ILUTE Agent-Based Microsimulation Modelling: Current Capabilities, Future Prospects Eric J. Miller, Ph.D. Professor, Department of Civil Engineering Director, Cities Centre University of Toronto EMME Users’ Conference Montreal, October 4, 2010

ILUTE Presentation Outline Criticisms of four-step, trip-based models Implications for improved models Categories of activity-based models Example of an activity-based model: TASHA Lessons learned

ILUTE Population & Employment Forecasts Trip GenerationTrip DistributionMode Split Trip Assignment Transportation Network & Service Attributes Link & O-D Flows, Times, Costs, Etc. Four-Step, Trip-Based Models The four-step, trip-based approach has been the standard paradigm for urban travel demand modelling since the 1950’s. While clearly useful, it has been criticized from many perspectives, especially over the past 20 years.

ILUTE Criticisms of the four-step, trip- based approach Aggregation issues Trips versus tours Trips versus activities Individuals versus households (Lack of) connection to auto ownership (Lack of) connection to residential & employment location choice

ILUTE Aggregation issues Spatial aggregation: trips are point-to-point, not centroid-to-centroid. Temporal aggregation: trips occur continuously over time, not chunked within arbitrary time periods. Socio-economic aggregation: trip-makers are very heterogeneous; their behaviour varies accordingly.

ILUTE Spatial aggregation Zone-based travel models have difficulty dealing with: –Short-distance trips – Walk/bicycle trips – Transit access/egress Transit LOS calculations are particularly sensitive to zone system design, etc.

ILUTE Transit access Traffic Zone Bus line 20 min 5 min10 min15 min Bus stop P(transit) Walk time (min) 5 min10 min Zone Centroid

ILUTE Temporal aggregation Four-step models aggregate time into one or more discrete time periods. Each time period is analyzed separately. Difficult to: –Model peak-spreading –Handle “spill-over” between time periods Computationally burdensome to model 24-hour travel.

ILUTE Socio-economic aggregation Trip-maker attributes have a major affect on trip rates by purpose, destination choice and mode choice. Matrix/zone/trip-based methods cannot keep track of individual trip-makers. Accounting for socio-economic heterogeneity using standard matrix-based data structures can add immensely to data requirements & computational effort.

ILUTE Travel behaviour varies with S-E attributes

ILUTE Socio-Economic Aggregation Error Income P(transit|Income) I1I1 I2I2 I avg P1P1 P2P2 P avg (I avg, P avg ) Aggregate models estimated using average values will be biased (B1). Inserting average values into a disaggregate model will yield biased results (B2). P(transit|I avg ) B2 B1

ILUTE Trips versus tours We do not organize our lives in terms of individual trips but in terms of sequences/patterns of activities; these linkages can be important in determining travel behaviour. Non-work/school, non-home-based and ride- sharing are all difficult to model in a trip-based approach.

ILUTE time of day x y home work shop time of day x y home work shop travel between activity locations by transit travel between activity locations by auto Base Case 1: Shopping episode on the way home from work. Auto used for the daily activity pattern. A transit improvement causes the person to shift to transit for the journey to/from work. In order to still go shopping, a new home-based auto-drive trip chain is generated. Auto usage & emissions will be under-estimated by a trip-based model.

ILUTE Base Case 2: Drop child at daycare on the way to/from work. Auto used for the daily activity pattern. The transit improvement in this case does not result in a shift to transit for the journey to/from work, despite the good service provided for this journey, since the need to drop-off/pickup the child dominates the mode choice. Auto usage & emissions will be under-estimated by a trip-based model. time of day x y home work drop child at daycare pickup child from daycare time of day x y home work drop child at daycare pickup child from daycare

ILUTE Trips versus activities We travel to participate in activities, not for the sake of travel per se. It is argued that in order to understand travel we need to understand activity participation: –Timing/frequency of activities by type (trip generation). –Location of activities (trip destination).

ILUTE Household-Level Models Household-level models are required to “properly” deal with many system components: housing location/type choice automobile ownership demographics/household structure/lifecycle stage activity/travel scheduling Households: share resources among household members constrain member behavior condition member decision-making generate activities Household Person 1Person 2 Requests for resources, availability for tasks Allocation of resources, assignment of tasks Pers1Pers 2Car 1 Request for car Time Allocation of the car to a given person

ILUTE Auto ownership Auto ownership/availability has a major impact on mode choice. 4-step models often assume that auto ownership is an exogenous input. Improved urban form and improved transit services can reduce auto ownership and, hence, auto usage. Incorporating auto ownership in a trip-based, individual-based model system is difficult.

ILUTE Dynamics VKT Time Base Year Forecast Horizon Historical Trend Projection Dynamic, path-dependent response to policy initiatives Static equilibrium projection Future system states are emergent outcomes of path-dependent processes. System equilibrium rarely exists.

ILUTE Summary of 4-step critique Virtually all of the criticisms of the 4-step paradigm involve recognition that the context within which travel decisions are made is critical: the details concerning who is deciding to travel under what circumstances determine the travel outcomes.

ILUTE Critique summary, cont’d Key issues that must be addressed: –Heterogenity –Non-linearity –Tour-based –Dynamics, path dependencies –Behavioural processes

ILUTE Heterogeneity Heterogeneity of decisions-makers and context combined with non-linearity of decision processes requires a disaggregate approach. Heterogeneity cannot be efficiently handled by matrix-based approaches – at some point, list- based approaches become much more efficient.

ILUTE Matrices Vs. Lists The current 4-stage model for the Greater Toronto-Hamilton Area (GTHA) has 1717 zones, 5 worker classes, 4 occupation groups and 4 age categories for HBW mode choice calculations. The HBW matrix-based AM-peak mode choice model therefore requires : 1717x1717x5x4x4 = 235,847,120 evaluations of a nested logit model. In 2006 there were approximately 2 million workers making morning peak-period trips within the GTHA. It is far more efficient to construct a list of the 2 million workers and their attributes and compute their mode choices directly.

ILUTE Lists & Disaggregate Models Adopting a list-based approach can significantly improve computational efficiency. It also represents a major step towards an agent- based approach. Efficiency grows as the complexity of the system/agents being modelled grows (e.g., modelling both persons and households).

ILUTE Person List Person Age Sex Educ. Occ. Emp. …. ID Code Code Status … M 4 1 FT …. … F 5 2 PT …. … F …. …. Household List Hhld No. of No. of …. ID Persons Cars … …. …. Job List Job Occ. Salary …. ID Code … $50K …. … $65K …. …. Dwelling Unit List DU Zone Price …. ID … $245K …. …. School List Sch Type Zone …. ID …. 23Primary 2669 …. …. Partial View of the ILUTE System State, Time T

ILUTE Tour-based models The only way to understand the relationship of one trip to another (and how one trip decision conditions/influences another) is by adopting a tour- based approach. NHB trips start to make some sense, conditioning of mode choices (e.g., cars need to return to the driveway), etc. Tours, in turn can only be explained in a behavioural sense by framing them within an activity-based, activity pattern/scheduling framework (tours to do what/when?).

ILUTE Dynamics Activity patterns and tours are conditioned by the order/timing in which decisions are made about what to do when. “High priority” / “inflexible” activities tend to be scheduled first, thereby providing constraints on subsequent decisions/trips. Travel occurs in “real time”; decisions at each point in time determine what actually happens at that point in time and condition subsequent decisions/actions.

ILUTE Dynamics, cont’d Learning / adaptation Memory / history Social networks / social influence Habit / experimentation All are important in determining travel behaviour at any point in time.

ILUTE Activity-based process models The need to participate in activities is what drives travel demand: the primary benefit of travel is the accomplishment of out-of-home activities that generate income, utility, benefits, etc. Travel itself is an activity that requires time & money to undertake (like other activities), and, like other activities, generates (dis)utility in its execution.

ILUTE Implications Need a disaggregate, tour-based model of activity participation and travel. Implementation mechanisms: –Microsimulation –Agent / object based models

ILUTE Microsimulation “Micro” implies a highly disaggregated model: spatially socio-economically (representation of actors) representation of processes “Simulation” implies: numerical dynamic (time dimension explicit) stochastic end state is “evolved” rather than “solved for” t = t 0 Synthesis of Base Sample For t = t 0 Endogenous Changes to Sample during this  t Disaggregate Behavioral Model Behavior/System State at (t +  t) Exogenous Inputs this  t t = t +  t

ILUTE Object-Oriented, Agent-Based Models The Object-Oriented Programming (OOP) paradigm is the industry standard for modelling complex process. Objects have identity, state and behaviour. OOP ideal for microsimulation applications. OOP is more than a programming style, it provides a “language” for conceptualizing activity/travel models and a clear development process for moving from conceptualization to implementation.

ILUTE Agent-Based Modelling Person 1 AgendaSchedule Person 1 AgendaSchedule Household Dwelling Unit Zone Worker Job Firm Building Agenda Vehicle Agenda Schedule An intelligent object is an agent. (“an object with attitude” – Paul Waddell). Agents: perceive the world around them make autonomous decisions act into the world Agents provide an efficient, highly extensible framework for modelling human socio-economic activity.

ILUTE Modelling Daily Activity & Travel Many “activity-based” travel models currently exist worldwide. These can be loosely divided into two primary types: Tour-based models Activity-scheduling models

ILUTE Tour-based models Tour-based models are the most common form of currently operational models. Key characteristics: –Focus on predicting the most common forms of daily tours. –Heavy reliance on deeply nested logit models (RUM). –Disaggregate, microsimulation based –Generally not truly agent-based (or truly activity-based).

ILUTE Tour-based models cont’d Several examples exist in the US and elsewhere. Originally derived from the seminal work of Bowman & Ben- Akiva in Portland (although the Portland model currently is not in operational use). Operational or near-operational models include: –San Francisco –Columbus, Ohio –New York City (NYMTC) (trip-, not tour-based, but microsimulation) –Atlanta –Denver –Los Angeles (under development)

ILUTE Activity-scheduling models This class of models focuses on predicting out-of-home activities and the associated travel required to execute these activities. More truly “activity-based”. Tours emerge out of the scheduling of activities. Typically (but not always) explicitly agent-based. Microsimulation-based. A variety of modelling methods used (RUM, rule-based, etc.). Typically quasi-operational (used in various policy studies but not yet generally mainstream operational).

ILUTE Activity-scheduling models, cont’d Examples include: –ALBATROSS (Arentze & Timmermans, The Netherlands) –TASHA (Roorda & Miller, Canada) –PCATS (Kitamura, Japan) –FAMOS (Pendyala, USA) –CEMDAP (Bhat, USA)

ILUTE TASHA TASHA (Travel/Activity Scheduler for Household Agents) has been developed at the University of Toronto. A validated version of the model is now operational. It is an activity-based, agent-based, microsimulation model of weekday activity/travel in the Greater Toronto-Hamilton Area (GTHA). Key features include: Household-based Activity scheduling Treatment of tours and modes Treatment of time Flexibility in development and application

ILUTE Key Features 1: Household-Based World Households Episode Distributions Spatial Representation Persons Person Projects Person Project Agenda Individual Activity Episodes Person Schedule Household Project Agenda Joint Activity Episodes Zones Distance Matrix Travel Time Matrices Household Projects Travel Episodes Individual & Joint Activity Episodes Persons exist within households. This allows TASHA to deal explicitly with: Vehicle allocation Ridesharing Joint activities/trips Serve-dependent activities/trips

ILUTE Vehicle Allocation within TASHA TASHA assigns household vehicles to drivers based on overall household utility derived from the vehicle usage. Drivers not allocated a car must take their second-best mode of travel.

ILUTE Household Ridesharing Options in TASHA Within-household ridesharing is explicitly handled within TASHA. Drivers will “offer” rides to household members if a net gain in household utility is obtained and feasibility criteria are met.

ILUTE Joint Activities …. Day n Person 1 …. Day n Person 2 Joint Shopping Activity: Duration: 2 hrs Location: The Mall Search for feasible joint time slot

ILUTE Serve Dependents Daycare At-Home Child’s Schedule At-Home Adult 1 Schedule At-Home Adult 2 Schedule Work Shopping Take child to/from daycare

ILUTE Key Features 2: Activity Scheduling Project 1 episode 1.1 episode 1.2 …. Project 2 episode 2.1 episode 2.2 …. Project N episode N.1 episode N.2 …. Day 1Day 2Day 3Day 4Day 5Day 6Day 7 … TASHA is an activity scheduling model in which individual activity episodes are generated and then explicitly scheduled. Out-of-home activity patterns and their associated trip-chains (tours) are thus “built from scratch” rather than selected from a pre-specified set of feasible patterns. Thus, travel patterns dynamically adjust to changes in transportation level of service, activity system “supply”, changes in household and personal constraints and needs, etc.

ILUTE PDF Activity Frequency Activity Frequency Joint PDF Start Time Feasible Start Times Start Time Joint PDF Duration Feasible Durations (a) Draw activity frequency from marginal PDF (b) Draw activity start time from feasible region in joint PDF (c) Draw activity Duration from feasible region in joint PDF Activity Episode Frequency, Start Time and Duration Generation At – Home Work Shop 1Shop 2 Other Work Project School Project Other Project Shopping Project Shop 1At-homeOtherShop 2 Person Schedule = “Gap” in Project Agenda= Activity Episode= Travel Episode At-home : : Scheduling Activity Episodes into a Daily Schedule TASHA generates the number of activity episodes from a set of “projects” that a person (or household) might engage in during a typical weekday. It also generates the desired start time and duration of each episode. It then builds each person’s daily schedule, adjusting start times and durations to ensure feasibility. Travel episodes are inserted as part of the scheduling process.

Key Feature 3: Tour-Based Mode Choice Chain c: 1. Home-Work 2. Work-Lunch 3. Lunch-Meeting 4. Meeting-Work 5. Work-Home m1 m2 m3 m4 m5 Non-drive option for Chain c m1 = drive Sub-Chain s: 2. Work-Lunch 3. Lunch-Meeting 4. Meeting-Work m2 m3 m4 Non-drive for Sub-chain s m2 = drive m3 = drive m4 = drive Drive for Sub-chain s m5 = drive Drive Option for Chain c mN = mode chosen for trip N TASHA’s tour-based mode choice model: Handles arbitrarily complex tours and sub-tours. without needing to pre-specify the tours Dynamically determine feasible combinations of modes available to use on tours. Modes can be added without changing the model structure. Cars automatically are used on all trips of a drive tour.

ILUTE Key Feature 4: Treatment of Time Models all out-of-home activities and trips for a 24-hour typical weekday 5 minute time increments are used for start times and durations/travel times –Provides great temporal detail but is computationally very efficient (integer storage & calculations) Trips can be aggregated to whatever level of temporal detail/categorization is required by the network assignment model Deals naturally with “peak-spreading”, etc. Provides excellent detail for environmental impact analysis

ILUTE Key Feature 5: Flexibility TASHA has been designed to be very flexible in terms of its development and its application. It has been developed using ordinary trip-based survey data for the GTA (but it could also exploit activity-based survey data). It can be used as a direct replacement for the first 3 stages in a 4- step system, or integrated within a full microsimulation model system. The data requirements for model development are no greater than other current models, including conventional trip-based models. Usable in a variety of contexts, and facilitates the evolution of the model system over time from aggregate to microsimulation.

ILUTE Application in a conventional setting Pop & Emp by zone Synthesize persons, hhlds & work/school locations TASHA Standard zone-based, static road & transit assignment Standard 4-step zone-based inputs Standard network assignment package TASHA contains its own synthesis procedures to convert aggregate, zone-based inputs into disaggregated persons, etc. required for microsimulation

ILUTE Base Year Census Data, Other Aggregate Data Synthesize Base Year Population, Employment, Dwellings, etc. ILUTE Evolutionary Engine For T = T0+1,T0+NT do: Demographic Update Building Stock Update Residential Housing Commercial Floorspace Firm/Job Location Update Household Composition Update Work/school Participation & Location Update Residential Location Update Auto Ownership Update Exogenous Inputs, Time T In-migration Policy changes … Dynamic Network Assignment Model (meso- or micro-scopic) T0 = Base time point T = Current time point being simulated NT= Number of simulation time steps Travel Models Commercial Vehicle Movement Update Activity/Travel Update (TASHA) Application in an full microsimulation setting

ILUTE Current Status TASHA was developed using 1996 travel survey data for the GTHA. The activity scheduler has been validated against 2001 survey data. Interfaces with both EMME and MATSIM.

ILUTE 2001 Validation Results

ILUTE Current status, cont’d TASHA has been used for several environmentally related studies in the GTHA. Currently it is being used to predict transportation energy use for the West Don Lands development in the Toronto Waterfront. Currently being tested for transferability to Montreal and London, UK. A joint UofT – Georgia Tech study is using TASHA to investigate the statistical properties of large microsimulation models.

ILUTE In a “Business as Usual” scenario with respect to GTA growth and transit system investment, auto usage is projected to grow faster than population; transit usage will grow at about half the rate of population. Pop. Growth Rate TASHA Application: GTHA Growth & Transportation Impacts

ILUTE Environmental Modeling with TASHA TASHA has been connected with: –EMME/2 road & transit network assignment model (link speeds & volumes by hour of day) –MOBILE6.2C emissions model (link emissions by type by link by time of day) –CALMET meteorological model –CALPUFF dispersion model (pollutant concentrations by zone by time of day) Dynamic population exposure to pollution by zone by time of day.

ILUTE Persons & Households Auto & Transit Travel Times/Costs TASHA Activity/Travel Scheduler Activity Patterns & Trip Chains Trips By Mode, Vehicle Type & Time of Day Transportation Network Model VKT by Facility Type, etc. Hot/Cold Soaks, Cold Starts, etc. Emissions Model Mobile Source Emissions Dispersion Model Locations of People by Time of Day Exposure to Pollution Household Auto Ownership Model Vehicle Allocation Model Land Use Policies Vehicle Technology Transportation Policies (Road pricing, carbon taxes, transit investment, etc.) EXAMPLE INTERVENTIONS

ILUTE Auto Emissions by location and time of day Link-based running emissions by time of day Zone-based soak emissions by time of day

ILUTE Dispersion of Emission Concentrations

ILUTE Zone NO2 Exposures

ILUTE TASHA-MATSIM More recently TASHA has been linked with MATSIM, an agent- based micro/meso-scopic network simulator. MATSIM allows us to keep track of individual agents as they travel through the network so we can accumulate their emissions (and, eventually, their exposure to pollutants). It also provides us with rudimentary vehicle dynamics, allowing a more detailed calculation of vehicle emissions.

ILUTE Emissions by hour ILUTE

Emissions by hour ILUTE

Evolution of idle emissions ILUTE

Lessons Learned Microsimulation simplifies many calculations, since one has a much more precise representation of the problem being modeled. The household-based approach also both simplifies and facilitates many calculations: –Auto-availability, auto allocation –Ridesharing –Joint travel –Serve-dependents

ILUTE Lessons Learned, cont’d: Tour-based models provide the “proper” framework for: –Handling non-home-based travel –Modeling travel across time periods (peak / off-peak) –Consistent handling of auto usage Activity-based, microsimulation is ideal for environmental impact modeling

ILUTE Lessons Learned, cont’d These gains in model performance can come at very little additional complexity, data requirements or computational burden. 4-step models are complex, messy and computationally burdensome – we are just used to them! Agent-based microsimulation models are easier to explain to policy-makers and are both behaviourally and computationally “cleaner” in their structure.

ILUTE Lessons Learned, cont’d Activity-based models, however, do not “automatically solve” all problems: –Activity/trip generation still typically crude –Non-work/school destination choice a weak link in all model systems, trip- or activity-based –Network assignment is still typically computationally burdensome for 24-hour modeling BUT, the activity-based paradigm provides a much more suitable framework for addressing these issues than the 4-step approach!

ILUTE THANK YOU!