The integration of Machine Learning (ML) in network modeling and simulations is key to evaluating ML-based solutions and algorithms used to configure and optimize networks. In addition data generated from simulations can be used to train and evaluate ML models, thus accelerating the design process and ensuring reliable comparisons with proposed solutions, whether they are based on ML or not.
The ability to access, manipulate, and process data has allowed network researchers to focus on many network optimization problems that were previously intractable due to complexity and scale. Solutions that make use of Machine Learning (ML) techniques are becoming increasingly popular. However, the integration and utilization of ML in current networking research and development workflows is still cumbersome. The goals of this project are two-fold. The first goal is to integrate ML and network simulations, so that ML-based algorithms and optimizations can be evaluated and developed with minimal overhead. The second goal is to enable the use of simulated (or synthetic) data to train and evaluate ML models before they can be applied in real networks.
The technical approach consists of two major components. The first one (illustration on the left) is the use of network simulators to generate synthetic datasets that can be used by the research community for training and evaluation of novel ML-based solutions. Hence, network simulations can be configured to generate data for different configurations and topologies, even when actual hardware is not yet available. The second component (illustration on the right) is the integration of ML-based solutions in network simulations in order to enable the development of new network functionalities.
We have defined a process for incorporating ML models available in major frameworks such as TensorFlow and Pytorch in ns-3 simulations without the need for re-implementing and re-training said models. Using this process, we have developed a proof of concept based on an architecture similar to what is proposed by the O-RAN Alliance. This includes a Radio Access Network (RAN) Intelligent Controller (RIC) that contains a data repository where all relevant simulation information reported by the simulation nodes is stored, and Logic Modules that act based on the information in the repository and generate commands to alter the network’s configuration and operational parameters.
The main benefits of this platform are to speed up the development and evaluation of ML algorithms in O-RAN-like network implementations in addition to supporting the generation of large datasets from different network configurations for training and evaluating ML models. Additional uses include the design of new scenarios and configurations, the validation of testbed setups and assumptions, and the evaluation with well-known reference data.