Mixed Traffic Simulator

The problem

Indonesian road planners size roads with a 1997 capacity manual, MKJI. It converts every vehicle into car equivalents with a fixed factor, and a motorcycle counts as 0.25 or 0.5 of a car depending on road type. Two decades of Indonesian field studies have measured that "fixed" factor at 0.198, 0.32, 0.35, 0.4 and 0.84. One regression study in Denpasar got −0.11 at one-hour aggregation, which means nothing physically.

This decides budgets. At one roundabout, the same peak hour rates level of service C by one method and F by another. Acceptable or failed depends on which method someone chose, and that decides whether an intervention gets funded. A field study can't settle it, because nobody can replay the same hour of traffic with the motorcycles taken out.

What I built

A controlled experiment no street allows

The simulator runs a traffic stream with a known share of motorcycles. Then it replays the identical scenario with every motorcycle swapped for a car, and compares capacity. That comparison is the true answer. Next, it applies each published estimation method to the same stream and shows the methods disagreeing with each other and with the answer already in hand.

The estimators only see what a roadside observer sees

The argument falls apart if the estimation methods can cheat. I walled them off from the simulation code, and the build fails if they try to reach into it. They receive only what a field engineer could record: when a vehicle crossed a line, its type judged by eye, its spot speed, how long it occupied the detector and where it was across the road. A method that knows true speeds and identities isn't the method engineers used. The only exception is the car-swap experiment, because ground truth is exactly what an observer can't get.

Any result can be shared and replayed exactly

The physics runs in fixed 0.05-second steps, separate from drawing, so results never depend on the viewer's hardware. All randomness comes from one seeded source. The whole state, including scenario, parameters, seed and aggregation window, fits in the link. A surprising result becomes a URL that reproduces it, and a batch sweep can be regenerated from its seed.

The modelling choice is named on screen

There is no agreed model for traffic that ignores lane lines, as Indonesian motorcycles do. So I shipped three rules for sideways movement: strict lanes, gap-seeking and a social-force model. The active one is always named on screen. The social-force rule has no citation in the traffic literature, and the interface says so as a caveat.

Impossible results stay on the chart

When an estimator returns a negative motorcycle factor, I plot it below the axis with a warning. I don't hide it or clamp it to zero. That number is the most persuasive result the app has.

Result

The simulator is live and public. The phantom-jam scenario puts 22 vehicles on a 230-metre ring with no bottleneck. A jam still forms from reaction delays alone and travels backward, measured in the app at −8.6 km/h, reproducing a published experiment. The equivalence bench plots five methods against the motorcycle proportion in traffic. Ground truth stays nearly flat, while the regression and speed methods swing from +16 to −30 and reproduce the published −0.11 at wide aggregation. An inspector shows the live driving arithmetic of any vehicle you select. I built the first version alone over two days in September 2026, then kept working through the rest of the month on a redesign, an accessibility pass, and mobile and keyboard support.

The app states plainly that it is not a calibrated model of any specific road and not a replacement for professional traffic software. Its job is to show an assumption failing.