Methods and systems for network self-optimization using deep learning

Patent metadata

Field	Value
Publication number	US-11477666-B2
Application number	US-201816486430-A
Country	US
Kind code	B2
Filing date	Feb 15, 2018
Priority date	Feb 16, 2017
Publication date	Oct 18, 2022
Grant date	Oct 18, 2022

Abstract

In one example embodiment, a method includes obtaining sets of input data and an objective, the sets of input data including values of wireless network communication parameters, generating sets of output data for the sets of input data in accordance with the objective, generating a mapping of the sets of input data to the sets of output data and training a network controller using the mapping.

First claim

The invention claimed is:

1. An automated optimization and machine-learning method for a network controller of a heterogeneous network, comprising:
obtaining sets of input data and an objective, the sets of input data including values of wireless network communication parameters comprising at least one of channel gains, measurements of received power, measurements of interference, power-related information, an available bandwidth, connectivity, user equipment (UE) and base station associations, quality of service (QoS) parameters, congestion and load information;
generating sets of output data for the sets of input data in accordance with the objective;
generating a mapping of the sets of input data to the sets of output data; and
training a network controller using the mapping;
wherein the network controller comprises a neural network configured to extract an underlying model describing the mapping, the neural network comprising a plurality of individual Feed-Forward neural networks each of which is trained on the set of input data and one individual bit of the output data, wherein said individual bit is a schedulable element representing an actual resource element in an available signal bandwidth.

2. The method of claim 1 , wherein the objective is a QoS objective including at least one of targeted data rates, a latency, a reliability, a data loss rate, an out-of-order delivery and round-trip time.

3. The method of claim 1 , wherein the obtaining sets of input data includes, obtaining the power-related information from a plurality of user equipments (UEs), the power-related information including a power-related value during almost blank subframes (ABS) and a power-related value during non-ABS subframes, the generating sets of output data generates the sets of output data based on the power-related information.

4. The method of claim 3 , wherein the sets of output data are optimized transmission schedules.

5. The method of claim 1 , wherein the generating sets of output data generates the sets of output data using at least one of genetic programming and a genetic algorithm.

6. The method of claim 1 , wherein each individual Feed-Forward neural network provides a schedulable element of a downlink transmission schedule.

7. The method of claim 6 , wherein the training includes, providing the sets of input data to each individual Feed-Forward neural network.

8. The method of claim 1 , wherein each individual Feed-Forward neural network represents a different number of user equipments (UEs) in a communication system.

9. The method of claim 1 , wherein the training includes, training individual networks, determining first layers of the individual Feed-Forward neural networks, the first layers associated with a higher metric relative to other layers of the individual networks, and combining the first layers to form the neural network.

10. The method of claim 1 , further comprising: deploying the network controller in a self-organizing network (SON) controller.

11. The method of claim 1 , further comprising: collecting measurement data from a plurality of UEs in a communications network; and generating, by the network controller, an optimized transmission schedule based on the measurement data.

12. A communication system comprising:
a trained network controller of a heterogeneous network, the trained network controller including, a memory storing computer-readable instructions, at least one processor configured to execute the computer-readable instructions to,
obtain input network measurements comprising at least one of channel gains, measurements of received power, measurements of interference, power-related information, an available bandwidth, connectivity, user equipment (UE) and base station associations, quality of service (QoS) parameters, congestion and load information, and
generate output network parameters based on the input network measurements; wherein the network controller comprises a Deep Feed Forward neural network comprising a plurality of individual Feed-Forward neural networks each of which receives the input network measurements and outputs an individual bit of output data, wherein said individual bit is a schedulable element representing an actual resource element in an available signal bandwidth.

13. The communication system of claim 12 , wherein the trained network controller is configured to execute the computer-readable instructions to optimize the output network parameters.

14. A non-transitory computer--readable medium carrying instructions which, when executed by a processor, cause the processor to,
obtain sets of input data and an objective, the sets of input data including values of wireless network communication parameters comprising at least one of channel gains, measurements of received power, measurements of interference, power-related information, an available bandwidth, connectivity, user equipment (UE) and base station associations, quality of service (QoS) parameters, congestion and load information;
generate sets of output data for the sets of input data in accordance with the objective;
generate a mapping of the sets of input data to the sets of output data; and
train a network controller using the mapping;
wherein the network controller comprises a neural network configured to extract an underlying model describing the mapping, the neural network comprising a plurality of individual Feed-Forward neural networks each of which is trained on the set of input data and one individual bit of the output data, wherein said individual bit is a schedulable element representing an actual resource element in an available signal bandwidth.

Assignees

• Univ College Dublin Nat Univ Ireland Dublin

Inventors

• Kucera Stepan
• Claussen Holger
• Lynch David
• Fagan David
• Fenton Michael
• O'Neill Michael

Classifications

• H04W72/542

  using measured or perceived quality · CPC title

• H04W24/02 Primary

  Arrangements for optimising operational condition · CPC title

• H04W24/06

  Testing, {supervising or monitoring} using simulated traffic · CPC title

• G06N3/02

  Neural networks · CPC title

• H04W52/54

  Signalisation aspects of the TPC commands, e.g. frame structure · CPC title