Exploring Heuristics Underlying Pedestrian Shopping Decision Processes

1
Exploring Heuristics Underlying Pedestrian
Shopping Decision Processes

• An application of gene expression programming

Ph.D. candidate Wei Zhu Professor Harry
Timmermans
2
Introduction

• Modeling pedestrian behavior has concentrated on
  individual level
• Decision processes only receive scant attention
• As the core of DDSS, are current models
  appropriate?
• Introducing a modeling platform, GEPAT
• Comparing models of go home decision

3
Random utility model

• Discrete choice models have been dominantly used
• Question 1 Too simple
• Only choice behavior is modeled, ignoring other
  mental activities such as information search,
  learning
• Question 2 Too complex
• Perfect knowledge about choice options is assumed
• Utility maximization is assumed
• Degree of appropriateness?

4
Heuristic model

• Simple decision rules
• E.g., one-reason decision, EBA, LEX, satificing
• Human rationality is bounded, bounded rationality
  theory
• Searching informationStopping searchDeciding by
  heuristics
• Degree of appropriateness?

5
Difficulties in heuristic model

• Implicit mental activities
• Test different models
• Structurally more complicated
• Get simultaneous solutions
• Irregular function landscape
• Effective, efficient numerical estimation
  algorithm

Bettman, 1979
6
The program--GEPAT

• Gene Expression Programming as an Adaptive
  Toolbox
• Gene expression programming (Candida Ferreira
  2001) as the core estimation algorithm
• Two features
• Get simultaneous solutions for inter-related
  functions
• Model complex systems through organizing simple
  building blocks

7
Genetic algorithm

• GA is a computational algorithm analogous to the
  biological evolutionary process
• It can search in a wide solutions space and find
  the good solution through exchanging information
  among solutions
• It has been proven powerful for problems which
  are nonlinear, non-deterministic, hard to be
  optimized by analytical algorithms

8
Get simultaneous solutions

• The chromosome structure in GEP
• Only one function can be estimated

-b2bbd-c
9
Get simultaneous solutions

• The chromosome structure in GEPAT
• Parallel functions can be estimated
  simultaneously.

10
Test different models

• Facilitate testing different models through
  organizing building blocks--processors
• Each processor is a simple information processing
  node (mental operator) in charge of a specific
  task

11
Parallel computing

• Message Passing Interface (MPI)
• Distribute computation by chromosome or record

Slave
12
Model comparison
Shall I go home?

• Go home decision
• Data Wang Fujing Street, Beijing, China, 2004
• Assumption The pedestrian thought about whether
  to go home at every stop.
• Observations 2741

Shall I go home?
Shall I go home?
13
Reason for going home

• Which are difficult to observe
• Using substitute factors
• Relative time
• Absolute time

14
Time estimation

• Estimate time based on spatial information
• Grid space
• Assumption
• Preference on types of the street
• Walking speed 1 m/s

15
Multinomial logit model

• Choice between shopping and going home

Go home
Shopping
16
Hard cut-off model

• Satisficing heuristic
• Lower and higher cut-offs for RT and AT

Go home
LCRT
HCRT
PNS
LCAT
HCAT
17
Soft cut-off model

• Heterogeneity, taste variation

LCMRT LCSDRT
HCMRT HCSDRT
PNS
LCMAT LCSDAT
HCMAT HCSDAT
18
Hybrid model

• When the decision is hard to be made, more
  complex rules are applied

19
Model calibrations
MNL MNL Hard Cut-off Hard Cut-off Soft Cut-off Soft Cut-off Hybrid Hybrid
P Value P Value P Value P Value
ß1 -0.007 LCRT 29.797 LCMRT 132.048 LCMRT 0.000
ß2 -0.008 - - LCSDRT 83.976 LCSDRT 327.290
ß3 -10.501 HCRT 674.966 HCMRT 676.000 HCMRT 676.992
- - - - HCSDRT 0.010 HCSDRT 0.010
- - LCAT 809.840 LCMAT 927.851 LCMAT 916.544
- - - - LCSDAT 87.422 LCSDAT 85.820
- - HCAT 1313.169 HCMAT 1305.591 HCMAT 1377.659
- - - - HCSDAT 104.161 HCSDAT 230.719
- - PhNS 0.308 PhNS 0.752 ß1 -0.047
- - - - - - ß2 0.000
- - - - - - ß3 -3.502
ML -1121.200 -1121.200 -1381.830 -1381.830 -1070.599 -1070.599 -1077.843 -1077.843
AIC 2248.400 2248.400 2773.660 2773.660 2159.199 2159.199 2177.687 2177.687
Sim 0.546 0.546 0.656 0.656 0.743 0.743 0.744 0.744
20
Discussion

• The satisficing heuristic fits the data better
  than the utility-maximizing rule, suggesting
  bounded rational behavior of pedestrians
• Introducing the soft cut-off model is appropriate
  and effective pedestrian behavior is
  heterogeneous
• Lower cut-offs, as the baseline of decision, are
  much more effective than high cut-offs in
  explaining data, suggesting that pedestrians
  rarely put themselves to the limit in practice

21
Future research

• Model other behaviors, e.g., direction choice,
  store patronage, environmental learning
• Compare models
• Improve GEPAT

22
Thank you

• Wei Zhu
• w.zhu_at_tue.nl
• Harry Timmermans
• h.j.p.timmermans_at_bwk.tue.nl