Destination Choice Modeling
of Discretionary Activities in
Transport Microsimulations
Andreas Horni
destination choice modeling for transport microsimulations
This Thesis
problem: implementation of a MATSim
destination choice module for shopping
and leisure activities efficiently applicable
for large-scale scenarios and easily
adoptable by other simulation models
contribute to microsimulation destination
choice modeling
•
efficiency and consistency
•
consistent and efficient computation
of quenched randomness
•
destination choice utility
function estimation
•
•
choice sets specification
analysis
•
results variability
•
analysis of temporal variability and
aggregation and variability
•
agent interactions
•
•
infrastructure competition modeling
CA cruising-for-parking simulation
Basic Procedure
instantiation
census
travel surveys
e.g., sociodemographcis
input
infrastructure data
estimation
e.g., network constraints,
opening hours
microsimulation core
Output
constraints
population
choice model
feedback
situation
(e.g. season, weather)
generalized
costs
network load
simulation
Umax (day chains)
Basic Procedure
(usually non-linear)
system of equations
microsimulation core
choice model

fixed point problem
(== UE)
feedback
network load simulation

initial
population
initial plans
initial
population
Initial plans
offsprings
offsprings
mutation
interaction
mutation
recombination
parents
execution
fitness
recombination
evaluation
parents
parent selection
execution
fitness
evaluation
replanning
replanning
scoring
survivor selection parent selection
scoring
survivor selection
optimized
population
species0
optimized
population
interaction
optimized
plans
agent0
agent1..n
species1..n
Co-Evolutionary algorithm
optimized
plans
MATSim
Destination Choice & Other Frameworks
search
space
space
TRANSIMS
hierarchical destination choice
(zone and intra-zonal choice)
draw from discrete choice model
ALBATROSS
various constraints
draw from decision trees
PCATS
time geography
draw from discrete choice model
MATSim Destination Choice Approaches
time-geographic space-time prisms
hollow prisms
time
distance
t1
destination
rin,out = f(act dur)
t0
origin
e
PPA
min (ctravel)
space
min (ctravel) with e - Dr < ctravel< e + Dr
ei
Unobserved Heterogeneity
discrete choice modeling:
MATSim:
adding heterogeneity: conceptually easy, full compatibility with DCM framework
but: technically tricky for large-scale application
Repeated Draws: Quenched Randomness
destinations
e00
e10
i
•
storing all eij
i,j ~ O(106) -> 4x1012Byte (4TByte)
eij
•
•
fixed initial random seed
freezing the generating order of eij
one additional random number
can destroy «quench»
enn
persons
personi
store seed ki
(actq)
alternativej
store seed kj
regenerate eij on the fly with
random seed f(ki,kj)
Search for Umax
U
global optimum
local optimum
ei,j
space
travel
disutility
exhaustive search → restrain search space
Search for Umax : Search Space Boundary
search space boundary dmax := ?
dmax := distance to destination with emax
A
A3 = π(4r)2 - 9πr2 = 7πr2
A2 = π(3r)2 - 4πr2 = 5πr2
A1 = π(2r)2 - πr2 = 3πr2
A0 = πr2
r
emax– bttravel = 0
pre-process once for
every person
approximate
by distance
realized
utilities with
Gumbel
distribution
Search for Umax in Search Space
work
shopping
home
tdeparture
tarrival
search space
Dijkstra forwards 1-n
Dijkstra backwards 1-n
approximation
probabilistic choice
exact calculation of tt for choice
Results 10% Zurich Scenario
70K agents
iteration: 10 days  5 minutes
shopping
link volumes
leisure
Conclusions
ZH scenario: 10 days  5 minutes (iteration)
but: module still needs to be faster for CH scenario
improve sampling, sample correction factor
more validation data with more degrees of freedom
procedure for quenched randomness important in all
iterative stochastic frameworks