutive passengers are deliberately two rows apart. Where the pattern looks smudged rather than crisp, that is queue compliance: people not standing in the order they were called. Under each strip is what a gate agent would actually have to say to produce it, which is usually the fastest way to see why a method is or is not realistic.
Click any card to open that cabin in 3D.
Discrete-time agent-based model: 50 ms steps on screen, 100 ms in the replay. Common random numbers — all fourteen methods receive the identical manifest, so differences are attributable to the ordering and not the draw. Fuel and CO₂ use a 175 kg/h narrow-body APU and 3.15 kg CO₂ per kg; cost uses $100.76 per block minute. Run with back → front the model boards at ~10.7 pax/min against ~9 reported in industry data: right magnitude, roughly 19% optimistic, so read differences between methods rather than absolute times. Full method, sources and limitations in the README.