Acceleration of applying AI to open intelligent network using parallel simulation for RL training
Minha Lee, Hyunsung Cho, Hun-je Yeon, Sukhdeep Singh, Hoejoo Lee · 2022 IEEE Globecom Workshops (GC Wkshps) · 2022
The era of Beyond 5G and 6G networks will further accelerate the development of intelligent networks with network virtualization and the introduction of open networks. This is because, in the next-generation networks, AI-based network operation is essential to manage explosively increasing network devices and increasingly complex services. Network intelligence is built through several stages, among which the stage of training AI models until they guarantee plausible performance consumes the most time and resources of all stages. In this paper, we explain our experiments to shorten model training time and accomplish cost reduction. We build a training host compatible with O-RAN specifications and create pipelines for parallel simulation to train Reinforcement Learning (RL) models. For RL model training, when it is difficult to interact with the real environment, the model interacts with simulators that mimic the environment. Simulation parallelization is performed to reduce the time consumed in the step of collecting the experiences from the simulations. The parallel simulation requires more resources than sequential training procedures, but can be performed with less time for training. According to the public cloud usage rate policy, the usage time of resources has a greater impact on the total usage fee than the total amount of resources used. Therefore, it is important to find an optimal pipeline configuration that can reduce training time without increasing overall cost. In conclusion, we have confirmed the effect of reducing training time by up to 51% and cost up to 80% through experiments using sample scenarios.