5th edition Newsletter – CENTRIC Open-Source Repositories for ML Community

🕒 8 mins

 

CENTRIC Open-Source Repositories for ML Community

The CENTRIC project has published several open datasets and training environments for researchers and developers, as early adopters of CENTRIC results, working on emerging ML techniques for broad application in wireless communications.

On its way towards the goal to enable sustainable user-centric 6G networks through an AI-native Air Interface, the CENTRIC project is leveraging Artificial Intelligence (AI) and Machine Learning (ML) techniques through a top-down, modular approach to wireless connectivity focusing on users’ communication needs and environmental constraints at the center of the network stack design. To achieve this, the CENTRIC project team worked on establishment and investigation of numerous ML techniques for their application in different aspects of network design, covering both physical and MAC networking layers, resulting with a number of open source repositories and data sets, allowing also wide research community to (re-)use the CENTRIC results, but also to investigate and validate own emerging solutions.

The newsletter summarizes the provided open-source repositories, which are now available for the wide researcher community. Detailed descriptions of the repositories, including relevant use cases, results, and further useful information for users of the repositories can be found in the CENTRIC project deliverables D3.4 and D4.1, which are available on our website at https://centric-sns.eu/public-deliverables/ 

CENTRIC’s AI-based MIMO Toolset links:

•  RL-Based Beam Management in ISAC Scenarios
•  Multiuser MIMO Neural Receiver
•  Transfer Learning Techniques for Neural Receivers
•  Narrow Beam Prediction Using NN Decoder

Repositories for Protocol Learning and Emergence Challenges links:

•  Multiple access with MuJoCo robots
•  Random access with MARL
•  DCI learning
•  6G in-factory subnetworks

 All are available on our CENTRIC’s MIMO AI Toolset & CENTRIC WP4 GitHubs.

 

---

---

The CENTRIC’s AI-based MIMO Toolset 

The CENTRIC’S MIMO toolset has been implemented by establishing several open source repositories, including software implementations and documentation of the related simulation environments and developed AI-based MIMO algorithms. Each repository focuses on a specific MIMO-AI algorithm, such as reinforcement learning for beam management in Integrated Sensing and Communication (ISAC) scenarios, AI techniques for wide-to-narrow beam prediction, multi-user MIMO neural network-based receiver, transfer learning techniques for neural receiver, and learning-based beam alignment.

---

RL Based Beam Management in ISAC Scenarios

The repository contains Reinforcement Learning (RL) based solutions for resource allocation in millimetre wave (mmWave) Integrated Sensing and Communication (ISAC) scenarios. The repository provides a simulation environment for beam management in a vehicular wireless network comprising of a base station (BS) serving multiple vehicles on appropriately allocated beam (Figure 1).

Figure 1: System model for vehicular wireless network

CENTRIC proposes a Proximal Policy Optimization solution for joint resource allocation and beam tracking in a mmWave wireless communication context. By coordinating the optimization of the resource allocation and beam tracking parameters, our system intends to mitigate the mmWave channel’s dynamic nature and its fluctuating channel quality, including blockage events. Python implementation of a novel reinforcement learning algorithm is also included in the repository, allowing evaluation and benchmarking of the included methods as well as easy integration of new and efficient AI algorithms for resource allocation in mmWave ISAC scenarios.

The public repository is available at https://github.com/CENTRIC-WP3/RL-Based-Beam-Management-in-ISAC-Scenarios

Multiuser MIMO Neural Receiver

The repository presents real time multi-user MIMO neural network (NN)-based receivers that comply with 5G New Radio (5G NR), positioning them as enabling technology for novel applications such as re-trainable site-specific base stations and pilotless communications. The neural receiver (NRX), which replaces significant portions of traditional physical layer receiver algorithms with neural networks, as illustrated in Figure 2.

Figure 2: A neural receiver replaces channel estimation, equalization and demapping

The current version of the receiver developed in CENTRIC is a flexible multi-user MIMO (MU-MIMO) receiver with 5G NR physical uplink shared channel (PUSCH) compatibility. The receiver architecture combines graph neural networks (GNN) and convolutional neural networks (CNN), allowing flexibility in handling varying user numbers and sub-carrier configurations without retraining. Also, the steps required to deploy a MU-MIMO neural receiver (NRX) in actual cellular communication systems are detailed, addressing challenges such as real-time inference and 5G NR standard compatibility.

The repository, managed by NVIDIA, provides comprehensive documentation, including simulation examples and tutorials, is available at https://github.com/CENTRIC-WP3/neural_rx

Transfer Learning Techniques for Neural Receivers

Training deep neural network models typically requires large datasets. Due to the dynamic nature of wireless communication setups, generating sufficient dataset for each configuration of wireless communication deployment can be cumbersome. To overcome this hurdle, transfer learning can be exploited wherein the weights of a source model trained on a large dataset can be modified fully or partially by the relatively smaller dataset of the target model.

In this repository, implementation of transfer learning techniques applied to the neural receiver is provided. The provided codes allow evaluation of the transferability of neural receivers trained in specific environments or for specific tasks to other environments or tasks and the potential of various transfer learning techniques to minimize the amount of data required for retraining of neural receivers for new environments or tasks. The repository, available at https://github.com/CENTRIC-WP3/Transfer-Learning-Techniques-for-MIMO-Neural-Receivers, makes it possible to easily integrate and evaluate new transfer learning techniques.

Narrow Beam Prediction Using NN Decoder

The repository presents implementation of an AI method for wide beam codebook design. The solution allows for exploration of the benefits of refined beam prediction and estimation. A neural network is applied as a decoder, where the measurement of signal powers acquired from the designed wide beams is the input, and the output is an estimate of the best UE refined beam.

The current beam prediction procedures in 5G consist of three phases. With the proposed idea, one can leverage the wide beam measurements from phase 1 to predict the refined beam having the highest reference signal received power (RSRP) for data transmission without needing the overhead from the phase 2 beam refinement procedure.

The repository is available at https://github.com/CENTRIC-WP3/Narrow-Beam-Prediction-using-Neural-Network-Decoder to allow reproducibility of the results and possible extension of the proposed AI algorithm for refined beam prediction.

 

---

---

Repositories for Protocol Learning and Emergence Challenges

These repositories address various challenges in protocol learning and emergence. Each repository focuses on a specific problem, such as multiple access with MuJoCo robots, random channel access with MARL, DCI learning for reducing the length of control messages, and 6G in-factory subnetworks for industrial applications. The repositories presented below offer a starting point for exploring alternative solutions, benchmarking, and customization, and can be integrated with machine learning algorithms for further research and development.

---

 

Multiple access with MuJoCo robots

The purpose of the repository is to provide a minimal working example of a communication network between a base station (BS) and several user equipments (UEs), instantiated as gymnasium environments (CartPole and MuJoCo). The model implemented in the repository (Figure 3) may be employed as a simple environment for learning and testing novel communication protocols.

Figure 3: Wireless communication model

The CartPole and the MuJoCo robots, together with the communication network orchestrated by the BS, provide toy-model scenarios for various 6G use-cases, such as collaborative robots on a smart factory floor, networks of heterogeneous URLLCs (ultra-low-latency) entities, massive machine-type communications (mMTC), Communication and Control Co-design applications etc. Here, the user may design, implement, optimize existing protocols or emerge new ones via AI techniques in general, and reinforcement learning in particular.

The repository is available at https://github.com/CENTRIC-WP4/Multiple-access-with-MuJoCo-robots

Random access with MARL

The repository addresses random access with MARL (Multi-Agent Reinforcement Learning), offering a platform for studying the behavior of multiple agents in a random access scenario. By utilizing reinforcement learning techniques, researchers can investigate how agents can effectively share limited resources and optimize their access strategies.

The repository considers in particular the massive machine type communications (mMTC), where the traffic consists of a rather sporadic transmissions by machine-type devices in the uplink. Application of Random Access (RA) makes more sense in this case, as scheduled access incurs signaling overhead even the traffic is mostly sporadic in nature, but on the other hand the RA is less efficient when the traffic load is high, caused by collisions.

This repository is available at https://github.com/CENTRIC-WP4/Random-Access-with-MARL may be used to extend the environment of RA using MARL. Scalability is a big issue in MARL and this repository may be used to design scalable solutions for MARL for RA. Moreover, users may also use this repository to compare different MARL algorithms for centralized and decentralized training and decentralized execution.

DCI learning

The repository focuses on DCI (Downlink Control Information) learning for reducing the length of 5GNR control messages. This repository provides a DCI simulator, that can be used to generate artificial DCI messages for protocol model training. By reducing the DCI length, researchers can improve the efficiency and scalability of communication protocols, leading to more efficient resource utilization by increasing the control channels capacity without occupying additional radio resources.

Potential reduction of the length of a DCI message can be achieved by applying and training ML models to understand temporary correlations among the control bits, which can be used for the control messages length reduction. A lossless compressor (Figure 4) can accomplish the task of reducing the length of a DCI message.

Figure 4: Implementation of a lossless compressor to reduce the DCI length

The repository is available at https://github.com/CENTRIC-WP4/DCI-Learning and it generates temporal correlated binary bits for the interested developer to design lossless compression algorithms that reduce the length of DCI messages.

6G in-factory subnetworks

The repository explores 6G in-factory subnetworks for industrial applications. It offers a framework for simulating and analyzing the performance of subnetworks in industrial and other relevant settings, so-called X subnetworks, as shown in Figure 5. Researchers can use this repository to evaluate the effectiveness of different communication protocols and network configurations, enabling them to design more reliable and efficient subnetworks for industrial applications.

Figure 5: Selected use cases for 6G in-X subnetworks

In this repository, we present software implementations of 6G in-X subnetworks for industrial applications in a multi-agent reinforcement learning environment to drive research into the challenges of developing RL techniques that solve both the problem of learning a protocol for wireless driving heterogeneous control services and radio resource allocation. The repository provides a tool for the development and evaluation of layer 2 and RRM algorithms for 6G in-X subnetworks supporting industrial control operations and it also allows seamless integration and evaluation of new machine-learning algorithms.

The repository is available at https://github.com/CENTRIC-WP4/6G-infactory-subnetworks

 

---

---

 

Instead of Conclusion

We wish the wide research community working on solutions for wireless networks based on AI and ML techniques lots of success while using and further exploring the CENTRIC open-source repositories!

We would, of course, appreciate feedback from the research community on the usability of our repositories, further results achieved by using the CENTRIC open source, and any other message on the experience while using our repositories.

 

Read it in its original Newsletter #5 format here