Using Deep Reinforcement Learning Model to Design Sustainable Bicycle Mobility Infrastructure رسالة ماجستير

Abstract

Nowadays, the whole world is concerned with the increasing environmental issues, and many countries are working towards reducing the impact of humans on the environment by adopting various sustainable development strategies. One of the promoted actions to face this issue is encouraging the use of bicycles as the primary mean of transportation. If cycling becomes the primary mean of transportation, there will be a need for new and suitable routes and paths that suit the needs of the bicycles’ riders. In this thesis, we will tackle the problems and propose solutions to the issues that cyclists may face concerning the city’s topography (e.g., types of road, road surface, and their slope). This thesis proposes a solution that promotes using an AI agent that utilizes reinforcement learning and neural network to find the best path in a way that is customized by user preferences. We first presented the data collection process and how these data will be used in a readily available way by the agent. Then, we tested several reinforcement learning algorithms to find the most suit- able method to be used in our challenging scenario. We have also converted the map into a graph which represents the deep reinforcement learning environment, and converted each feature into a sub-reward in our complex reward system. Finally, we trained multiple reinforcement learning models. The results show that Dual Deep Q Network has the best outcome; we achieved 7500 cumulative rewards in less than 5 hours of training time, and our agent was able to design a route based on the end-user specification and overpass all the roads that do not meet the criteria.