Modeling Destination Choice in MATSim Andreas Horni IVT ETH Zürich July 2011
Modeling Destination Choice in MATSim
Andreas Horni
IVTETHZürich
July 2011
Destination Choice in MATSim
2
initial demand
analysesexecution scoring
replanning
I. Search Method & Capacity Restraints II. Adding Unobserved Heterogeneity -> Adapted Search Method
III. Variability Analyses
IV. Model Estimation
3
I. Local Search in Our Coevolutionary System
day plansfixed and discretionary activities
travel time budget
relatively small set of locations per iteration step
time geography Hägerstrand
4
r = tbudget/2 * v
check all locationsttravel ≤ tbudget
→ choice set
check ∑ttravel ≤ tbudget
random choice
510 % ZH Scenario: 60K agents
6
I. Competition on the Activites Infrastructureload-dependent decrease of score reduces number of implausibly
overloaded facilities
0
5000
10000
15000
20000
25000
1 2 3 4
Load category
Vis
ito
rs it_0_config2/3
it500_config2
it500_config3
Load category1: 0 – 33 %2: 33 - 66 %3: 66 - 100 %4: > 100%10 % ZH Scenario: 60K agents
Realism
Stability of algorithm
7
II. Adding Unobserved Heterogeneity: Scoring Function
V + eimplicit
+ eexplicit
II. Adding Unobserved Heterogeneity: Search Space
II. Adding Unobserved Heterogeneity: Search Space
costs(location(emax))
estimate by distance
realized utilities
preprocess once for every person
emax– bttravel = 0
search space boundary dmax = …distance to loc with emax
dmax
10
shopping leisure
II. Adding Unobserved Heterogeneity: Results
ei
j
personi alternativej
seed seed Random draws from DCM
microsimulation results = random variables X
III. Variability Analysis
-> microsimulations are a sampling tool
estimation of parameters for X (=statistic) with random sampling
Results should be given as interval estimation
Standard error of sampling distribution= sampling error
# of runs
20 runs (goal: 30 for TRB)time, route, destination choice, 200 iterations
Comparison OVER runs -> var = random var
Scores
population level
average executed plans:184.36: avg0.166: s0.09: s in percent of avg
III. Variability Analysis
agent level
Aggregation reduces var (applied in different fieldse.g. filtering (moving average))
III. Variability Analysis
daily link volumes
Previous studies confirmed (...?)
Castiglione TAZ levelup to 6% std.dev.
III. Variability Analysis
Hourly link volumes
high!Var = f(spatial resolutionstemporal resolutionchoice dimensions)
Previous studies at lowerres levels or less choicedimensions!-> difficult to compare
high!
III. Variability Analysis
Large intra-run var
1. Large var through replanning modules (20% replanner)(iteration 201: 100% select best)
2. Not in equilibrium3. Utility plateau -> genetic drift
ideas?
-> future research
16
Modeling Temporal Variability in MATSim
intrapersonal (temporal) variability
UMATSim = V + e MATSim cross-sectional model(average working day)
Correlations!
drawing from
?
(best case)
17
Correlations
t
y
General rhythm of life
Avoidance behavior
+
18
model estimation
IV. Model Estimation
+ eexplicit – penaltycap
correlations tastescoefficients b