# Reinforcement Learning for Dynamic Resource Optimization in 5G Radio Access Network Slicing

**Authors:** Yi Shi, Yalin E. Sagduyu, Tugba Erpek  
**arXiv:** [2009.06579](https://arxiv.org/abs/2009.06579)  
**Date:** 2020-09-14  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2009.06579v1/

## Key points

- The paper presents a novel solution to resource allocation in 5G radio access network slicing using reinforcement learning.
- Network slice requests arrive with specific requirements including priority, throughput, computational resource, and latency.
- Traditional myopic solutions become ineffective in the context of network slicing.
- The authors propose a Q-learning solution to maximize overall network utility while considering communication and computational constraints.
- Reinforcement learning provides significant improvements in 5G network utility compared to other solutions.
- It exhibits scalable performance even with an increasing number of served users and can effectively allocate resources when spectrum needs to be shared with incumbent users.
- The proposed approach enhances performance and scalability while accommodating dynamic requirements and constraints.

## Summary

The paper "Reinforcement Learning for Dynamic Resource Optimization in 5G Radio Access Network Slicing" presents a novel solution to the problem of resource allocation in 5G radio access network slicing using reinforcement learning. In this context, available communication resources such as frequency-time blocks and transmit powers, as well as computational resources like processor usage, need to be allocated to stochastic arrivals of network slice requests. Each network slice request arrives with specific requirements, including priority (weight), throughput, computational resource, and latency (deadline). If feasible, the request is served by allocating the available communication and computational resources over its requested duration. However, traditional myopic solutions that optimize only the current resource allocation become ineffective in the context of network slicing. To address this challenge, the authors propose a Q-learning solution that aims to maximize the overall network utility over a time horizon while considering communication and computational constraints. The network utility is measured in terms of the total weight of granted network slicing requests. The results demonstrate that reinforcement learning provides significant improvements in 5G network utility compared to myopic, random and first-come-first-served solutions. Moreover, it exhibits scalable performance even as the number of served users increases and can effectively allocate resources when spectrum needs to be shared with incumbent users who may dynamically occupy some frequency-time blocks. Overall, this paper contributes significantly to advancing resource allocation techniques in 5G radio access network slicing by leveraging reinforcement learning. The proposed approach offers enhanced performance and scalability while accommodating dynamic requirements and constraints imposed by different network slice requests.