Predicting network communication performance using federated learning

Patent metadata

Field	Value
Publication number	US-12052145-B2
Application number	US-201917299152-A
Country	US
Kind code	B2
Filing date	Dec 6, 2019
Priority date	Dec 7, 2018
Publication date	Jul 30, 2024
Grant date	Jul 30, 2024

Abstract

Federated learning is used to predict network communication performance at an edge node (10-1). The edge node (10-1) trains a local model (14-1) of network communication performance over one or more rounds of training, based on a local training dataset (22-1) and based on multi-node training information (20-1) received in each round of training. This information (20-1) comprises information about local models (14-2, . . . 14-N) at other respective edge nodes (10-2, . . . 10-N) as trained based on local training datasets at the other edge nodes (10-2, . . . 10-N). After or as part of each round of training, the edge node (10-1) transmits control signaling that indicates an accuracy of the local model (14-1) as trained through that round of training, that indicates whether another round of training is needed or desired, and/or that indicates whether any further information (20-1) is needed or desired. The edge node (10-1) predicts network communication performance at the edge node (10-1) based on the trained local model (14-1).

First claim

What is claimed is:

1. An edge node configured for use in an edge communication network and for using federated learning to predict network communication performance at the edge node, the edge node comprising:
communication circuitry; and
processing circuitry configured to:
train a local model of network communication performance over multiple rounds of training at the edge node, based on a local training dataset at the edge node and based on multi-node training information received in each round of training, wherein the multi-node training information comprises information about local models at other respective edge nodes as trained based on local training datasets at the other edge nodes;
after or as part of each round of training, transmit control signaling that indicates an accuracy of the local model as trained by the edge node through that round of training, that indicates whether another round of training is needed or desired at the edge node, and/or that indicates whether any further multi-node training information is needed or desired at the edge node;
predict network communication performance at the edge node based on the trained local model; and
perform one or more remedial or preventative measures to account for the network communication performance at the edge node being predicted to decrease.

2. The edge node of claim 1 , wherein the control signaling indicates whether another round of training is needed or desired at the edge node and/or indicates whether any further update information is needed or desired at the edge node.

3. The edge node of claim 2 , the processing circuitry further configured to, for each of the multiple rounds of training:
determine, based on at least the accuracy of the local model as trained by the edge node through that round of training, whether one or more conditions are met for stopping training of the local model at the edge node; and
generate the control signaling to indicate that another round of training and/or further update information is not, or is, needed or desired at the edge node, depending respectively on whether or not the one or more conditions are met.

4. The edge node of claim 3 , wherein the one or more conditions include the accuracy of the local model reaching an accuracy threshold and/or improving by less than a minimum incremental improvement threshold since one or more previous rounds of training.

5. The edge node of claim 2 , wherein the control signaling indicates that the edge node does not need or desire another round of training and/or any further update information and further indicates for how long the edge node does not need or desire another round of training and/or any further update information.

6. The edge node of claim 1 , wherein the control signaling indicates an accuracy of the local model as trained by the edge node through that round of training, wherein the control signaling indicates the accuracy of the local model with respect to a local test dataset at the edge node.

7. The edge node of claim 1 , wherein the multi-node training information includes a combination of local updates that the other edge nodes respectively made to local models at the other edge nodes.

8. The edge node of claim 1 , wherein the multi-node training information includes, for each of multiple other edge nodes, a local update that the other edge node made to a local model at the other edge node, and wherein, for each of the multiple rounds of training, the processing circuitry is configured to:
decide which one or more of the local updates to combine with one another, and/or with a local update determined by the edge node based on the local training dataset at the edge node, into a combined update, and
update the local model at the edge node based on the combined update.

9. The edge node of claim 1 , wherein the processing circuitry is configured to perform one or more remedial or preventative measures by adjusting a transmission power of the edge node.

10. The edge node of claim 1 , wherein the local model at the edge node is a model of a predicted level of degradation in one or more key performance indicators that indicate network communication performance in the edge communication network.

11. The edge node of claim 10 , wherein the local model at the edge node maps the one or more key performance indicators as input to an output in the form of a multiclass label that represents the predicted level of degradation in the one or more key performance indicators, wherein the multiclass label has multiple possible values associated with different predicted levels of degradation in the one or more key performance indicators.

12. The edge node of claim 1 , wherein the local model at the edge node and the local models at the other edge nodes are each a neural network model, and wherein a local update to a neural network model includes an updated weight matrix.

13. A server for coordinating training of local models of network communication performance at respective edge nodes, the server comprising:
communication circuitry; and
processing circuitry configured to, for each of multiple rounds of training:
transmit, to one or more of the edge nodes, multi-node training information that comprises information about local models at respective edge nodes as trained based on local training datasets at the edge nodes; and
receive, from one or more of the edge nodes, control signaling that indicates an accuracy of the local model at the edge node as trained through the round of training, that indicates whether another round of training is needed or desired at the edge node, and/or that indicates whether any further multi-node training information is needed or desired at the edge node; and
control generation of or transmission of multi-node training information in any next round of training based on the received control signaling.

14. The server of claim 13 , wherein the control signaling indicates whether another round of training is needed or desired at the edge node and/or indicates whether any further update information is needed or desired at the edge node.

15. The server of claim 13 , wherein the control signaling indicates that the edge node does not need or desire another round of training and/or any further update information and further indicates for how long the edge node does not need or desire another round of training and/or any further update information.

16. The server of claim 13 , wherein the processing circuitry is configured to control generation of or transmission of multi-node training information by transmitting or not transmitting multi-node training information to an edge node in a next round of training, depending respectively on whether or not the control signaling from the edge node indicates that the edge node needs or desires the further multi-node training information or another round of training.

17. The server of claim 13 , wherein the control signaling indicates the accuracy of the local model at the edge node as trained through the round of training, and wherein the processing circuitry is configured to control generation of or transmission of multi-node training information by:
determining, based on at least the accuracy indicated by the control signaling received from an edge node, whether one or more conditions are met for stopping training of the local model at the edge node; and
refraining from transmitting, or transmitting, further multi-node training information to the edge node in a next round of training, depending respectively on whether or not the one or more conditions are met.

Assignees

• Ericsson Telefon Ab L M

Inventors

• Vandikas Konstantinos
• Ickin Selim

Classifications

• G06N3/092

  Reinforcement learning · CPC title

• G06N3/098

  Distributed learning, e.g. federated learning · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0499

  Feedforward networks · CPC title

• H04L41/147 Primary

  for predicting network behaviour · CPC title