WO2024092755A1

WO2024092755A1 - Management of machine learning models in communication systems

Info

Publication number: WO2024092755A1
Application number: PCT/CN2022/130001
Authority: WO
Inventors: Salah Eddine HAJRI; Qiaolin SHI
Original assignee: Huawei Technologies Co., Ltd.
Priority date: 2022-11-04
Filing date: 2022-11-04
Publication date: 2024-05-10

Abstract

Embodiments of the invention relate to management of machine learning (ML) models in a communication system based on an identity for a ML model. A first communication device (100) determines an identity for a ML model for a communication session over a radio channel between the first communication device (100) and a second communication device (300) and transmits the identity for the ML model to the second communication device (300). The identity indicates a configuration of the ML model and can thereby provide information about the ML model and/or conditions supported by the ML model to the second communication device (300). Furthermore, the invention also relates to corresponding methods and a computer program.

Description

MANAGEMENT OF MACHINE LEARNING MODELS IN COMMUNICATION SYSTEMS

Technical Field

Embodiments of the invention relate to management of machine learning (ML) models for a first communication device and a second communication device in a communication system. Furthermore, embodiments of the invention also relate to corresponding methods and a computer program.

Background

Artificial intelligence (AI) /machine learning (ML) methods have proven their worth in a multitude of fields, covering different problems, including classification, regression, and interaction with dynamic environments. Given the ability of some ML models to understand patterns and dependencies in the data, which are not typically captured by conventional signal processing techniques, exploiting their potential in the air interface of wireless communication networks could deliver non-negligible performance gains.

In this context, 3GPP agreed a study item in Rel-18, “Study on Artificial Intelligence (AI) /Machine learning (ML) for NR air interface” , to study the potential enhancements, performance gain, general framework and standard impact that AI/ML methods would entail on the air interface in several important use cases. Beam management, channel state information (CSI) reporting and positioning performance enhancements are considered the main use cases of interest in Rel-18, which does not preclude work on other use cases relevant to the air interface of cellular wireless networks.

The Rel-18 study item is not the first effort from 3GPP to study the impact of AI/ML in wireless communication networks. A higher layer-focused study item, “Study on enhancement for data collection for NR and ENDC” , was already part of the agenda of Rel-17.

While several elements of the agreed functional framework for higher layers can, at least partly, be reused for the new radio (NR) air interface, the particular constraints of the physical layer mandates a proper adaptation to guarantee performance in terms of throughput, capacity, latency and reliability.

While AI/ML methods could deliver substantial performance gains, some hurdles need to be considered, including, among others, complexity, power consumption and latency. Indeed, the operations in the physical layer of cellular networks are subject to stringent latency and reliability requirements, which are among the most critical aspects of high performance air interfaces.

Summary

An objective of embodiments of the invention is to provide a solution which mitigates or solves the drawbacks and problems of conventional solutions.

Another objective of embodiments of the invention is to provide a solution which enables communication devices to exchange information about ML models in an efficient way.

The above and further objectives are solved by the subject matter of the independent claims. Further embodiments of the invention can be found in the dependent claims.

According to a first aspect of the invention, the above mentioned and other objectives are achieved with a first communication device, the first communication device being configured to:

determine an identity for a machine learning, ML, model for a communication session over a radio channel between the first communication device and a second communication device, the identity indicating a configuration of the ML model; and

transmit a first control signal to the second communication device, the first control signal comprising the identity.

An advantage of the first communication device according to the first aspect is that the determined identity conveys information about the configuration of the ML model and can be used to identify ML models during signaling for all actions of ML model life cycle management. As the identity itself is informative, information about conditions at the first communication device and suitable ML models can be exchanged in an efficient way, without having to the explicitly exchange the ML model. Ultimately, different levels of coordination between communication devices for ML operations over the air interface, can be supported.

In an implementation form of a first communication device according to the first aspect, the first communication device is configured to

determine the identity based on one or more of: a channel characteristic of the radio channel, a traffic characteristic of the communication session, a target performance for the communication session, a use case for the ML model, a target latency for the ML model, a complexity of the ML model, and a model version of the ML model.

An advantage with this implementation form is that the identity of the ML model is informative and a ML model suitable to conditions perceived by the first communication device can be determined. Indeed, as the identity is determined based on the characteristics of the radio channel and/or traffic, target performance of the communication session, and/or ML model characteristics, and the identity itself conveys information, the identity can be used as means to convey information on the ongoing or upcoming conditions, in addition to denoting specific ML models and their associated configurations.

In an implementation form of a first communication device according to the first aspect, the configuration of the ML model comprises one or more information elements associated with one or more of: a channel characteristic of the radio channel, a traffic characteristic of the communication session, a target performance for the communication session, a use case for the ML model, a target latency for the ML model, a complexity of the ML model, and a model version of the ML model.

An advantage with this implementation form is that the identity is associated with a configuration of the ML model that can be used during ML model life cycle management signaling without having to exchange the explicit ML models, this includes version control. As the identity is constructed based on the characteristics of the radio channel and/or traffic, target performance of the communication session, and/or ML model characteristics, the identity itself conveys information, even when the ML model is not known, at either end of the communication link. Additionally, while the ML model itself can be changed, as long as it satisfies the information elements in the associated configuration, the same identity can be used. This enables to avoid unnecessary signaling.

perform one or more of: activate the ML model, update the ML model, train the ML model, and register the ML model based on the determined identity.

An advantage with this implementation form is that the identity can be used to perform multiple functions related to the ML model life cycle management and operations. The proposed identity can be used in the signaling between two communication devices without forcibly requiring explicit sharing of the ML model.

In an implementation form of a first communication device according to the first aspect, the first control signal further indicates one or more of: an activation request for the ML model, an activation notification for the ML model and an activation command for the ML model.

An advantage with this implementation form is that the identity can be used to perform multiple function related to the ML model life cycle management and operations. In case a new ML model needs to be activated, different options are available. The first communication device may activate a ML model autonomously or request a new ML model activation. Furthermore, the first communication device may transmit the determined identity as a command for activating a ML model by the second communication device.

In an implementation form of a first communication device according to the first aspect, the first communication device is configured to when the first control signal indicates the activation request for the ML model:

receive a second control signal from the second communication device, the second control signal indicating an activation response; and

perform one or more of: activate a ML model, update a ML model, train a ML model, and register a ML model based on the second control signal.

An advantage with this implementation form is that the first communication device can propose a ML model adaptation action, based on the conditions it perceives, and the second communication device can respond to the proposed action, e.g., allow or correct the action. As the first communication device may ignore or be unaware of other conditions that could impact the ML model operations, the second communication device can use the second control signal to correct the proposed action and perception of the ongoing or upcoming conditions, in which the ML model is expected to operate.

receive a third control signal from the second communication device, the third control signal indicating support for one or more ML models; and

determine the identity further based on the third control signal.

An advantage with this implementation form is that the proposed identity can be used in ML capability enquiry and ML capability information transfer between the first communication device and the second communication device. By indicating an identity during ML capability information transfer, a communication device can convey the scenarios that it can support.

Additionally, by indicating an identity in an ML capability enquiry, a communication device can request an ML capability information transfer, regarding the information elements comprised in the identity.

In an implementation form of a first communication device according to the first aspect, the identity comprises a sequence of bit strings, wherein each bit string indicates an information element in the configuration of the ML model.

An advantage with this implementation form is that the identity can have a variable length and be tailored to the exact use case and supported scenario. Depending on the implementation, the degree to which ML models are specialized may vary. Consequently, having a modular structure of the identity enables the identity to be adapt to each scenario and implementation.

In an implementation form of a first communication device according to the first aspect, the identity is associated with one or more of: a radio resource configuration for the ML model, a model parameter configuration of the ML model and a model coefficient of the ML model. The radio resource configuration for the ML model may include reporting resources and radio measurement resources.

An advantage with this implementation form is that the exchange of the identity can adapt the active ML model and the radio resources needed to operate the ML model either for reporting or measurements. Consequently, there is no need for further signaling and the exchange of the identity could be sufficient to adapt the overall operation relevant to the ML model, over the air interface.

According to a second aspect of the invention, the above mentioned and other objectives are achieved with a second communication device, the second communication device being configured to:

receive a first control signal from a first communication device, the first control signal comprising an identity for a ML model for a communication session over a radio channel between the first communication device and the second communication device, the identity indicating a configuration of the ML model.

An advantage of the second communication device according to the second aspect is that the determined identity conveys information about the configuration ML model and can be used to identify ML models during signaling for all actions of ML model life cycle management. As the identity itself is informative, information about conditions at the first communication device and suitable ML models can be exchanged in an efficient way, without having to the explicitly exchange the ML model.

In an implementation form of a second communication device according to the second aspect, the configuration of the ML model comprises one or more information elements associated with one or more of: a channel characteristic of the radio channel, a traffic characteristic of the communication session, a target performance for the communication session, a use case for the ML model, a target latency for the ML model, a complexity of the ML model, and a model version of the ML model.

An advantage with this implementation form is that the identity of the ML model is informative. Indeed, as the identity is constructed to indicate the characteristics of the radio channel and/or traffic, target performance of the communication session and/or ML model characteristics, the identity itself conveys information, even when the ML model is not known, at either end of the communication link. Consequently, the model identity can be used as means to convey information on the ongoing or upcoming conditions, in addition to denoting specific ML models and their associated configurations.

In an implementation form of a second communication device according to the second aspect, the second communication device is configured to

perform one or more of: activate the ML model, update the ML model, train the ML model, and register the ML model based on the first control signal.

An advantage with this implementation form is that the identity can be used to perform multiple function related to the ML model life cycle management and operations. The proposed identity can be used in the signaling between two communication devices without requiring explicit sharing of the ML model.

In an implementation form of a second communication device according to the second aspect, the first control signal further indicates one or more of: an activation request for the ML model, an activation notification for the ML model and an activation command for the ML model.

An advantage with this implementation form is that the identity can be used to perform multiple function related to the ML model life cycle management and operations. The proposed identity can be used in the signaling between the communication devices to e.g., support different types of activations of ML models such as autonomous or network controlled.

In an implementation form of a second communication device according to the second aspect, the second communication device is configured to when the first control signal indicates the activation request for the ML model

transmit a second control signal to the first communication device, the second control signal indicating an activation response.

An advantage with this implementation form is that the first communication device can propose a ML model adaptation action, based on the conditions it perceives, and the second communication device can respond to the proposed action, e.g., allow or correct the action. As the first communication device may ignore or be unaware of other conditions that could impact the model operations, the second control signal can be used, by the second communication device, to correct the action and perception of the first communication device regarding the ongoing or upcoming conditions, in which the ML model is expected to operate.

transmit a third control signal to the first communication device previous to receiving the first control signal, the third control signal indicating support for one or more ML models.

An advantage with this implementation form is that the proposed model identity can be used in ML capability enquiry and ML capability information transfer between the first communication device and the second communication device. By indicating an identity during ML capability information transfer, a communication device can convey the scenarios that it can support. Additionally, by indicating an identity in an ML capability enquiry, a communication device can request an ML capability information transfer, regarding the information elements comprised in the identity.

In an implementation form of a second communication device according to the second aspect, the identity comprises a sequence of bit strings, wherein each bit string indicates an information element in the configuration of the ML model.

An advantage with this implementation form is that the identity can have a variable length and be tailored to the exact use case and supported scenario. Depending on how much the ML models are specialized, the conditions and characteristics that are supported by the ML models can be quantized differently. The modular structure of the identity enables the identity to be adapt to different communication device capabilities and implementations.

In an implementation form of a second communication device according to the second aspect, the identity is associated with one or more of: a radio resource configuration for the ML model, a model parameter configuration of the ML model and a model coefficient of the ML model. The radio resource configuration for the ML model may include reporting resources and radio measurement resources.

According to a third aspect of the invention, the above mentioned and other objectives are achieved with a method for a first communication device, the method comprises

determining an identity for a machine learning, ML, model for a communication session over a radio channel between the first communication device and a second communication device, the identity indicating a configuration of the ML model; and

transmitting a first control signal to the second communication device, the first control signal comprising the identity.

The method according to the third aspect can be extended into implementation forms corresponding to the implementation forms of the first communication device according to the first aspect. Hence, an implementation form of the method comprises the feature (s) of the corresponding implementation form of the first communication device.

The advantages of the methods according to the third aspect are the same as those for the corresponding implementation forms of the first communication device according to the first aspect.

According to a fourth aspect of the invention, the above mentioned and other objectives are achieved with a method for a second communication device, the method comprises

receiving a first control signal from a first communication device, the first control signal comprising an identity for a ML model for a communication session over a radio channel between the first communication device and the second communication device, the identity indicating a configuration of the ML model.

The method according to the fourth aspect can be extended into implementation forms corresponding to the implementation forms of the second communication device according to the second aspect. Hence, an implementation form of the method comprises the feature (s) of the corresponding implementation form of the second communication device.

The advantages of the methods according to the fourth aspect are the same as those for the corresponding implementation forms of the second communication device according to the second aspect.

Embodiments of the invention also relate to a computer program, characterized in program code, which when run by at least one processor causes the at least one processor to execute any method according to embodiments of the invention. Further, embodiments of the invention also relate to a computer program product comprising a computer readable medium and the mentioned computer program, wherein the computer program is included in the computer readable medium, and may comprises one or more from the group of: read-only memory (ROM) , programmable ROM (PROM) , erasable PROM (EPROM) , flash memory, electrically erasable PROM (EEPROM) , hard disk drive, etc.

Further applications and advantages of embodiments of the invention will be apparent from the following detailed description.

Brief Description of the Drawings

The appended drawings are intended to clarify and explain different embodiments of the invention, in which:

- Fig. 1 shows a first communication device according to an embodiment of the invention;

- Fig. 2 shows a flow chart of a method for a first communication device according to an embodiment of the invention;

- Fig. 3 shows a second communication device according to an embodiment of the invention;

- Fig. 4 shows a flow chart of a method for a second communication device according to an embodiment of the invention;

- Fig. 5 shows a communication system according to an embodiment of the invention; and

- Fig. 6 shows signaling for ML model management according to an embodiment of the invention;

- Fig. 7 shows signaling for ML model management according to an embodiment of the invention;

- Fig. 8 shows signaling for ML model management according to an embodiment of the invention;

- Fig. 9 shows a structure of an identity according to an embodiment of the invention;

- Fig. 10 shows ML model adaptation based on channel characteristics according to an embodiment of the invention; and

- Fig. 11 shows signaling for ML model adaptation based on channel characteristics according to an embodiment of the invention.

Detailed Description

To optimize the performance gain possible with AI/ML methods in the NR air interface, several aspects related to the general AI/ML framework need to be considered, including but not limited to, model training, model registration, model adaptation, model performance monitoring, training and inference data collection, model transfer, user equipment (UE) capability transfer, model update, model selection, model activation/deactivation, model switching and fallback operation.

Apart from studying the characteristics of the general ML/AI framework in the air interface, three main use cases were considered for study in the 3GPP study item, i.e., beam management, CSI reporting and positioning performance enhancements.

For a given use case, the UE and/or the network may prepare and train a ML model with all available relevant data to suit all possible scenarios/conditions or prepare and train multiple ML models, each suited for a given scenario. A hybrid approach is also possible, where a main ML model is used as backbone for different other ML models, each dedicated to a scenario or feature.

From a training perspective, one can consider training a ML model at the network and/or at the UE, separately or jointly. When multiple ML models are available for use, at either side of the link, a proper management framework is needed for ML model life cycle management.

Furthermore, given the highly dynamic environment in which NR and next generation radio access network (NG-RAN) operate, ML models may need to be adapted to guarantee performance and to avoid drifting between inference data and the data that was used for training of the ML model. Consequently, a framework for ML model adaptation is also needed.

According to embodiments of the invention an identity for ML models is therefore introduced which can be used as a tool in several aspects of air interface ML model life cycle management and ML model adaptation.

Fig. 1 shows a first communication device 100 according to an embodiment of the invention where the first communication device 100 is a client device. However, the first communication device 100 is not limited thereto and may in embodiments instead be a network access node, such as e.g., the network access node shown in Fig. 3. In the embodiment shown in Fig. 1, the first communication device 100 comprises a processor 102, a transceiver 104 and a memory 106. The processor 102 is coupled to the transceiver 104 and the memory 106 by communication means 108 known in the art. The first communication device 100 further comprises an antenna or antenna array 110 coupled to the transceiver 104, which means that the first communication device 100 is configured for wireless communications in a communication system.

The processor 102 may be referred to as one or more general-purpose central processing units (CPUs) , one or more digital signal processors (DSPs) , one or more application-specific integrated circuits (ASICs) , one or more field programmable gate arrays (FPGAs) , one or more programmable logic devices, one or more discrete gates, one or more transistor logic devices, one or more discrete hardware components, or one or more chipsets. The memory 106 may be a read-only memory, a random access memory (RAM) , or a non-volatile RAM (NVRAM) . The transceiver 304 may be a transceiver circuit, a power controller, or an interface providing capability to communicate with other communication modules or communication devices, such as network nodes and network servers. The transceiver 104, memory 106 and/or processor 102 may be implemented in separate chipsets or may be implemented in a common chipset.

That the first communication device 100 is configured to perform certain actions can in this disclosure be understood to mean that the first communication device 100 comprises suitable means, such as e.g., the processor 102 and the transceiver 104, configured to perform the actions.

According to embodiments of the invention the first communication device 100 is configured to determine an identity for a ML model for a communication session over a radio channel between the first communication device 100 and a second communication device 300, the identity indicating a configuration of the ML model. The first communication device 100 is further configured to transmit a first control signal 510 to the second communication device 300, the first control signal 510 comprising the identity.

Furthermore, in an embodiment of the invention, the first communication device 100 for a communication system 500 comprises a processor configured to determine an identity for a ML model for a communication session over a radio channel between the first communication device 100 and a second communication device 300, the identity indicating a configuration of the ML model. The first communication device comprises a transceiver configured to transmit a first control signal 510 to the second communication device 300, the first control signal 510 comprising the identity.

Moreover, in yet another embodiment of the invention, the first communication 100 for a communication system 500 comprises a processor and a memory having computer readable instructions stored thereon which, when executed by the processor, cause the processor to: determine an identity for a ML model for a communication session over a radio channel between the first communication device 100 and a second communication device 300, the identity indicating a configuration of the ML model; and transmit a first control signal 510 to the second communication device 300, the first control signal 510 comprising the identity.

Fig. 2 shows a flow chart of a corresponding method 200 which may be executed in a first communication device 100, such as the one shown in Fig. 1. The method 200 comprises determining 202 an identity for a ML model for a communication session over a radio channel between the first communication device 100 and a second communication device 300, the identity indicating a configuration of the ML model. The method 200 further comprises transmitting 204 a first control signal 510 to the second communication device 300, the first control signal 510 comprising the identity.

Fig. 3 shows a second communication device 300 according to an embodiment of the invention where the second communication device 300 is a network access node. However, the second communication device 300 is not limited thereto and may in embodiments instead be a client device, such as e.g., the client device shown in Fig. 1. In the embodiment shown in Fig. 3, the second communication device 300 comprises a processor 302, a transceiver 304 and a memory 306. The processor 302 is coupled to the transceiver 304 and the memory 306 by communication means 308 known in the art. The second communication device 300 may be configured for wireless and/or wired communications in a communication system. The wireless communication capability may be provided with an antenna or antenna array 310 coupled to the transceiver 304, while the wired communication capability may be provided with a wired communication interface 312 e.g., coupled to the transceiver 304.

The processor 302 may be referred to as one or more general-purpose CPUs, one or more DSPs, one or more ASICs, one or more FPGAs, one or more programmable logic devices, one or more discrete gates, one or more transistor logic devices, one or more discrete hardware components, one or more chipsets. The memory 306 may be a read-only memory, a RAM, or a NVRAM. The transceiver 104 may be a transceiver circuit, a power controller, or an interface providing capability to communicate with other communication modules or communication devices. The transceiver 304, the memory 306 and/or the processor 302 may be implemented in separate chipsets or may be implemented in a common chipset.

That the second communication device 300 is configured to perform certain actions can in this disclosure be understood to mean that the second communication device 300 comprises suitable means, such as e.g., the processor 302 and the transceiver 304, configured to perform the actions.

According to embodiments of the invention the second communication device 300 is configured to receive a first control signal 510 from a first communication device 100, the first control signal 510 comprising an identity for a ML model for a communication session over a radio channel between the first communication device 100 and the second communication device 300, the identity indicating a configuration of the ML model.

Furthermore, in an embodiment of the invention, the second communication device 300 for a communication system 500 comprises a transceiver configured to receive a first control signal 510 from a first communication device 100, the first control signal 510 comprising an identity for a ML model for a communication session over a radio channel between the first communication device 100 and the second communication device 300, the identity indicating a configuration of the ML model.

Moreover, in yet another embodiment of the invention, the second communication device 300 for a communication system 500 comprises a processor and a memory having computer readable instructions stored thereon which, when executed by the processor, cause the processor to: receive a first control signal 510 from a first communication device 100, the first control signal 510 comprising an identity for a ML model for a communication session over a radio channel between the first communication device 100 and the second communication device 300, the identity indicating a configuration of the ML model.

Fig. 4 shows a flow chart of a corresponding method 400 which may be executed in a second communication device 300, such as the one shown in Fig. 3. The method 400 comprises receiving 402 a first control signal 510 from a first communication device 100, the first control signal 510 comprising an identity for a ML model for a communication session over a radio channel between the first communication device 100 and the second communication device 300, the identity indicating a configuration of the ML model.

Fig. 5 shows a communication system 500 according to an embodiment of the invention. The communication system 500 in the disclosed embodiment comprises two first communication devices 100 and two second communication devices 300 configured to communicate and operate in the communication system 500. However, the communication system 500 may comprise any number of first communication devices 100 and any number of second communication devices 300 without deviating from the scope of the invention.

As indicated in Fig. 5, both the first communication devices 100 and the two second communication devices 300 may be either a client device or a network access node. The network access nodes may be connected to a network NW such as e.g., a core network over a communication interface. The communication system 500 may be a communication system according to the 3GPP standard such as e.g., a 5G system in which case the client devices may be UEs and the network access nodes may be next generation node Bs (gNBs) but the invention is not limited thereto.

The first communication devices 100 and the second communication devices 300 communicate with each other over radio channels. With reference to Fig. 5, the radio channels may be used for one or more of uplink, downlink, and sidelink communication. In case of a 5G system, the uplink/downlink communication may be performed over the Uu interface and the and sidelink communication over the PC5 interface.

ML models may be used for the communication sessions over the radio channels between the first communication devices 100 and the second communication devices 300. The ML models may e.g., be used to perform physical layer operations such as CSI reporting, radio resource measurements enhancements, power control, etc. Given client device mobility and the dynamic nature of radio channels and traffic characteristics, the performance of a ML model used to perform physical layer operations may degrade over time. Input drift due to change in the large-scale parameters of the radio channel or change in the traffic requirements, e.g., due to a change from an ultra-reliable low latency communication (URLLC) service to an enhanced mobile broadband (eMBB) service, may necessitate a change in link adaptation, radio resources measurements, resource allocation policies, among others, and subsequently a change in active ML models.

An active ML model may hence need to be adapted in order to guarantee performance and to avoid drifting between inference data and the data that was used for training of the ML model. ML model adaptation may entail the employment of a totally different ML model or changing one or more ML model parameters such as e.g., training data, input features, prediction targets, prediction space constraints, ML model architecture, weights, version, action space, and state space.

Consequently, means to efficiently register and trigger new ML models or adapt previous ML models for the communication sessions between the first communication devices 100 and the second communication devices 300 would be beneficial.

According to embodiments of the invention an identity for ML models capable of providing information about the ML model and scenarios supported by the ML model is therefore provided. Exchange of the identity according to the invention enables the first communication device 100 and the second communication device 300 to obtain information about ML models used by the other communication device and supported scenarios for ML operations, even when the ML models used by the other communication device are not known to the first communication device 100 and/or the second communication device 300.

With reference to Fig. 5, the first communication devices 100 determines an identity for a ML model for a communication session with the second communication device 300 and informs the second communication device 300 about the identity by transmitting a first control signal 510 to the second communication devices 300. The identity indicates a configuration of the ML model and hence enables the second communication device 300 to obtain information related to the ML model determined or selected by the first communication device 100.

Fig. 6 shows signaling for ML model management between the first communication device 100 and the second communication device 300 according to an embodiment of the invention. In the shown embodiment, it is assumed that the first communication device 100 is allowed to autonomously perform ML model adaptation.

In step I in Fig. 6, the first communication device 100 determines an identity for a ML model for a communication session over a radio channel between the first communication device 100 and the second communication device 300. The first communication device 100 may determine the identity for the ML model upon detecting a need for ML model activation or adaption, e.g., based on a change in one or more conditions or characteristics related to the radio channel, the communication session and/or the ML model. Hence, a change in these conditions and/or characteristics may trigger the first communication device 100 to determine a ML model which is suitable for the new conditions and/or characteristics and the identity for that ML model.

In embodiments, the first communication device 100 determines the identity based on one or more of: a channel characteristic of the radio channel, a traffic characteristic of the communication session, a target performance for the communication session, a use case for the ML model, a target latency for the ML model, a complexity of the ML model, and a model version of the ML model. The information related to channel and/or traffic characteristics may be obtained from measurements by the first communication device 100 and/or reported by the second communication device 300. In addition, the information related to the ML model may be obtained from e.g., network indications, configuration in the first communication device 100 and/or configuration received from the second communication device 300.

Examples of channel characteristics of the radio channel which may be used to determine a suitable ML model and its associated identity may be line-of-sight (LoS) /none line-of-sight (NLoS) conditions, Doppler and/or delay spread, sparsity of the channel (number of multipath clusters) , maximum/minimum Doppler shift, coverage region, interference power, signal to interference noise ratio (SINR) levels, coding rate, modulation coding scheme (MCS) etc. Examples of traffic characteristics of the communication session may be type of service such as e.g., URLLC, eMBB, extended reality (XR) , massive machine-type communication (mMTC) . Examples of target performances for the communication session may be key performance indicators (KPIs) related to e.g., throughput, block error rate (BLER) , bit error rate, communication latency, coverage, resource usage, positioning, energy efficiency and power savings. Examples of use case for the ML model may be CSI reporting, beam management, positioning, mobility prediction, power control, radio resource management and measurements. Examples of target latency for the ML model may be a maximum time to train or make prediction with the ML model. Examples of complexity of the ML model may be number of operations needed to train the ML model or to obtain a prediction from the ML model, model coefficients, and model algorithm. Examples of model versions of the ML model may be version number (1, 2, …, n) or version time stamp.

The determined identity indicates a configuration of the ML model and hence provides information about the ML model, i.e., information about the configuration of the ML model can be obtained from the identity. The configuration of the ML model may convey information about the scenarios and/or conditions which the ML model may be used for, as well as characteristics of the ML model itself. In embodiments, the configuration of the ML model comprises one or more information elements associated with one or more of: a channel characteristic of the radio channel, a traffic characteristic of the communication session, a target performance for the communication session, a use case for the ML model, a target latency for the ML model, a complexity of the ML model, and a model version of the ML model. The information elements in the configuration of the ML model may hence provide information about characteristics, conditions and use cases supported by the ML model and characteristics of the ML model. As the identity according to the invention is constructed to indicate the configuration of the ML model, the identity may indicate one or more of these information elements, as will be further described with below with reference to Fig. 9. The identity is hence not only identifying the ML model but also providing information about the ML model.

In step II in Fig. 6, the first communication device 100 transmits a first control signal 510 to the second communication device 300, the first control signal 510 comprising the identity. The second communication device 300 receives the first control signal 510 from the first communication device 100 and hence the identity comprised in the first control signal 510. From the identity for the ML model which indicates the configuration of the ML model, the second communication device 300 may then obtain information about the ML model determined by first communication device 100.

In embodiments, the first control signal 510 further indicates one or more of: an activation request for the ML model, an activation notification for the ML model and an activation command for the ML model. The activation request, the activation notification and the activation command may be implicitly indicated by the identity itself or explicitly indicated in the first control signal 510. The first communication device 100 may further transmits a first control signal 510 comprising the identity and then transmit transmits another first control signal 510′ indicating an activation request, an activation notification and/or an activation command for the ML model associated with the determined identity.

In step III in Fig. 6, the first communication device 100 performs one or more of: activate the ML model, update the ML model, train the ML model, and register the ML model based on the determined identity. Thus, the first communication device 100 may based on the determined identity e.g., activate the ML model for use for the communication session with second communication device 300. The ML model may be activated for training or for inference. Hence, the first communication device 100 may start to train the ML model or start using the ML model for deriving predictions/inference results. The activation may include switching from an earlier ML model to the ML model associated with the determined identity. The first communication device 100 may further register the ML model for later use or update the ML model, e.g., change one or more model parameters, based on the determined identity.

Although step III is shown after step II in Fig. 6, step III may in embodiments instead be performed before or at the same time as step II, i.e., previous to or in parallel with the transmission of the first control signal 510. For example, the first communication device 100 may first activate the ML model and then transmit a first control signal 510 comprising the identity and indicating an activation notification for the ML model to the second communication device 300. In this way, the second communication device 300 can be informed that the ML model has been activated by the first communication device 100 and further provided with information about the ML model, enabling the second communication device 300 to also activate the ML model or adapt its transmissions to the ML model. The second communication device 300 may e.g., use the obtained identity for the ML model to adapt reference signal transmissions, transmit power, etc., without knowing the exact architecture or implementation of the ML model.

The second communication device 300 may further perform one or more actions/operations based on the obtained first control signal 510. In embodiments, the second communication device 300 may perform one or more of: activate the ML model, update the ML model, train the ML model, and register the ML model based on the first control signal 510, as indicated by the optional step IV in Fig. 6. The second communication device 300 may hence use the identity obtained from the first control signal 510 to determine a ML model and use it for the communication session with the first communication device 100, e.g., for interference or training or for updating the ML model for future use.

When the first control signal 510 indicates an activation command for the ML model, the second communication device 300 may activate the ML model based on the identity and the activation command. The ML model may in this case be activated by the first communication device 100 and the second communication device 300 at the same or different time instances. Furthermore, the first communication device 100 may in embodiments only transmit an activation command for the ML model and not activate the ML model itself such that the ML model is activated only on the second communication device 300 side of the communication session.

The exchange of the identity according to the invention hence enables actions associated with the ML model to be performed at the first communication device 100 and/or at the second communication device 300. For example, an adaptation of the ML model may be performed only at one side of the radio channel between the first communication device 100 and the second communication device 300 or at both sides. In the latter case, the adaption may be performed synchronously or at different time instances.

In embodiments, the identity is associated with one or more of: a radio resource configuration for the ML model, a model parameter configuration of the ML model and a model coefficient of the ML model. The first communication device 100 and the second communication device 300 may hence from the determined or obtained identity determine a radio resource configuration, a model parameter configuration and/or a model coefficient to use with the ML model associated with the identity. For example, when the ML model is activated, the associated radio resource configuration, model parameter configuration and/or a model coefficient may also be activated. Furthermore, the second communication device 300 receiving an identity for a ML model and a notification that the ML model has been activated by the first communication device 100 may, from the identity, determine an associated radio resource configuration to use towards the first communication device 100 even if the ML model itself is not activated by the second communication device 300. The model parameter configuration may be changed based on a parameter codebook e.g., number of layers, wherein each set of parameters is associated with a configured identity. Additionally, reporting resources, when reporting is needed, can be adapted. The adaptation of the ML model may further require an adaptation in the reporting format or inference format, which is performed by either communication devices. The possible formats may be comprised in the model parameter configuration.

Fig. 7 shows signaling for ML model management between the first communication device 100 and the second communication device 300 according to an embodiment of the invention. In the shown embodiment, the first communication device 100 determines an identity for a ML model to be used for a communication session with the second communication device 300 and then requests activation of the ML model from the second communication device 300.

In a similar way as in Fig. 6 and with reference to step I and II in Fig. 7, the first communication device 100 determines an identity for a ML model for a communication session over a radio channel between the first communication device 100 and the second communication device 300 and then transmits a first control signal 510 comprising the identity to the second communication device 300. In Fig. 7, it is assumed that the first control signal 510 further indicates an activation request for the ML model.

Based on the activation request indicated in the first control signal 510, the second communication device 300 transmits a second control signal 520 to the first communication device 100, in step III in Fig. 7. The second control signal 520 indicates an activation response. The activation response may be positive or negative, i.e., indicate whether the ML model can be activated or not. When the activation response is negative, the activation response may further indicate another ML model than the requested one. Thus, the second communication device 300 may in embodiments determine that the ML model indicated by the first communication device 100 should not be used and instead determine another ML model to use and indicate the identity associated with the other ML model in the second control signal 520.

In step IV in Fig. 7, the first communication device 100 performs one or more of: activate a ML model, update a ML model, train a ML model, and register a ML model based on the received second control signal 520. For example, when the activation response is positive, the first communication device 100 may activate or update the ML model, i.e., the ML model associated with the identity determined in step I in Fig. 7. When the activation response is negative but indicates another ML model to be activated, the first communication device 100 may activate or update the other ML model.

Optionally, the second communication device 300 may also performs an action associated with the ML model determined by the first communication device 100 or by the second communication device 300, as indicated by optional step V in Fig. 7. The second communication device 300 may hence perform one or more of: activate the ML model, update the ML model, train the ML model, and register the ML model based on the activation response in the second control signal 520.

According to embodiments of the invention, the first communication device 100 and the second communication device 300 may exchange capabilities related to ML models, e.g., during an initial configuration procedure. Fig. 8 shows signaling between the first communication device 100 and the second communication device 300 according to such an embodiment.

In step I in Fig. 8, the second communication device 300 transmit a third control signal 530 to the first communication device 100 previous to receiving the first control signal 510, the third control signal 530 indicating support for one or more ML models. The third control signal 530 may hence indicate the ML models supported by the second communication device 300. In embodiments, the support for one or more ML models may be indicated by comprising the identities associated with the supported ML models in the third control signal 530.

The first communication device 100 receives the third control signal 530 from the second communication device 300 and hence the support for one or more ML models indicated in the third control signal 530. In step II in Fig. 8, the first communication device 100 then determine the identity further based on the third control signal 530. In other words, in addition to the type of information described above with reference to step I in Fig. 6, the first communication device 100 may further consider the ML models supported by the second communication device 300 when determining the identity for the ML model.

In step III in Fig. 8, the first communication device 100 transmits a first control signal 510 to the second communication device 300, the first control signal 510 comprising the identity. In step IV and V in Fig. 8, the first communication device 100 and/or the second communication device 300 may then perform one or more actions based on the determined or received identity, respectively, as described with reference to Fig. 6 and 7.

According to embodiments of the invention the identity may comprise a sequence of bit strings, wherein each bit string indicates an information element in the configuration of the ML model. Each bit string may comprise one or more bits and may indicate a value of one information element. Thus, one or more bits in the identity may be used to indicate each information element in the configuration of the ML model. The number of bits in each bit string may depend on the type of the information element and the number of different values the information element can have. For example, a bit string with one bit may be used to indicate an information element which can have only two values, while two or more bits are needed for information elements which can have more than two values. The length of the identity may hence depend on the number of information elements to be indicated by the identity and the type of information elements, i.e., the length of each bit string.

The order of the bit strings in the sequence may be configured and may further be arranged to create multiple level of the information elements in the configuration of the ML model, e.g., in a tree like structure. Fig. 9 shows an example of an identity for a ML model indicating channel characteristics of the radio channel supported by the ML model, where the identity is arranged in a tree like structure. At a first level, an information element associated with LoS/NLoS is indicated. Under each value of the LoS/NLoS information element, delay spread ΔT _DS and/or Doppler shift f _d conditions are indicated with a respective bit string. Table 1 below show the possible values of the identity constructed according to the embodiment shown in Fig. 9, where the sequence for the LoS cases have been adapted to 4 bits to align the length of the sequences for the LoS and NLoS cases.

Identity	Associated conditions
0000	LoS-dominated+ [f _d0, f _d1 [
0001	LoS-dominated+ [f _d1, f _d2 [
0010	LoS-dominated+ [f _d2, f _d3 [
1000	NLoS-dominated+ [ΔT _DS1, ΔT _DS2 [+ [f _d0, f _d1 [
1001	NLoS-dominated+ [ΔT _DS1, ΔT _DS2 [+ [f _d1, f _d2 [
1010	NLoS-dominated+ [ΔT _DS1, ΔT _DS2 [+ [f _d2, f _d3 [
1100	NLoS-dominated+ [ΔT _DS0, ΔT _DS1 [+ [f _d0, f _d1 [
1101	NLoS-dominated+ [ΔT _DS0, ΔT _DS1 [+ [f _d1, f _d2 [
1110	NLoS-dominated+ [ΔT _DS0, ΔT _DS1 [+ [f _d2, f _d3 [

Table 1

Further details related to embodiments of the invention will now be described in a 3GPP 5G context with reference to Fig. 10 and 11. Thus, 3GPP 5G terminology, definitions, expressions and system architecture will be used. Especially, the first communication device 100 according to the invention may in these embodiments be configured to perform any of the described functions of a 3GPP UE and will be referred to as a UE 100. The second communication device 300 according to the invention may in these embodiments be configured to perform any of the described functions of a 3GPP gNB and will be referred to as a gNB 300. It may however be noted that embodiments of the invention are not limited thereto.

Fig. 10 and 11 show the management of ML models for CSI reporting according to an embodiment of the invention. In case of CSI compression based on an encoder-decoder model as defined by 3GPP, large scale channel properties may be used to discriminate data, e.g., LoS/NLoS, frequency selectivity, delay spread, etc. Multiple ML models for CSI compression suitable for different propagation scenarios may be prepared. In this way, the learning dimensions can be reduced such that complexity of the ML models for CSI reporting can be distributed. Inference time and/or complexity can also be reduced, which in turn impacts CSI reporting delay and could lead to reducing corresponding computations and power consumption at the UE 100.

With reference to Fig. 10, the UE 100 is at a first time instance t1 configured with or trains multiple ML models for CSI reporting, each adapted for specific propagation conditions, e.g. different LoS/NLoS power ratios, delay spreads, etc.

During a first time interval T1, the UE 100 uses a CSI compression model, e.g. encoder, adapted to detected propagation conditions. In the shown embodiment, it is assumed that during the first time interval T1 the conditions are such that the radio channel between the UE 100 and the gNB 300 is NLoS dominated.

At a second time instance t2, the UE 100 detects a change of LoS/NLoS power ratios indicating that the radio channel is no longer a NLoS dominated radio channel but a LoS dominated radio channel. Note that this detection can be based on measurements obtained at second time instance t2 only or at second time instance t2 and during the first time interval T1.

During a second time interval T2, the UE 100 determines the identity of a new or adapted CSI compression model more suitable for the LOS dominated radio channel based on the detected change of LoS/NLoS power ratios and adapts or switches to the determined CSI compression model. The UE 100 further transmits a first control signal 510 comprising the identity to indicate that the CSI compression model has been adapted or switched. The identity indicating the new propagation conditions in its sequence of bit strings e.g., LoS and delay spread.

Fig. 11 shows signaling between the UE 100 and the gNB 300 for the case shown in Fig. 10. In step I in Fig. 11, the UE 100 and the gNB 300 performs initial access, UE capability transfer and radio resource control (RRC) configuration procedures, including configuration of multiple ML models for CSI reporting, each ML model being adapted for specific propagation conditions and associated with an identity indicating its configuration and supported scenarios. In optional step II in Fig. 11, the UE 100 and the gNB 300 may perform over the air training of the ML models for CSI reporting. In some embodiments, over the air training may be used to train ML models from scratch or to further tune pre-trained ML models.

In step III in Fig. 11, the UE 100 detect propagation conditions and select a ML model for CSI reporting based on downlink reference signal (DL RS) measurements on DL RSs received from the gNB 300. The propagation conditions may e.g., reflect LoS or NLoS conditions, delay spread and/or frequency selectivity of the radio channel.

In step IV in Fig. 11, the UE 100 transmits a first control message 510 to the gNB 300 to inform the gNB 300 about the selected ML model. The first control message 510 comprises the identity for the selected ML model and may be an uplink control information (UCI) or an uplink medium access control (MAC) control element (CE) . The identity may in this case indicate NLoS dominated channel and a first Doppler shift range.

The UE 100 starts using the selected ML model for CSI reporting and continues to perform DL RS measurements to identify propagation conditions, as indicated in step V in Fig. 11.

In step VI in Fig. 11, the UE 100 detects a change in propagation conditions which satisfy a criterion for ML model adaptation and selects a new ML model for CSI reporting. The new ML model being more suitable for the new propagation conditions.

In step VII in Fig. 11, the UE 100 transmits another first control message 510 to the gNB 300, the first control message 510 comprising the identity for the new ML model. The identity may in this case indicate LoS dominated channel, a first delay spread range and a second Doppler shift range. As previously described, the UE 100 may switch ML model autonomously or request activation of the ML model from the gNB 300. The first control message 510 may hence further indicate an activation notification or an activation request for the new ML model.

If the UE 100 switch ML model autonomously, the explicit ML model may be transparent to the gNB 300. However, as the gNB 300 receives the identity of the ML model which indicates the configuration of the ML model, the gNB 300 obtains information about the current propagation conditions, as seen by the UE 100, without requiring other reporting quantities. For example, in the shown embodiment the identity comprises an indication on the LoS/NLoS condition of the radio channel.

In step VII in Fig. 11, the UE 100 starts using the new ML model for CSI reporting and continues to perform DL RS measurements to identify propagation conditions. As previously described, the gNB 300 may further adapt the transmission of DL RS based on the propagation conditions indicated in the received identity.

A first communication device and a second communication device herein may be denoted as a client device which in turn may be denoted as a user device, a user equipment (UE) , a mobile station, an internet of things (IoT) device, a sensor device, a wireless terminal and/or a mobile terminal, and is enabled to communicate wirelessly in a wireless communication system, sometimes also referred to as a cellular radio system. The UEs may further be referred to as mobile telephones, cellular telephones, computer tablets or laptops with wireless capability. The UEs in this context may be, for example, portable, pocket-storable, hand-held, computer-comprised, or vehicle-mounted mobile devices, enabled to communicate voice and/or data, via a radio access network (RAN) , with another communication entity, such as another receiver or a server. The UE may further be a station, which is any device that contains an IEEE 802.11-conformant MAC and PHY interface to the WM. The UE may be configured for communication in 3GPP related LTE, LTE-advanced, 5G wireless systems, such as NR, and their evolutions, as well as in IEEE related Wi-Fi, WiMAX and their evolutions.

A first communication device and a second communication device herein may be denoted as a network access node which in turn may be denoted as a radio network access node, an access network access node, an access point (AP) , or a base station (BS) , e.g., a radio base station (RBS) , which in some networks may be referred to as transmitter, “gNB” , “gNodeB” , “eNB” , “eNodeB” , “NodeB” or “B node” , depending on the standard, technology and terminology used. The radio network access node may be of different classes or types such as e.g., macro eNodeB, home eNodeB or pico base station, based on transmission power and thereby the cell size. The radio network access node may further be a station, which is any device that contains an IEEE 802.11-conformant media access control (MAC) and physical layer (PHY) interface to the wireless medium (WM) . The radio network access node may be configured for communication in 3GPP related long term evolution (LTE) , LTE-advanced, fifth generation (5G) wireless systems, such as new radio (NR) and their evolutions, as well as in IEEE related Wi-Fi, worldwide interoperability for microwave access (WiMAX) and their evolutions.

Furthermore, any method according to embodiments of the invention may be implemented in a computer program, having code means, which when run by processing means causes the processing means to execute the steps of the method. The computer program is included in a computer readable medium of a computer program product. The computer readable medium may comprise essentially any memory, such as previously mentioned a ROM, a PROM, an EPROM, a flash memory, an EEPROM, or a hard disk drive.

Moreover, it should be realized that the first communication device and the second communication device comprise the necessary communication capabilities in the form of e.g., functions, means, units, elements, etc., for performing or implementing embodiments of the invention. Examples of other such means, units, elements and functions are: processors, memory, buffers, control logic, encoders, decoders, rate matchers, de-rate matchers, mapping units, multipliers, decision units, selecting units, switches, interleavers, de-interleavers, modulators, demodulators, inputs, outputs, antennas, amplifiers, receiver units, transmitter units, DSPs, TCM encoder, TCM decoder, power supply units, power feeders, communication interfaces, communication protocols, etc. which are suitably arranged together for performing the solution.

Therefore, the processor (s) of the first communication device and the second communication device may comprise, e.g., one or more instances of a CPU, a processing unit, a processing circuit, a processor, an ASIC, a microprocessor, or other processing logic that may interpret and execute instructions. The expression “processor” may thus represent a processing circuitry comprising a plurality of processing circuits, such as e.g., any, some or all of the ones mentioned above. The processing circuitry may further perform data processing functions for inputting, outputting, and processing of data comprising data buffering and device control functions, such as call processing control, user interface control, or the like.

Finally, it should be understood that the invention is not limited to the embodiments described above, but also relates to and incorporates all embodiments within the scope of the appended independent claims.

Claims

A first communication device (100) for a communication system (500) , the first communication device (100) being configured to:

determine an identity for a machine learning, ML, model for a communication session over a radio channel between the first communication device (100) and a second communication device (300) , the identity indicating a configuration of the ML model; and

transmit a first control signal (510) to the second communication device (300) , the first control signal (510) comprising the identity.
The first communication device (100) according to claim 1, configured to

determine the identity based on one or more of: a channel characteristic of the radio channel, a traffic characteristic of the communication session, a target performance for the communication session, a use case for the ML model, a target latency for the ML model, a complexity of the ML model, and a model version of the ML model.
The first communication device (100) according to claim 1 or 2, wherein the configuration of the ML model comprises one or more information elements associated with one or more of: a channel characteristic of the radio channel, a traffic characteristic of the communication session, a target performance for the communication session, a use case for the ML model, a target latency for the ML model, a complexity of the ML model, and a model version of the ML model.
The first communication device (100) according to any one of the preceding claims, configured to

perform one or more of: activate the ML model, update the ML model, train the ML model, and register the ML model based on the determined identity.
The first communication device (100) according to any one of the preceding claims, wherein the first control signal (510) further indicates one or more of: an activation request for the ML model, an activation notification for the ML model and an activation command for the ML model.
The first communication device (100) according to claim 5, configured to when the first control signal (510) indicates the activation request for the ML model:

receive a second control signal (520) from the second communication device (300) , the second control signal (520) indicating an activation response; and

perform one or more of: activate a ML model, update a ML model, train a ML model, and register a ML model based on the second control signal (520) .
The first communication device (100) according to any one of the preceding claims, configured to

receive a third control signal (530) from the second communication device (300) , the third control signal (530) indicating support for one or more ML models; and

determine the identity further based on the third control signal (530) .
The first communication device (100) according to any one of the preceding claims, wherein the identity comprises a sequence of bit strings, wherein each bit string indicates an information element in the configuration of the ML model.
The first communication device (100) according to any one of the preceding claims, wherein the identity is associated with one or more of: a radio resource configuration for the ML model, a model parameter configuration of the ML model and a model coefficient of the ML model.
A second communication device (300) for a communication system (500) , the second communication device (300) being configured to:

receive a first control signal (510) from a first communication device (100) , the first control signal (510) comprising an identity for a ML model for a communication session over a radio channel between the first communication device (100) and the second communication device (300) , the identity indicating a configuration of the ML model.
The second communication device (300) according to claim 10, wherein the configuration of the ML model comprises one or more information elements associated with one or more of: a channel characteristic of the radio channel, a traffic characteristic of the communication session, a target performance for the communication session, a use case for the ML model, a target latency for the ML model, a complexity of the ML model, and a model version of the ML model.
The second communication device (300) according to claim 10 or 11, configured to

perform one or more of: activate the ML model, update the ML model, train the ML model, and register the ML model based on the first control signal (510) .
The second communication device (300) according to any one of claim 10 to 12, wherein the first control signal (510) further indicates one or more of: an activation request for the ML model, an activation notification for the ML model and an activation command for the ML model.
The second communication device (300) according to claim 13, configured to when the first control signal (510) indicates the activation request for the ML model:

transmit a second control signal (520) to the first communication device (100) , the second control signal (520) indicating an activation response.
The second communication device (300) according to any one of claim 10 to 14, configured to

transmit a third control signal (530) to the first communication device (100) previous to receiving the first control signal (510) , the third control signal (530) indicating support for one or more ML models.
The second communication device (300) according to any one of claim 10 to 15, wherein the identity comprises a sequence of bit strings, wherein each bit string indicates an information element in the configuration of the ML model.
The second communication device (300) according to any one of claim 10 to 16, wherein the identity is associated with one or more of: a radio resource configuration for the ML model, a model parameter configuration of the ML model and a model coefficient of the ML model.
A method (200) for a first communication device (100) , the method (200) comprising

determining (202) an identity for a machine learning, ML, model for a communication session over a radio channel between the first communication device (100) and a second communication device (300) , the identity indicating a configuration of the ML model; and

transmitting (204) a first control signal (510) to the second communication device (300) , the first control signal (510) comprising the identity.
A method (400) for a second communication device (300) , the method (400) comprising

receiving (404) a first control signal (510) from a first communication device (100) , the first control signal (510) comprising an identity for a ML model for a communication session over a radio channel between the first communication device (100) and the second communication device (300) , the identity indicating a configuration of the ML model.
A computer program with a program code for performing a method according to claim 18 or 19 when the computer program runs on a computer.