The first day of the peer review was spent by BKR discussing plans for a new BKR model. During this discussion, many topics came up which initiated discussion among peer review panel members and between panel members, BKR, and RSG. This section documents the key points that arose during these discussions.
BKR consulted with RSG to explore two options for improving the BKR Model. The first option was to develop a hybrid/enhanced trip-based model and the second option was to adapt PSRC’s activity-based model (ABM), SoundCast, to the BKR modeling region. BKR directed the panel’s attention toward reviewing the hybrid/enhanced trip-based model design, for the following reasons:
There are several key features of the hybrid/enhanced trip-based model, compared to their existing trip-based model, that will be discussed in detail in the subsequent sections:
The BKR model will be enhanced by doubling the number of TAZs from the current model to a total of 1500 TAZs. The zones would be very disaggregate within the three city region and include TAZs outside of the three cities but within the PSRC region at a more aggregate level. Figure 2, below, shows the proposed BKR model boundary delineations and TAZ system. Light orange are TAZs within the cities of Bellevue, Kirkland, and Redmond and their surrounding/buffer areas. Dark orange TAZs are outside of the three city boundaries but within the BKR model. Areas in gray are within the PSRC region, but will not be modeled directly within the BKR model.
The panel agreed that making BKR regional boundary smaller than PSRC region is a good idea. However, the panel suggested expanding the north and south boundaries to ensure that most residential locations of people who work within the BKR region are included as internal zones within the BKR model.
The enhanced BKR model will take in as input ten supply-side employment categories:
The categories were designed to be consistent with the PSRC model categories.
The panel stated that the employment categories should line up with the trip purpose categories as much as possible. For example, government should be separated from K-12 education, since school trip destinations will be correlated with K-12 education and not government. The panel also suggested that BKR model should consider separate categories for university employment, hospitality employment, and healthcare related employment.
Socioeconomic data will be obtained from PSRC’s population synthesis by aggregating from parcel and block groups to BKR TAZs. BKR mentioned concerns with issues involving random sampling of the population synthesis, especially when analyzing a very small geographic area. The panel stated that the same random seed should be used when running the population synthesizer and that when BKR is concerned with exploring socio-demographic changes for only a single zone or small sub-area, the analysts should run the population synthesizer on only the zone or sub-area of interest and combine the new data with the original socio-economic data for all other TAZs.
BKR plans to develop an extensive bicycle network and has identified the following components for inclusion on the network links:
The panel stated that the list may be more detailed than necessary. Vehicle facility type can be used as a proxy for vehicle volumes. Elevation gain is a better measure than slope. The panel stressed that the bicycle network must be an all-streets network. The panel confirmed that bicycle assignment should be handled in separately than other network supply models because of increased network detail and different relevant link attributes. The panel indicated that accessibilities from bicycle improvements may primarily result in route choice changes rather than shifts in mode choice. The bicycle assignment model takes in as input bicycle trip outputs from the demand model. The state-of-the-practice bicycle assignment models tend to be logit models and are fundamentally different from vehicle assignment models, especially since bicycle lanes do not have capacity constraints. The results of bicycle assignment model do not need to be fed back into the demand model because the bicycle impedances do not change due to the lack of capacity constraints.
The enhanced/hybrid trip-based model includes two sub-modes within the transit: bus and LRT. Transit travel times are linked to the auto network. BKR was concerned that adding too much complexity into the mode choice model via additional sub-modes would add to time needed to run the model. The panel assured them that mode choice runs very quickly. Additional time will be needed to set-up the skims, but once they are in place running the skims should not take too long.
The panel commented that from a political standpoint it may be important to model LRT separately from bus within mode choice, but from a modeling standpoint this is less important. The panel also suggested that BKR consider adding additional nests for access type, such as park and ride or kiss and ride.
BKR is considering adding network capabilities so that the assignment is sensitive to intersection delay as well as designing the generalized cost function such that delay, or congested time, is weighted more heavily than free flow time. This methodology is based on current research under the SHRP2 L04 project that is investigating a means of incorporating reliability in static assignment.
The panel cautioned BKR to not incorporate intersection delay functionality into the network. They stated that it was a burden to maintain and that there was no evidence that is helped with calibration and validation. They stated that there are fundamental problems with static assignment that cannot be solved with intersection delay.
With regard to using perceived time within the generalized cost function, the panel cautioned against weighting congested time differently from free flow time; however they did not have detailed knowledge of the SHRP2 L04 project and thus could not comment directly on the methodology. The panel did suggest to incorporate a separate mode for tolled facilities within the network since they exist in the area.
BKR wants a coherent treatment of accessibility throughout the model, and have put together a set of seven accessibility measures that they are considering including in their model:
They were interested in hearing the panel’s recommendations on how to incorporate accessibility measures into the model.
The panel commented that different accessibility measures are used for different purposes. Logsums are desirable for informing models, but are difficult for modelers to explain to themselves and to clients and the public. The panelists suggested to not abandon logsums altogether.
With regard to the other accessibility measures, the panel suggested using network-based measures (i.e. time and distance) rather than using crow fly distance. They also suggested adding a decay function to calculate the accessibility measure. This will make locations farther away less accessible than those closer in, while eliminating “cliff effects” where a location just outside of the radius is not included while a location just inside the measure is included.
The panel also said that it is important for TAZs to be properly designed so that accessibility measures perform optimally. Other accessibility measures they suggested considering were number of intersections by type and employment within 30 minutes of transit.
The enhanced BKR model will include an auto ownership model that will take in as input the following population synthesis variables:
In addition they are considering using variables that measure accessibility to work, school, retail, and service locations.
The panel cautioned that using accessibility measures in auto ownership model could result in a model system that is too sensitive to accessibility. The panel said that when running the modeling system, turn on and off certain components to control the sensitivity. Hold the auto ownership model constant for most model runs.
The panel suggested including accessibility to usual work place location within the auto ownership model. They also suggested modeling auto availability instead of auto ownership, if the household travel survey differentiates between the two. They also suggested calculating accessibility based on the auto ownership level (e.g. zero vehicles vs. one or more vehicles).
BKR is considering the following home-based trip purposes:
The panel suggested adding personal business, social, and recreational trips as trip purposes. They highly recommended incorporating a usual work location model before trip generation. The usual work location model is a better way to handle the modeling of high-income high-tech jobs rather than trying to generate high-tech work trips directly within trip generation and then modeling them as a separate trip purpose within the destination choice model.
BKR identified a set of modes they are interested in assessing. These include:
The model would be a nested logit model with non-motorized, auto, and transit at the upper-level nest. Bike and walk would be nested under non-motorized. The transit nest would be divided into walk-to-transit and drive-to-transit. The auto mode would include SOV and HOV as the next level nest, with HOV 2 and HOV 3+ further nested under HOV. Finally, toll and non-toll would be nested under each occupancy-level auto mode.
The panel agreed with BKR that freight will be borrowed from the PSRC model and not directly modeled within the BKR model. The panel stated that telecommute/work at home is a work location choice decision and not a mode. Activities performed online or at-home are not modeled in the state-of-the-practice activity-based models.
The panel suggested including taxi/ride-sharing as its own mode. The latest household travel survey did include questions asking about taxi/ride-sharing, although very few individuals reported making trips by this mode. However, including taxi/ride-sharing directly as its own mode will give the analysts flexibility later on to do scenario testing. Specifically, it will help with autonomous vehicle modeling. BKR can assert the constant now since the household travel survey probably does not contain enough data to properly estimate the constant.
Similarly, the panel recommended nesting each individual transit mode within the transit nest. They suggested including all modes into the structure that may occur in the future. This would include bus, LRT, and BRT. FTA has accepted BRT as a mode, whose alternative-specific constant may be asserted, for example as ½ of the rail constant. It is easier to have a placeholder for each mode now rather than trying to add it in later.
The panel also recommended including more detailed modeling of park and ride and kiss and ride modes. One suggestion made by a panelist was to add informal park and ride which would allow most TAZs that don’t have parking charges or restrictions to accept the ability to park and ride in non-formal park and ride locations.
The panel agreed with modeling toll modes directly within the mode choice model, since the mode choice model can capture implied value of time. Doing it this way is more difficult to estimate, but it does not add much additional computation time during model application. The panel did caution, that if toll is included in mode choice then it will also need to be included as a separate mode in the assignment model.
The trip distribution model would distribute trips separately by trip purpose. Depending on the trip purpose being modeled, various multimodal accessibility variables and land use size variables (e.g. employees by category, park acres, number of households, number of hotel rooms) would be included in the model.
The panel agreed with the procedure laid out in the modeling plan. The panel was supportive of the use of shadow pricing techniques for, e.g., constraining workplace choices to actual jobs available. They stated that the model will not work well without shadow pricing and that any locations that have capacity constraints, such as park and ride locations must have shadow pricing.
The panel suggested to make sure the accessibility measures and choice set reflects auto availability. They also suggested including dummy variables where needed during model calibration. They cautioned that big datasets generated from passive data sources may be inadequate for detailed model calibration, but supported using the data at the district-to-district level as an additional source for model calibration and validation.
BKR is contemplating modeling and assigning trips at the daily level and then allocating to AM and PM peak periods. This methodology would decrease run times and ensure consistency between more disaggregate time periods and daily totals.
Time-of-day issues are paramount to many policy decisions that the agencies will need to analyze. The area has an influx of workers from outside of the modeling region during working hours; thus there is a lot of directional congestion within the region that would not be captured appropriately if trips were modeled at the daily level. Given the region’s emphasis on tolling by time-of-day, the model and skims need to capture varying price and travel times by different times of day. In addition, if the agencies are interested in examining peak spreading then they will need to include disaggregate time periods. On the other hand, the panel was okay with maintaining only peak and off-peak transit skims.
The panel stated that BKR can always sum up more disaggregate time periods to obtain the daily total, and thus inconsistency between finer time periods and daily totals should not be a concern. The panel stressed that modeling at least four to five time periods is very important, and BKR may want to consider skimming and assigning up to nine time periods (e.g. three 1-hour periods in the AM peak, three 1-hour periods in the PM peak, mid-day, evening, and night). All panel members cautioned against skimming at a finer disaggregation than one-hour increments, but most panel members supported modeling as low as half-hour increments if development of an ABM was pursued.
The panel did agree that it is a run-time trade-off, in that the more time periods that the model includes the greater the run-time of the model. However, the panel cautioned that time-of-day modeling should not be compromised and that BKR should find other ways to manage and decrease run-time issues.
The hybrid modeling approach would begin by modeling home-based trip generation at the disaggregate person-level. The disaggregate trip generation would replace cross-classification tables that are used in traditional four-step models. These trips at the disaggregate level would then feed into mode choice, which would also be handled at the person-level. Trips would then be aggregated immediately after mode choice and fed into trip distribution. Based on the mode, purpose, and trip destination, non-home based trips would be generated. These non-home based trips would be categorized into work-based trips if they are associated with a work trip or non-work-based trips if they are associated with a non-work-based trip. The main benefit of this approach is that the model can capture the modal and spatial consistency between the non-home-based trip and the home-based trip it is linked to.
This modeling approach has been adopted by the Salt Lake City and Anchorage MPOs and has been included within the Tennessee and Iowa statewide models. These agencies chose to adopt this approach since they were relatively satisfied with their four-step model and wanted to adopt incremental improvements without starting over with a new model.
The panel’s main concern with this approach is the model’s lack of sensitivity to the policies that the agencies would be interested in analyzing, especially those related to time-of-day. Even though the model is at a disaggregate level for trip generation and mode choice, non-home based trips represent approximately 40% of all trips. Given that non-home based trips are represented at the aggregate-level, the model would still be missing valuable information for a large percentage of trips. Time-of-day consistency between home-based and non-home based trips is also lost with this approach. The model would not be able to guarantee a continuous path through both space and time. If BKR is interested in analyzing demand management strategies, then having a means to analyze finer time-of-day disaggregation is necessary for modeling the appropriate sensitivities to these strategies. The panel stressed the importance of BKR to really consider the policies they are interested in analyzing and consider whether the enhanced/hybrid model will be able to address these analyses to the extent desired.
BKR is interested in designing the enhanced model so that it can interface with PSRC’s ABM. For trips generated from the area outside of BKR, one option would be to take the results from the regional model and aggregate them to BKR zones.
The panel was not convinced that the enhanced BKR model would ensure that trips coming into BKR from outside the area would be appropriately sensitive to travel demand management polices that the agencies may be interested in testing out (e.g. providing employees bus passes). They did agree that large-scale regional projects, such as adding in rail line into the BKR region, might be better handled by PSRC’s ABM directly rather than using the enhanced/hybrid trip-based model.
The panel stated that there are a lot of details that still need to be worked out and a lot of issues that will need to be addressed with regard to integration of the enhanced/hybrid BKR model with the PSRC ABM. The panel agreed that trying to integrate the enhanced/hybrid BKR model may not be as desirable and will be much more complicated than either adopting PSRC’s ABM or not integrating the BKR model with the PSRC model at all.
BKR identified a number of concerns with adopting PSRC’s activity-based model to the BKR area including:
The panel’s response to these concerns are discussed in the next section along with their assessment of the benefits of adopting PSRC’s activity-based model.
BKR is concerned that ABM results will differ significantly by model run when analyzing local projects and thus will require conducting multiple time intensive model runs to get presentable results. They prefer that the model results be relatively easily reproducible.
The panel agreed that when analyzing one particular zone, simulation variation can be an issue. The larger the area that is analyzed, the less simulation variation is an issue. However, the panel argued that simulation variation can be managed and controlled. In addition, the panel pointed out that simulation variation is attenuated by static assignment procedures, and if the results are fed into DTA then the issue with ABM simulation variation is a minor issue compared to the simulation variation that exists within DTA models.
They also noted that when analyzing individual local projects or polices both the hybrid/ enhanced trip-based model and an ABM are going to have issues in analyzing the results, though they may are different types of issues. With the hybrid/enhanced trip-based model the issues may be more related to the lack of sensitivity to a particular analysis, while with the ABM the issues may be due to simulation variation problems.
PSRC’s ABM takes 30 hours to run, while the existing BKR model takes only 45 minutes including assignment. Thus, there is legitimate concern that moving toward an ABM would significantly increase the time needed to run the model. However, PSRC is working on improvements to the model to decrease run-times.
The panel noted that depending on the zone system and time-of-day periods chosen for the enhanced/hybrid trip-based model, run times may not be significantly different from those that would exist in an ABM developed for the area. The panel also noted that run-times can be managed through improved hardware and multi-threading capabilities, decisions regarding time-of-day disaggregation, proper project management, and through simulating a smaller sub-area population or sampling the population at different rates.
The enhanced/hybrid trip-based model is a simpler model than an ABM. Some of the panel members agree that given BKR’ familiarity with trip-based model the enhanced/trip-based model may be easier to understand than an ABM. In addition, there are less data needs for a trip-based model and trip-based models are easier to calibrate.
The panel did agree with BKR that data and input requirements of ABMs can be burdensome. The panel did not believe that it would be more difficult to calibrate the ABM then the enhanced/hybrid trip-based model. They believed calibrating an ABM is more intuitive than calibrating a trip-based model. They also thought that adopting the PSRC model will be less complicated than developing the enhanced/hybrid model from scratch and then finding a way to integrate the model with PSRC’s ABM. PSRC is willing to work with BKR on ABM training and data sharing.
Throughout the Peer Review meeting, the panel questioned BKR on why they devoted so much attention during the peer review to presenting the enhanced/hybrid trip-based model over discussion of adopting PSRC’s ABM. Given the existence of PSRC’s ABM and the desire of the BKR model to be sensitive to changes in land-use, accessibility, transportation policies, the panel thought that adopting PSRC’s model was the prudent option. There were differing opinions on whether BKR should begin adoption of the PSRC model immediately or whether it should be a longer term plan. Some of the panel members thought that it was best to wait until the PSRC ABM has been used by PSRC for a longer period of time prior to adopting the model for BKR to ensure the model has been fully tested and vetted.
Since BKR would develop the enhanced/trip-based model from scratch rather than incrementally improving their existing trip-based model, most of the panel members believe that adoption of PSRC’s model will be less complicated, cost less, and take a shorter time to develop. BKR has a strong interest in being able to properly examine a variety of complex policy strategies. An ABM is the most appropriate tool for examining these scenarios, especially since it can handle analysis related to time-of-day.
