Practical Application of Activity-Based Models Tour-based Activity-based Micro-simulation Aggregate 4-step
4-Step is a “top down” approach Divide population by zone / income / hh size (maybe also number of workers, car ownership, age group) More segments would be better, but there is a practical problem…
4-Step is a “top down” approach Divide population by zone / income / hh size TRIP GENERATION Adds trip purpose dimension For example, Home-based work, Home-based school, Home based shopping Home-based other Work-based Other Non-home-based (NHB)
4-Step is a “top down” approach Divide population by zone / income / hh size TRIP GENERATION Adds purpose dimension TRIP DISTRIBUTION Adds origin-destination dimension Output is many trip matrices
4-Step is a “top down” approach Divide population by zone / income / hh size TRIP GENERATION Adds purpose dimension TRIP DISTRIBUTION Adds origin-destination dimension MODE CHOICE Adds mode dimension Output is even more trip matrices
4-Step is a “top down” approach Divide population by zone / income / hh size TRIP GENERATION Adds purpose dimension TRIP DISTRIBUTION Adds origin-destination dimension MODE CHOICE Adds mode dimension NETWORK ASSIGNMENT Adds time of day and route dimensions DIMENSIONALITY CRISIS!
ACTIVITY SCHEDULE HOME 5 SHOP START 7:00 AM 7 MOVIE 2 & 4 SCHOOL 550 min 10 min 90 min HOME 5 5 SHOP 25 min 1 Eat, 6 Eat, 8 Sleep 30 min START 7:00 AM 25 min 20 min 7 6 4 15 min 1 15 min 240 min 7 MOVIE 2 & 4 SCHOOL 240 min 3 120 min 2 10 min 10 min 3 LUNCH 40 min
Major problems with aggregate trip-based approaches Non-home-based trips! Mode choice not consistent with adjacent trips Destination choice not consistent with next trip Time of day not constrained by adjacent trips No substitution between tours No interactions between household members Aggregation errors/biases
Trip-Based to Tour-Based Trip generation Tour generation (fixed rates?) Trip time period Tour time periods (fixed factors?) Trip distribution Tour destination choice (gravity model?) Trip mode choice Tour mode choice Intermediate stops
Tour-based to person-day-based Tour generation Day-pattern choice Activity generation Trip chaining Tour time periods Tour sequencing and time periods Tour destinations Tour destinations Tour mode choice Tour mode choice Intermediate stops Intermediate stops
Person-day to Household-day Day pattern choice Day patterns linked across HH members Activity generation Joint HH activities Linked HH activities (escorting) Allocated HH activities (maintenance tasks) Individual activities All tours individual Some tours joint/linked
Geography of New Generation Developed & Used Portland (METRO) San Francisco County (SFCTA) New York (NYMTC) Columbus (MORPC) Started: Atlanta (ARC) Denver (DRCOG) Dallas (NCTCOG) Tampa Bay (FDOT) Considering: Houston (HGCOG) Raleigh-Durham (CAMPO) Sacramento (SACOG) Kansas City (MARC) Seattle (PSRC) San Diego (SANDAG)
Geography of New Generation Seattle Portland NY Denver Columbus Kansas Sacramento SF Raleigh Dallas Atlanta San Diego Houston Tampa
Main Features Already in earlier designs (Portland, San Francisco, New York): Tour as unit of modeling Consistent generation of all tours made during a person-day Stochastic micro-simulation application framework Added in later designs (Columbus, Atlanta, Denver): Explicit modeling of intra-household interactions Greater temporal detail (1 hour or less) and consistency in time use and activity / travel scheduling Greater spatial detail (10,000-20,000 grid cells) for LU and walk / bike / transit accessibility
Microsimulation is a bottom-up approach POPULATION SYNTHESIZER Create a synthetic population by sampling from actual households to matches control statistics or forecasts by zone Output is a full list of households/persons (like census data)
Microsimulation is a bottom-up approach ACTIVITY AND TRAVEL SIMULATOR Uses similar models to 4-step (activity generation, destination choice, mode choice) but uses the Monte Carlo method to simulate discrete choices from probabilities Also considers trip-chaining (tours) and scheduling (time-of-day) Output is a list of trips and activities (like household travel survey data) POPULATION SYNTHESIZER
Microsimulation is a bottom-up approach AGGREGATOR Compile trip matrices for network assignment or simulation. Can also produce reports to look at travel by specific population segments. ACTIVITY AND TRAVEL SIMULATOR POPULATION SYNTHESIZER
Microsimulation is a bottom-up approach NETWORK ASSIGNMENT/SIMULATION AGGREGATOR ACTIVITY AND TRAVEL SIMULATOR POPULATION SYNTHESIZER
“Continuous” space Use very small units – GIS parcels or grid cells (e.g. 200 meter squares) Very good for modeling transit accessibility and activity attractions. Density variables used to capture surrounding land uses. Matrix-based measures such as in-vehicle times remain at zonal level.
Benefits of using grid cell data Walk access time to transit based on grid cell GIS measures – much better results Intra-zonal walk times based on distance between O and D grid cells - intrazonal dummy variable becomes insignificant Grid cell-based measure of percent of streets with sidewalks gives better explanation of walk/bike share than CBD dummy or other zone-based measures.
“Continuous” time Use small time periods- 1 hour or half-hour Model activity or tour start and end times simultaneously, conditional on time remaining after higher priority activities. Better to capture interactions between tours and activities. Better for modeling peak-spreading More accurate input to traffic simulation
Important Policy Areas Congestion pricing / time-of-day incentives Policies affecting work or business hours Parking policies Ridesharing policies Demographic shifts (aging, household composition)
How should models be judged? Ability to predict future changes Sensitivity to a wide range of policies Ability to match current data
How are models typically judged? Ability to match data on current situation Simplicity of models, data, and forecasts Predictability of forecasts Replicability of forecasts
Issues in simulation error Stochastic models do not necessarily converge Need to separate real variability from simulation error. Simulation error decreases with square root of iterations. Stability of results depends on level of resolution (TAZ, county, etc.) Simulation errors do not multiply – compensation is more likely.
Tests of Random Simulation Error Ran the model system (except for assignment) 100 times Changed the random seed for each model for each run. Analyzed the variability in results obtained from each model in the system. Main questions: What is the range of results obtained? How fast do the results converge toward the mean? How is the variability related to the level of aggregation?
Trips per Person % Difference from Final Mean
Tours by Mode from a Single Origin TAZ % Difference from Final Mean
Conclusions Regarding Simulation Error For region-wide results, a single run is adequate. For corridor-level or neighborhood-level results, 5 to 10 runs should be adequate. Looking at very small areas (TAZ’s), rare sub-populations (e.g. single parents) or rare behavior (e.g. transit use in some regions) requires more runs to reach stable results. We have not yet looked at results with full equilibration with assignment. The feedback from level-of-service should dampen the variation even further.
Further Conceptual Evolution Intra-person integrity Activity & travel pattern configuration Time use & activity generation Time-space constraints on activity location Feedback through individual time budgets Inter-person intra-household integrity Coordinated daily patterns Episodic joint activity & travel Maintenance task allocation Car allocation
Simultaneous vs. sequential choices At the tour or trip level – sequence of Mode choices Destination choices Scheduling/sequencing choices Trip chaining decisions Empirical question, may vary by purpose. More data on constraints and flexibility would be useful Use different sequences for different types of situations or individuals? Need a more flexible modeling framework.
Need dynamic models to deal with … Advance vs. real-time planning Simultaneous vs. sequential processes Learning and information acquisition Feedback processes over time Direction of causality Location vs. travel (induced demand) Supply vs. demand (peak spreading)
Dynamic models will require … Different types of data Panels (?) Before and after surveys Retrospective surveys Hypothetical choice contexts Different types of models (?) Strict adherence to econometric choice theory has prevented the use of non-static models