At the heart of a traffic assignment model is its assignment algorithm, the process by which it establishes an equilibrium between alternative routes. This needs to balance congestion around the network and at different flow levels replicate the route choices made by drivers.

For most situations the incremental assignment method is a fair approximation of reality and will give reasonable results. There will be occasions however when the model will be misleading because of some important differences between the algorithm and real behaviour. Modellers need to clearly understand the reasons why distorted route choices can occur.

Figure 1 illustrates how the algorithm works for a typical traffic system in which there is a choice between a fast and slow route. At low traffic levels the faster route offers quicker travel times, however as the network becomes busier an intersection congests to the point where the slow route becomes preferable.

The model loads traffic in perhaps five increments. The first two or three increments see an empty or near empty network, and follow the faster left-hand route. For the final steps however an intersection becomes badly congested and traffic switches to the alternative route.

The result is certainly plausible and will often be a reasonable estimate of the route choice split. Some points should be noted however:

each increment sees the network in a different condition, and this is the basis for making different route choices

although none of the traffic has selected a route on the basis of a full network, all of it is subject to the delays applying when 100% full

regardless how serious the congestion becomes, early increments remain in place.

There are variations on the mathematical algorithm described above. As far as the writer is aware however all start with an empty network and all achieve equilibrium by varying network conditions not the drivers.

The process is in fact quite different from real life. In the real world every driver sees him or herself as the last to enter the network. For a given time period nobody experiences an empty network. Route choices are made on the basis of essentially constant network conditions. The comparison is set out in Table 1.

If every driver was faced with exactly the same network conditions and behaved in exactly the same way, one would expect them to make exactly the same route choices. Although occasionally drivers can view adverse conditions on the preferred route in enough time to select an alternative, in general their behaviour is governed by their own nature and by conditions in the immediate vicinity. Even small differences between routes would cause "all or nothing" preferences.

The fact that this is not the case suggests that driver variability is an important factor in route choice. There are of course wide differences in how drivers use the network. A tourist from out of town will behave differently from a taxi-driver who knows all the back streets and alleyways; a truckdriver will have quite a different attitude to steep grades and narrow lanes than a motorcyclist.

The discrepancy between real and model processes will have its greatest effect where alternative routes are very different in character:

high-speed versus slow-speed roads

routes with signalised intersections versus those with priority control

arterial roads versus local road network.

In these situations it may be difficult to reproduce actual route choice. The modeller should not be tempted to alter the network description to improve agreement. The problem arises through route choices established when the network is empty or near empty so manipulating volume-delay functions will be ineffective. Changing the network description will improve agreement for the base case but only by compromising the model's ability to predict anything else.

To summarise:

Model assignment algorithms achieve equilibrium by progressively varying the network from 0-100% full, whereas in reality every driver sees him or herself as the last to enter the system.

The variability of drivers and their vehicles should be recognised as an important factor in how traffic exploits a network.

In calibrating a model some errors cannot and should not be fixed. Sometimes they will result from route choices made on the basis of an empty or near-empty network, so manipulating delay functions will be ineffective. ì

Finally, the phenomenon is a reminder to maintain sensible traffic management. A model showing severe distortion usually indicates unresolved conflict, major bottlenecks or a roading hierarchy which is not clear to users. These should be addressed before more rational performance can be expected, either in the model or in reality.

Bill Barclay BARCLAY TRAFFIC PLANNING Lower Hutt, New Zealand