A Review of Reinforcement Learning: A Tripartite Framework of Environment Design, Algorithmic Innovation, and Application Scenarios

Reinforcement learning is gradually shifting from a research paradigm dominated by games and simulations toward real complex scenarios with high safety requirements and high costs, such as energy systems, industrial control, robotics, medical decision-making, and AI assistants. However, existing surveys are mostly centered on algorithms or applications, lacking a systematic analysis of the interactions among environment modeling, algorithm selection, and real deployment constraints. This paper proposes a ternary collaborative framework composed of environment design, algorithmic innovation, and application scenarios, to systematically sort out the development path of reinforcement learning. From the environment design dimension, it focuses on analyzing the influence of feature distribution, reward mechanism, dynamic uncertainty, and scalability on learning stability and generalization ability; from the algorithm dimension, it compares value-based, policy-based, model-based, and Actor–Critic frameworks as well as representative methods in recent years, emphasizing the trade-off among stability, sample efficiency, and deployability; from the application dimension, it summarizes the common design patterns, evaluation metrics, and key challenges of reinforcement learning in fields such as games, robotics and autonomous driving, energy systems and smart grids, industrial process control, healthcare, and AI assistants. It further summarizes cross-domain problems such as high-dimensional decision-making, sparse and delayed feedback, parameter sensitivity, and long-term credit assignment, and discusses potential solution directions such as world models, safe and constrained reinforcement learning, offline reinforcement learning, and integration with large language models. This paper aims to provide a unified analytical perspective and practical reference for the engineering implementation of reinforcement learning in real complex systems, in order to support engineering implementation in real complex systems.