# Deep Reinforcement Learning for Distributed Dynamic Power Allocation in Wireless Networks

**Authors:** Yasar Sinan Nasir, Dongning Guo  
**arXiv:** [1808.00490](https://arxiv.org/abs/1808.00490)  
**Date:** 2018-08-01  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/1808.00490v1/

## Key points

- Authors Nasir and Guo explore deep reinforcement learning for power allocation in wireless networks
- Existing methods for power allocation are not easily scalable to real-world networks due to computational demands and need for instantaneous CSI
- Proposed model-free distributed dynamic power allocation scheme based on deep reinforcement learning using delayed CSI measurements
- Approach leverages deep Q-learning to handle random variations effectively
- Each transmitter adjusts transmit power based on gathered CSI and QoS data from neighboring nodes to maximize a weighted sum-rate utility function
- Objective can be tailored to prioritize maximum sum-rate or proportionally fair scheduling with time-varying weights
- Near-optimal power allocation achieved in real-time within typical network architecture
- Deep reinforcement learning-based radio resource management shows rapid performance improvements, especially in scenarios with inaccurate system models and significant CSI delay.

## Summary

In their paper titled "Deep Reinforcement Learning for Distributed Dynamic Power Allocation in Wireless Networks," authors Yasar Sinan Nasir and Dongning Guo explore the potential of deep reinforcement learning techniques for optimizing transmit power control in emerging and future wireless networks. The existing literature offers various methods to achieve near-optimal power allocations, typically through solving complex optimization problems. However, many of these algorithms are not easily scalable to large real-world networks due to their computational demands and the need for instantaneous cross-cell channel state information (CSI). To address these limitations, the authors propose a model-free distributed dynamic power allocation scheme based on deep reinforcement learning. This approach utilizes delayed CSI measurements available to the agents and leverages deep Q-learning to handle random variations effectively. By gathering CSI and quality of service (QoS) data from neighboring nodes, each transmitter adjusts its transmit power accordingly with the primary objective of maximizing a weighted sum-rate utility function that can be tailored to prioritize either maximum sum-rate or proportionally fair scheduling with time-varying weights. Through this approach, the system achieves near-optimal power allocation in real-time within a typical network architecture. This study demonstrates that deep reinforcement learning-based radio resource management can deliver rapid performance improvements, particularly in scenarios where system models are inaccurate and CSI delay cannot be ignored. Overall, Nasir and Guo's work highlights the promising capabilities of deep reinforcement learning techniques in enhancing radio resource management efficiency in wireless networks, paving the way for more adaptive and scalable solutions in future communication systems.