O-RAN AI/ML Workflow Implementation of Personalized Network Optimization via Reinforcement Learning

In this paper, we study AI-based RAN technology for 5G and 6G networks that are more complex and difficult to analyze than previous generations to make the network more intelligent. We implement a reference AI/ML workflow using RAN Intelligent Controller (RIC) by referring to the AI/ML workflow architecture of O-RAN. We focus on the establishment of an online training environment based on the RIC platform. We use various open-source platforms to serve the ML model as an inference service and to build a Machine Learning (ML) training pipeline for online training. We train our own Reinforcement Learning (RL) model which controls function parameters in Distributed Unit (DU) to maximize total cell throughput. After training the model with data from a specific cell, it is deployed in a different environment. We demonstrate the effectiveness of our proposal by optimizing the model performance and executing the training pipeline for retraining the model using online workflow. As compared to the model before retraining, the total cell throughput has increased by 19.4% when controlled using the retrained model.