Devices and methods for negotiation and signalling for collaborative control of power amplifier non-linearities

Abstract

Devices and methods for negotiation and signalling for collaborative control of power amplifier non-linearities are disclosed. At least some example embodiments may allow signalling to support design and deployment of the collaborative control of the power amplifier non-linearities for digital pre-distortion (DPD) application in a user device and digital post-distortion (DPoD) application in a network node where the digital pre-distortion may be used to control out-of- band emissions (e.g., adjacent channel leakage ratio (ACLR)) and the digital post-distortion may be used to control in-band emissions (e.g., error vector magnitude (EVM)).

Description

[0001] DEVICES AND METHODS FOR NEGOTIATION AND SIGNALLING FOR COLLABORATIVE CONTROL OF POWER AMPLIFIER NON-LINEARITIES

[0002] TECHNICAL FIELD

[0003] The disclosure relates generally to communications and, more particularly but not exclusively, to negotiation and signalling for collaborative control of power ampli fier non-linearities , as well as related devices and methods .

[0004] BACKGROUND

[0005] Power ampli fiers used in radio transmitters included, e . g . , in user devices , such as mobile phones and the like , may have non-linearity or distortion that needs to be controlled to improve power e f ficiency and overall performance .

[0006] Furthermore , strict emission requirements have been introduced for the user devices in fi fth generation ( 5G) and sixth generation ( 6G) wireless networks , for example .

[0007] Accordingly, at least in some situations , there may be a need for controlling the non-linearity in the power ampli fiers used in the user device radio transmitters in way that also takes into account the emiss ion requirements .

[0008] BRIEF SUMMARY

[0009] The scope of protection sought for various example embodiments of the invention is set out by the independent claims . The example embodiments and features , i f any, described in this speci fication that do not fall under the scope o f the independent claims are to be interpreted as examples useful for understanding various example embodiments of the invention . An example embodiment of a user device comprises at least one processor, and at least one memory storing instructions that , when executed by the at least one processor, cause the user device at least to carry out , for collaborative power amplifier non-linearity control with a network node , a negotiation procedure with the network node . The negotiation procedure comprises negotiating whether the user device carries out training of a first machine learning model applicable to digital pre-distortion for controlling out-of-band emissions in advance or subsequently to training of a second machine learning model applicable to digital post-distortion for controlling in-band emissions . In the case the training of the first machine learning model is carried out in advance to the training of the second machine learning model , the instructions , when executed by the at least one processor, further cause the user device at least to transmit , to the network node , information on the completion of the training of the first machine learning model along with information for labelling input data for the training of the second machine learning model . In the case the training of the first machine learning model is carried out subsequently to the training of the second machine learning model , the instructions , when executed by the at least one processor, further cause the user device at least to receive , from the network node , information on output parameters of the training of the second machine learning model , and carry out the training o f the first machine learning model utili zing the information .

[0010] In an example embodiment , alternatively or in addition to the above-described example embodiments , the negotiation procedure further comprises transmitting, to the network node , a request for the training of the first machine learning model and / or of the second machine learning model , or receiving, from the network node a request for the training of the first machine learning model and / or the second machine learning model .

[0011] In an example embodiment , alternatively or in addition to the above-described example embodiments , the negotiation procedure further comprises coordination signalling for using a common cost function .

[0012] In an example embodiment , alternatively or in addition to the above-described example embodiments , the negotiation procedure further comprises control of a duration of the training o f the first machine learning model and / or the training of the second machine learning model .

[0013] In an example embodiment , alternatively or in addition to the above-described example embodiments , the first machine learning model is applicable to control an adj acent channel leakage ratio and the second machine learning model is applicable to control an error vector magnitude .

[0014] In an example embodiment , alternatively or in addition to the above-described example embodiments , the instructions , when executed by the at least one processor, further cause the user device to determine to use another first machine learning model . The instructions , when executed by the at least one processor, further cause the user device to transmit an indication of the another first machine learning model to the network node . I f training is needed, the instructions , when executed by the at least one processor, further cause the user device to carry out the negotiation procedure in relation to the another first machine learning model and a second machine learning model related to the another first machine learning model .

[0015] An example embodiment of a network node comprises at least one processor, and at least one memory storing instructions that , when executed by the at least one processor, cause the network node at least to carry out , for collaborative power amplifier non-linearity control with a user device , a negotiation procedure with the user device . The negotiation procedure comprises negotiating whether the user device carries out training of a first machine learning model applicable to digital pre-distortion for controlling out-of-band emissions in advance or subsequently to training of a second machine learning model applicable to digital post-distortion for controlling in-band emissions . In the case the training of the first machine learning model is carried out in advance to the training of the second machine learning model , the instructions , when executed by the at least one processor, further cause the network node at least to receive , from the user device , information on the completion of the training of the first machine learning model along with information for labelling input data for the training o f the second machine learning model , and carry out the training of the second machine learning model . In the case the training of the first machine learning model is carried out subsequently to the training o f the second machine learning model , the instructions , when executed by the at least one processor, further cause the network node at least to transmit , to the user device , information on output parameters of the training of the second machine learning model for the training of the first machine learning model .

[0016] In an example embodiment , alternatively or in addition to the above-described example embodiments , the negotiation procedure further comprises receiving, from the user device , a request for the training of the first machine learning model and / or of the second machine learning model , or transmitting, to the user device a request for the training o f the first machine learning model and / or the second machine learning model .

[0017] In an example embodiment , alternatively or in addition to the above-described example embodiments , the negotiation procedure further comprises coordination signalling for using a common cost function . In an example embodiment , alternatively or in addition to the above-described example embodiments , the negotiation procedure further comprises control of a duration of the training o f the first machine learning model and / or the training of the second machine learning model .

[0018] In an example embodiment , alternatively or in addition to the above-described example embodiments , the first machine learning model is applicable to control an adj acent channel leakage ratio and the second machine learning model is applicable to control an error vector magnitude .

[0019] In an example embodiment , alternatively or in addition to the above-described example embodiments , the instructions , when executed by the at least one processor, further cause the network node to receive from the user device an indication of another first machine learning model determined to be used . I f training is needed, the instructions , when executed by the at least one processor, further cause the network node to carry out the negotiation procedure in relation to the another first machine learning model and a second machine learning model related to the another first machine learning model .

[0020] An example embodiment of a method comprises carrying out , by a user device for collaborative power ampli fier non-linearity control with a network node , a negotiation procedure with the network node . The negotiation procedure comprises negotiating whether the user device carries out training of a first machine learning model applicable to digital pre-distortion for controlling out-of-band emissions in advance or subsequently to training of a second machine learning model applicable to digital post-distortion for controlling in-band emissions . In the case the training of the first machine learning model is carried out in advance to the training of the second machine learning model , the method further comprises transmitting, from the user device to the network node , information on the completion of the training of the first machine learning model along with information for labelling input data for the training of the second machine learning model . In the case the training of the first machine learning model is carried out subsequently to the training of the second machine learning model , the method further comprises receiving, at the user device from the network node , information on output parameters of the training of the second machine learning model , and carrying out, by the user device , the training of the first machine learning model utilizing the information .

[0021] In an example embodiment , alternatively or in addition to the above-described example embodiments , the negotiation procedure further comprises transmitting, to the network node , a request for the training of the first machine learning model and / or of the second machine learning model , or receiving, from the network node a request for the training of the first machine learning model and / or the second machine learning model .

[0022] In an example embodiment , alternatively or in addition to the above-described example embodiments , the negotiation procedure further comprises coordination signalling for using a common cost function .

[0023] In an example embodiment , alternatively or in addition to the above-described example embodiments , the negotiation procedure further comprises control of a duration of the training o f the first machine learning model and / or the training of the second machine learning model .

[0024] In an example embodiment , alternatively or in addition to the above-described example embodiments , the first machine learning model is applicable to control an adj acent channel leakage ratio and the second machine learning model is applicable to control an error vector magnitude . In an example embodiment , alternatively or in addition to the above-described example embodiments , the method further comprises determining, by the user device , to use another first machine learning model . The method further comprises transmitting an indication of the another f irst machine learning model from the user device to the network node . I f training is needed, the method further comprises carrying out , by the user device , the negotiation procedure in relation to the another first machine learning model and a second machine learning model related to the another first machine learning model .

[0025] An example embodiment of a method comprises carrying out , by a network node for collaborative power ampli fier non-linearity control with a user device , a negotiation procedure with the user device . The negotiation procedure comprises negotiating whether the user device carries out training of a first machine learning model applicable to digital pre-distortion for controlling out-of-band emissions in advance or subsequently to training of a second machine learning model applicable to digital post-distortion for controlling in-band emissions . In the case the training of the first machine learning model is carried out in advance to the training of the second machine learning model , the method further comprises receiving, at the network node from the user device , information on the completion of the training of the first machine learning model along with information for labelling input data for the training of the second machine learning model , and carrying out , by the network node , the training of the second machine learning model . In the case the training of the first machine learning model is carried out subsequently to the training of the second machine learning model , the method further comprises transmitting, from the network node to the user device , information on output parameters of the training of the second machine learning model for the training of the first machine learning model .

[0026] In an example embodiment , alternatively or in addition to the above-described example embodiments , the negotiation procedure further comprises receiving, from the user device , a request for the training of the first machine learning model and / or of the second machine learning model , or transmitting, to the user device a request for the training o f the first machine learning model and / or the second machine learning model .

[0027] In an example embodiment , alternatively or in addition to the above-described example embodiments , the negotiation procedure further comprises coordination signalling for using a common cost function .

[0028] In an example embodiment , alternatively or in addition to the above-described example embodiments , the negotiation procedure further comprises control of a duration of the training o f the first machine learning model and / or the training of the second machine learning model .

[0029] In an example embodiment , alternatively or in addition to the above-described example embodiments , the first machine learning model is applicable to control an adj acent channel leakage ratio and the second machine learning model is applicable to control an error vector magnitude .

[0030] In an example embodiment , alternatively or in addition to the above-described example embodiments , the method further comprises receiving at the network node from the user device an indication of another first machine learning model determined to be used . I f training is needed, the method further comprises carrying out , by the network node , the negotiation procedure in relation to the another first machine learning model and a second machine learning model related to the another first machine learning model . An example embodiment of an apparatus comprises means for carrying out a method according to any of the above-described example embodiments .

[0031] An example embodiment of a computer program comprises instructions for causing a user device to perform at least the following : carrying out , for collaborative power ampli fier non-linearity control with a network node , a negotiation procedure with the network node , the negotiation procedure comprising negotiating whether the user device carries out training of a first machine learning model applicable to digital pre-dis- tortion for controlling out-of-band emissions in advance or subsequently to training of a second machine learning model applicable to digital post-distortion for controlling in-band emissions ; in the case the training of the f irst machine learning model is carried out in advance to the training of the second machine learning model , transmitting to the network node information on the completion of the training of the first machine learning model along with information for labelling input data for the training of the second machine learning model ; and in the case the training of the first machine learning model is carried out subsequently to the training of the second machine learning model , receiving from the network node information on output parameters of the training of the second machine learning model , and carrying out the training of the first machine learning model utili zing the information .

[0032] An example embodiment of a computer program comprises instructions for causing a network node to perform at least the following : carrying out , for collaborative power amplifier non-linearity control with a user device , a negotiation procedure with the user device , the negotiation procedure comprising negotiating whether the user device carries out training of a first machine learning model applicable to digital pre-dis- tortion for controlling out-of-band emissions in advance or subsequently to training of a second machine learning model applicable to digital post-distortion for controlling in-band emissions ; in the case the training of the f irst machine learning model is carried out in advance to the training of the second machine learning model , receiving from the user device information on the completion of the training of the first machine learning model along with information for labelling input data for the training of the second machine learning model , and carrying out the training of the second machine learning model ; and in the case the training of the first machine learning model is carried out subsequently to the training of the second machine learning model , transmitting to the user device information on output parameters of the training of the second machine learning model for the training of the first machine learning model .

[0033] DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings , which are included to provide a further understanding of the embodiments and constitute a part of this specification, illustrate embodiments and together with the description help to explain the principles of the embodiments . In the drawings :

[0035] FIG . 1 shows an example embodiment of the subj ect matter described herein illustrating an example system, where various embodiments of the present disclosure may be implemented;

[0036] FIG . 2A shows an example embodiment of the subj ect matter described herein illustrating a user device ;

[0037] FIG . 2B shows an example embodiment of the subj ect matter described herein illustrating a network node ; FIG . 3A shows an example embodiment of the subj ect matter described herein illustrating an architecture for the disclosed collaborative control of power ampli fier non-linearities ;

[0038] FIG . 3B shows an example embodiment of the subj ect matter described herein illustrating an architecture for the disclosed collaborative training between a user device and a network node for the control of the power ampli fier non-linearities ;

[0039] FIG . 4 shows an example embodiment of the subj ect matter described herein illustrating the disclosed collaborative training between a user device and a network node ;

[0040] FIG . 5 shows another example embodiment of the subj ect matter described herein illustrating the disclosed collaborative training between a user device and a network node ;

[0041] FIG . 6 shows an example embodiment of the subj ect matter described herein illustrating disclosed synchroni zation of machine learning models when switching to new machine learning models ;

[0042] FIG . 7 shows an example embodiment of the subj ect matter described herein illustrating a disclosed method for a user device ; and

[0043] FIG . 8 shows an example embodiment of the subj ect matter described herein illustrating a disclosed method for a network node .

[0044] Like reference numerals are used to designate like parts in the accompanying drawings .

[0045] DETAILED DESCRIPTION

[0046] Reference will now be made in detail to embodiments , examples of which are illustrated in the accompanying drawings . The detailed description provided below in connection with the appended drawings is intended as a description of the present examples and is not intended to represent the only forms in which the present example may be constructed or utilized. The description sets forth the functions of the example and the sequence of steps for constructing and operating the example. However, the same or equivalent functions and sequences may be accomplished by different examples.

[0047] Fig. 1 illustrates an example system 100, where various embodiments of the present disclosure may be implemented. The system 100 may comprise, e.g., a fifth generation (5G) or sixth generation (6G) communications network 110. An example representation of the system 100 is shown depicting a user device 200 and a network node 210. At least in some embodiments, the communications network 110 may comprise one or more massive machine- to-machine (M2M) network (s) , massive machine type communications (mMTC) network(s) , internet of things (ToT) network(s) , industrial internet-of-things (IIoT) network^) , enhanced mobile broadband (eMBB) network(s) , ultra-reliable low-latency communication (URLLC) network^) , and / or the like. In other words, the communications network 110 may be configured to serve diverse service types and / or use cases, and it may logically be seen as comprising one or more networks.

[0048] The user device 200 may include, e.g., a mobile phone, a smartphone, a tablet computer, a smart watch, or any hand-held, portable and / or wearable device. The user device 200 may also be referred to as a user equipment (UE) . The network node 210 may comprise, e.g., a base station or an access node. The base station may include, e.g., a 5G or 6G base station (gNB) or any device suitable for providing an air interface for client devices to connect to a wireless network via wireless transmissions.

[0049] In the following, various example embodiments will be discussed. At least some of these example embodiments described herein may allow negotiation and signalling for collaborative control of power amplifier non-linearities .

[0050] Furthermore, at least some of the example embodiments described herein may allow signalling to support design and deployment of the collaborative control of the power amplifier non-linearities for digital predistortion (DPD) application in a user device and digital post-distortion (DPoD) application in a network node where the digital pre-distortion may be used to control out-of-band emissions (e.g., adjacent channel leakage ratio (ACLR) ) and the digital post-distortion may be used to control in-band emissions (e.g., error vector magnitude (EVM) ) .

[0051] Furthermore, at least some of the example embodiments described herein may allow reducing computational load at the user device 200.

[0052] Furthermore, at least some of the example embodiments described herein may allow providing robustness to channel conditions when training accounts for different channel and signal-to-interf erence-plus-noise ratio (SINR) regimes.

[0053] Diagram 300 of Fig. 3A illustrates an example of an architecture or structure that may be used for the disclosed collaborative control of power amplifier nonlinearities, and diagram 350 of Fig. 3B illustrates an example of an architecture or structure that may be used for the disclosed collaborative training between the user device 200 and the network node 210 for the control of the power amplifier non-linearities.

[0054] In the examples of diagrams 300 and 350, blocks 301. 302, 303, 304, 309, 208 may be implemented in the user device 200 (TX side) , and blocks 306, 307, 308 may be implemented in the network node 210 (RX side) .

[0055] The "partially learned" in diagrams 300, 350 indicates that the respective block implements a combination between digital signal processing and machine learning models. Moreover, in some embodiments, the observation receiver 309 may be replaced by a power detector configured to detect power of out-of-band emissions, as at least in some embodiments this may be sufficient for training a fully learned DPD.

[0056] In the examples of diagrams 300 and 350, the user device 200 may generate a bit stream b (block 301) which may be encoded and modulated into a complex symbol vector s (block 302) that may be orthogonal frequencydivision multiplexing (OFDM) -modulated (block 303) to yield an OFDM signal ri . The signal ri may be passed through a first machine learning (ML) model 251 (in block 304) whose objective is to control (e.g., suppress or reduce) the ACLR of an amplified signal at the output of the power amplifier 208. Also, a vector 0Xmay include trainable parameters of the first ML model 251 as the DPD at the user device 200. Thus, the first ML model 251 may be modelled as a function 01 (0!, ri) which may be implemented as a first neural network with N>1 hidden layers that takes as input the vector ri and produces the vector ro = G1 (0X, ri) . The first neural network may comprise a deep neural network, a convolutional neural network, or any other suitable neural network, or any combination thereof. The output signal of the first ML model 251 / DPD may be fed to the power amplifier 208 which may generate a corresponding output signal ro. The signal ro may be processed by further TX blocks and the resulting signal at the output of a transmit antenna array 206 is denoted as y. The output signal ro may be expected to be affected by high EVM but low ACLR as a result of the DPD used before the power amplifier 208.

[0057] The signal y may be sent over the wireless channel 305 with an impulse response h and the network node 210 may receive a signal u which is a convolution between the transmitted signal y and the channel h, corrupted by additive white Gaussian noise (AWGN) n. The signal u may be OFDM demodulated (block 306) into a signal u' and then input to a second ML model 252 (in block 307) whose objective is to remove EVM. Trainable parameters of the second ML model as the DPoD at the network node 210 may be collected in a vector 02. The second ML 252 may thus be seen as implementing a function G2 (02, u' ) which may be implemented as a second neural network with N2>1 hidden layers. The second neural network may comprise a deep neural network, a convolutional neural network, or any other suitable neural network, or any combination thereof. Next, channel estimation, equalization and detection may be performed, and the estimated symbols s' may further be de-mapped and decoded to RX bits b' (block 308) . At least in some embodiments, the second ML model 252 may also learn to do other baseband (BB) processing tasks, feeding its output directly to a channel decoder (i.e., a fully learned radio receiver may be implemented) .

[0058] In the examples of diagrams 300 and 350, block 310 represents a first intermediate cost function fl (described in more detail below) , block 311 represents a second intermediate cost function f2 (described in more detail below) , and block 312 represents a total or combined cost function f (described in more detail below) .

[0059] The DPD in the user device 200 may have various different architectures. For example, the user device 200 may comprise a full observation receiver 309 if the user device 200 is relying on a conventional DPD, or the user device 200 may have a power detector for measuring the out-of-band emissions, if the DPD is implemented using a neural network or ML model -based approach (fully learned) . In the latter case, one may learn the DPD parameters 0Xby observing power of out-of-band emissions, for example.

[0060] Fig. 2A is a block diagram of the user device 200, in accordance with an example embodiment. The user device 200 comprises one or more processors 202 and one or more memories 204 that comprise computer program code. The user device 200 may be configured to transmit information to and receive information from other devices. In one example, the user device 200 may transmit / receive signalling information and data in accordance with at least one cellular communication protocol. The user device 200 may be configured to provide at least one wireless radio connection, such as for example a 3GPP mobile broadband connection (e.g., 5G or 6G) . The user device 200 may comprise, or be configured to be coupled to, at least one antenna 206 to transmit / receive radio frequency signals. The user device 200 may further comprise a power amplifier 208. The user device 200 may also include other elements not shown in Fig. 2A, such as a transceiver.

[0061] Although the user device 200 is depicted to include only one processor 202, the user device 200 may include more processors. In an embodiment, the memory 204 is capable of storing instructions, such as an operating system and / or various applications. Furthermore, the memory 204 may include a storage that may be used to store, e.g., at least some of the information and data used in the disclosed embodiments, such as a first machine learning model 251 described in more detail below.

[0062] Furthermore, the processor 202 is capable of executing the stored instructions. In an embodiment, the processor 202 may be embodied as a multi-core processor, a single core processor, or a combination of one or more multi-core processors and one or more single core processors. For example, the processor 202 may be embodied as one or more of various processing devices, such as a coprocessor, a microprocessor, a controller, a digital signal processor (DSP) , a processing circuitry with or without an accompanying DSP, or various other processing devices including integrated circuits such as , for example , an application specific integrated circuit (AS IC ) , a field programmable gate array ( FPGA) , a microcontroller unit (MCU) , a hardware accelerator, a special-purpose computer chip, a neural network (NN) chip, an arti ficial intelligence (Al ) accelerator, a tensor processing unit ( TPU) , a neural processing unit (NPU) , or the like . In an embodiment , the processor 202 may be configured to execute hard-coded functionality . In an embodiment , the processor 202 is embodied as an executor of software instructions , wherein the instructions may speci fically configure the processor 202 to perform the algorithms and / or operations described herein when the instructions are executed .

[0063] The memory 204 may be embodied as one or more volatile memory devices , one or more non-volatile memory devices , and / or a combination of one or more volatile memory devices and non-volatile memory devices . For example , the memory 204 may be embodied as semiconductor memories ( such as mask ROM, PROM (programmable ROM) , EPROM ( erasable PROM) , flash ROM, RAM ( random access memory) , etc . ) .

[0064] When executed by the at least one processor 202 , instructions stored in the at least one memory 204 cause the user device 200 at least to carry out , for collaborative power ampli fier non-linearity control with the network node 210 , a negotiation procedure with the network node 210 .

[0065] The negotiation procedure comprises negotiating whether the user device 200 carries out training of a first machine learning model 251 applicable to digital pre-distortion ( DPD) for controlling out-of-band emissions in advance or subsequently to training of a second machine learning model 252 applicable to digital postdistortion ( DPoP ) for controlling in-band emissions . At least in some embodiments , the second machine learning model 252 may be comprised in the network node 210 as shown in Figs. 2B and 3A-3B, for example.

[0066] For example, the first machine learning model 251 may be applicable to control an adjacent channel leakage ratio (ACLR) , and the second machine learning model 252 may be applicable to control an error vector magnitude (EVR) . The ACLR refers to a ratio of a filtered mean power centered on an assigned channel frequency to a filtered mean power centered on an adjacent channel frequency. The EVM refers to a modulation quality metric used in digital communications. More specifically, the EVM is a measure of a deviation of actual modulation constellation points from their ideal locations in a modulation constellation diagram.

[0067] At least in some embodiments, the negotiation procedure may further comprise transmitting to the network node 210 a request for the training of the first machine learning model 251 and / or of the second machine learning model 252, or receiving, from the network node 210 a request for the training of the first machine learning model 251 and / or the second machine learning model 252.

[0068] At least in some embodiments, the negotiation procedure may further comprise coordination signalling for using a common (or total or combined) cost function.

[0069] For example, to train the ML models 251, 252, the combined cost function may be used as follows, targeting the two ML models 251, 252 separately:

[0070] - a first intermediate cost function fl (block

[0071] 310 in diagrams 300, 350) may target the first ML model

[0072] 251 and thus aim to minimize the ACLR of the signal y, i.e., aim to minimize the cost function fl = ACLR(y,01) by adjusting the parameters of the first ML model 251; and

[0073] - a second intermediate cost function f2 (block

[0074] 311 in diagrams 300, 350) may target the second ML model

[0075] 252 and thus aim to minimize EVM. That is, the second ML model 252 may learn such parameters 02which result in the distance between the estimated symbols s' and the original symbols s being minimized, e.g., f2 = MSE(s' ,s, 02) • As an alternative, f2 may be defined over the transmitted and detected bits as f2 = CE (b' ,b, 02) , where CE denotes a cross-entropy loss function.

[0076] At least in some embodiments, the first ML model 251 may also target to minimize the second intermediate cost function f2.

[0077] The total or combined cost function f (block 312 in diagrams 300, 350) may use fl and f2, e.g., as follows :

[0078] - in an embodiment, the cost function used to train both the first and second ML models 251, 252 may be formulated, e.g., as: f = f2 constrained by fl< threshold; in another embodiment, the combined cost function may include, e.g. : f = a i + a2f2, where al and a2 weigh the importance of fl vs. f 2 and ax+ a2= 1.

[0079] The trainable parameters of the DPD and DPoD may be obtained, e.g., by minimizing the total / combined loss function:

[0080] 01(02= argmin f 01,02

[0081] At least in some embodiments, the negotiation procedure may further comprise control of a duration of the training of the first machine learning model 251 and / or the training of the second machine learning model 252.

[0082] In the case the training of the first machine learning model 251 is carried out in advance to the training of the second machine learning model 252, the instructions, when executed by the at least one processor 202, further cause the user device 200 at least to transmit, to the network node 210, information on the completion of the training of the first machine learning model 251 along with information for labelling input data for the training of the second machine learning model 252.

[0083] In machine learning, labels in a dataset may refer to, e.g., predefined tags or annotations assigned to the data. Labelling the data may allow to train models to recognize patterns and make predictions based on the provided labels.

[0084] For example, the labelling input data for the training of the second machine learning model 252 may comprise bits (e.g., the transmission (TX) bits 301 or label bits or bits b described above in more detail) generated by the user device 200.

[0085] In the case the training of the first machine learning model 251 is carried out subsequently to the training of the second machine learning model 252, the instructions, when executed by the at least one processor 202, further cause the user device 200 at least to receive, from the network node 210, information on output parameters of the training of the second machine learning model 252, and carry out the training of the first machine learning model 251 utilizing the information .

[0086] For example, the output parameters of the training of the second machine learning model 252 may comprise gradient or parameter values as adjusted.

[0087] In other words, the information on the output parameters of the training of the second machine learning model may be related to using a gradient descent algorithm in training a deep learning model. Gradient descent (GD) is an iterative optimization algorithm commonly used to train machine learning models and neural networks .

[0088] A gradient descent targets to minimize a used cost function, or an error between a predicted outcome and a related actual outcome. The algorithm uses information on a direction of change and a learning rate. These parameters are used to determine partial derivative calculations of future iterations, allowing the algorithm to gradually approach a local or global minimum (i.e., a point of convergence) . A learning rate (also referred to as a step size or alpha) describes a size of steps taken to reach or approach the minimum.

[0089] The cost (or loss) function measures the difference, or error, between an actual outcome and a predicted outcome. This improves machine learning model's efficacy by providing feedback to the model so that it can adjust the parameters to minimize the error and find the local or global minimum. The algorithm is iterative moving along the direction of steepest descent (or negative gradient) until the cost function is close to or at zero. At this point, the training of the model is ended .

[0090] First the gradient that indicates the direction of the steepest increase in the function's value is calculated, then parameter values are adjusted by moving in the opposite direction of the gradient. The learning rate determines the size of the steps taken in each iteration. Iteration of the parameter values is carried out until convergence or until a stopping condition is met. The outcome is optimized parameter values.

[0091] In deep learning, a variant called stochastic gradient descent (SGD) may be used. It updates the parameters based on a randomly selected subset of training samples in each iteration, rather than the entire dataset. This helps in speeding up the training process and making it feasible for large-scale problems. Another method is called mini-batch learning GD.

[0092] Thus, the information on the output parameters of the training of the second machine learning model may comprise gradient or parameter values as adjusted.

[0093] At least in some embodiments, the instructions, when executed by the at least one processor 202, may further cause the user device 200 to determine to use another first machine learning model . The instructions , when executed by the at least one processor 202 , may further cause the user device 200 to transmit an indication of the another first machine learning model to the network node 210 . I f training is needed, the instructions , when executed by the at least one processor 202 , may further cause the user device 200 to carry out the negotiation procedure in relation to the another first machine learning model and a second machine learning model related to the another first machine learning model .

[0094] In other words , at least some of the disclosed embodiments may allow model synchroni zation ( discussed in more detail below in connection with Fig . 6 .

[0095] Fig . 2B is a block diagram of the network node 210 , in accordance with an example embodiment .

[0096] The network node 210 comprises one or more processors 212 and one or more memories 214 that comprise computer program code . The network node 210 may be configured to receive information from and transmit information to other devices . In one example , the network node 210 may receive / transmit signalling information and data in accordance with at least one cellular communication protocol . The network node 210 may be configured to provide at least one wireless radio connection, such as for example a 3GPP mobile broadband connection ( e . g . , 5G or 6G) . The network node 210 may comprise , or be configured to be coupled to , at least one antenna 216 to receive / transmit radio frequency signals . The network node 210 may also include other elements not shown in Fig . 2B, such as a transceiver .

[0097] Although the network node 210 is depicted to include only one processor 212 , the network node 210 may include more processors . In an embodiment , the memory 214 i s capable of storing instructions , such as an operating system and / or various applications . Furthermore , the memory 214 may include a storage that may be used to store, e.g., at least some of the information and data used in the disclosed embodiments, such as a second machine learning model 252 described in more detail below.

[0098] Furthermore, the processor 212 is capable of executing the stored instructions. In an embodiment, the processor 212 may be embodied as a multi-core processor, a single core processor, or a combination of one or more multi-core processors and one or more single core processors. For example, the processor 212 may be embodied as one or more of various processing devices, such as a coprocessor, a microprocessor, a controller, a digital signal processor (DSP) , a processing circuitry with or without an accompanying DSP, or various other processing devices including integrated circuits such as, for example, an application specific integrated circuit (ASIC) , a field programmable gate array (FPGA) , a microcontroller unit (MCU) , a hardware accelerator, a special-purpose computer chip, a neural network (NN) chip, an artificial intelligence (Al) accelerator, a tensor processing unit (TPU) , a neural processing unit (NPU) , or the like. In an embodiment, the processor 212 may be configured to execute hard-coded functionality. In an embodiment, the processor 212 is embodied as an executor of software instructions, wherein the instructions may specifically configure the processor 212 to perform the algorithms and / or operations described herein when the instructions are executed.

[0099] The memory 214 may be embodied as one or more volatile memory devices, one or more non-volatile memory devices, and / or a combination of one or more volatile memory devices and non-volatile memory devices. For example, the memory 214 may be embodied as semiconductor memories (such as mask ROM, PROM (programmable ROM) , EPROM (erasable PROM) , flash ROM, RAM (random access memory) , etc . ) . When executed by the at least one processor 212 , instructions stored in the at least one memory 214 cause the network node 210 at least to carry out , for collaborative power ampli fier non-linearity control with a user device 200 , a negotiation procedure with the user device 200 .

[0100] The negotiation procedure comprises negotiating whether the user device 200 carries out training of a first machine learning model 251 applicable to digital pre-distortion for controlling out-of-band emissions in advance or subsequently to training of a second machine learning model 252 applicable to digital post-distortion for controlling in-band emissions .

[0101] At least in some embodiments , the negotiation procedure may further comprise receiving, from the user device 200 , a request for the training of the first machine learning model 251 and / or of the second machine learning model 252 , or transmitting, to the user device 200 a request for the training of the first machine learning model 251 and / or the second machine learning model 252 .

[0102] In the case the training of the first machine learning model 251 is carried out in advance to the training of the second machine learning model 252 , the instructions , when executed by the at least one processor 212 , further cause the network node 210 at least to receive , from the user device 200 , information on the completion of the training of the first machine learning model 251 along with information for labelling input data for the training of the second machine learning model 252 , and carry out the training of the second machine learning model 252 .

[0103] In the case the training of the first machine learning model 251 is carried out subsequently to the training of the second machine learning model 252 , the instructions , when executed by the at least one processor 212 , further cause the network node 210 at least to transmit, to the user device 200, information on output parameters of the training of the second machine learning model 252 for the training of the first machine learning model 251.

[0104] At least in some embodiments, the instructions, when executed by the at least one processor 212, may further cause the network node 210 to receive from the user device 200 an indication of another first machine learning model determined to be used. If training is needed, the instructions, when executed by the at least one processor 212, may further cause the network node 210 to carry out the negotiation procedure in relation to the another first machine learning model and a second machine learning model related to the another first machine learning model.

[0105] Further features of the network node 210 directly result from the functionalities and parameters of the user device 200 and thus are not repeated here.

[0106] Diagram 400 of Fig. 4 illustrates an example of the disclosed collaborative training between the user device 200 and the network node 210.

[0107] As discussed above, a training entity (residing in the user device 200 or the network node 210) may use a combined cost function f, where the combination contains contributions of two independent or intermediate cost functions fl and f2, each targeting the two ML models 251, 252 separately. Alternatively, the cost functions fl and f2 may be used by themselves.

[0108] The cost functions fl, f2, f may be agreed between the user device 200 and the network node 210 at operation 401.

[0109] The training entity may transfer the trained blocks to the corresponding entities, i.e., ML-based DPD to the user device 200 and ML-based DPoD to the network node 210. Depending on where the trainer resides, the model transfer may be realized, e.g., via radio resource control (RRC) or medium access control (MAC) signalling and / or LTE positioning protocol (LPP) / NR positioning protocol A (NRPPa) , if the trainer is, e.g., in a location server.

[0110] Operation 401 may further include the user device 200 signalling (over the air) to the network node 210 information about its hardware, e.g., model numbers (which the network node 210 may map to, e.g., hardware models and / or ML model requirements) . The user device 200 may also provide information on its operating requirements or training conditions, e.g., in terms of mobility and / or expected signal-to-noise ratio (SNR) and / or signal-to-interf erence-plus-noise ratio (SINR) , as well as identifying preferred cost function (s) . This information may be transferred, for example, over MAC control elements (CEs) .

[0111] At operation 402, the network node 210 may run nominal user device and channel models, collect the RX signal, and / or perform decoding to obtain the estimated bits b' , as described in more detail above in connection with diagrams 300, 350.

[0112] After obtaining the necessary information, the network node 210 may proceed with training the second ML model 252 based on the information provided by the user device 200 using the cost function f2, operation 403.

[0113] At operation 404, the network node 210 may transmit gradients (over the air) to the user device 200 as output parameters .

[0114] At operation 405, the user device 200 may proceed with training the first ML model 251 based on the gradients / other information provided by the network node 210 using the cost function fl.

[0115] Utilized ML model specifications may include, e.g., neural network (NN) depth, weights, biases, activation functions, and / or custom or special functions inside the respective ML model. Diagram 500 of Fig . 5 further illustrates another example of the disclosed collaborative training between the user device 200 and the network node 210 .

[0116] In the example embodiment of diagram 500 , the DPD and DPoD respectively at the user device 200 and network node 210 may be trained sequentially / itera- tively .

[0117] For example , phases DPD_Tr and DPD_Inf may be defined for DPD . In the DPD_Tr phase , the DPD may be trained / updated . However, in the DPD_Inf phase , the DPD parameters may be frozen and the trained DPD may be used to reduce the ACLR . E . g . , periodic or aperiodic triggering mechanisms may be used to switch between the training and inference phases of the DPD .

[0118] Similarly, DPoD_Tr and DPoD_Inf phases may be defined for the DPoD at the network node 210 . In parallel or sequential modes to the DPD training procedure , the DPoD at the network node 210 may be trained by switching to the DPoD_Tr phase , where the trainable parameters of the DPoD block may be trained / updated to reduce the EVM of the received signal .

[0119] At least in some embodiments , the training / in- ference phase at one entity may be transparent to the other entity . Alternatively, the operations happening during these phases may be transparent , but the other entity may know when the first entity stopped training, so that their DPD and DPoD versions may remain synchroni zed .

[0120] The entity that trains first may be commonly agreed between the user device 200 and the network node 210 , operation 501 . Then, upon completion of the training of operation 502 , the first entity may freeze its architecture and inform at operation 503 the second entity that it may start its own training . I f the second entity is the network node 210 , it may need in addition the raw bits ( TX bits 301 ) that are being proces sed by the user device 200 , as these bits may be used as the labels for training the DPoD module at the network node 210.

[0121] For example, at least some of the following information may be transferred via RRC or MAC CE signaling :

[0122] - the user device TX bits b;

[0123] - a handshake or the like between the user device 200 and the network node 210 to choose the first trainer.

[0124] - a signal or the like from the first trainer to the second trainer when the first training is complete.

[0125] As discussed above, the bits b are generated by the user device 200 and these bits may be the labels that the network node 210 is using for training. The bits may encoded using a code C() , modulated, using a modulation function M() and OFDM modulated DFT, so signal y = DFT (M (C (b) ) ) . Then, u = y * h + n.

[0126] The example procedure shown in diagram 500, involves first identifying the initial conditions, e.g., which entity trains its model first, what are the cost functions, and what are the assumptions for training conditions in terms of SNR, mobility, and other relevant channel parameters (operation 501) . In this example, the user device 200 trains its model first using the cost function fl and a fixed network node 210 -side model (operation 502) , after which it may inform the network node 210 when the training is completed (operation 503) . The completion of the training indicates at which point the user device 200 model is again fixed, and the network node 210 model may be released for training. Then, the network node 210 may start to train its DPoD function (operation 505) using the cost function f2, and the previously agreed TX bits b 301 (transmitted at operation 504, e.g., as waveform encoded) so that it can generate proper labels for the received signal. The number of training data slots transmitted by the user device 200 may be agreed in operation 501 so that the user device 200 knows when it may start transmitting useful data. If more training data is needed by the network node 210, a new training step may be initiated. At least in some embodiments, the user device 200 may be configured to send the known TX bits b 301 in a given time interval, and then send data. The user device 200 may not know what the network node 210 does in that time interval with the bits b, e.g., the network node 210 may buffer them and train models periodically, or the network node 210 may assess that the channel conditions are too poor to start any training, or the like.

[0127] For example, different learning schemes may be considered for training of the DPD and DPoD blocks. For example, supervised learning may be used for both the DPD and DPoD blocks. Alternatively, reinforcement learning (RL) may be used for training the DPD block (a reward may be defined based on the cost functions of fl and / or f2) , while supervised learning may be used for training the DPoD block.

[0128] Diagram 600 of Fig. 6 illustrates an example of the disclosed synchronization of the machine learning models when switching to new machine learning models.

[0129] After the two models 251, 252 have been trained, the user device 200 and the network node 210 may, independently from each other, monitor model usage and current conditions (operation 601) and decide to switch among its own available models. When one entity decides to switch from model X to model Y, it must inform the other entity (operation 602) about, e.g. :

[0130] - the new model Y. At least in some embodiments, the DPD and DPoD models may use same identifications (IDs) and the model switching based on an ID may be sufficient. Alternatively, each model may be associated with a unique ID . I f the network node 210 maintains multiple DPoD sub-models that match a given DPD model ID, then it may be up to the network node 210 implementation which of the DPoD sub-models to switch on . The network node 210 and the user device 200 may be synchroni zed with the models supported by each node but the network node 210 may have a model that fits with multiple models in the user device 200 and vice versa, so the ID mapping may not necessarily be one-to-one ;

[0131] - a switch time , e . g . , how many subframes after the transmission of the message the switch occurs ; and / or

[0132] - a minimum time usage of the new model Y .

[0133] At operation 603 , the other entity performs the switch, and sends an acknowledgement at operation 604 .

[0134] Alternatively at operation 604 , each entity may decline the switch to a speci fic model and propose another model ID .

[0135] Fig . 7 il lustrates an example flow chart o f a method 700 for the user device 200 , in accordance with an example embodiment .

[0136] At operations 701-705 , the user device 200 carries out the negotiation procedure with the network node 210 for the collaborative power amplifier non-linearity control with the network node 210 .

[0137] At optional operation 701 , the negotiation procedure with the network node 210 may be initiated .

[0138] At optional operation 702 , the user device 200 may transmit to the network node 210 or receive from the network node 210 the request for the training of the first machine learning model 251 and / or of the second machine learning model 252 .

[0139] At optional operation 703 , the coordination signalling for using the common cost function may be exchanged with the network node 210 . At optional operation 704 , the control of the duration of the training o f the first machine learning model 251 and / or the training of the second machine learning model 252 may be establ ished with the network node 210 .

[0140] Any of the optional operations 701 -704 may alternatively be performed in another order, e . g . , after operation 705 .

[0141] At operation 705 , the user device 200 negotiates whether the user device 200 carries out the training o f the first machine learning model 251 applicable to the digital pre-distortion for controlling the out- of-band emissions in advance or subsequently to the training of the second machine learning model 252 applicable to the digital post-distortion for controlling the in-band emissions .

[0142] When the training of the first machine learning model 251 i s carried out in advance to the training of the second machine learning model 252 , the method 700 may proceed to optional operation 706 in which the user device 200 may carry out the training o f the first machine learning model 251 .

[0143] At operation 707 , the user device 200 transmits to the network node 210 the information on the completion of the training of the first machine learning model 251 at operation 706 along with the information for the labelling input data for the training of the second machine learning model 252 .

[0144] When the training of the first machine learning model 251 is carried out subsequently to the training of the second machine learning model 252 , the method 700 proceeds to operation 708 in which the user device 200 receives from the network node 210 the information on the output parameters of the training of the second machine learning model 252 . At operation 709 , the user device 200 carries out the training of the first machine learning model 251 utili zing the information received at operation 708 .

[0145] At optional operation 710 , the user device 200 may determine to use another first machine learning model . The method 700 may then proceed to optional operation 711 in which the user device 200 may transmit the indication of the another first machine learning model to the network node 210 . I f training is determined to be needed at optional operation 712 , the method 700 may return to , e . g . , optional operation 701 to carry out the negotiation procedure in relation to the another first machine learning model and the second machine learning model related to the another first machine learning model . I f training is determined not to be needed at optional operation 712 , the method 700 may, e . g . , exit at optional operation 713 .

[0146] Embodiments and examples with regard to Fig . 7 may be carried out by user device 200 of Fig . 2A. Operations 701-713 may, for example , be carried out by at least one processor 202 and at least one memory 204 . Further features of method 700 directly resulting from the functionalities and parameters of user device 200 are not repeated here . Method 700 can be carried out by computer program ( s ) or portions thereof .

[0147] Another example of an apparatus suitable for carrying out the embodiments and examples with regard to Fig . 7 comprises means for : carrying out , at operations 701-705 , for collaborative power ampli fier non-linearity control with network node 210 , a negotiation procedure with network node 210 , wherein the negotiation procedure comprises negotiating, at operation 705 , whether user device 200 carries out training of first machine learning model 251 applicable to digital pre-distortion for controlling out-of-band emissions in advance or subsequently to training of second machine learning model 252 applicable to digital post-distortion for controlling in-band emissions ; in the case the training of first machine learning model 251 is carried out in advance to the training of second machine learning model 252 : transmitting, at operation 707 , to network node 210 information on the completion of the training of first machine learning model 251 along with information for labelling input data for the training of second machine learning model 252 ; and in the case the training of first machine learning model 251 is carried out subsequently to the training of second machine learning model 252 : receiving, at operation 708 , from network node 210 information on output parameters of the training of second machine learning model 252 , and carrying out , at operation 709 , the training of first machine learning model 251 utili zing the information .

[0148] Fig . 8 il lustrates an example flow chart o f a method 800 for the network node 210 , in accordance with an example embodiment .

[0149] At operations 801- 805 , the network node 210 carries out the negotiation procedure with the user device 200 for the collaborative power ampli fier non-linearity control with the user device 200 .

[0150] At optional operation 801 , the negotiation procedure with the user device 200 may be initiated .

[0151] At optional operation 802 , the network node 210 may receive from the user device 200 or transmit to the user device 200 the request for the training of the first machine learning model 251 and / or of the second machine learning model 252 .

[0152] At optional operation 803 , the coordination signalling for using the common cost function may be exchanged with the user device 200 .

[0153] At optional operation 804 , the control of the duration of the training o f the first machine learning model 251 and / or the training of the second machine learning model 252 may be established with the user device 200 .

[0154] Any of the optional operations 801 - 804 may alternatively be performed in another order, e . g . , after operation 805 .

[0155] At operation 805, the network node 210 negotiates whether the user device 200 carries out the training o f the first machine learning model 251 appl icable to the digital pre-distortion for controlling the out- of-band emissions in advance or subsequently to the training of the second machine learning model 252 applicable to the digital post-distortion for controlling the in-band emissions .

[0156] When the training of the first machine learning model 251 i s carried out in advance to the training of the second machine learning model 252 , the method 800 proceeds to operation 806 in which the network node 210 receives from the user device 200 information on the completion of the training of the first machine learning model 251 along with the information for the labell ing input data for the training of the second machine learning model 252 .

[0157] At operation 807 , the network node 210 carries out the training of the second machine learning model 252 .

[0158] When the training of the first machine learning model 251 is carried out subsequently to the training of the second machine learning model 252 , the method 800 may proceed to optional operation 808 in which the network node 210 may carry out the training o f the second machine learning model 252 .

[0159] At operation 809 , the network node 210 transmits to the user device 200 the information on the output parameters of the training of the second machine learning model 252 performed at optional operation 808 for the training of the first machine learning model 251 . At optional operation 810 , the network node 210 may receive from the user device 200 the indication of the another first machine learning model determined to be used . I f training is determined to be needed at optional operation 8112 , the method 800 may return to, e . g . , optional operation 801 to carry out the negotiation procedure in relation to the another first machine learning model and the second machine learning model related to the another first machine learning model . I f training is determined not to be needed at optional operation 811 , the method 800 may, e . g . , exit at optional operation 812 .

[0160] Embodiments and examples with regard to Fig . 8 may be carried out by network node 210 of Fig . 2B . Operations 801- 812 may, for example , be carried out by at least one processor 212 and at least one memory 214 . Further features of method 800 directly resulting from the functionalities and parameters of network node 210 are not repeated here . Method 800 can be carried out by computer program ( s ) or portions thereof .

[0161] Another example of an apparatus suitable for carrying out the embodiments and examples with regard to Fig . 8 comprises means for : carrying out , at operations 801- 805 , for collaborative power ampli fier non-linearity control with user device 200 , a negotiation procedure with user device 200 , wherein the negotiation procedure comprises negotiating, at operation 805 , whether user device 200 carries out training of first machine learning model 251 applicable to digital pre-distortion for controlling out-of-band emissions in advance or subsequently to training of second machine learning model 252 applicable to digital post-distortion for controlling in-band emissions ; in the case the training of first machine learning model 251 is carried out in advance to the training of second machine learning model 252 : receiving, at operation 806 , from user device 200 , information on the completion of the training of first machine learning model 251 along with information for labelling input data for the training of second machine learning model 252 , and carrying out, at operation 807 , the training of second machine learning model 252 ; and in the case the training of first machine learning model 251 is carried out subsequently to the training of second machine learning model 252 : transmitting, at operation 809, to user device 200 information on output parameters of the training of second machine learning model 252 for the training of first machine learning model 251 .

[0162] The user device 200 may comprise means for performing at least one method described herein . In one example , the means may comprise the at least one processor 202 , and the at least one memory 204 storing instructions that , when executed by the at least one processor, cause the user device 200 to perform the method .

[0163] The network node 210 may comprise means for performing at least one method described herein . In one example , the means may comprise the at least one processor 212 , and the at least one memory 214 storing instructions that , when executed by the at least one processor, cause the network node 210 to perform the method .

[0164] The functionality described herein can be performed, at least in part , by one or more computer program product components such as software components . According to an embodiment , the user device 200 and / or the network node 210 may compri se a processor or processor circuitry, such as for example a microcontroller, configured by the program code when executed to execute the embodiments of the operations and functionality de- scribed. Alternatively, or in addition, the functionality described herein can be performed, at least in part, by one or more hardware logic components. For example, and without limitation, illustrative types of hardware logic components that can be used include Field-programmable Gate Arrays (FPGAs) , Application-specific Integrated Circuits (ASICs) , Application-specific Standard Products (ASSPs) , System-on-a-chip systems (SOCs) , Complex Programmable Logic Devices (CPLDs) , Tensor Processing Units (TPUs) , and Graphics Processing Units (GPUs) .

[0165] In the disclosed example embodiments, it may be possible to train one machine learning (ML) model with a specific architecture, then derive another ML model from that using processes such as compilation, pruning, quantization or distillation. The ML model may be executed using any suitable apparatus, for example a CPU, GPU, ASIC, FPGA, compute-in-memory, analog, or digital, or optical apparatus. It is also possible to execute the ML model in an apparatus that combines features from any number of these, for instance digital- optical or analog-digital hybrids. In some examples, weights and required computations in these systems may be programmed to correspond to the ML model. In some examples, the apparatus may be designed and manufactured so as to perform the task defined by the ML model so that the apparatus is configured to perform the task when it is manufactured without the apparatus being programmable as such.

[0166] Any range or device value given herein may be extended or altered without losing the effect sought. Also, any embodiment may be combined with another embodiment unless explicitly disallowed.

[0167] Although the subject matter has been described in language specific to structural features and / or acts, it is to be understood that the subject matter defined in the appended claims is not necessarily limited to the speci fic features or acts described above . Rather, the speci fic features and acts described above are disclosed as examples of implementing the claims and other equivalent features and acts are intended to be within the scope of the claims .

[0168] It will be understood that the benefits and advantages described above may relate to one embodiment or may relate to several embodiments . The embodiments are not limited to those that solve any or all of the stated problems or those that have any or all of the stated benef its and advantages . It will further be understood that reference to ' an ' item may refer to one or more of those items .

[0169] The steps of the methods described herein may be carried out in any suitable order, or simultaneously where appropriate . Additionally, individual blocks may be deleted from any of the methods without departing from the spirit and scope of the subj ect matter described herein . Aspects of any of the embodiments described above may be combined with aspects of any of the other embodiments described to form further embodiments without losing the ef fect sought .

[0170] The term ' comprising ' is used herein to mean including the method, blocks or elements identi fied, but that such blocks or elements do not comprise an exclusive list and a method or apparatus may contain additional blocks or elements .

[0171] It will be understood that the above description is given by way of example only and that various modi f ications may be made by those s kil led in the art . The above speci fication, examples and data provide a complete description o f the structure and use o f exemplary embodiments . Although various embodiments have been described above with a certain degree of particularity, or with reference to one or more individual embodiments , those skilled in the art could make numerous alterations to the disclosed embodiments without departing from the spirit or scope of this speci fication .

Claims

CLAIMS :

1. A user device (200) , comprising: at least one processor (202) ; and at least one memory (204) storing instructions that, when executed by the at least one processor (202) , cause the user device (200) at least to: carry out, for collaborative power amplifier non-linearity control with a network node (210) , a negotiation procedure with the network node (210) , wherein the negotiation procedure comprises negotiating whether the user device (200) carries out training of a first machine learning model (251) applicable to digital predistortion for controlling out-of-band emissions in advance or subsequently to training of a second machine learning model (252) applicable to digital post-distortion for controlling in-band emissions; in the case the training of the first machine learning model (251) is carried out in advance to the training of the second machine learning model (252) : transmit, to the network node (210) , information on the completion of the training of the first machine learning model (251) along with information for labelling input data for the training of the second machine learning model (252) ; and in the case the training of the first machine learning model (251) is carried out subsequently to the training of the second machine learning model (252) : receive, from the network node (210) , information on output parameters of the training of the second machine learning model (252) , and carry out the training of the first machine learning model (251) utilizing the information.

2. The user device (200) according to claim 1, wherein the negotiation procedure further comprises transmitting, to the network node (210) , a request for the training of the first machine learning model (251)and / or of the second machine learning model (252) , or receiving, from the network node (210) a request for the training of the first machine learning model (251) and / or the second machine learning model (252) .

3. The user device (200) according to claim 1 or 2, wherein the negotiation procedure further comprises coordination signalling for using a common cost function .

4. The user device (200) according to any of claims 1 to 3, wherein the negotiation procedure further comprises control of a duration of the training of the first machine learning model (251) and / or the training of the second machine learning model (252) .

5. The user device (200) according to any of claims 1 to 4, wherein the first machine learning model (251) is applicable to control an adjacent channel leakage ratio and the second machine learning model (252) is applicable to control an error vector magnitude.

6. The user device (200) according to any of claims 1 to 5, wherein the instructions, when executed by the at least one processor (202) , further cause the user device (200) to: determine to use another first machine learning model ; transmit an indication of the another first machine learning model to the network node (210) , and if training is needed, carry out the negotiation procedure in relation to the another first machine learning model and a second machine learning model related to the another first machine learning model.

7. A network node (210) , comprising:at least one processor (212) ; and at least one memory (214) storing instructions that, when executed by the at least one processor (212) , cause the network node (210) at least to: carry out, for collaborative power amplifier non-linearity control with a user device (200) , a negotiation procedure with the user device (200) , wherein the negotiation procedure comprises negotiating whether the user device (200) carries out training of a first machine learning model (251) applicable to digital predistortion for controlling out-of-band emissions in advance or subsequently to training of a second machine learning model (252) applicable to digital post-distortion for controlling in-band emissions; in the case the training of the first machine learning model (251) is carried out in advance to the training of the second machine learning model (252) : receive, from the user device (200) , information on the completion of the training of the first machine learning model (251) along with information for labelling input data for the training of the second machine learning model (252) , and carry out the training of the second machine learning model (252) ; and in the case the training of the first machine learning model (251) is carried out subsequently to the training of the second machine learning model (252) : transmit, to the user device (200) , information on output parameters of the training of the second machine learning model (252) for the training of the first machine learning model (251) .

8. The network node (210) according to claim 7, wherein the negotiation procedure further comprises receiving, from the user device (200) , a request for the training of the first machine learning model (251) and / or of the second machine learning model (252) , or transmitting, to the user device (200) a request for thetraining of the first machine learning model (251) and / or the second machine learning model (252) .

9. The network node (210) according to claim 7 or 8, wherein the negotiation procedure further comprises coordination signalling for using a common cost function .

10. The network node (210) according to any of claims 7 to 9, wherein the negotiation procedure further comprises control of a duration of the training of the first machine learning model (251) and / or the training of the second machine learning model (252) .

11. The network node (210) according to any of claims 7 to 10, wherein the first machine learning model (251) is applicable to control an adjacent channel leakage ratio and the second machine learning model (252) is applicable to control an error vector magnitude.

12. The network node (210) according to any of claims 7 to 11, wherein the instructions, when executed by the at least one processor (212) , further cause the network node (210) to: receive from the user device (200) an indication of another first machine learning model determined to be used, and if training is needed, carry out the negotiation procedure in relation to the another first machine learning model and a second machine learning model related to the another first machine learning model.

13. A method (700) , comprising: carrying out (701-705) , by a user device (200) for collaborative power amplifier non-linearity control with a network node (210) , a negotiation procedure withthe network node (210) , wherein the negotiation procedure comprises negotiating (705) whether the user device (200) carries out training of a first machine learning model (251) applicable to digital pre-distortion for controlling out-of-band emissions in advance or subsequently to training of a second machine learning model (252) applicable to digital post-distortion for controlling in-band emissions; in the case the training of the first machine learning model (251) is carried out in advance to the training of the second machine learning model (252) : transmitting (707) , from the user device (200) to the network node (210) , information on the completion of the training of the first machine learning model (251) along with information for labelling input data for the training of the second machine learning model (252) ; and in the case the training of the first machine learning model (251) is carried out subsequently to the training of the second machine learning model (252) : receiving (708) , at the user device (200) from the network node (210) , information on output parameters of the training of the second machine learning model (252) , and carrying out (709) , by the user device (200) , the training of the first machine learning model (251) utilizing the information.

14. The method (700) according to claim 13, wherein the negotiation procedure further comprises transmitting, from the user device (200) to the network node (210) , a request for the training of the first machine learning model (251) and / or of the second machine learning model (252) , or receiving, at the user device (200) from the network node (210) a request for the training of the first machine learning model (251) and / or the second machine learning model (252) .

15. The method (700) according to claim 13 or 14, wherein the negotiation procedure further comprises coordination signalling for using a common cost function .

16. The method (700) according to any of claims 13 to 15, wherein the negotiation procedure further comprises control of a duration of the training of the first machine learning model (251) and / or the training of the second machine learning model (252) .

17. The method (700) according to any of claims 13 to 16, wherein the first machine learning model (251) is applicable to control an adjacent channel leakage ratio and the second machine learning model (252) is applicable to control an error vector magnitude.

18. The method (700) according to any of claims 13 to 17, further comprising: determining (710) , by the user device (200) to use another first machine learning model; transmitting (711) an indication of the another first machine learning model from the user device (200) to the network node (210) , and if training is needed, carrying out (701-705) , by the user device (200) , the negotiation procedure in relation to the another first machine learning model and a second machine learning model related to the another first machine learning model.

19. A method (800) , comprising: carrying out (801-805) , by a network node (210) for collaborative power amplifier non-linearity control with a user device (200) , a negotiation procedure with the user device (200) , wherein the negotiation procedure comprises negotiating (805) whether the user device(200) carries out training of a first machine learning model (251) applicable to digital pre-distortion for controlling out-of-band emissions in advance or subsequently to training of a second machine learning model (252) applicable to digital post-distortion for controlling in-band emissions; in the case the training of the first machine learning model (251) is carried out in advance to the training of the second machine learning model (252) : receiving (806) , at the network node (210) from the user device (200) , information on the completion of the training of the first machine learning model (251) along with information for labelling input data for the training of the second machine learning model (252) , and carrying out (807) , by the network node (210) , the training of the second machine learning model (252) ; and in the case the training of the first machine learning model (251) is carried out subsequently to the training of the second machine learning model (252) : transmitting (809) , from the network node (210) to the user device (200) , information on output parameters of the training of the second machine learning model (252) for the training of the first machine learning model (251) .

20. The method (800) according to claim 19, wherein the negotiation procedure further comprises receiving, at the network node (210) from the user device (200) , a request for the training of the first machine learning model (251) and / or of the second machine learning model (252) , or transmitting, from the network node (210) to the user device (200) a request for the training of the first machine learning model (251) and / or the second machine learning model (252) .

21. The method (800) according to claim 19 or 20, wherein the negotiation procedure further comprises coordination signalling for using a common cost function .

22. The method (800) according to any of claims 19 to 21, wherein the negotiation procedure further comprises control of a duration of the training of the first machine learning model (251) and / or the training of the second machine learning model (252) .

23. The method (800) according to any of claims 19 to 22, wherein the first machine learning model (251) is applicable to control an adjacent channel leakage ratio and the second machine learning model (252) is applicable to control an error vector magnitude.

24. The method (800) according to any of claims 19 to 23, wherein the method (800) further comprises: receiving (810) , at the network node (210) from the user device (200) , an indication of another first machine learning model determined to be used, and if training is needed, carrying out (801-805) , by the network node (210) , the negotiation procedure in relation to the another first machine learning model and a second machine learning model related to the another first machine learning model.

25. An apparatus, comprising means for carrying out the method according to any of claims 13 to 18 or 19 to 24.

Images