Experimental setup and simulation results

The study described in this paper was carried out as part of the Transims project (Transims, 1992), which was at that time aimed at simulating the whole city of Portland microscopically (i.e., with resolution down to single individuals) under consideration of activity generation, modal choice and route planning, and transportation dynamics. The simulations described in this paper were run on a road network consisting of 8,564 nodes and 20,024 links representing a subset of the real network.

Traffic counts for validation are available for 495 links comprising flow data for the morning peak from 7:15am to 8:15am. Data are available for the years 1992 and 1994. Data for 1992 is used for those links for which no 1994 data are available (68 links); for all other links, the counts of 1994 are used.

The data were collected using pneumatic road tubes and averaged over two or three weekdays; mostly on Tuesdays, Wednesdays, and Thursdays outside of holiday periods and while school was in session. The counts are not seasonally adjusted. Axle adjustment factors are applied to account for trucks, which are not explicitely counted. The accuracy of the counts is considered to be $80-85\%$ (Bill Stein, Portland Metro, personal communication).

Another set of data available are the results of assignment runs by Portland Metro. These runs use their own demand generation, and the EMME/2 assignment algorithm (Babin, 1982). Note that ``EMME/2'' results in this paper will refer to results of that particular study by Portland Metro including its demand generation.

One problem with our census based assignment approach is that trip times are overestimated for at least two reasons:

(1) First, when people are asked for the time they spend for their trip to work they usually report the total door to door time including the time to get to the car or park the car. On top of that, people tend to overestimate the time they spend driving especially in stop-and-go traffic (K. Lawton, personal communication).

(2) Second, the road network used for our simulation does not cover most minor streets. That means the time people spend on these roads should be taken out of the distribution.

The amounts of those times can however not be estimated without further information. To get an idea whether a trip time distribution which is shifted to lower trip times yields more realistic results, two different workplace assignment iterations were done: One with the original census distribution, and another with all desired travel times reduced to $80\%$ of the original value. In the following we refer to these runs as run sim-100 and sim-80, respectively.

In Fig. 36.2 the total trip time is plotted for both series, sim-100 and sim-80. Each simulation run refers to running the queue simulation for the morning (from 4am till 12pm). After every 5 iterations in which people are rerouted only, people are assigned to new workplaces. This can be seen as a sudden, normally upward jump of the total trip time in the plot. The reason for the jump is that it takes some reroute iterations to adjust the routes to the changes in the trip demand pattern. We ran 20 route iterations after the last workplace assignment to make sure that the routes are actually relaxed.

As expected, the total trip times are lower for sim-80 (Fig. 36.2). Yet, it is striking that a decrease in desired trip times by $20\%$ results in actual trip times which are about $50\%$ lower. The reason will be explained in the next paragraph.

By looking at the trip time distributions in the simulation (Fig. 36.3), it can be seen that the resulting distribution for sim-80 is closer to the corresponding census distribution than it is for sim-100. Even after assignment and route relaxation, there are still a lot of unrealistically high trip times for sim-100. This results from the fact that the overall traffic demand is more than the network can carry, leading to a lot of congestion. It is well known that large fluctuations occur when transportation systems are operated with demands that exceed capacities (Kelly, 1997; Nagel and Rasmussen, 1994). Actually, detailed investigation shows that in each simulation run different people account for the very high trip times, which underlines the influence of large fluctuations. Also for sim-80, the distribution resulting from the simulation does not perfectly match the corresponding modified census distribution. Nevertheless, the effect of large fluctuations due to congestion is smaller than for sim-100. These erratic occurrences of large trip times are also the reason why the reduction of the desired trip times by 20% leads to a decrease in actual trip times by 50%: In sim-100, the system is simply not capable to find a solution that is able to match the demand, and thus has too few contributions at trip times around 500 secs while it has too many contributions at trip times above 3000 secs.

As mentioned above, we do not claim that the $80\%$ census trip time distribution leads to a realistic representation of the real traffic flows in the study area. The idea is just to check the assumption that a reduced distribution leads to more realistic traffic flow patterns. The comparison with the field data is topic of the following section.

Figure 36.2: Total trip time in the simulation during the iterative assignment with the original census trip time distribution (sim-100) and the census distribution with trip times reduced to $80\%$ (sim-80).

Figure 36.3: Trip time distributions in the queuing simulation at the 70th iteration in comparison to the 100% and the $80\%$ census trip time distribution. Only completed trips contribute to the distribution.