碩士 / 國立雲林科技大學 / 電機工程系 / 107 / Autonomous robotic navigation in a complex environment has been one of the most practical applications in real life. To this end, this thesis proposes a feasible solution reflective of deep reinforcement learning that enables a robot with sufficient training to navigate a given space in an autonomous manner. Deep reinforcement learning is poised to advance the field of machine learning by relieving conventional techniques of formidable state explosion problems, i.e., settings with high-dimensional state and action spaces. As a step toward constructing autonomous systems with a higher-level understanding of the environment, this study aims to implement an Asynchronous Advantage Actor Critic (A3C) algorithm, allowing control policies for a robot to be learned directly from LiDAR inputs on site. A3C not only combines advantage updates with the actor-critic formulation but also involves asynchronously updated policy and value function networks trained in parallel over multiple threads. While stabilizing improvements in parameters, the use of multiple agents allows of more effective exploration to occur. Overall, this research consists of three major aspects: 1) The development of an A3C algorithm that governs updates to knowledge in response to dynamics in the sensed surroundings, 2) the development of software over the Robot Operating System for mobile robot obstacle avoidance until reaching the intended location, and 3) extensive experiments for finding how system performance varies under different numbers of agents, different entropy settings, and different numbers of neural network layers. Experimental results suggest that a model composed of 3 agents, setting entropy to 5, and 2 network layers best serve the purpose of autonomous navigation. Simulations indicate that the model achieves an accuracy of 92% for reaching target areas.
| Identifer | oai:union.ndltd.org:TW/107YUNT0441052 |
| Date | January 2019 |
| Creators | Liang-Yu Chang, 張良宇 |
| Contributors | Kuang-Hui Chi, 紀光輝 |
| Source Sets | National Digital Library of Theses and Dissertations in Taiwan |
| Language | zh-TW |
| Detected Language | English |
| Type | 學位論文 ; thesis |
| Format | 53 |
Page generated in 0.015 seconds