A self-driving system using soft actor-critic deep reinforcement learning | |
| Author | Witoon Wiphusitphunpol |
| Call Number | AIT Thesis no.DSAI-23-01 |
| Subject(s) | Reinforcement learning Automated vehicles--Data processing Machine learning |
| Note | A thesis submitted in partial fulfillment of the requirements for the degree of Master of Science in Data Science and Artificial Intelligence |
| Publisher | Asian Institute of Technology |
| Abstract | A self-driving vehicle agent using machine learning approach can be categorized into two categories, Imitation learning and Reinforcement learning. The former approach re quires large amount of driving data, whereas the latter does not require. Transferring the agent from a simulator to the real world can help speed up training and avoid dangerous situations when training in the real world. However, the agent that has been trained in simulator performs poorly in real-world task due to differences in representations of the environment. Nevertheless, in the self-driving vehicle case, some sensors can be reason ably simulated such as LIDAR. However, it is more expensive than a camera and adds more complexity from fusing multiple sensors’ data. Transferring self-driving vehicle agent trained in simulator to the real-world environment has been done before using Double Deep Q Network and applying semantic segmentation model to the observa tion of simulator and real-world environment so that they have similar representations, but it has never been done with continuous-action off-policy algorithm nor demonstrate the ability to turn at an intersection. This thesis aims to use Soft Actor-Critic, which is continuous-action, off-policy deep reinforcement learning algorithm, to train end-to end self-driving agents to drive and turn at an intersection in the simulator. It uses only one front-facing camera to perceive the environment. Later, it is transferred to the real-world environment with the help of semantic segmentation model that makes both environments look alike. Training the agent to turn is unsuccessful due to insufficient information in the observation, and the agent cannot remember the turning behavior to demonstrate during evaluation. The real-world evaluation is not entirely successful be cause the agent drives nearer to the edge of the road than in the simulator, but it can still drive without going out of a lane. It has been identified that the poorer quality of semantic segmentation images and receiving a slower input rate contributed to the lower performance. |
| Year | 2023 |
| Type | Thesis |
| School | School of Engineering and Technology |
| Department | Department of Information and Communications Technologies (DICT) |
| Academic Program/FoS | Data Science and Artificial Intelligence (DSAI) |
| Chairperson(s) | Dailey, Matthew N.; |
| Examination Committee(s) | Mongkol Ekpanyapong;Chaklam Silpasuwanchai; |
| Scholarship Donor(s) | His Majesty the King’s Scholarships (Thailand); |
| Degree | Thesis (M. Sc.) - Asian Institute of Technology, 2023 |