Multi-agent reinforcement learning for cooperative vehicle task offloading

The rise of software applications in vehicles has transformed the automotive industry, leading to significant advancements in vehicle functionality and user experience. These advancements, however, are not for free and come at the cost of increased computational requirements. Vehicles nowadays perform many tasks, from critical tasks such as the ones related to accident prevention and monitoring, to other high priority and low priority tasks. Being able to perform all tasks on a timely manner in a vehicle requires enough computational resources, which sometimes may not be readily available to provide fast response times. Offloading the tasks is a possible approach to solving this problem, but this requires a common system of communication and critical safety measures. Furthermore, leveraging the distribution of the tasks in real time is not trivial, since many factors such as vehicle to vehicle distance and speed need to be taken into account, and system performance can vary a lot depending on the scheduling policy chosen. Addressing this challenge, utilising a simulator to get real life traffic data, this master's thesis builds upon the AVE framework to present a novel method that leverages asynchronous Multi-Agent Deep Reinforcement Learning (MADRL) for efficient task scheduling among nearby vehicles, in order to achieve higher software responsiveness in a multi-agent system without the need of a centralised system.