Wireless communication network resource allocation method with dynamic adjustment on demand

Patent metadata

Field	Value
Publication number	US-12279158-B2
Application number	US-202217857114-A
Country	US
Kind code	B2
Filing date	Jul 4, 2022
Priority date	Dec 14, 2021
Publication date	Apr 15, 2025
Grant date	Apr 15, 2025

Abstract

A wireless communication network resource allocation method implemented in a server in a wireless communication network, includes: obtaining task feature information of each user device and a CPU frequency of the server in each time slot; obtaining a task data volume average value; determining, based on a knowledge base including sample data groups and optimal resource allocation models, a target optimal resource allocation model matched with the task data volume average value and the CPU frequency of the server; obtaining, based on the task feature information of the user devices in the time slot and the target optimal resource allocation model, resource allocation results of the user devices, and transmitting task data to the user devices based on the results. A width of a dynamic neural network can be automatically adjusted according to task features and computational capacity, and on-demand adjustment of decision speed and resource optimality can be realized.

First claim

What is claimed is:

1. A wireless communication network resource allocation method implemented in a server in a wireless communication network comprising user devices, the method comprising:
sending task feature information of each of the user devices for each of a plurality of time slots to the server, wherein:
the task feature information of each of the user devices comprises: a task data volume, a task transmission distance, and a task importance weight;
a task data volume average value is based on the task data volumes of the user devices;
a target optimal resource allocation model matched with the task data volume average value and a CPU frequency of the server is based on a knowledge base;
the knowledge base comprises: a plurality of sample data groups composed of different task data volume average values, CPU frequencies of the server, and optimal resource allocation models corresponding to the plurality of sample data groups respectively;
the optimal resource allocation model corresponding to each of the plurality of sample data groups is based on a minimum of total delays in resource allocation as an optimization objective, optimal widths of dynamic neural networks in trained resource allocation models, the task feature information of the user devices of the sample data group, and the CPU frequency of the server of the sample data group;
a resource to be allocated in each time slot is one of a bandwidth and a power; and

transmitting task data to the user devices based on resource allocation results, wherein resource allocation results are based on the task feature information of the user devices in each time slot and the target optimal resource allocation model.

2. The wireless communication network resource allocation method according to claim 1 , wherein:
the resource allocation results are further based on resource allocation values of the user devices in each time slot;
the resource allocation values are based on input vectors of the user devices and the target optimal resource allocation model; and
the input vectors are based on the task feature information of each of the user devices in each time slot.

3. The wireless communication network resource allocation method according to claim 2 , wherein:
the optimal resource allocation model corresponding to each of the plurality of sample data groups in the knowledge base is further based on a plurality of sample data groups composed of the different task data volume average values and the CPU frequencies of the server;
each of the plurality of sample data groups comprises one of the task data volume average values and one of the CPU frequencies of the server;
for each of the plurality of sample data groups, the task feature information of the user devices is based on the task data volume average value of the sample data group, and allocated resource data of each of the user devices;
the allocated resource data is the other one of the bandwidth and the power;
the trained resource allocation models have fixed widths corresponding to the dynamic neural networks having different widths;
the total delay of each of the trained resource allocation models with the fixed widths in the resource allocation is based on the task feature information of the user devices of the sample data group, the CPU frequency of the server of the sample data group, and the allocated resource data of the user devices of the sample data group; and
one of the trained resource allocation models with the fixed width corresponds to the minimum of the total delays of the sample data group as the optimal resource allocation model corresponding to the sample data group.

4. The wireless communication network resource allocation method according to claim 3 , wherein:
the trained resource allocation models with fixed widths are based on initial resource allocation models; and
the initial resource allocation models are based on a dynamic neural network with an adjustable width; and
for each of the initial resource allocation models, training the initial resource allocation model through back propagation based on the task feature information of the user devices and a preset loss function.

5. The wireless communication network resource allocation method according to claim 4 , wherein a structure of each of the initial resource allocation models comprises:
an input network, a plurality of expert layers, and an output network connected in sequential order;
wherein each of the plurality of expert layers comprises: an input sublayer, a gating function layer connected to an output end of the input sublayer, a softmax layer connected to an output end of the gating function layer, M number of expert sublayers in parallel with each other connected to another output end of the input sublayer and an output end of the softmax layer, and an output sublayer connected to output ends of the M number of expert sublayers; and
wherein the width of the dynamic neural network in the initial resource allocation model is based on a number of used expert sublayers of the M number of expert sublayers, and whether one of the M number of expert sublayers is used is controlled by a gating value.

6. The wireless communication network resource allocation method according to claim 5 , wherein an input-output mapping function of a gating function network is that: G(X)=RemainK(H (X), K);
where H(X)=X·Wg+Normal·Softplus(X·W noise ), RemainK() represents a sparce function, K represents the width of the dynamic neural network, X represents an input matrix, Wg represents a weight coefficient matrix of the gating function network, Normal represents a standard normal white noise, Softplus() presents an activation function, and W noise represents a noise matrix.

7. The wireless communication network resource allocation method according to claim 5 , wherein the preset loss function is that:

Loss
=

∑
N

i
=
1

w
i

⁢

T

tra
,
i

+

L
balance

,

L
balance

=

w

balance

·

CV
⁡
(

∑

j
=
1

M

g
j

Assignees

• Univ Xidian

Inventors

• Cheng Nan
• Li Changle
• Ma Longfei
• Wang Xiucheng
• Sun Ruijin

Classifications

• Y02D30/70

  in wireless communication networks · CPC title

• H04L67/1008 Primary

  based on parameters of servers, e.g. available memory or workload (monitoring of computer activity G06F11/30) · CPC title

• G06F9/505

  considering the load · CPC title

• G06F9/5044

  considering hardware capabilities · CPC title

• G06F9/5094

  where the allocation takes into account power or heat criteria (power management in computers in general G06F1/3203; thermal management in computers in general G06F1/206) · CPC title