Discusssion

The purpose of this study was to couple a simple demand generation method with route assignment and transportation micro-simulation via a computational feedback procedure. We wanted to explore in how far such an approach is feasible, and then out of scientific curiosity and as a benchmark we compared the results with real world data and with existing EMME/2 study results for the same problem. What can one learn from this?

First, it is now indeed both methodologically and computationally possible to systematically couple demand generation, route selection, and transportation micro-simulation. Again, this does not automatically mean that this is always the best method; however, it can and thus should be explored as one of many alternatives. Also note again that practitioners have always done some version of this feedback: If an assignment did not generate plausible flows, it was common practice to adjust the trip matrix (K. Cervenka, personal communication). The main differences thus are that we do it systematically and computerized, and that we use a micro-simulation instead of a static assignment. -- The second result is that for the morning peak, extremely simple assumptions yield results which are comparable to results of an EMME/2 study.

An important task would be to separate the influences of the different modules. In addition to the input data, there are four computational modules involved in this study: demand generation, routing, traffic flow simulation, and feedback mechanism. All of these can contribute to variations in the volumes. A systematic study would vary or switch these modules one by one and establish the effect on the volumes. This was beyond the scope of this investigation; the following paragraphs will discuss some of the issues.

NETWORK DATA: We have used the same network input data as the EMME/2 studies. Errors here should, to a certain extent, show up similarly with both approaches. It seems that at the level of current accuracy, there are no major errors in these files. That belief is reinforced by the fact that Portland Metro has been using these files for many years.

DEMAND GENERATION INPUT DATA: The data used here was: household locations, workplace locations, and distributions of start times and trip times. The accuracy of these is unkown. With regard to trip times, it was already discussed earlier that the trip times from the census most probably over-estimate times on our network, for two reasons: (1) Travelers intuitively report the time from door to door, not the time actually on the road. (2) Since many local streets are missing in our network, the time spent in our network should be smaller than the complete time on the road. Indeed, reducing all trip times to 80% (``sim-80'') in our study did not lead to significant changes in volumes and even led to higher (and more realistic) volumes on the major streets, adding to the assumption that reported trip times are probably too high. Also, just looking at home-to-work trips is a simplification. Any traffic besides home-to-work trips is neglected, such as deliveries, people returning from night shifts, shopping, leisure, etc. All these will be indispensable in order to understand 24-hour traffic patterns.

VOLUME COUNT DATA: There is a slight inconsistency between the input data and the volume count data: Input relies on the census, which is from 1990, while the volume counts are from 1992 and 1994. In fact, the average change (mean bias; see above for definition) of traffic flows from 1992 to 1994 is $+$4%. A bigger challenge is the variability of the data. Fig. 36.5 shows, where available, the counts from 1992 against the counts from 1994. There is strong variability of the counts, and the average absolute difference (mean error, see above for definition) is in fact 31%.^36.4This indicates that in future two things need to be done: (1) Field data need to include a measure of variability; and (2) the corresponding variability measure needs to be obtained from simulations.

ROUTING: This study assumes fastest path routing. Most probably, this is only an approximation of what real people do. In fact, both our simulation results and the model results from the Portland Metro study over-state traffic on minor streets, indicating that the simulated travelers are more willing to accept complicated detours than real world travelers. Also, at the moment no other mode of transportation is included. For the Portland case, this should for example lead to an over-estimation of car traffic between downtown locations.

TRAFFIC FLOW SIMULATION (also called network loading): As discussed earlier, our traffic flow simulation (the queue model) underestimates volumes. In contrast, traditional assignment network loading usually over-estimates volumes (depending on the cost function).

A heuristic possibility for progress would be to design a traffic flow simulation with a behavior somewhere in between our queue model and the traditional assignment network loading. A more systematic approach would be to use a more realistic micro-simulation in order to exactly pin-point the deficiencies. In that context, it would be interesting to also look at link speeds in order to decide whether low counts are caused by low traffic or by congestion. This data is easy to extract from the simulations, but it typically does not exist for the field. ITS technology will have a significant impact here.

FEEDBACK: Our feedback method performs slow adaptation based on the previous iteration, similar to fictitious play in game theory. While the result of such an approach is not exactly a Nash Equilibrium, it is assumed to be close.^36.5 Two aspects need to be considered separately:

• Convergence/uniqueness: If one sees the second-by-second trajectory of the micro-simulation as a point in state space, then the iterations are mappings from that state space into itself (e.g. Bottom, 2000). The way our iterations are set up, they describe a Markov-process in that state space, which means that the iterations eventually reach a steady state with a corresponding steady state density in state space (e.g. Cantarella and Cascetta, 1995). Little is known about the characteristics of this steady state density distribution, for example if it is unique, or how many iterations one would need to be reasonably close to ergodicity. In practice, it seems that route iterations behave in a similar way as traditional steady state assignment, that is, they normally yield, within Gaussian fluctuations, unique results for the traffic on the link level (e.g. Bottom, personal communication; Nagel et al, 1999). We are not aware of results of how this extends to feedback iterations into the trip distribution as considered in this paper.

• Human behavior: It is well-known that convergence results are used only because they are scientifically well-defined, not because they are realistic. When comparing to field data, one should keep in mind that it is unclear how close real systems are to the converged result.

INHOMOGENEITIES: One aspect already mentioned earlier in the text but that should be stressed again is that our method unrealistically assumes homogeneity of all aspects of the scenario except for traffic. For example, it is assumed that the behavioral function $f_{ch}$ is the same for everybody, and that one can obtain it by averaging both the trip times and the accessibility over the whole population and the whole region. This is clearly a simplifying assumption -- for example, one might expect that people living downtown have a stronger dislike of long trip times than the average population.

Another inhomogeneity in the Portland situation stems from the fact that the part of the metro region which is north of the Columbia river, so-called Clark County, is part of the State of Washington, while the rest of Portland is part of the State of Oregon. Many Oregon workers choose to live in Clark County for the lower property taxes and cheaper large-lot housing (an effect of differences in land use policy), despite the congested commute and Oregon income tax. Oregon has one of the highest personal income taxes of the U.S. States, while Washington does not have a State tax on personal income. Oregon personal income tax is also paid by non-Oregon residents as long as they work in Oregon. Thus, there is a substantial tax incentive for those who live in Clark County to also work there. This, however, is often not possible due to a low jobs-housing ratio in Clark County. All this results in a relatively high split between peak and non-peak direction volumes on the Columbia River bridges. Sales tax is the opposite: There is no sales tax in Oregon while sales taxes in Clark county average 8%. In consequence, retail activity in Clark County is somewhat suppressed by residents' proximity to tax-free shopping in Oregon. For example, there is a major big-box retail area on the Oregon side of the I-5 bridge that owes its existence to the sales tax disparity. (Bill Stein, Portland Metro, personal communication)

This should result in less traffic northbound into Clark county in the morning peak in reality than in our model. This is easy to check since there are only two bridges across the Columbia river. Indeed, with sim-80 we obtain 7473 veh/hour northbound as opposed to 4650 in the field, while southbound the numbers are comparable: 10052 and 9740, respectively. Sim-100 numbers are lower than sim-80 numbers, due to congestion in the model, but have the same tendency.

Figure 36.5: Variability of field data. For some measurement locations, count data were available both for 1994 and 1992. For those locations, the 1992 value is plotted against the 1994 value. A better understanding of field data variability will be necessary for further progress.