This thesis explores the field of Multi-Objective Reinforcement Learning (MORL), a research area within AI and Data science, emphasizing the balancing of multiple, often conflicting objectives. Given the dynamic nature of real-world optimization problems, Reinforcement Learning (RL) approaches are given the difficult task to adapt to these changes to accurately address real-world objectives. However, this task is made more difficult by a lack of research surveying different MORL approaches. To alleviate this, this study aims to answer the research question: What MORL approaches with dynamic preferences and constraints are used in modern research, and what are their proposed use cases, strengths and weaknesses? This is achieved through a qualitative literature review to survey and analyze various MORL approaches that handle dynamic preferences and constraints.

Results from the thematic analysis identified themes of use cases, strengths, and weaknesses. The identified use cases span autonomous vehicle control, edge computing, and manufacturing. The strengths of MORL approaches include enhanced algorithm accuracy, optimized efficiency, operational enhancements, and dynamic adaptation. However, weaknesses such as real-world applicability, environmental assumptions, training difficulties, learning dynamics, and negative test results were also noted. Despite the implementation of dynamic preferences and constraints in all reviewed articles, many other codes were less frequently addressed.

The study concludes by highlighting the diverse application areas of MORL. The most frequently identified strengths were Effective Pareto frontier approximation and coverage, aligning with MORL's foundational aspects and measures of effectiveness, followed by a number of secondary strengths focusing on efficiency. Weaknesses were discussed briefly in only 11 articles, predominantly concerning real-world applicability, potentially due to bias or benchmarking approaches. No definitive conclusions could be drawn regarding the comparison between primary algorithms and their use cases, strengths, or weaknesses.