Neuro-Symbolic AI for Improving Energy Efficiency in 6G (AI6G)

The goal of this project is to develop novel neuro-symbolic reinforcement learning (RL) algorithms that efficiently learn safe and transferable behavior. We will use the resulting policies to improve energy efficiency in 6G telecom networks.

Traditionally, artificial intelligence (AI) systems for sequential decision making were limited to rule-based systems that reasoned by manipulating explicitly represented knowledge in the form of symbols. Symbolic AI systems have the advantage that they are interpretable and can thus be deployed safely. However, symbolic AI usually faces the problem of combinatorial explosion and therefore fails to scale to complex real-world scenarios. In contrast, non-symbolic deep learning systems can efficiently learn representations from high-dimensional data and scale to large industrial applications. However, the deep neural network models have the drawback that their black-box nature makes them uninterpretable and thus unsafe to deploy for high-stakes applications.

In this project, we will develop neuro-symbolic AI algorithms for sequential decision making and thus combine the transparency and safety of symbolic AI with the scalability and flexibility of deep learning. With this hybrid approach, we will be able to efficiently reason about abstract concepts and make inferences that are beyond the capabilities of either approach alone. Consequently, our project contributes towards making neuro-symbolic AI become the new state of the art for sequential decision planning. It also helps to bridge the gap between the research fields of learning and reasoning.

The main practical target for this work is to improve the energy efficiency of telecom networks. This is important from both environmental and economic perspectives, and indeed flattening the curve of energy consumption is a key goal for the current fifth generation of mobile networks (5G) and is also a sustained objective that also applies to the upcoming sixth generation of mobile networks (6G). Since radio base stations (RBSs) are responsible for about 60-80% of the total network energy consumption, making them more energy efficient will impact the total mobile network energy efficiency. To decrease the RBS energy consumption, we use RBS sleep modes. Deeper sleep modes deactivate more parts of the RBS and thus consume less energy, but they also have longer latencies compared to shallower sleep modes.

PI: Jendrik Seipp

Core team: Kristina Levina, Aneta Vulgarakis Feljan, Athanasios Karapantelakis

Funding: This project is a collaboration with Ericsson Research and it is partially supported by the Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation.