Radio Communication with Artificial Intelligence (FunKI)

Subproject: Research into efficient receiver procedures for radio communication using methods of artificial intelligence

FunKI Logo

Initial Situation

Progressive digitalization is leading to a rapidly increasing number of networked devices and requires modern wireless communication systems that are both powerful and resource-efficient. Methods of artificial intelligence (AI) can make a significant contribution to this and are already successfully used on the higher protocol layers of 4G/5G systems. In contrast, good and practicable mathematical-abstract models for the physical layer (PHY) and for media access (MAC) could be used up to now under idealized assumptions, which enabled the development of efficient procedures and algorithms. However, the increasing complexity of systems with a growing number of devices, the growing demands on data rates and latency as well as the diversity of services to be provided (broadband, sporadic/very reliable machine communication, emergency call, telemedicine, IoT etc.) are pushing the previous model-based approaches to system and technology design to their limits. It therefore seems inevitable that new concepts are required

Problem Statement

The sometimes conflicting requirements in terms of reliability, transmission rate, number of terminals per area, latency, available radio resources, energy efficiency, complexity, and hardware costs are making implementation with traditional methods increasingly difficult. If, for example, mathematical modelling of the problem is no longer possible (model deficit) or leads to very complex models (algorithm deficit), data-driven machine learning (ML) methods represent a promising approach for gaining an understanding of the model or for system and technology design.

Within the framework of FunKI, a fundamental investigation of AI-driven technologies for radio communication is therefore pursued, which is oriented towards the 5th mobile radio generation (5G) and its further development. In particular, concepts for improved model understanding based on measurement campaigns are developed, AI-based methods for parameter estimation are developed, data-driven optimization of transmission and reception procedures is carried out and AI-based training of hardware implementations is carried out. 


The overall objective of FunKI is the development and testing of learnable and adaptive communication systems that use existing resources efficiently and sustainably and are based on the 5G system specifications and 5G applications. Starting with the PHY and MAC layer up to the FPGA hardware implementation and ASIC design, the essential communication components are analyzed and optimized with different AI methods. The properties of known AI methods are identified problem-specifically and possible limitations in their applicability are analyzed.

Research Contribution of the University of Bremen

The aim of the subproject is to design very efficient receiver procedures for radio communication with machine learning methods. Starting from the theoretical concept development, the procedures will be researched and demonstrated by means of prototypical implementation, taking into account implementation issues and 5G-specific system parameters and scenarios. Two different approaches will be pursued.

The Information Bottleneck Method (IBM) is a very general approach motivated by information theory to learn information processing methods. This method has already been used successfully for the optimization of quantizers and for the design of discrete decoders where iterative decoding is implemented very efficiently with the help of learned lookup tables (LUTs). The goal is to extend this approach to more general code structures and to implement the hardware realization together with the project partners. This allows the analysis of the implementation efficiency and the corresponding messaging metrics under conditions of realistic hardware implementations.

In a second approach, concepts for MIMO equalization and channel decoding will be explored using machine learning techniques. In particular, the efficient decoding of short channel codes with high reliability is a challenge and shall be addressed both with learned neural networks and by ML-based adaptation of conventional decoding methods. In principle, the common MIMO equalization and channel decoding allows a reduction of latency due to the otherwise common sequential or iterative processing. Starting from hybrid approaches combining NN-based components with classical algorithms, a common NN-based receiver structure will be developed, learned and analyzed step by step with respect to performance/complexity tradeoff.

Based on the identified 5G use cases, an SDR-based 5G transmission link will be established and the AI-based receiver procedures will be implemented. Thus, the adaptation of the procedures to real transmission systems and scenarios is enabled and the practical suitability of AI-based transmission systems is proven.

Further Information

Project webpage 

BMBF announcement


Duration: 05/2020 - 05/2023
Funding:Federal Ministry of Education and Research
Partners:Creonic GmbH
German Research Center for Artificial Intelligence (DFKI), Research Department Intelligent Networks
Intel Deutschland GmbH
Motius GmbH
Nokia Solutions and Networks GmbH & Co. KG
Division of Mikroelectronic System Design, Technische Universität Kaiserslautern (TUK)
Institute of Telecommunications, University of Stuttgart
Precursor:Characteristic Information Processing on the Physical Layer for Massive Communications (CIPMaC)
Artificial Intelligence for Mobile Systems (AIMS)
Subsequent:Open 6G Hub


Involved Staff

Last change on 02.05.2022 by D. Wübben
AIT ieee GOC tzi ith Fachbereich 1
© Department of Communications Engineering - University of BremenImprint / Contact