This review highlights key applications of reinforcement learning in dialogue policy optimization, including task-oriented systems, open-domain dialogue agents, and personalized interaction models. Additionally, the paper discusses critical challenges such as the design of effective reward functions, sample efficiency, and generalization across domains, as well as the ethical considerations surrounding biased language generation. By evaluating state-of-the-art approaches, such as deep Q-learning, proximal policy optimization (PPO), and partially observable Markov decision processes (POMDPs), this review outlines the progress made and identifies potential future directions, including transfer learning and human-in-the-loop RL for more scalable, real-world implementations.