Machine learning for communications

   As 5G/6G networks become extremely complex and heterogeneous, the use of AI/ML is essential in fulfilling a large variety of requirements. ML algorithms will help the system to better adapt to the particular use case considering traffic patterns and quality requirements. In the previous research we considered the application of ML algorithms (deep neural networks, DNN) for the channel decoding problem. It is easy to see that the decoding task is a classification task: the channel output is to be assigned to one of the classes (codewords). The significant difference between this problem and a typical classification problem lies in the exponentially large number of classes. To deal with the curse of dimensionality problem it was suggested to combine deep learning methods with existing decoding approaches (see our papers [1,2]). As a result, we proposed a decoding algorithm based on DNN with a special architecture that allows for training on zero codeword only. 

 

   We plan to continue this line of research and extend it as follows:


  1. The DNN proposed for the channel decoding problem is constructed in accordance with the belief propagation (Min-Sum) algorithm. It means that it repeats the steps of the algorithm and adds trainable weights to deal with trapping sets and improve the performance. In the literature several different approaches were already proposed, namely hyper-networks [3], syndrome based NN and so on. The main goal is to find the requirements for the NN structure and propose better solutions.

  2. One another direction is to construct error-correcting codes or optimize existing codes (such as LDPC and polar codes) with use of ML techniques [4, 5].

  3. ML algorithms can be applied at different system levels. I also plan to consider wireless channel estimation problems as well as resource allocation and scheduling.

References:
  1. D. Artemasov, K. Andreev, P. Rybin and A. Frolov, Soft-Output Deep Neural Network-Based Decoding, 2023 IEEE Globecom Workshops (GC Wkshps), Kuala Lumpur, Malaysia, 2023, pp. 1692-1697.

  2. K. Andreev, A. Frolov, G. Svistunov, K. Wu and J. Liang, Deep Neural Network Based Decoding of Short 5G LDPC Codes, XVII International Symposium “Problems of Redundancy in Information and Control Systems” (REDUNDANCY), 2021, pp. 155-160.

  3. Eliya Nachmani, Lior Wolf: Hyper-Graph-Network Decoders for Block Codes. NeurIPS 2019: 2326-2336

  4. Y. Jiang, H. Kim, H. Asnani, S. Kannan, S. Oh and P. Viswanath, "LEARN Codes: Inventing Low-Latency Codes via Recurrent Neural Networks," in IEEE Journal on Selected Areas in Information Theory, vol. 1, no. 1, pp. 207-216, May 2020.

  5. Y. Liao, S. A. Hashemi, J. Cioffi and A. Goldsmith, "Construction of Polar Codes with Reinforcement Learning," GLOBECOM 2020 - 2020 IEEE Global Communications Conference, 2020.