Purgatory/RouteGuessing

=Why random Routes are evil and some Bullets= In the past, both route computation applications from the SUMO suite, namely DUAROUTER and JTRROUTER had the possibility to generate random routes. Some people were happy with this feature; we were not, especially because of the frequent questions on the usage of this feature.

This text explains why it is strongly NOT recommended to use random routes. It also shows some attempts for solving the problem of having no demand by some more sophisticated approaches for route generation.

As a base, we will use two example networks for which we know the demand:
 * The first is a network from the city of Bologna. It stems from the iTETRIS project and arrived us as a Vissim simulation scenario, including both the network and the routes. It is an urban scenario.
 * The second is a higway scenario. We have reconstructed the routes using information collected on observation points (induction loops) for the TrafficOnline project. We suppose the demand to be valid.

Ok, now we build random routes for both scenarios, choosing randomly a source edge ("road") and a destination edge. We generate as many vehicles as are simulated within the real, correct demand. Here are the results:

Ok, what has happened, what is different?
 * The probability to use a road differs between the original and the random demand
 * The mean velocities differ
 * In the original scenarios, there is less traffic during night, more at day, most during the peek hours.

We will now address the issues one by one.

Probability to use a Road
When using a random demand, we choose a small road with the same probability to be a starting/ending road as we do with a big one. One could argue that the capacity for standing vehicles of big roads are similar to those of small roads - may be, but I personally would assume that smaller roads are statistically less frequented during work time, mainly only during the rush hour where persons are leaving their homes or coming back. Shopping places are found rather on the major roads what makes them more probably to be used as origins/destinations during the day. Also, there are parking places etc, also mostly accessed from the major roads. Of course, this may be not true for shopping areas, where people try to find parking places in covered small streets. Nonetheless, this is supposed to be seldom - and, you don't know this from just looking at the road network.

As a conclusion, it is rather not assumed to choose minor and major roads for origins and sources with the same frequency. Still, we have no solution for this. But read further...

Transit Traffic
Choosing randomly an edge from within the network completely ignores the fact that the smaller the simulated area is, the more traffic is just transiting the network - enters the network at its boundaries and leaves the network at its boundaries.

Of course, for our highway scenario, this is probably the largest source of mismodelling - no one starts or ends his trip in at a place in the middle of the highway.

So for improving our method for route generation, we have to determine the network's boundaries and the edges which are incoming into it or outgoing from it. We can then try to use those as the major origins/destination of our routes.

Turning Directions
Some people who tried the random routes generator may have noticed that many vehicles were turning. This is BTW also the case with our first "sophisticated" approach. The reason is very simple: if we choose two roads randomly, the probability to choose an origin which points to the opposite direction then the destination is located in is about 50%; sure, in detail this depends on the network topology. The same counts for the destination; the probability that the route ends at the road into the opposite direction is also high, maybe not 50%, but still high.

In real life, you have to turn at least once, too. Nonetheless, due to not having transit traffic, the overall number of turnings is much larger in random routes than in reality.