WO2023097452A1

WO2023097452A1 - Methods, apparatuses, and computer readable media for radio resource allocation

Info

Publication number: WO2023097452A1
Application number: PCT/CN2021/134446
Authority: WO
Inventors: Jiefeng JIN
Original assignee: Nokia Shanghai Bell Co., Ltd.; Nokia Solutions And Networks Oy
Priority date: 2021-11-30
Filing date: 2021-11-30
Publication date: 2023-06-08

Abstract

Disclosed is a method for radio resource allocation. An example radio resource allocation method may include a process to develop a slice weight for scheduling comprising calculating a current deviation of a current allocation volume from a target allocation volume for respective at least one slice; calculating a learning rate and a gradient of a cost function of respective neural network coefficients by backpropagation for the respective at least one slice; updating the respective neural network coefficients based on the learning rate and the gradient for the respective at least one slice; and calculating the slice weight based on the current deviation, historical deviations, and the updated neural network coefficients for the respective at least one slice. Related radio resource allocation apparatuses and computer readable media are also disclosed.

Description

METHODS, APPARATUSES, AND COMPUTER READABLE MEDIA FOR RADIO RESOURCE ALLOCATION

TECHNICAL FIELD

Various embodiments relate to methods, apparatuses, and computer readable media for radio resource allocation.

BACKGROUND

In a telecommunication network such as a new radio (NR) system, a tenant may request to establish a network slice. The network slice may be allocated by, for example, a scheduler in a radio access network (RAN) with time and/or frequency resources such that the network may appear as dedicated for the tenant. A plurality of network slices may have to be implemented, completely or partially, on a common infrastructure layer, including assets such as the time and frequency resources, or throughput. The radio resources may be allocated by the scheduler according to a given target, and the target may be dynamically adjusted according to, for example, quality of service (QoS) or quality of experience (QoE) requirements.

SUMMARY

A brief summary of exemplary embodiments is provided below to provide basic understanding of some aspects of various embodiments. It should be noted that this summary is not intended to identify key features of essential elements or define scopes of the embodiments, and its sole purpose is to introduce some concepts in a simplified form as a preamble for a more detailed description provided below.

In a first aspect, disclosed is a radio resource allocation method. The radio resource allocation method may include a process to develop a slice weight for scheduling comprising: calculating a current deviation of a current allocation volume from a target allocation volume for respective at least one slice; calculating a learning rate and a gradient of a cost function of respective neural network coefficients by backpropagation for the respective at least one slice; updating the respective neural network coefficients based on the learning rate and the gradient for the respective at least one slice; and calculating the slice weight based on the current deviation, historical deviations, and the updated neural network coefficients for the respective at least one slice.

In some embodiments, the historical deviations may comprise a previous deviation and an accumulated deviation, and the accumulated deviation may be calculated based on exponential moving average.

In some embodiments, the cost function may be defined by the current deviation and a previous value of the cost function of the respective at least one slice.

In some embodiments, the cost function may be defined based on exponential moving average, and the respective neural network coefficients may be updated by gradient descent for minimizing the cost function.

In some embodiments, at least one derivative for calculating the gradient of the cost function may be embodied as a sign function and/or a Heaviside step function.

In some embodiments, the method may further include applying the calculated slice weight for the respective at least one slice to a scheduler to determine a next allocation volume for the respective at least one slice.

In some embodiments, the method may further include a step to develop a comprehensive slice weight for a plurality of slices comprising: integrating the calculated slice weight for each of the respective at least one slice based on a sum of a plurality of calculated slice weights for the plurality of slices.

In some embodiments, the method may further include applying the integrated calculated slice weight for the respective at least one slice to a scheduler to determine a next allocation volume for the respective at least one slice.

In some embodiments, the determined next allocation volume may be iteratively used as the current allocation volume after a period starting from calculating the current deviation.

In some embodiments, the radio resource for allocation may comprise time and frequency resource, or throughput.

In a second aspect, disclosed is a radio resource allocation apparatus. The radio resource allocation apparatus may include at least one processor and at least one memory. The at least one memory may include computer program code, and the at least one memory and the computer program code may be configured to, with the at least one processor, cause the apparatus to perform a process to develop a slice weight for scheduling comprising: calculating a current deviation of a current allocation volume from a target allocation volume for respective at least one slice; calculating a learning rate and a gradient of a cost function of respective neural network coefficients by backpropagation for the respective at least one slice; updating the respective neural network coefficients based on the learning rate and the gradient for the respective at least one slice; and calculating the slice weight based on the current deviation, historical deviations, and the updated neural network coefficients for the respective at least one slice.

In some embodiments, the at least one memory and the computer program code may be further configured to, with the at least one processor, cause the apparatus to further perform applying the calculated slice weight for the respective at least one slice to a scheduler to determine a next allocation volume for the respective at least one slice.

In some embodiments, the at least one memory and the computer program code may be further configured to, with the at least one processor, cause the apparatus to further perform a step to develop a comprehensive slice weight for a plurality of slices comprising: integrating the calculated slice weight for each of the respective at least one slice based on a plurality of calculated slice weights for the plurality of slices.

In some embodiments, the at least one memory and the computer program code may be further configured to, with the at least one processor, cause the apparatus to further perform applying the integrated calculated slice weight for the respective at least one slice to a scheduler to determine a next allocation volume for the respective at least one slice.

In a third aspect, disclosed is a radio resource allocation apparatus. The radio resource allocation apparatus may include means for developing a slice weight for scheduling comprising: means for calculating a current deviation of a current allocation volume from a target allocation volume for respective at least one slice; means for calculating a learning rate and a gradient of a cost function of respective neural network coefficients by backpropagation for the respective at least one slice; means for updating the respective neural network coefficients based on the learning rate and the gradient for the respective at least one slice; and means for calculating the slice weight based on the current deviation, historical deviations, and the updated neural network coefficients for the respective at least one slice.

In some embodiments, the apparatus may further include means for applying the calculated slice weight for the respective at least one slice to a scheduler to determine a next allocation volume for the respective at least one slice.

In some embodiments, the apparatus may further include means for developing a comprehensive slice weight for a plurality of slices comprising: means for integrating the calculated slice weight for each of the respective at least one slice based on a plurality of calculated slice weights for the plurality of slices.

In some embodiments, the apparatus may further include means for applying the integrated calculated slice weight for the respective at least one slice to a scheduler to determine a next allocation volume for the respective at least one slice.

In a fourth aspect, a computer readable medium is disclosed. The computer readable medium may include instructions stored thereon for causing a radio resource allocation apparatus to perform a process to develop a slice weight for scheduling comprising: calculating a current deviation of a current allocation volume from a target allocation volume for respective at least one slice; calculating a learning rate and a gradient of a cost function of respective neural network coefficients by backpropagation for the respective at least one slice; updating the respective neural network coefficients based on the learning rate and the gradient for the respective at least one slice; and calculating the slice weight based on the current deviation, historical deviations, and the updated neural network coefficients for the respective at least one slice.

In some embodiments, the gradient of a cost function may be defined by the current deviation and a previous value of the cost function of the respective at least one slice.

In some embodiments, the cost function may be defined based on exponential moving average, and the respective neural network coefficients may be updated by the gradient descent for minimizing the cost function.

In some embodiments, the computer readable medium may further include instructions stored thereon for causing the apparatus to further perform applying the calculated slice weight for the respective at least one slice to a scheduler to determine a next allocation volume for the respective at least one slice.

In some embodiments, the computer readable medium may further include instructions stored thereon for causing the apparatus to further perform a step to develop a comprehensive slice weight for a plurality of slices comprising: integrating the calculated slice weight for each of the respective at least one slice based on a sum of a plurality of calculated slice weights for the plurality of slices.

In some embodiments, the computer readable medium may further include instructions stored thereon for causing the apparatus to further perform applying the integrated calculated slice weight for the respective at least one slice to a scheduler to determine a next allocation volume for the respective at least one slice.

Other features and advantages of the example embodiments of the present disclosure will also be apparent from the following description of specific embodiments when read in conjunction with the accompanying drawings, which illustrate, by way of example, the principles of example embodiments of the present disclosure.

BRIEF DESCRIPTION OF THE DRAWINGS

Some example embodiments will now be described, by way of non-limiting examples, with reference to the accompanying drawings.

FIG. 1 shows a radio resource allocation system in which example embodiments of the present disclosure can be implemented.

FIG. 2 shows a radio resource allocation system in which example embodiments of the present disclosure can be implemented.

FIG. 3 shows a radio resource allocation system in which example embodiments of the present disclosure can be implemented.

FIG. 4 shows a radio resource allocation system in which example embodiments of the present disclosure can be implemented.

FIG. 5 shows a simulation result of radio resource allocation according to an example embodiment of the present disclosure.

FIG. 6 shows a flow chart illustrating an example method for radio resource allocation according to embodiments of the present disclosure.

FIG. 7 shows a block diagram illustrating an example apparatus for radio resource allocation according to embodiments of the present disclosure.

FIG. 8 shows a block diagram illustrating an example apparatus for radio resource allocation according to embodiments of the present disclosure.

Throughout the drawings, same or similar reference numbers indicate same or similar elements. A repetitive description on the same elements would be omitted.

DETAILED DESCRIPTION

Herein below, some example embodiments are described in detail with reference to the accompanying drawings. The following description includes specific details for the purpose of providing a thorough understanding of various concepts. However, it will be apparent to those skilled in the art that these concepts may be practiced without these specific details. In some instances, well known circuits, techniques and components are shown in block diagram form to avoid obscuring the described concepts and features.

Embodiments of the present disclosure provide a slice weight mechanism for radio resource allocation. A radio resource allocation with improved stability and precision is achieved according to the embodiments of the present disclosure.

FIG. 1 shows a radio resource allocation system 100 in which example embodiments of the present disclosure can be implemented. A scheduler 110, a module 120 for performing a process to develop a slice weight for scheduling, and a module 130 for allocation counting are shown as example modules of the system 100. The scheduler 110, the module 120, and the module 130 for allocation counting may be located in the RAN.

Referring to the FIG. 1, the scheduler 110 may determine the radio resource allocation according to a slice weight for respective slice together with other constraints. Different kinds of schedulers may have different principles to use the slice weight, but generally a scheduler decides the radio resource allocation according to the slice weight, and generally the slice with a bigger weight may have a higher priority to be allocated radio resources. Thus the slice with a bigger weight (higher priority) can get more resources in a round of resource allocation process. The slice weight may be a non-negative value, and a zero value means the slice shall not be scheduled.

According to the embodiments of the present disclosure, to achieve stable and precise control of radio resource allocation over a measurement window for the slices, the module 120 may adjust the slice weight for respective slice in real-time and notify the module 110 when the slice weight changes. The slice weight may be a dynamic variable that can be updated by the module 120 based on a slicing control process. The slice-weight-based RAN slicing control process cooperatively works with the scheduler 110 but may be independent to the processes performed by the scheduler 110.

The module 130 for allocation counting may perform an actual radio resource allocation based on the allocation decision output from the scheduler 110. The actual radio resource allocation performed by the module 130 for allocation counting may be identical to the allocation decision output from the scheduler 110. Alternatively or additionally, the module 130 for allocation counting may perform the actual radio resource allocation further based on an uplink data traffic statistic, such that the actual radio resource allocation result may be different from the allocation decision output from the scheduler 110. The scheduler 110, the module 130, or both may be located in a centralized entity, in distributed entities, or be realized by cloud computing, etc.

The actual radio resource allocation result may be input as feedback to the module 120, and this process may be done in an iteration way to minimize the computational cost.

The volume of radio resource allocated for respective slice may be referred as to an allocation volume for the respective slice. The allocation volume may be a share allocated for the slice of a total radio resource, alternatively may be an amount of radio resources, e.g. physical resource blocks (PRBs) . To decide the slice weight, a quota target of the radio resource may be determined for respective slice. The quota target can either be statically assigned according to the demand of the slice tenant or be dynamically adjusted according to the user equipment (UE) status in the slice.

Ideally, the scheduler 110 will allocate the radio resource according to the slice weight, so the quota target of the radio resource, which may also be termed as a target allocation volume of resource, can be used as an initial value of the slice weight. W _s (0) may denote the initial (time = 0) slice weight for a slice s, and may be determined by the formula (1) :

Where

denotes the target allocation volume required by the slice s.

Typically, the scheduler 110 may determine the radio resource allocation based on many factors, e.g. the number limitation of simultaneously-scheduled UEs, radio condition, buffer status of the UEs, beamforming condition, and UE’s limited transmitting power. However, a legitimate assumption is that the volume of radio resources allocated to the slice by the scheduler 110 is positively correlated with the slice weight of the slice in most cases.

Base on this assumption, an adaptive slicing control (ASC) process may be used to achieve the stable and precise radio resource allocation.

In order to achieve the stable and precise radio resource allocation, a control loop process based on feedbacks may be utilized. An improved proportional–integral–derivative (PID) controlling process as control loop may be used in the embodiments of the present disclosure. In the embodiments of the present disclosure, the radio resource for allocation may comprise time and frequency resource, or throughput.

FIG. 2 shows a radio resource allocation system 200 in which example embodiments of the present disclosure can be implemented. Referring to the FIG. 2,

Units

210, 220, 230, 240, and 250 may be included in the module 120, a unit 260 may function as the scheduler 110 and the module 130 for allocation counting.

In the unit 210, the module 120 may calculate a current deviation of a current allocation volume from a target allocation volume for respective at least one slice. A slice s may represent any one slice of the respective at least one slice. In an embodiment, for the slice s, at a current time t, the current deviation E _s (t) of the current allocation volume

from the target allocation volume

is calculated as the formula (2) :

The target allocation volume may be a constant

or may change with the time based on demands, and the target allocation volume at the current time t may be denoted as

The current allocation volume

may be output from the unit 260 for scheduling and allocation counting. The current deviation E _s (t) may also be termed as a current control error.

A changing value of the slice weight at the time t for the slice s may be denoted as ΔW _s (t) . The changing value ΔW _s (t) may be function of the current deviation E _s (t) , which may be denoted as the formula (3) :

ΔW _s (t) =f (E _s (t) ) (3)

In the

units

220, 230, and 240, the module 120 may calculate proportional term E _s (t) , integral term

and derivative term E _s (t) -E _s (t-1) , respectively. The E _s (t-1) is a previous deviation at time t-1. The

is an exponential moving average (EMA) item and may be calculated as the formula (4) :

where τ is a length of an averaging time window.

The initial value of

at time 0 may be calculated as the formula (5) :

Thus the

may represent an accumulated deviation. The previous deviation and the accumulated deviation may embody and be included in historical deviations.

The proportional term E _s (t) , integral term

and derivative term E _s (t) -E _s (t-1) have corresponding neural network coefficients denoted as K _s, 1, K _s, 2, K _s, 3, respectively. The module 120 may calculate a learning rate and a gradient of a cost function of respective neural network coefficients by backpropagation for the respective at least one slice, and update the respective neural network coefficients based on the learning rate and the gradient for the respective at least one slice, e.g., the slice s, which will be described later.

Then, the module 120 may calculate the slice weight based on the current deviation, the historical deviations, and the updated neural network coefficients. For example, the module 120 may calculate the changing value ΔW _s (t) by the formula (6) :

where K _s, 1, K _s, 2, K _s, 3 are the neural network coefficients of proportional, integral, and derivative terms, respectively.

The changing value ΔW _s (t) may be the input to the unit 250. In the unit 250, the module 120 may calculate the slice weight based on the changing value for the respective at least one slice. For example, for the slice s, the module 120 may calculate the slice weight by the formula (7) :

W _s (t) =W _s (t-1) +ΔW _s (t) (7)

where the W _s (t-1) is previous slice weight at time t-1, and W _s (t) is the calculated slice weight at time t.

Then, the module 120 may apply the calculated slice weight for the respective at least one slice to the scheduler 110 to determine a next allocation volume for the respective at least one slice. The next actual allocation volume output from the unit 260 may be measured and input to the unit 210 for the next loop. The determined next allocation volume may be iteratively used as the current allocation volume after a period starting from calculating the current deviation. In this case, the module 120 may calculate the slice weight per the pre-defined period. The period may be, e.g., predefined or configured dynamically. The predefined period may be, for example, a time slot.

Thus, the module 120 continuously calculates the deviation value as the difference between the target volume and the measured allocation volume and applies a correction based on the proportional, integral, and derivative terms. In the improved PID controlling process, the neural network coefficients K _s, 1, K _s, 2, K _s, 3 are tuned in order to make the PID controlling process work properly and minimize the converging time, which will be described with respect to the following embodiments.

FIG. 3 shows a radio resource allocation system 300 in which example embodiments of the present disclosure can be implemented. Referring to the FIG. 3,

Units

310, 312, 320, 322, 324, and 330 may be included in the module 120, a unit 340 may function as the scheduler 110 and the module 130 for allocation counting. The radio resource allocation system 300 may be implemented as another form of the radio resource allocation system 200. In the radio resource allocation system 300, the module 120 may be implemented as a PID neural network. The

units

310, 312, 320, 322, 324, and 330 may be neurons. The improved PID controlling process may constantly learn from the recent deviations and be adaptive to different radio environments and configurations.

In the description with respect to the FIG. 3, it is assumed that the improved PID controlling process is for the slice s, and thus label “s” may be omitted. And

and

are the input and output of the ith node in the lth layer respectively.

The input layer, which may be denoted as 1 ^st layer or layer 1, includes units 310 and 312, and the status transfer function of the input layer may be determined as the formulae (8) and (9):

In the layer 1, the module 120 may calculating the current deviation of the current allocation volume from the target allocation volume.

The hidden layer, which may be denoted as 2 ^nd layer or layer 2, includes units 320, 322 and 324. In order to simplify the computation, the neural network weight between the input layer and the hidden layer may be fixed to (1, -1) , so the status transfer function of the hidden layer may be determined as the formulae (10) and (11) :

The

is the current deviation. The

is be the previous deviation such that the

reflect the previous deviation. The

may be the accumulated deviation calculated based on EMA. w ₁, w ₂, and w ₃ are the neural network coefficients of terms

and

respectively. The w ₁, w ₂, and w ₃ may be updated by the module 120. The previous deviation and the accumulated deviation may embody and be included in the historical deviations.

Thus, the module 120 may calculate the slice weight based on the current deviation, the historical deviations and the updated neural network coefficients in an output layer.

The output layer, which may be denoted as 3 ^rd layer or layer 3, includes unit 330. The status transfer function of the output layer may be determined as the formula (12) :

The input

to the unit 330 may be the changing value, and in the unit 330, the module 120 may calculate the slice weight based on the changing value. The

may be the calculated slice weight W _s (t) .

Then, the module 120 may apply the calculated slice weight for the slice s to the scheduler 110 to determine a next allocation volume for the respective at least one slice. The next actual allocation volume output from the unit 340 may be measured and input to the unit 312 for the next loop. The determined next allocation volume may be iteratively used as the current allocation volume after a period starting from calculating the current deviation. The period may be, e.g., predefined or configured dynamically.

The neural network coefficients w ₁, w ₂, and w ₃ may correspond to the the neural network coefficients K _s, 1, K _s, 2, and K _s, 3, respectively. The module 120 may calculate a learning rate and a gradient of a cost function of respective neural network coefficients by backpropagation for the respective at least one slice and update the respective neural network coefficients based on the learning rate and the gradient for the respective at least one slice, e.g., the slice s. In an embodiment, the cost function may be defined by the current deviation and a previous value of the cost function of the respective at least one slice. In an embodiment, the module 120 may update the respective neural network coefficients by, for example, gradient descent.

In order to avoid the impact of long-term accumulation deviation, the cost function may be defined based on EMA. For example, the cost function at time t may be the denoted as the formula (13) :

where τ is the length of the averaging time window.

The initial value of J _s (t) at time 0 may be calculated as the formula (14) :

Based on the backpropagation, the gradient descent may be used for updating the neural network coefficients to minimize the cost function. In this case, the module 120 may update the respective neural network coefficients by gradient descent based on the learning rate and the gradient. For example, the neural network coefficient, which may be referred to as weight w _j, may be updated based on the formula (15) :

where j is the index of the connection of the neurons in the layer, for example, w ₁, w ₂, and w ₃ in the layer 2, η _j is the learning rate, and

is the gradient of the cost function J with respect to the neural network coefficients which may be referred to as weights w _j.

The gradient of the cost function of respective neural network coefficients may be calculated by backpropagation For example, according to the neural network above, the gradient of cost function may be denoted as the formula (16) :

where s is the index of the slice.

The differential items in the formula (16) can be derived as the formulae (17) to (20) :

Because there is no precise knowledge about the scheduler 110 in a working RAN system, the differential item

cannot be defined accurately. At least one derivative for calculating the gradient of the cost function may be embodied as a sign function and/or a Heaviside step function. The sign function may be, for example, the function sgn () in the above formulae, and it may be appreciated that the Heaviside step function may also apply to derive differential items in the formula (16) . The sign function and/or the Heaviside step function are good enough to generate an estimation. Such embodiment can also be applied to

to avoid 0 as denominator issue.

FIG. 4 shows a radio resource allocation system 400 in which example embodiments of the present disclosure can be implemented. Referring to the FIG. 4, there are a plurality of slices labeled as 0, 1, …, s, which correspond to target allocation volumes

and allocation volume

respectively. The

units

411 and 412, the

units

413 and 414, and the

units

415 and 416 may be similar to the

units

310 and 312, respectively. The

units

421, 422 and 423, the

units

424, 425 and 426, and the

units

427, 428 and 429 may be similar to the

units

320, 322 and 324, respectively. The unit 431, the unit 433, and the unit 435 may be similar to the unit 330. So repetitive descriptions thereof may be omitted. Referring to the FIG. 4, Units 411 to 416, 421 to 429, 431, 433, 435, 441, 443, and 445 may be included in the module 120, a unit 450 may function as the scheduler 110 and the module 130 for allocation counting. It may be appreciated that the module 120 may include further units not shown for the slices other than the

slices

0, 1, s.

In an embodiment, the module 120 may perform a step to develop a comprehensive slice weight for a plurality of slices. For example, the module 120 may integrate the calculated slice weight for each of the respective at least one slice based on a plurality of calculated slice weights for the plurality of slices. Compared to the neural network shown in the FIG. 3, the neural network shown in the FIG. 4 include an extra layer 4 to integrate the slice weights. Referring to the FIG. 4, the layer 4 includes

units

441, 443, and 445. In an embodiment, because the scheduler 110 uses the slice weight as a relative value among the plurality of slices, to limit the value of the slice weight in a reasonable range by normalization may avoid overflow.

In a case where the module 120 perform the integration for the plurality of slice weights by way of, for example, normalization or weighted normalization, the status transfer function of the layer 4 may be determined as the formulae (21) and (22) in case of normalization:

Or as the formulae (23) and (24) in case of weighted normalization, where k _s is the weight factor for different slices:

The

represent the integrated calculated slice weight for the slice n. The module 120 may apply the integrated calculated slice weight for the respective at least one slice to the scheduler 110 to determine a next allocation volume for the respective at least one slice. The next actual allocation volume output from the unit 450 may be measured and input to the

units

412, 414, and 416 for the next loop, respectively. The determined next allocation volume may be iteratively used as the current allocation volume after a period starting from calculating the current deviation.

The gradient of cost function may be denoted as the formulae (25) :

Comparing to the formula (16) , the additional differential items in the formula (25) can be extended as the formulae (26) and (27) :

The gradient may be may be denoted as the formula (28) :

The common part may be denoted as the formula (29) :

The iteration form of the gradient may be denoted as the formula (30) :

The gradient for each neural network weight at time t may be denoted as the formulae (31) to (33) :

G _s, 1 (t) =C _s (t) *E _s (t) (31)

G _s, 2 (t) =C _s (t) * (E _s (t) ) -E _s (t-1) ) (33)

The neural network coefficient w _s, j for slice s and neuron j may be updated based on the formula (34) :

w _s, j (t) =w _s, j (t-1) -η _s (t) *G _s, j (t) (34)

where the learning rate may be chosen according to Lyapunov stability theory.

The learning rate of respective neural network coefficients may be calculated by backpropagation. For example, the learning rate η satisfies the formula (35) :

The learning rate may be updated based on the formula (36) :

where α and β may be predefined parameters, which can affect the reaction and convergence time of the process. For example, α=0.2 and β=1000.

FIG. 5 shows a simulation result of radio resource allocation according to an example embodiment of the present disclosure. The performance evaluation shown in the FIG. 5 is done by MORSE 5G simulation tool. The simulation is carried out on the channel model specified in 3GPP specification TS 38.901, in a 5G time division duplexing (TDD) cell with a bandwidth of 100 MHz (273 PRBs) . The number of simultaneously scheduled UEs is 8. Suppose 3 slices are configured, and each slice has 5 UEs, which are randomly located in one cell. Slice target configuration is listed in Table 1, and 5G RAN configuration is listed in Table 2.

Table 1:

Slice identifier	Quota target
Slice
0	19.04 %
Slice
1	38.09 %
Slice
2	42.85 %

Table 2:

In the FIG. 5, the slice volume over time shows how the PRB share of respective slice changes over time. The FIG. 5 shows that a stable PRB resource allocation may be achieved and the constant slice weight is achieved in a very short convergence time.

FIG. 6 shows a flow chart illustrating an example method 600 for radio resource allocation according to embodiments of the present disclosure. The example method 600 may be performed for example at a network device comprising the module 120 above.

Referring to the FIG. 6, the example method 600 may include a process to develop a slice weight for scheduling comprising: an operation 610 of calculating a current deviation of a current allocation volume from a target allocation volume for respective at least one slice; an operation 620 of calculating a learning rate and a gradient of a cost function of respective neural network coefficients by backpropagation for the respective at least one slice; an operation 630 of updating the respective neural network coefficients based on the learning rate and the gradient for the respective at least one slice; and an operation 640 of calculating the slice weight based on the current deviation, historical deviations, and the updated neural network coefficients for the respective at least one slice.

Details of the operation 610 have been described in the above descriptions with respect to at least the unit 210, and repetitive descriptions thereof are omitted here.

Details of the operation 620 have been described in the above descriptions with respect to at least the learning rate and the gradient of the cost function, and repetitive descriptions thereof are omitted here.

Details of the operation 630 have been described in the above descriptions with respect to at least the neural network coefficients, and repetitive descriptions thereof are omitted here.

Details of the operation 640 have been described in the above descriptions with respect to at least the unit 250 and the unit 330, and repetitive descriptions thereof are omitted here.

In an embodiment, the historical deviations may comprise a previous deviation and an accumulated deviation, and the accumulated deviation may be calculated based on exponential moving average. The more details have been described in the above descriptions with respect to at least the unit 230, unit 240, the unit 322 and the unit 324, and repetitive descriptions thereof are omitted here.

In an embodiment, the cost function may be defined by the current deviation and a previous value of the cost function of the respective at least one slice. The more details have been described in the above descriptions with respect to at least the cost function, and repetitive descriptions thereof are omitted here.

In an embodiment, the cost function may be defined based on EMA, and the respective neural network coefficients may be updated by the gradient descent for minimizing the cost function. The more details have been described in the above descriptions with respect to at least the cost function and the gradient descent, and repetitive descriptions thereof are omitted here.

In an embodiment, at least one derivative for calculating the gradient of the cost function is embodied as a sign function and/or a Heaviside step function. The more details have been described in the above descriptions with respect to at least the sign function and the Heaviside step function, and repetitive descriptions thereof are omitted here.

In an embodiment, the example method 600 may further include an operation of applying the calculated slice weight for the respective at least one slice to a scheduler to determine a next allocation volume for the respective at least one slice. The more details have been described in the above descriptions with respect to at least the scheduler 110, the unit 260 and the unit 340, and repetitive descriptions thereof are omitted here.

In an embodiment, the example method 600 may further include a step to develop a comprehensive slice weight for a plurality of slices comprising: an operation of integrating the calculated slice weight for each of the respective at least one slice based on a plurality of calculated slice weights for the plurality of slices. The more details have been described in the above descriptions with respect to at least the unit 441, the unit 443 and the unit 445, and repetitive descriptions thereof are omitted here.

In an embodiment, the example method 600 may further include an operation of applying the integrated calculated slice weight for the respective at least one slice to a scheduler to determine a next allocation volume for the respective at least one slice. The more details have been described in the above descriptions with respect to at least the scheduler 110 and the unit 450, and repetitive descriptions thereof are omitted here.

In an embodiment, the determined next allocation volume may be iteratively used as the current allocation volume after a period starting from calculating the current deviation. The more details have been described in the above descriptions with respect to at least the period, and repetitive descriptions thereof are omitted here.

In an embodiment, the radio resource for allocation may comprise time and frequency resource, or throughput. The more details have been described in the above descriptions with respect to at least the time and frequency resource as well as the throughput, and repetitive descriptions thereof are omitted here.

FIG. 7 shows a block diagram illustrating an example apparatus 700 for radio resource allocation according to embodiments of the present disclosure. The apparatus, for example, may be a network device comprising the module 120 above.

As shown in the FIG. 7, the example apparatus 700 may include at least one processor 710 and at least one memory 720 that may include computer program code 730. The at least one memory 720 and the computer program code 730 may be configured to, with the at least one processor 710, cause the apparatus 700 at least to perform the example method 600 described above.

In various example embodiments, the at least one processor 710 in the example apparatus 700 may include, but not limited to, at least one hardware processor, including at least one microprocessor such as a central processing unit (CPU) , a portion of at least one hardware processor, and any other suitable dedicated processor such as those developed based on for example Field Programmable Gate Array (FPGA) and Application Specific Integrated Circuit (ASIC) . Further, the at least one processor 710 may also include at least one other circuitry or element not shown in the FIG. 7.

In various example embodiments, the at least one memory 720 in the example apparatus 700 may include at least one storage medium in various forms, such as a volatile memory and/or a non-volatile memory. The volatile memory may include, but not limited to, for example, a random-access memory (RAM) , a cache, and so on. The non-volatile memory may include, but not limited to, for example, a read only memory (ROM) , a hard disk, a flash memory, and so on. Further, the at least memory 820 may include, but are not limited to, an electric, a magnetic, an optical, an electromagnetic, an infrared, or a semiconductor system, apparatus, or device or any combination of the above.

Further, in various example embodiments, the example apparatus 700 may also include at least one other circuitry, element, and interface, for example at least one I/O interface, at least one antenna element, and the like.

In various example embodiments, the circuitries, parts, elements, and interfaces in the example apparatus 700, including the at least one processor 710 and the at least one memory 720, may be coupled together via any suitable connections including, but not limited to, buses, crossbars, wiring and/or wireless lines, in any suitable ways, for example electrically, magnetically, optically, electromagnetically, and the like.

It is appreciated that the structure of the apparatus as a network device comprising the module 120 above is not limited to the above example apparatus 700.

FIG. 8 shows a block diagram illustrating an example apparatus 800 for radio resource allocation according to embodiments of the present disclosure. The apparatus, for example, may be a network device comprising the module 120 above.

As shown in the FIG. 8, the example apparatus 800 may include means 810 for performing the operation 610 of the example method 600, means 820 for performing the operation 620 of the example method 600, means 830 for performing the operation 630 of the example method 600, and means 840 for performing the operation 640 of the example method 6008. In one or more another example embodiments, at least one I/O interface, at least one antenna element, and the like may also be included in the example apparatus 800.

In some example embodiments, examples of means in the example apparatus 800 may include circuitries. For example, an example of means 810 may include a circuitry configured to perform the operation 610 of the example method 600, an example of means 820 may include a circuitry configured to perform the operation 620 of the example method 600, an example of means 830 may include a circuitry configured to perform the operation 630 of the example method 600, and an example of means 840 may include a circuitry configured to perform the operation 640 of the example method 600. In some example embodiments, examples of means may also include software modules and any other suitable function entities.

The term “circuitry” throughout this disclosure may refer to one or more or all of the following: (a) hardware-only circuit implementations (such as implementations in only analog and/or digital circuitry) ; (b) combinations of hardware circuits and software, such as (as applicable) (i) a combination of analog and/or digital hardware circuit (s) with software/firmware and (ii) any portions of hardware processor (s) with software (including digital signal processor (s) ) , software, and memory (ies) that work together to cause an apparatus, such as a mobile phone or server, to perform various functions) ; and (c) hardware circuit (s) and or processor (s) , such as a microprocessor (s) or a portion of a microprocessor (s) , that requires software (e.g., firmware) for operation, but the software may not be present when it is not needed for operation. This definition of circuitry applies to one or all uses of this term in this disclosure, including in any claims. As a further example, as used in this disclosure, the term circuitry also covers an implementation of merely a hardware circuit or processor (or multiple processors) or portion of a hardware circuit or processor and its (or their) accompanying software and/or firmware. The term circuitry also covers, for example and if applicable to the claim element, a baseband integrated circuit or processor integrated circuit for a mobile device or a similar integrated circuit in server, a cellular network device, or other computing or network device.

Another example embodiment may relate to computer program codes or instructions which may cause an apparatus to perform at least respective methods described above. Another example embodiment may be related to a computer readable medium having such computer program codes or instructions stored thereon. In some embodiments, such a computer readable medium may include at least one storage medium in various forms such as a volatile memory and/or a non-volatile memory. The volatile memory may include, but not limited to, for example, a RAM, a cache, and so on. The non-volatile memory may include, but not limited to, a ROM, a hard disk, a flash memory, and so on. The non-volatile memory may also include, but are not limited to, an electric, a magnetic, an optical, an electromagnetic, an infrared, or a semiconductor system, apparatus, or device or any combination of the above.

Unless the context clearly requires otherwise, throughout the description and the claims, the words “comprise, ” “comprising, ” and the like are to be construed in an inclusive sense, as opposed to an exclusive or exhaustive sense; that is to say, in the sense of “including, but not limited to. ” The word “coupled” , as generally used herein, refers to two or more elements that may be either directly connected, or connected by way of one or more intermediate elements. Likewise, the word “connected” , as generally used herein, refers to two or more elements that may be either directly connected, or connected by way of one or more intermediate elements. Additionally, the words “herein, ” “above, ” “below, ” and words of similar import, when used in this application, shall refer to this application as a whole and not to any particular portions of this application. Where the context permits, words in the description using the singular or plural number may also include the plural or singular number respectively. The word “or” in reference to a list of two or more items, that word covers all of the following interpretations of the word: any of the items in the list, all of the items in the list, and any combination of the items in the list.

Moreover, conditional language used herein, such as, among others, “can, ” “could, ” “might, ” “may, ” “e.g., ” “for example, ” “such as” and the like, unless specifically stated otherwise, or otherwise understood within the context as used, is generally intended to convey that certain embodiments include, while other embodiments do not include, certain features, elements and/or states. Thus, such conditional language is not generally intended to imply that features, elements and/or states are in any way required for one or more embodiments or that one or more embodiments necessarily include logic for deciding, with or without author input or prompting, whether these features, elements and/or states are included or are to be performed in any particular embodiment.

As used herein, the term "determine/determining" (and grammatical variants thereof) can include, not least: calculating, computing, processing, deriving, measuring, investigating, looking up (for example, looking up in a table, a database or another data structure) , ascertaining and the like. Also, "determining" can include receiving (for example, receiving information) , accessing (for example, accessing data in a memory) , obtaining and the like. Also, "determine/determining" can include resolving, selecting, choosing, establishing, and the like.

While some embodiments have been described, these embodiments have been presented by way of example, and are not intended to limit the scope of the disclosure. Indeed, the apparatus, methods, and systems described herein may be embodied in a variety of other forms; furthermore, various omissions, substitutions and changes in the form of the methods and systems described herein may be made without departing from the spirit of the disclosure. For example, while blocks are presented in a given arrangement, alternative embodiments may perform similar functionalities with different components and/or circuit topologies, and some blocks may be deleted, moved, added, subdivided, combined, and/or modified. At least one of these blocks may be implemented in a variety of different ways. The order of these blocks may also be changed. Any suitable combination of the elements and acts of the some embodiments described above can be combined to provide further embodiments. The accompanying claims and their equivalents are intended to cover such forms or modifications as would fall within the scope and spirit of the disclosure.

Abbreviations used in the description and/or in the figures are defined as follows:

ASC adaptive slicing control

BS base station

EMA exponential moving average

NR new radio

PID proportional–integral–derivative

PRB physical resource block

QoE quality of experience

QoS quality of service

RAN radio access network

SCS subcarrier spacing

TDD time division duplexing

UE user equipment

Claims

A radio resource allocation method, comprising a process to develop a slice weight for scheduling comprising:

calculating a current deviation of a current allocation volume from a target allocation volume for respective at least one slice;

calculating a learning rate and a gradient of a cost function of respective neural network coefficients by backpropagation for the respective at least one slice;

updating the respective neural network coefficients based on the learning rate and the gradient for the respective at least one slice; and

calculating the slice weight based on the current deviation, historical deviations, and the updated neural network coefficients for the respective at least one slice.
The method of claim 1, wherein the historical deviations comprises a previous deviation and an accumulated deviation, and the accumulated deviation is calculated based on exponential moving average.
The method of claim 1 or 2, wherein the cost function is defined by the current deviation and a previous value of the cost function of the respective at least one slice.
The method of claim 3, wherein the cost function is defined based on exponential moving average, and the respective neural network coefficients are updated by gradient descent for minimizing the cost function.
The method of claim 4, wherein at least one derivative for calculating the gradient of the cost function is embodied as a sign function and/or a Heaviside step function.
The method of any of claims 1 to 5, further comprising:

applying the calculated slice weight for the respective at least one slice to a scheduler to determine a next allocation volume for the respective at least one slice.
The method of any of claims 1 to 5, further comprising a step to develop a comprehensive slice weight for a plurality of slices comprising:

integrating the calculated slice weight for each of the respective at least one slice based on a plurality of calculated slice weights for the plurality of slices.
The method of claim 7, further comprising:

applying the integrated calculated slice weight for the respective at least one slice to a scheduler to determine a next allocation volume for the respective at least one slice.
The method of claim 6 or 8, wherein the determined next allocation volume is iteratively used as the current allocation volume after a period starting from calculating the current deviation.
The method of any of claims 1 to 9, wherein the radio resource for allocation comprises time and frequency resource, or throughput.
A radio resource allocation apparatus comprising:

at least one processor; and

at least one memory comprising computer program code, the at least one memory and the computer program code being configured to, with the at least one processor, cause the apparatus to perform a process to develop a slice weight for scheduling comprising:

calculating a current deviation of a current allocation volume from a target allocation volume for respective at least one slice;

calculating a learning rate and a gradient of a cost function of respective neural network coefficients by backpropagation for the respective at least one slice;

updating the respective neural network coefficients based on the learning rate and the gradient for the respective at least one slice; and

calculating the slice weight based on the current deviation, historical deviations, and the updated neural network coefficients for the respective at least one slice.
The apparatus of claim 11, wherein the historical deviations comprises a previous deviation and an accumulated deviation, and the accumulated deviation is calculated based on exponential moving average.
The apparatus of claim 11 or 12, wherein the cost function is defined by the current deviation and a previous value of the cost function of the respective at least one slice.
The apparatus of claim 13, wherein the cost function is defined based on exponential moving average, and the respective neural network coefficients are updated by gradient descent for minimizing the cost function.
The apparatus of claim 14, wherein at least one derivative for calculating the gradient of the cost function is embodied as a sign function and/or a Heaviside step function.
The apparatus of any of claims 11 to 15, wherein the at least one memory and the computer program code are further configured to, with the at least one processor, cause the apparatus to further perform:

applying the calculated slice weight for the respective at least one slice to a scheduler to determine a next allocation volume for the respective at least one slice.
The apparatus of any of claims 11 to 15, wherein the at least one memory and the computer program code are further configured to, with the at least one processor, cause the apparatus to further perform a step to develop a comprehensive slice weight for a plurality of slices comprising:

integrating the calculated slice weight for each of the respective at least one slice based on a plurality of calculated slice weights for the plurality of slices.
The apparatus of claim 17, wherein the at least one memory and the computer program code are further configured to, with the at least one processor, cause the apparatus to further perform:

applying the integrated calculated slice weight for the respective at least one slice to a scheduler to determine a next allocation volume for the respective at least one slice.
The apparatus of claim 16 or 18, wherein the determined next allocation volume is iteratively used as the current allocation volume after a period starting from calculating the current deviation.
The apparatus of any of claims 11 to 19, wherein the radio resource for allocation comprises time and frequency resource, or throughput.
A radio resource allocation apparatus, comprising means for developing a slice weight for scheduling comprising:

means for calculating a current deviation of a current allocation volume from a target allocation volume for respective at least one slice;

means for calculating a learning rate and a gradient of a cost function of respective neural network coefficients by backpropagation for the respective at least one slice;

means for updating the respective neural network coefficients based on the learning rate and the gradient for the respective at least one slice; and

means for calculating the slice weight based on the current deviation, historical deviations, and the updated neural network coefficients for the respective at least one slice.
A computer readable medium comprising program instructions for causing a radio resource allocation apparatus to perform a process to develop a slice weight for scheduling comprising:

calculating a current deviation of a current allocation volume from a target allocation volume for respective at least one slice;

calculating a learning rate and a gradient of a cost function of respective neural network coefficients by backpropagation for the respective at least one slice;

updating the respective neural network coefficients based on the learning rate and the gradient for the respective at least one slice; and

calculating the slice weight based on the current deviation, historical deviations, and the updated neural network coefficients for the respective at least one slice.