1. ADAPTS: Agent-based Dynamic Activity Planning and Travel Scheduling Update on Model Development and Data Collection Joshua Auld CTS IGERT Seminar Presentation February 26, 2009

2. Overview  Accomplishments/IGERT Requirements  Introduction: Activity-Based Modeling  ADAPTS Framework (mostly complete)  Population Synthesis (complete)  Activity Generation (in progress)  Activity Scheduling (complete)  GPS Travel Survey / Activity Planning (in progress)

3. Update on Accomplishments and IGERT Requirements

4. IGERT Requirements  Requirements completed: –All coursework –Preliminary qualifications –Proposal defense (09/08) –International internship:  One month at University of Toronto  Work with Eric Miller, Matt Roorda, and others  One publication, advising on thesis proposal, future work  Remaining –Domestic internship –Finish dissertation

5. Potential Collaboration Opportunities  Kostas Goulias, UCSB  Ram Pendyala, ASU  Harry Timmerman, Eindhoven  All working on variations of dynamic activity based models or GPS data collection for models

6. Publications  Auld, J.A., A. Mohammadian and M.J. Roorda (2009). Implementation of a Scheduling Conflict Resolution Model in an Activity Scheduling System. Forthcoming in Transportation Research Record: Journal of the Transportation Research Board  Auld, J.A., A. Mohammadian and K. Wies (2009). Population Synthesis with Region-Level Control Variable Aggregation. Forthcoming in Journal of Transportation Engineering.  Auld, J.A., A. Mohammadian and S.T. Doherty (2009). Modeling Activity Conflict Resolution Strategies Using Scheduling Process Data. Forthcoming in Transportation Research Part A: Policy and Practice. (available online December 2008)  Auld, J. A., C. Williams, A. Mohammadian and P. Nelson (2009). An Automated GPS-Based Prompted Recall Survey With Learning Algorithms. Journal of Transportation Letters, 1 (1), 59-79  Auld, J.A., A. Mohammadian and S.T. Doherty (2008). Analysis of Activity Conflict Resolution Strategies. Transportation Research Record: Journal of the Transportation Research Board. 2054, 10-19

7. Presentations  AATT08 (10 th International Conference on Applications of Advanced Technology in Transportation) –Conflict resolution –Population Synthesis  TRB09 (88 th Annual Meeting of the TRB) –ADAPTS framework –Scheduling rules model  Transport Chicago –GPS Survey  UPCOMING: –TRB Planning Applications – Population Synthesis Forecasting –IATBR (potentially) – Dynamic Activity Planning –TRB 2010

8. Introduction

9. Need for Travel Demand Modeling  TDMs are used in many policy and planning analyses –Impacts of construction –Location/necessity of new construction –Congestion pricing –Impacts of other transportation demand policies:  HOV lanes  Telecommuting, flex-time shifts  Transit oriented development  Land-use policies 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Activity Scheduler 6.Survey Results

10. Why do we need travel demand model for ITA development? ITA Implementation and usage Changes in: travel planning, travel behavior (encourage rideshare, efficient trip planning, schedule optimization, the list goes on….) Changes in: travel demand, transportation network utilization Costs and benefits to society -Need to be evaluated (initially and on a continuing basis) - Essential in order for public/private implementation to succeed) How do we evaluate behavioral changes, travel demand changes and hence costs/benefits? ACTIVITY-BASED TRAVEL DEMAND MODEL! 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results

11. Activity based modeling  Use of activity-based modeling –Microsimulation models, which develop an activity schedule for modeled individuals –Usually at the household or individual level –Pattern of activities and travel explicitly developed for entire population  Can represent time very accurately –Time of day choice often a core model component  Have a behavioral basis –Can represent response to policy changes very well –Location choice, time of day choice, mode choice utility based –Explicitly captures trip chaining response  Currently lacking: –Representation of planning dynamics –Realistic activity planning –Integration with traffic simulation – usually done through feedback 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results

12. Issues in Activity-Based Modeling  Fixed order of priority of activities: –Activities added to schedule and attributes picked in fixed order –In other models: activities added in order of assumed priority –Does not match observations from data (Roorda et al. 2005)  Fixed order of attribute scheduling: –In ALBATROSS: Party > Duration > Time > Mode > Location –In other models: nesting structure fixed, calling order fixed –Again, does not match actual scheduling process  Scheduling planning dynamics –Order of decisions can impact subsequent decisions –Impulsive/unexpected events in simulation or scenarios –Currently, entire schedule generated then executed  May lead to erroneous results, especially with behavioral- based demand management strategies 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results

13. ADAPTS Framework

14. Framework - Introduction  ADAPTS scheduling process model: –Simulation of how activities are planned and scheduled –Extends concept of “planning horizon” to activity attributes –Time-of-day, location, mode, party composition  Fits within overall framework of activity-based microsimulation model –Constraints from long-term simulation (land-use model) –Combined with route choice and traffic simulation  Models being generated for Chicago region –Datasources: CHASE planning data, CMAP household travel survey, CMAP land-use database, Census 2000  Auld, J.A. and A. Mohammadian. ADAPTS: Agent-based Dynamic Activity Planning and Travel Scheduling Model – A Framework. Proceedings of the 88 th Annual Meeting of the Transportation Research Board (DVD), January 11-15, 2009, Washington, D.C. 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results

15. Overall Integrated Land-Use Transportation Model Framework Population Synthesis Home/Work Location choice Vehicle Ownership Vehicle Transaction Model Work/Home Change and Choice Model Household Composition Land Use Patterns Transportation System Household Long-Term Context Long-term Decision Making Short-term Simulation Activity/Travel Model Traffic Simulation 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results

16. Framework: ADAPTS model Current Focus Waiting on GPS data Mostly complete 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results

17. Decision Example: T1 Plan new activity -T time -T loc -T who-with -T mode T T time T loc T mode/who T time Plan time-of-day T loc Plan location T who = T mode Plan mode and who-with T execute TTTT T exec Execute Activity Simulation Time Schedule At Home Time: 12:00 AM – 8:00 AM Loc: Home Mode: None Work Time: 8:00 AM – 4:00 PM Loc: ? Mode: ? Shop Time: ? Loc: ? Mode: ? Shop Time: 4:00 – 5:00 Loc: Mall Mode: Auto ? ? Work Time: 8:00 AM – 4:00 PM Loc: HOME Mode: None 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results

18. Framework: Simulation Objects 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results Entity Methods GenerateActivity() AddActivity() RemoveActivity() SetPlanTimes() PlanStart() … PlanMode() ScheduleActivity() ResolveConflicts() isOccupied?() isTraveling?() World Attributes: Zonelist[1,2,…,Z] Time Methods: Run Simulation() Zone Attributes: ZoneData HHList[1,2,…,H]

19. Remaining work  Attribute planning order model –When to run each attribute sub-model –Need to collect planning data –GPS activity planning survey – starting soon  Time-of-day, mode choice, party composition, etc. –Model from CMAP travel data, GPS survey and other sources –Combination of model types, logit, decision tree, etc.  Incorporate traffic simulation – work with VISTA  Fit all models into overall framework 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results

20. Population Synthesis

21.  Presented at AATT09  Upcoming presentation at TRB Planning Applications Conference in Houston  Published: –Auld, J.A., A. Mohammadian and K. Wies (2009). Population Synthesis with Region-Level Control Variable Aggregation. Forthcoming in Journal of Transportation Engineering. 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results

22. Population Synthesis - overview  Generating Synthetic individuals for the simulated region –Using Census or HH survey data –Generate all individuals in region  GOAL: transfer joint distribution and sample household to small geographies (usually PUMS to Census Tract/BG –Detailed samples (joint-distributions) given at large geographies (PUMS) –Marginal distributions found at small geographies (CT/BG) –Want to transfer joint-distribution to small area then draw from samples  Two stages: –IPF: generate joint distribution across several control variables from sample –Selection: selecting households from sample data to build population  New features: –Marginal constraints in household selection –Customizable – no fixed geography/variables –Subregional control variable aggregation – combine infrequent marginal categories at subregion level –Built-in scenario evaluater/forecast tool 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results

23. Population Synthesis - IPF  IPF algorithm: –Iteratively update seed matrix to match one each control variables –Continue until convergence (or iteration limit) is reached –Assumption: Correlation structure (odds-ratios) remains the same for each zone in the region Updating Factors at each stage Continue until (F-1) < e (convergence threshold) 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results

24. Population Synthesis – HH selection  After fitting distribution for zone:  For each household in sample data –Calculate selection probability Ph –Determine if household to be added – marginal constraint –If added  Update number of households required for zone Nz  Update number of households of type (Mc) for zone –Continue until no more households needed 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results

25. Population Synthesis – Base Procedure Results  Base Procedure Validation:  Validation of category aggregation routine: 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results

26. Population Synthesis – SURE forecasting model  SURE marginal changes forecasting model: –System of linear regression equations –Related only through correlated error terms –Accounts for cross equation correlations –d(hh,pop,emp) -› dhhsize=1, dhhsize=2, etc. –Estimate change in hhsize and num workers categories

27. Population Synthesis - Forecasting  Use SURE model to estimate marginal changes:  Update marginals  Run popsyn with new marginals and base sample –Generates forecast population –Closest distribution to base sample that satisfies forecast marginals –Other categories (Race, Age, Income), can be adjusted through scenario analyzer 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results

28. Activity Generation

29. Activity Generation: Overview  First step in activity-travel simulation  Current focus of work – very preliminary  Generate activities randomly –Monte carlo simulation at each timestep –Drawn from probability distribution for each activity type  Example: 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results

30. Activity Generation: Correction Factors  Using observed generation rates gives incorrect results –Due to collisions (i.e. activity conflicts) –Activities split, postponed, deleted, etc.  Unobserved planned activity generation  Try to correct generation distributions through simulation: –f i * = S( i f i ), minimize (f i * - f i )  i  activity types – i f i approximates unobserved planned activity generation –Must be solved simultaneously  Example: mean-fitting technique, t = t-1 (  * t-1 /  ); 1 = 1.0 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results I = 1 I = 2.86 I = 4.03 I = 3.88

31. Activity Scheduling

32. Activity Scheduler  Rules for adding activities to the planned schedule  Conflicts arise due to random generation of activities/unexpected acts  Scheduler resolves conflicts to create feasible schedule  Activity Scheduling Combines: –Conflict Resolution Model –Scheduling Rules  Related publications: –Auld, J.A., A. Mohammadian and M.J. Roorda (2009). Implementation of a Scheduling Conflict Resolution Model in an Activity Scheduling System. Forthcoming in Transportation Research Record: Journal of the Transportation Research Board –Auld, J.A., A. Mohammadian and S.T. Doherty (2009). Modeling Activity Conflict Resolution Strategies Using Scheduling Process Data. Forthcoming in Transportation Research Part A: Policy and Practice. (available online December 2008) –Auld, J.A., A. Mohammadian and S.T. Doherty (2008). Analysis of Activity Conflict Resolution Strategies. Transportation Research Record: Journal of the Transportation Research Board. 2054, 10-19

33. Conflict Resolution - Introduction  Conflict resolution in previous models –Assumed priority rules –Simple heuristics –Not very realistic –Usually based on travel survey, NOT Process data  Conflict resolution in scheduling process data –Look at how conflicts actually resolved during scheduling –Empirical observations (Roorda et al. 2005, Ruiz et al. 2005) –Conflict resolution models (Ruiz and Timmermans 2006)  Based on actual scheduling data  Modify preplanned activities surrounding conflicting activity

34. Conflict Resolution Model  Due to dynamic nature of scheduling, conflicts naturally arise –Timing, location, resource  Conflict resolution model chooses strategy for resolving conflict –Currently only for timing –Uses decision trees –Strategies based on demographics, constraints, schedule characteristics, etc. Time

35. Conflict Resolution Model  Decision Tree model –Represent rule-based conflict solving –Evaluated using Exhaustive CHAID (Biggs et al. 1991) decision tree  Need to be Manually optimized  Discrete choice models –Utility-based conflict resolution solving –Multinomial logit –Nested logit (potential correlation for modify choices)  Dependent variable –Four resolution strategies: RS1-RS4 –Out-of-home and in-home modeled separate for all

36. Model Performance Comparison  Similar performance for all models –Approximately 73% correct predictions –Less accurate prediction of type 3 (modify both activities) resolutions  Typical problem in any classifier model due to low observation –Predict type 4 resolutions (delete original) well –26% improvement over null model

37. Conflict Model Discussion  In-home conflict resolution –Similar for decision tree and logit models –Travel requirements most highly significant –Duration, personal fixity, overlap in both –In DT model: conflict type significant –In logit models: time fixity, original duration  Out-of-home conflict resolution –Again, similar for both models –Plan horizon is most significant – preplanned more likely to be deleted –Other significant variables: conflict type, overlap, duration  In conclusion: –Activity, conflict and fixity attributes most important –Sociodemographic do not matter much –Similar to observations in other studies – Ruiz, Timmermans –Choice of model does not have much impact on outcome

38. Scheduling Rules - Overview  Set of rules for scheduling randomly generated activities  Attempts to resolve conflicts by modifying each activity – series of rules determine how modifications are made  System based on the scheduling rules found in TASHA model  Includes results of conflict resolution model: –TASHA – conflict resolution based on ad hoc logical rules –New rules – ad hoc logical rules determine how conflict resolution strategy is implemented –Possible resolutions for two activities in conflict: delete original activity, modify original, modify conflicting, modify both  New rules allow for the consideration of more complicated conflict types and deletion operations  When activities can be truncated, each activity assumed to be truncated proportionally to duration

39. Scheduling Rules – Comparison to TASHA

40. Scheduling Rules - Example  Under the TASHA rules:  i.Move Activity A, align end of Activity A with start of Activity B  ii.Move Activity B backward  iii.Truncate Activity A and Activity B proportionally to their durations  iv.Insertion is not feasible.  Under the new rules, situation handled as follows:  i.If resolution type is ‘Delete Original’  a.Remove Activity B from schedule, add Activity A  ii.If resolution type is ‘Modify Original’  a.Move Activity B, align start of Activity B with end of Activity A  b.Truncate Activity B  c.Insertion is not feasible  iii.If resolution type is ‘Modify Conflicting’  a.Move Activity A, align end of Activity A with start of Activity B  b.Truncate Activity A  c.Insertion is not feasible  iv.If resolution type is ‘Modify Both’  a.Move Activity A, align end of Activity A with start of Activity B  b.Move Activity B backward  c.Truncate Activity A and Activity B proportional to durations;  d.Insertion is not feasible.

41. Scheduling Rules - Validation  Actual CHASE activities scheduled with TASHA and ADAPTS  Compare results v. actual schedule with sequence alignment measure –Align schedules activity type by activity type –Weight insertion/deletion and move operations separately

42. Scheduling Rules - Validation

43. GPS Data Collection Update

44. Background: GPS-enabled surveys  Published in: –Auld, J. A., C. Williams, A. Mohammadian and P. Nelson (2009). An Automated GPS-Based Prompted Recall Survey With Learning Algorithms. Journal of Transportation Letters, 1 (1), 59-79  Currently focusing on replacing activity diary –Lower respondent burden –Capture more accurate trip/activity attributes –Longer range/panel studies –Gain more detailed information, esp. for route selection  Enhanced by technological progress –Person-based, wearable GPS loggers –Increased battery life –Differential / Assisted GPS 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results

45. New GPS Survey: Key features  Internet enabled and entirely automated –Participants upload data to central server –Survey completed on same day as data acquisition  Scans data to generate interactive PR survey –Utilize Google Maps API –Activity timeline  Participants validate activity/travel episodes  Survey activity-travel attributes –Who with, planning horizons, location choices, route and mode choice decisions  Incorporate learning algorithms to reduce survey burden –Suggest answers known with some confidence –Remove questions when answers known with high confidence –Proactively identify likely upcoming activities and prompt for planning data –Pre-populate planning items for learned recurrent activities 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results

46.  Designed to overcome issues regarding person- based tracking  Track all modes and indoor/outdoor activities  Activity-location finding: –Distance and time thresholds –Heuristics to determine threshold values –Distance threshold varies with land-use pattern, travel mode, etc. –Time threshold varies with travel mode Design of GPS survey: Activity location finding 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results

47. Demonstration: Activity-travel verification 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results

48. GPS Survey: Activity-travel prompted recall survey 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results

49. GPS Survey: Activity Patterns 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results

50. GPS Survey – Planning Results 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results

51. GPS Survey: Planning Order Results 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results

52. GPS Survey - Current Status  Pilot test results: –10 individuals for avg. of 6.8 days –210 total out-of-home activities (3.1 per day) –Avg. completion time: 23.5 min/day  Location finding algorithm works well –97% recall with 87% precision Design of survey captures dynamics of activity planning Verifying Activity-Travel Pattern  Total person-days 20  Average completion time 0:14:00  St. dev. of completion time 0:09:32 Answering Survey Questions (daily average)  Total person-days68  Activities per day3.9  Trips per day3.3  Acts & Trips per day7.2  Activity question answer time0:03:42  Trip question answer time 0:02:45  Completion time0:23:37 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results

53. GPS Survey – Next Steps  Survey starting next week –Currently recruiting participants/training survey assistants –Targeting 50 elderly, 50 non-elderly households –Attempt to collect 2 weeks of data per individual  Potential collaborations: –McMaster University – testing survey –University of Toronto GPS Survey 1.Intro 2.Framework 3.Population Synthesis 4.Activity Generation 5.Survey Results

54. The End Questions?

55. Conflict Resolution Decision Tree Mod. Orig. Mod. Conf. Mod. Both Delete Orig.

56. Discrete choice results Out-of-Home Conflicts

57. Discrete choice results In-Home Conflicts

58. Discrete choice model Fit statistics