I am trying to build models for Open Radio Access Network (O-RAN) use cases such as traffic steering, load balancing, and Quality of Service (QoS) prediction. My aim is to monitor network, cell, and User Equipment (UE) performance continuously and adapt policies using AI/ML algorithms. For real-time acquisition of Key Performance Indicators (KPIs) and sending back control inputs, what steps should I follow? I am presently utilizing NetSim but am open to suggestions for any other effective tool or simulation software. Thank you.
NetSim incorporates O-RU, O-DU, O-CU (CP and DP) functions within its PHY, MAC, and RLC layers.
You would need to use the messaging interface to send KPIs (or measurements) from NetSim to your AI/ML module; this should then send back the control actions to NetSim. NetSim will then execute these actions and send the next set of KPIs (or measurements)
Depending on the type of control you plan to use it could be near RT (>10ms, < 1s) or it could be non-RT (> 1s) control loop. Near RT control is akin to xApp, while non-real-time control resembles rApp.
Yes, traffic steering would be a good use case. For example: the UE-gNB association need not be just based on highest SINR but can be based on gNB loads (sometimes called load balancing) or can be based on service type (Voice or Video).
These could be useful discussions:
You may consider ns-O-RAN framework in NS3; see e.g. https://openrangym.com/tutorials/ns-o-ran
See also the IEEE article "Programmable and Customized Intelligence for Traffic Steering in 5G Networks Using Open RAN Architectures" at https://arxiv.org/pdf/2209.14171v3.pdf
Sai R N Thank you for your response.
I am thinking of using use an RL algorithm (like multi-armed bandit) for traffic steering. The idea is to couple MATLAB and NetSim, with the MAB algorithm running in MATLAB.
If I were to take a case of say 3 gNBs and 9 UEs, the each set of UE-gNB associations would be an arm. For each arm NetSim would return the per UE (or sum) throughput. Then the epsilon-greedy MAB algorithm would explore different associations and exploit the ones that have shown better throughput in the past. The challenge is that there are 9^3 associations which is time consuming to evaluate.
Would this work? Please let me know your suggestions.
These paper might help you:
1. Energy-Aware Dynamic VNF Splitting in O-RAN Using Deep Reinforcement Learning
2. Optimizing Virtual Network Function Splitting in Open-RAN Environments
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Programming Languages
- Python