Combinatorial Semi-Bandit Methods for Navigation of Electric Vehicles
Doktorsavhandling, 2024
Climate change is one of the most urgent global challenges humanity is currently facing. As major contributors of greenhouse gas emissions, the transport and automotive sectors have crucial roles to play in solving the problem. To reduce the usage of fossil fuels, electric vehicles need to become more attractive as alternatives to conventional vehicles. Concerns like range anxiety can be mitigated with more accurate navigation systems, especially if such systems are able to sequentially and adaptively collect data to improve their knowledge of the environment.
Hence, this thesis explores a number of different perspectives, settings and methods relating to navigation problems for electric vehicles in uncertain traffic environments. In particular, we focus on a combinatorial multi-armed bandit perspective, since it allows us to adapt and utilize efficient methods for targeted data collection within the navigation setting. Such methods include Bayesian bandit algorithms like Thompson sampling and BayesUCB, which can be used together with prior beliefs informed by domain-specific knowledge to efficiently explore the traffic environment while simultaneously solving the navigation problem.
Throughout the thesis, we apply these kinds of perspectives and methods to various problem settings, including both city-sized and country-sized road networks, relating to online versions of combinatorial optimization problems connected to navigation tasks. Within the appended works, we study the minimization of both expected energy consumption and travel time (including the time required for charging sessions). To show the efficiency of our proposed methods, we perform multiple thorough empirical studies with simulation experiments on realistic problem instances. We also analyze the methods by deriving theoretical upper bounds on their expected regret. With these performance guarantees and results, we aim to demonstrate the utility of the methods for real-world problems and applications.
Hence, this thesis explores a number of different perspectives, settings and methods relating to navigation problems for electric vehicles in uncertain traffic environments. In particular, we focus on a combinatorial multi-armed bandit perspective, since it allows us to adapt and utilize efficient methods for targeted data collection within the navigation setting. Such methods include Bayesian bandit algorithms like Thompson sampling and BayesUCB, which can be used together with prior beliefs informed by domain-specific knowledge to efficiently explore the traffic environment while simultaneously solving the navigation problem.
Throughout the thesis, we apply these kinds of perspectives and methods to various problem settings, including both city-sized and country-sized road networks, relating to online versions of combinatorial optimization problems connected to navigation tasks. Within the appended works, we study the minimization of both expected energy consumption and travel time (including the time required for charging sessions). To show the efficiency of our proposed methods, we perform multiple thorough empirical studies with simulation experiments on realistic problem instances. We also analyze the methods by deriving theoretical upper bounds on their expected regret. With these performance guarantees and results, we aim to demonstrate the utility of the methods for real-world problems and applications.
energy-efficient navigation
online learning
multi-armed bandit problem
Thompson sampling
combinatorial semi-bandit problem
Författare
Niklas Åkerblom
Chalmers, Data- och informationsteknik, Data Science och AI
Online Learning of Energy Consumption for Navigation of Electric Vehicles
Artificial Intelligence,;Vol. 317(2023)
Artikel i vetenskaplig tidskrift
Online Learning of Network Bottlenecks via Minimax Paths
Machine Learning,;Vol. 112(2023)p. 131-150
Artikel i vetenskaplig tidskrift
A Combinatorial Semi-Bandit Approach to Charging Station Selection for Electric Vehicles
Transactions on Machine Learning Research,;(2023)
Artikel i vetenskaplig tidskrift
Climate change caused by greenhouse gas emissions is one of the most important and difficult problems facing humanity. The automotive and transport sectors can help to improve the situation through a more rapid transition from vehicles using fossil fuels to, for example, battery-electric vehicles. Since batteries still have relatively low energy storage capacity, and charging might be slow, a good navigation system can be a cost-efficient way of improving trust in the technology, if it is robust and reliable. This, however, requires accurate information about the road network and the traffic environment. While such information can be collected by vehicles with sensors and cellular connections, there is still a risk that an insufficient amount of data is collected for infrequently visited parts of the road network, especially if the fleet of available vehicles is small.
This thesis proposes methods for selecting routes in an efficient and adaptive way, balancing between choosing the best routes according to information already possessed or choosing routes for the purpose of collecting useful new information. This is achieved through a class of machine learning methods, called combinatorial semi-bandit algorithms, specifically focused on complex decision-making problems in uncertain environments. The proposed methods are applied to both city-sized and country-sized road networks, under various assumptions, and their performance is demonstrated through empirical and theoretical results.
This thesis proposes methods for selecting routes in an efficient and adaptive way, balancing between choosing the best routes according to information already possessed or choosing routes for the purpose of collecting useful new information. This is achieved through a class of machine learning methods, called combinatorial semi-bandit algorithms, specifically focused on complex decision-making problems in uncertain environments. The proposed methods are applied to both city-sized and country-sized road networks, under various assumptions, and their performance is demonstrated through empirical and theoretical results.
EENE: Energieffektiv Navigering för Elfordon
FFI - Fordonsstrategisk forskning och innovation (2018-01937), 2019-01-01 -- 2022-12-31.
Styrkeområden
Transport
Ämneskategorier
Datavetenskap (datalogi)
ISBN
978-91-8103-006-8
Doktorsavhandlingar vid Chalmers tekniska högskola. Ny serie: 5464
Utgivare
Chalmers