The impact of error-tolerance using customized DIJKSTRA on large scale parallel crowd simulations / Luuk Sterke
This Dissertation aims at exploring and quantifying the errors that occur in parallel crowd simulation when not implementing measures to prevent these errors. Consequently, the simulation should run faster, as less code needs to be executed. A fully functional crowd simulator has been developed a...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Published: |
2022
|
| Subjects: | |
| Online Access: | http://studentsrepo.um.edu.my/14439/1/Luuk_Sterke.pdf http://studentsrepo.um.edu.my/14439/2/Luuk_Sterke.pdf http://studentsrepo.um.edu.my/14439/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | This Dissertation aims at exploring and quantifying the errors that occur in parallel crowd
simulation when not implementing measures to prevent these errors. Consequently, the
simulation should run faster, as less code needs to be executed. A fully functional crowd
simulator has been developed and various features for measuring accuracy and performance
have been implemented. Part of the research was identifying the right metrics for measuring
these quantities. It turns out that a heat map in combination with waiting times, walking
times and local flow measuring entities together create a bijection between this data and
the simulation producing it. The impact of allowing errors on the simulator’s accuracy can
be measured quite well by comparing these statistics. The performance is then measured
using different internal stopwatches keeping track of the time needed to simulate a fixed
amount of agents. From many simulation environments and settings, it becomes clear
that significant speedups can be achieved using the proposed techniques. These speedups
of up to 15% are achieved at the expense of the simulator’s accuracy, where flow and
waiting times are off by single-digit percentages. Heat maps deviate more percentage-wise
because they are more sensitive to small changes. In one case, error tolerance was not
faster and less accurate since error tolerance was not applied on the scale of all agents
but on reducing internal overhead time. Allowing errors to examine many simulations
quickly and then perform precise simulations on just the promising ones are recommended.
Whether the errors observed are large or small is not conclusive as that depends on the
simulated event and because such conclusions are outside of this dissertation’s scope.
|
|---|