Online Learning for Energy Efficient Navigation using Contextual Information

Abstract

Accurately predicting the energy consumption of road segments is an important topic in electric vehicles that might alleviate the range concerns if it is addressed properly. We employ a contextual combinatorial multi-armed bandit framework to learn the unknown parameters of an energy consumption model which is necessary for energy-efficient navigation. Four different agents: Thompson Sampling, Disjoint LinUCB, Hybrid LinUCB, and greedy algorithms are implemented to observe their performance. All experiments are conducted on the output of a Luxembourg SUMO traffic simulation. The main finding of this research is that contextual information such as speed and acceleration data contributes to better learning of parameters. Although the contextual combinatorial algorithms seem promising for addressing the energy-efficient shortest path problem, none of the agents achieve zero regret consistently which indicates that further improvements are necessary to obtain the desired results.