A Survey of Deep Reinforcement Learning for Intelligent Decision-Making Systems

Abstract

Deep Reinforcement Learning (DRL) has become a strong architecture that integrates decision-making with the capabilities of Reinforcement Learning (RL) with the representational capacity of deep neural networks. DRL has been used to break new ground in robotics, autonomous systems, games, health, finance, and communication networks by allowing agents to acquire complex behaviors by taking in high-dimensional inputs and through sequentially with the environment. This survey gives an overview of the basic concepts of DRL, the core algorithms, including value-based, policy-based, actor-critic, model-based, and multi-agent methods, the training methods, optimization methods, and neural network architectures. The paper also examines practical uses with an emphasis on the transformative value of DRL in various fields. Major challenges, such as sample inefficiency, unstable training, poor generalization, sparse rewards, and safety concerns, as well as interpretability and scalability limitations are discussed as well. Lastly, the survey provides the open research directions and future trends, such as safe and explainable DRL, multi-agent coordination, model-based and edge-native learning, and integration with new AI technologies.