Deep Reinforcement Learning Based Resource Allocation for D2D Communications Underlay Cellular Networks.

Area of Science:

• Wireless Communications
• Network Resource Management
• Artificial Intelligence

Background:

• Device-to-device (D2D) communications offer enhanced data rates and reduced latency in cellular networks.
• Resource allocation (RA) in underlay cellular networks with shared channels presents significant computational challenges.
• Optimizing RA for numerous D2D links over multiple time steps is often infeasible due to high system overhead.

Purpose of the Study:

• To design an efficient resource allocation (RA) scheme for device-to-device (D2D) communications within cellular networks.
• To address the complexity of optimal RA in scenarios with shared cellular channels and multiple D2D links.
• To maximize the sum of average effective throughput for both cellular and D2D links.

Main Methods:

• A sub-optimal resource allocation (RA) scheme utilizing multi-agent deep reinforcement learning (DRL) is proposed.
• Each D2D link is represented by an agent, learning collaboratively with shared device information (locations, resources).
• Agents learn in a staggered and cyclic manner to adapt to dynamic network conditions.

Main Results:

• The proposed DRL-based RA scheme enables prompt resource allocation for D2D devices.
• The scheme effectively handles dynamically changing network configurations, including device locations.
• The sub-optimal RA scheme demonstrates superior performance compared to existing methods, especially in high-density scenarios.

Conclusions:

• The multi-agent DRL approach provides an effective solution for resource allocation in D2D underlay cellular networks.
• The proposed scheme offers a practical and efficient method for managing resources in complex, dynamic wireless environments.
• Significant performance gains are observed with increasing device density, highlighting the scheme's scalability.