Adaptive e-learning systems have become the backbone of personalized digital e-learning systems, which make use of data-driven intelligence and adjust personal learning experience. However, despite many advances in educational techniques, majority of adaptive systems rely on static heuristics, fixed rule sets and restricted personalization models. In this review, the increasing importance of Reinforcement Learning (RL) to become a disruptive education paradigm for adaptive e-learning is examined. By formulating learning as a sequential decision problem, RL enables the systems to make moment-by-moment adaptations to learners’ own performance, cognitive engagement, and mastery. Moreover, the combination of IRL and preference-based feedback allows for fairness-aware reward design that increases fairness considerations when designing adaptive systems. The paper reviews significant advancements in RL-based educational systems, including algorithmic advances such as the Deep Q-Network (DQN) and Proximal Policy Optimization (PPO), the knowledge graph-augmented state representation. It also makes the case for using explainable AI techniques to allow educators to make sense of, and have trust in model decision making. The review highlights several ongoing remaining challenges toward both fairness, interpretability and cross-domain generalization that hamper the practical scalability of current frameworks. Moreover, the paper describes a proposed architecture driven by RL that combines learner modelling, reward learning, fairness constraints and explainable feedback. Projected - Highlights Outcomes from these posited frameworks could potentially increase learner engagement rates, learning strings and STEM equity for many groups of learners, while lowering bias and increasing transparency.