000			03459nam a22003137a 4500
008			260608t2023 bx a|||g |||| 00| 0 eng d
020			_qhardback
040			_aUniversiti Teknologi Brunei _beng _cUTB
084			_aUTB 120 REPORT, THESIS & DISSERTATION _aRTDS 410
100	1		_aNurulhidayati Haji Mohd Sani _eauthor.
245	1	0	_aA study of feedback signal calibration in reinforcement learning with sub-goals / _cNurulhidayati Haji Mohd Sani
260			_aBandar Seri Begawan : _bUniversiti Teknologi Brunei, _c2023.
300			_axvi, 148 pages : _billustrations ; _c30 cm.
500			_aSubmitted in fulfillment of the requirements for the degree of Doctor of Philosophy.
500			_aABSTRACT Reinforcement learning (RL) is a machine learning technique that allows intelligent agents to learn a new task without being explicitly supervised. RL agents learn to perform a new task based on reinforcement signals. Traditional RL enumerates possible states (S) and associates actions (policy) to each state. The main disadvantage of traditional RL is that the process is slower in a larger and more complex environment. Also, many decision-making processes are involved in a complex environment instead of focusing on just one goal. In contrast to the traditional position-based approach, we investigate a visual-based Q-learning agent that uses the projection of rays to perceive its environ-ment. Despite the increasing number of possible state value inputs due to the number of angles between rays that the agent can perceive, or the increasing number of objects in the environment when this approach is used, it allows the agent greater flexibility in reusing its strategy in different environments. This flexibility is very useful, especially in a real-world application where the environment is known to be very dynamic. Our preliminary study allows our agent to use visual perception to navigate different environment sizes and settings. In our thesis, we also studied a Q-learning agent in a navigation problem with sub-goals using amplified feedback signals to determine the most effective strategies for amplification signals to solve the problem. We investigated these signals using two different problem configurations: sequential sub-goals and non-sequential sub-goals. In the problem with sequential sub-goals, the agent is forced to reach the goal in a specific order to achieve the goal. In the problem with non-sequential sub-goals, the order of the goal is irrelevant but necessary to achieve the optimal reward. The results show that although the agent can learn and achieve the goal in most of the feedback signals, having consistent, incremental rewards and immediate rewards contribute most to the agent's performance in achieving the goal.
502			_aDissertation (Doctor of Philosophy) - Universiti Teknologi Brunei (2032)
504			_aIncludes bibliographical references.
610		4	_aUniversiti Teknologi Brunei _vThesis
610		4	_aUniversiti Teknologi Brunei _vFinal Year Report
650		4	_aDissertation, Academic
650		4	_aThesis writing
650		4	_aDissertation Universiti Teknologi Brunei
650		4	_aComputing and Informatics
700	1		_aSomnuk Phon-Amnuaisuk, Prof _eadvisors.
700	1		_aThien Wan Au, Dr. _eadvisors.
710			_aUniversiti Teknologi Brunei
942			_2lc _cRTDS
998			_eReports, Thesis & Dissertation _s850582 : 002454 c.1 UTB _xUniversiti Teknologi Brunei
999			_c24218 _d24218