|
" Study of Low Discrepancy Parameter Sweep in Pedestrian Dynamics Modeling "
Islam, Tasvirul
Srinivasan, Ashok AS
| Document Type
|
:
|
Latin Dissertation
|
| Language of Document
|
:
|
English
|
| Record Number
|
:
|
1057059
|
| Doc. No
|
:
|
TL56176
|
| Main Entry
|
:
|
Islam, Tasvirul
|
| Title & Author
|
:
|
Study of Low Discrepancy Parameter Sweep in Pedestrian Dynamics Modeling\ Islam, TasvirulSrinivasan, Ashok AS
|
| College
|
:
|
The University of West Florida
|
| Date
|
:
|
2020
|
| Degree
|
:
|
M.S.
|
| student score
|
:
|
2020
|
| Note
|
:
|
55 p.
|
| Abstract
|
:
|
Pedestrian dynamics models are used to simulate the walking movement of people. A multi-dimensional parameter sweep in these simulations can account for the variability in human movement patterns. A conventional lattice-based parameter sweep suffers from the following limitations that leads to a large computational effort. (i) it does not explore the parameter space efficiently, and (ii) it leads to inefficient convergence checks in high dimensions, which are required for robust scientific results. Low discrepancy parameter sequences (LDS) have asymptotically better uniformity than a lattice and overcome both the above limitations. However, they have the following limitations: (i) a parallelization leads to significant load imbalance, and (ii) the convergence rate worsens with dimension. In this thesis, we examine whether pseudorandom sequences can overcome these defects. While pseudorandom sequences have asymptotically worse convergence rates, the rate is independent of dimension. Consequently, it is possible that they have lower error for realistic sample sizes. In addition, we explore the potential of (i) a hybrid parameter sweep, where a small number of important dimensions use LDS while the other dimensions use a pseudorandom sequence and (ii) a parallel LDS, which can potentially yield good load balance. We also examine whether the convergence criteria used in previous work are necessary. Prior work shows that the existing convergence criteria are sufficient for accurate results. However, if they are not necessary, then one may perform a smaller number of simulations to obtain accurate results. Our empirical results show that while a parallel LDS has low load imbalance, its convergence rate is too poor to be beneficial. Hybrid and pseudorandom parameter sweep, on the other hand, show moderately worse convergence than LDS for high accuracies, but good load balance. We also find that the prior convergence criteria are too strict, with two of the four criteria not being necessary. We use these insights to evaluate efficiently new boarding procedures introduced in airplanes during the COVID-19 pandemic. We show that these procedures actually increase infection risk. Our results promise to help a variety of other applications that employ large parameter sweeps.
|
| Descriptor
|
:
|
Computer science
|
|
|
:
|
COVID-19
|
| Added Entry
|
:
|
Srinivasan, Ashok AS
|
| Added Entry
|
:
|
The University of West Florida
|
| |