SORA: Energy-Efficient Resource Allocation in Open Radio Access Network With Online Learning

Open Radio Access Network architecture, which leverages virtualized base stations deployable on diverse commodity hardware, are set to enhance the flexibility and cost-effectiveness of next-generation cellular networks. However, their widespread adoption faces challenges due to the complexity of configuration choices, which can greatly affect both network performance and energy efficiency. Achieving the right balance requires smart resource allocation—optimizing Modulation and Coding Scheme selection and Physical Resource Block assignment to make the most of available spectrum and processing power. In this work, we propose a Thompson Sampling-based online learning framework for energy-efficient resource allocation in virtualized base stations. Our approach is designed to operate with minimal assumptions about the network environment, making it highly adaptable to both stationary and dynamic network conditions. By integrating variational inference with Thompson Sampling, the framework efficiently balances exploration and exploitation, allowing it to dynamically adjust MCS and PRB allocations in response to changing network states. The proposed algorithm achieves sub-linear regret, ensuring that it converges towards an optimal policy over time while maintaining robust adaptation capabilities. We validate our method using open-source platforms, OpenAirInterface and FlexRIC, demonstrating its practical effectiveness. Our results show that the proposed approach can achieve energy savings of up to 15 percent while improving resource block utilization efficiency by up to 31 percent compared to the default resource scheduling policy. These findings underscore the potential of the proposed method to significantly enhance the sustainability and efficiency of Open Radio Access Network deployments, paving the way for smarter, more adaptive next-generation cellular networks.