Edge intelligence aims to bring artificial intelligence (AI) capabilities to edge computing environments, enabling smart services with low latency. However, realizing this vision requires optimizing computation offloading, resource allocation, and communication in dynamic conditions. Deep Reinforcement Learning (DRL) has emerged as a powerful approach to make intelligent decisions under uncertainty in edge networks. This survey highlights the role of DRL in optimizing computation, communication, and decision-making in edge intelligence systems, including scenarios with aerial access nodes (e.g., UAV base stations). We review how DRL techniques are applied to key challenges such as task offloading, resource management, and semantic-aware networking. Recent advances show that DRL agents can significantly improve latency, energy efficiency, and Quality of Service by learning policies that adapt to network states and user demands. We discuss fundamental DRL algorithms (DQN, PPO, DDPG, multi-agent methods) and their suitability for edge scenarios. We also analyze state-of-the-art works from semantic-aware networks and mobile edge computing domains, identifying their approaches and contributions. A comparative assessment of DRL methods is provided, including a table contrasting their stability, sample efficiency, scalability, action support (discrete/continuous), and multi-agent capabilities. Finally, we outline open issues and future research directions, such as hybrid model-based learning, explainable agents, real-world deployment challenges, and deeper integration of semantic understanding into edge intelligence.