Abu Bakar, Mohamad Hafiz (2023) Fusion sparse and shaping reward function in soft actor-critic deep reinforcement learning for mobile robot navigation. Masters thesis, Universiti Tun Hussein Onn Malaysia.

Preview	Text 24p MOHAMAD HAFIZ ABU BAKAR.pdf Download (493kB) | Preview
	Text (Copyright Declaration) MOHAMAD HAFIZ ABU BAKAR COPYRIGHT DECLARATION.pdf Restricted to Repository staff only Download (342kB) | Request a copy
	Text (Full Text) MOHAMAD HAFIZ ABU BAKAR WATERMARK.pdf Restricted to Registered users only Download (20MB) | Request a copy

Abstract

Nowadays, the progress in autonomous robots is being driven by the advancements in new technologies, particularly Deep Reinforcement Learning (DRL). DRL facilitates the autonomous navigation of robots by enabling them to interact with their environment and navigate automatically. Achieving accurate navigation is crucial, and the utilization of Soft Actor-Critic Deep Reinforcement Learning (SAC DRL) offers the most effective solution based on the principles of Reinforcement Learning (RL). However, certain weaknesses in SAC DRL have been identified, particularly in the exploration process for accurate learning with faster maturity. To address this issue, this research has designed and developed a solution based on an appropriate reward function to guide the learning process. Several types of reward functions based on sparse and shaping rewards in the SAC method have been proposed in this research. These include the reward function with angle correction (RFAC), the reward function without angle correction (RFWAC), the reward function without sparse reward (RFWSR), and the reward function without sparse reward and angle correction (RFWSRAC). These reward functions aim to investigate the effectiveness of mobile robot navigation learning. Through a series of experiments, the results demonstrate that the fusion of sparse and shaping rewards in the SAC DRL facilitates successful navigation of the robot to the target position, while also enhancing accuracy and maturity. Specifically, the incorporation of sparse rewards in the reward function leads to a significant improvement. The system with the sparse reward achieves the lowest average error of 4.989%, outperforming the system without sparse rewards, which exhibits the highest average error of 99.252%

Actions (login required)

View Item