Application of Heterogeneous CPU-GPU Hardware for Large-Scale Agent-Based Mobility Simulations
![Thumbnail](/bitstream/handle/20.500.11850/601899/saprykin_phd_thesis.pdf.jpg?sequence=5&isAllowed=y)
Open access
Author
Date
2022Type
- Doctoral Thesis
ETH Bibliography
yes
Altmetrics
Abstract
The complexity of modern multi-modal transport systems increases along with the spatial scale of many high-density urban and metropolitan areas. These changes make the existing agent-based modelling tools impractical and lead to long runtimes. This work attempts to close this research gap by developing an agent-based GPU-enhanced mobility simulator (GEMSim) framework with a co-evolutionary learning loop. GEMSim incorporates a high-resolution, multi-modal and massively parallel GPU-accelerated mobility model capable of running large-scale scenarios with millions of agents in a matter of minutes. Unlike other, relatively simple and limited GPU-accelerated traffic models, GEMSim provides a modular and extensible foundation that can be used to run ever more complex future scenarios. The framework was extended to simulate public transit, on-demand coordinated fleets, and integrated Mobility-as-a-Service (MaaS) packages. The flexibility of framework’s architecture allowed it to adapt and efficiently run the same traffic model on heterogeneous CPU-GPU hardware with many-core CPUs. The results show that the developed GPU-accelerated traffic model runs up to 100 times faster than one of the existing state-of-the-art mobility simulators, while the whole learning loop runs more than 22 times faster. GEMSim was applied to quantify the uncertainties of simulation outputs when running scenarios with samples of population and downscaled infrastructure. Results show that a sample of at least 30% is recommended in scenarios with cars and public transit to accurately reproduce disaggregated outputs, and a sample of 100% is recommended for scenarios with on-demand coordinated fleets. These findings justify the need for improvements in the runtime performance of existing tools to be able to obtain accurate results. Furthermore, GEMSim was used in multiple case studies to evaluate the potential impacts of deploying coordinated fleets in cities. First, the replacement of cars with coordinated taxi fleets in the Zurich area (Switzerland) can reduce the car ownership at the cost of the increased traffic congestion due to longer total driven distances. Second, integrated MaaS packages with public transit and coordinated fleets in the Munich area (Germany) can attract substantial numbers of public transit riders living in the areas with a relatively high connectivity; car users, however, are unlikely to switch to MaaS without push and pull measures as no significant benefits in travel time are provided. Both scenarios were created and validated using a novel agent-based demand modelling pipeline, developed as part of this thesis. Finally, a GPU-accelerated visual analytics framework for large-scale and agent-based scenarios, Quartz, was developed and evaluated using GEMSim. Quartz is capable of performing spatio-temporal analytics on disaggregated simulation outputs with millions of agents, as well as rendering real-time visualizations of the agents’ dynamics. Show more
Permanent link
https://doi.org/10.3929/ethz-b-000601899Publication status
publishedExternal links
Search print copy at ETH Library
Contributors
Examiner: Abhari, Reza
Examiner: Cai, Wentong
Examiner: Klumpner, Hubert
Examiner: Chokani, Ndaona
Publisher
ETH ZurichSubject
agent-based model; mobility; GPU programming; daily-activity model; fleet; MaaS; traffic models; uncertainty quantification; visualizationOrganisational unit
03548 - Abhari, Reza S. / Abhari, Reza S.
02890 - Inst. of Science, Technology and Policy / Inst. of Science, Technology and Policy
Funding
ISTP-RI-02 - Swiss Mobility Lab (ETHZ)
More
Show all metadata
ETH Bibliography
yes
Altmetrics