1. Location Choice Modeling for Shopping and Leisure Activities with MATSim: Status Update & Next Steps A. Horni IVT, ETH Zurich

2. …Next Steps Done … Local search based on time geography First validation steps Competition on activity infrastructure Disaggregation level of multi-agent models vs. data base General predictability of leisure activities f (person attributes) Estimate  Choice set generation (& F.P.?) Existence & Uniqueness of scheduling equilibrium (& D.C.?) Leisure: Integrate … - Social networks - Detailed psychological models Activity differentiation combined w/ random assignment Ring-shaped PPA (leisure) Shopping UTF extensions  arbitrary Further measures (e.g. link speeds ← GPS) TRB 08/09 (TRR)TRB 09/10? Computational issuesRealism of planning tool MATSim Theoretical fundament + realism of planning tool MATSim Intro

3. 3 Modify activity timing, routes and activity locations of agents‘ plans initial demand analysesexecutionscoring replanning Trip generation/attraction Trip distribution Location choice Location Choice in MATSim crucial! > 1 million facilities!

4. 4 Location Choice in MATSim: Local Search – WHY? Relaxed state (i.e. scheduling equilibrium … (not network eqilibrium (Wardrop I/II), Nash? ) Huge search space prohibitively large to be searched exhaustively or even worse by global random search Dimensions (LC): # (Shopping, Leisure) alternatives (facilities) # Agents + Time dimension → agent interactions Local search + escape local optima Existence and uniqueness of equilibrium

5. 5 Local Search in Our Coevolutionary System – HOW? Day plans Fixed and flexible activities Travel time budget Relatively small set of locations per iteration step Time Geography Hägerstrand

6. 6 10 % ZH Scenario: 60K agents

7. 7 Competition on the Activites Infrastructure Load-dependent decrease of score Reduces number of implausibly overloaded facilities Load category 1: 0 – 33 % 2: 33 - 66 % 3: 66 - 100 % 4: > 100% 10 % ZH Scenario: 60K agents Realism Stability of algorithm

8. 8 First Validation Steps Count data (avg. working day) Micro census (shopping and leisure) Starting point Larger volume of more disaggregated data necessary … - GPS - FCD - M Cumulus, Supercard, … - License plate - GSM - …

9. 9 Leisure location choice modeling – ring-shaped PPA Leisure travel <= models of social interaction and sophisticated utility function Not yet productive MATSim longterm goal First goal: model shopping location choice => Activity-based models (chains) → reasonable shopping location choice model requires sound leisure location choice modeling (aggregate level) trip generation/distribution → activity-based multi-agent framework Trip distance distribution MC → act chains (ring-shaped potential path area) Agent population Assignment of travel distances crucial and non-trivial for multi-agent models! Leisure Predictability of leisure travel based on f(agent attributes)? Leisure trip distance ↔ -desired leisure activity duration -working activity activity chains ← f(agent attributes)

10. 10 Utility Function Extension Consider potential for application/testing of estimated utility maximization models → hypothesis testing w/ data basis ≠ used for model estimation MATSim utility maximization framework Improve simulation results Store size Stores density SituationAlternativePerson

11. 11 Results – Avg. Trip Distances Config 0: base case Config 1: leisure PPA Config 2: + shopping activity differentiation (grocery – non-grocery; random assignment) Config 3.1: config 2 + store size Config 3.2: config 2 + stores density Shopping trips (car) Leisure trips (car)

12. 12 Results – Avg. Trip Durations Strong underestimation in general! -Missing intersection dynamics -Access to (coarse) network (parking lots etc) -Freight traffic essentially missing Shopping trips (car)

13. 13 Microcensus bin size ratio (bin 0 / bin 1 ) = 4.22 Config 0 bin size ratio (bin 0 / bin 1 ) = 19.41 Config 1 bin size ratio (bin 0 / bin 1 ) = 7.08 Config 2 bin size ratio (bin 0 / bin 1 ) = 7.00 Config 3.1 bin size ratio (bin 0 / bin 1 ) = 6.41 Config 3.2 bin size ratio (bin 0 / bin 1 ) = 6.44 Results – Shopping Trip Distance Distributions (Car)

14. Config 0 Config 1 Config 3.1 Results – Count Data – 18-19 h

15. 15 Results – Count Data – 24 h Config  daily [%] 0 1 2 3.1 3.2 Weighting by shopping traffic work: (#trips * trip length) ≈ 7 % (excl. back to home trips) Small effects  (i,j) [%] 23.82 0.07 0.46 0.45 -60.25 -36.43 -36.36 -35.90 -35.91 Works aggregated model No improvement w/ respect to spatial distribution of trips Retest: -... more disaggregated data! -... more stations (now 300 stations for CH) - … time dimension - … compare with variance(year) - … Reject hypothesis

16. Estimate (Shopping) Utility Function Parameters r ho = f(r observed ) ? Shopping round trips by car → mode, → chaining, … Choice set generation & F.P. dominance attributes r observed r ho cs real (t)  distance Model cs real ~ cs ho ?  = f(r ho ) r ho arbitrary →  i arbitrary

17. Estimate (Shopping) Utility Function Parameters ii Unawareness set cs real Awareness set = cs real (t –  t) Inept set (-) Bias? cs real (t) Where is the relevant cut? choice(t) Narayana and Markin 1975 Evoked set (+) Inert set (0) Survey(s) in 2010?

18. MATSim measures? Travel distance distribution Travel time distribution Link loads Winner-loser statistics (WU) Number of visitors of type xy … 18 Activity-based Demand Modeling Problem to solve Activity differentiation (shopping → grocery ↔ non-grocery) + random assignment Neglectable effect Facilities info Model InputOutput I input I model (+ I emergence ) I output ~ I input × (I model + I emergence ) no info! I output = level ×  Level: e.g. count data vs. avg. trip length The closer we look the larger the error (I output fixed) ? our hope! define level and  Necessary information (data) Research … I output = level ×  for MATSim Structure of data (variance of behav.) (explicable + random part) → reachable level and  in principle → range of solvable problems little info!

19. Activity-based Demand Modeling – e.g., Location Choice SSH Uni HB Same flow, different people Facilities information Errors at different levels Different flowComparison w/ aggregated models: Gravity models: trip length distribution → information about heterogeneity superior? Agent attributes information (e.g. income) Our hope: Reduction of error at „coarser“ level? „Averaging“ of local decisions and effects (traffic jams)?

20. Activity-based demand modeling Model quality (level × error) Data (volume and level) Aggregated models Disaggregated models GSM? Always superior? Saturation behavior? Is there error propagation and thus error accumulation in the chains?

21. Predictability of leisure activities Life path: Reducing leisure travel to a cross- sectional sample (e.g. 1 day)? Leisure behavior Constraints Possibilities ← Environment Person attributesUnobservable personal life path (friends etc.) Shopping behavior Descr. statistics Reduction of complexity (by statistics)? Integration of Social Networks and Detailed Psych. Models of Individuals Starting w/ combining MATSim with rule-based models etc.

24. 24 Results – Count Data – 24 h (i, j)  (i,j) [%]  dist (i,j) [%] 2, 3.10.46 2, 3.20.45 Car shopping trips Retest: -... more disaggregated data! -... more stations (now 300 stations for CH) - … time dimension - … compare with variance(year) - … H 0 General underestimation of traffic volume  dist = upper bound for reduction of error due to increased traffic volume (increased avg. distances) Utf. extensions productive → spatial distribution of trips Reject hypothesis No improvement w/ respect to spatial distribution of trips 0.62 0.39

25. Activity-based Demand Modeling – e.g., Initial Demand Census: Population h, w (chain anchors) Micro census: Chains (chain structure) Assignment of chains → population f(agent attributes)

26. Activity-based Demand Modeling – e.g., Initial Demand 7‘500‘000 chains Sample „inflating“? MC: 30‘000 chains representative sample (of persons and also chains?) 126 99 18  = … - level 1: 0 - level 0: 3+3 = 6 Real chain distribution Random assignment Initial demand: Assignment of chains → population f(agent attributes) Region 1Region 2 Missing information at level 0: Systematic part explicable by f(agent attributes) Random part observable but not explicable Underlying distribution? Interpolation?