Scheduling and Power Control for Wireless Multicast Systems via Deep Reinforcement Learning.

Area of Science:

• Wireless communication networks
• Content-centric networking
• Machine learning applications

Background:

• Wireless multicasting leverages user request redundancy in content-centric networks.
• Power control and scheduling enhance multicast performance but traditional methods lack scalability.
• Existing model-based approaches struggle with large state spaces and dynamic system changes.

Purpose of the Study:

• To develop scalable power control and scheduling policies for wireless multicast systems using deep reinforcement learning.
• To address limitations of model-based approaches in large and dynamic wireless environments.
• To enable cross-layer optimization for improved network performance.

Main Methods:

• Utilized deep reinforcement learning with deep neural network function approximation for Q-function.
• Employed multi-timescale stochastic optimization to manage average power constraints.
• Modified learning algorithms to track time-varying system statistics and integrated queuing strategy learning.

Main Results:

• Demonstrated the ability to learn effective power control policies for large-scale wireless systems.
• Showcased the scalability and tracking capabilities of the proposed multi-timescale learning algorithms.
• Validated cross-layer optimization by simultaneously learning power control and queuing strategies.

Conclusions:

• Deep reinforcement learning offers a scalable solution for power control and scheduling in wireless multicast networks.
• The multi-timescale approach effectively handles dynamic system conditions and constraints.
• The framework is adaptable for general large state-space dynamical systems with multiple objectives.