Date of Graduation

5-2023

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Engineering (PhD)

Degree Level

Graduate

Department

Industrial Engineering

Advisor/Mentor

Ashlea Bennett Milburn

Committee Member

Haitao Liao

Second Committee Member

Chase Rainwater

Keywords

A*, Adaptive Algorithms, Canadian Traveler Problem, Path Planning, Stochastic, Uncertain Networks

Abstract

The primary objective of this research is to develop adaptive online algorithms for solving the Canadian Traveler Problem (CTP), which is a well-studied problem in the literature that has important applications in disaster scenarios. To this end, we propose two novel approaches, namely Maximum Likely Node (MLN) and Maximum Likely Path (MLP), to address the single-agent single-destination variant of the CTP. Our computational experiments demonstrate that the MLN and MLP algorithms together achieve new best-known solutions for 10,715 instances. In the context of disaster scenarios, the CTP can be extended to the multiple-agent multiple-destination variant, which we refer to as MAD-CTP. We propose two approaches, namely MAD-OMT and MAD-HOP, to solve this variant. We evaluate the performance of these algorithms on Delaunay and Euclidean graphs of varying sizes, ranging from 20 nodes with 49 edges to 500 nodes with 1500 edges. Our results demonstrate that MAD-HOP outperforms MAD-OMT by a considerable margin, achieving a replan time of under 9 seconds for all instances. Furthermore, we extend the existing state-of-the-art algorithm, UCT, which was previously shown by Eyerich et al. (2010) to be effective for solving the single-source single-destination variant of the CTP, to address the MAD-CTP problem. We compare the performance of UCT and MAD-HOP on a range of instances, and our results indicate that MAD-HOP offers better performance than UCT on most instances. In addition, UCT exhibited a very high replan time of around 10 minutes. The inferior results of UCT may be attributed to the number of rollouts used in the experiments but increasing the number of rollouts did not conclusively demonstrate whether UCT could outperform MAD-HOP. This may be due to the benefits obtained from using multiple agents, as MAD-HOP appears to benefit to a greater extent than UCT when information is shared among agents.

Download

Share

COinS