Reinforcement Learning for Autonomous Navigation: A Case Study in Structured Environment
Name
Anne Ott
Abstract
Over the past years, the field of autonomous driving has known immense progress.
Solutions based on machine learning advancements have been used to speed up the
development of highly automated driving. Significant progress has been made in solving complex problems using a machine learning’s technique deep reinforcement learning. That field deals with agents navigating in an environment to maximize total reward by completing tasks. It has become a robust learning framework now capable of learning complicated behaviours in high dimensional environments. One of the essential aspects of reinforcement algorithms is to ensure safety, as any mistake can have life-threatening consequences.
This thesis aims to provide a standardized environment and an upgraded architecture for training reinforcement learning agents in a structured world. Algorithms based on Trust Region Policy Optimization and Proximal Policy Optimization are used for the training. The results are evaluated by how successfully the agent completes its goal and how well the safety of the agent and its surroundings is preserved. In addition, an evaluation of the proposed models is conducted by comparing them with baseline models.
Solutions based on machine learning advancements have been used to speed up the
development of highly automated driving. Significant progress has been made in solving complex problems using a machine learning’s technique deep reinforcement learning. That field deals with agents navigating in an environment to maximize total reward by completing tasks. It has become a robust learning framework now capable of learning complicated behaviours in high dimensional environments. One of the essential aspects of reinforcement algorithms is to ensure safety, as any mistake can have life-threatening consequences.
This thesis aims to provide a standardized environment and an upgraded architecture for training reinforcement learning agents in a structured world. Algorithms based on Trust Region Policy Optimization and Proximal Policy Optimization are used for the training. The results are evaluated by how successfully the agent completes its goal and how well the safety of the agent and its surroundings is preserved. In addition, an evaluation of the proposed models is conducted by comparing them with baseline models.
Graduation Thesis language
English
Graduation Thesis type
Master - Computer Science
Supervisor(s)
Amnir Hadachi, Shan Wu
Defence year
2021