Single-Agent Deep Reinforcement Learning

To optimize key factors of communication systems in centralized manner, we propose single-agent deep reinforcement learning framework.

Single-Agent Deep Reinforcement Learning (SADRL) serves as a foundational framework for centralized optimization in complex communication systems. In this paradigm, a single intelligent agent observes the entire state of the network and makes global decisions to maximize a unified objective, such as total network throughput, energy efficiency, or fairness among users.

Centralized Deep Reinforcement Learning architecture for global network optimization.

The primary advantage of SADRL is its ability to find globally optimal policies by considering the complex interdependencies between different network components. For instance, an SADRL agent can jointly optimize power control and subchannel allocation across multiple nodes. This holistic view eliminates the coordination overhead inherent in decentralized systems.

Our research in SADRL focuses on addressing the stability and efficiency of learning in high-dimensional environments. We investigate techniques to ensure stable convergence and efficient state representation, allowing the agent to find reliable solutions for large-scale network optimization problems. We also emphasize the integration of domain-specific knowledge into the reward function design.