Evaluation of features in ai air interface architecture

By evaluating AI air interface features with quality indicators and dynamically changing states, the solution addresses the challenge of maintaining optimal performance in AI air interface architectures, ensuring consistent QoS and adapting to network changes.

WO2026238221A1PCT designated stage Publication Date: 2026-11-19HUAWEI TECH CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/EP2025/062870
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Filing Date
2025-05-12
Publication Date
2026-11-19

AI Technical Summary

Technical Problem

Conventional AI air interface architectures face challenges in maintaining optimal performance over time due to the need for continuous monitoring and adaptation of AI models, especially in Al feature-based architectures where feature extraction models can degrade and impact quality-of-service (QoS) without continuous evaluation and management.

Method used

A network device determines quality indicators for AI air interface features using an evaluation function, comparing them with thresholds to assess validity, enabling dynamic state changes and model switching to maintain consistent performance.

Benefits of technology

This approach ensures constant performance efficiency by adapting to changing conditions, reducing signaling overhead, and promptly responding to performance degradation, thus maintaining QoS in AI air interface architectures.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure EP2025062870_19112026_PF_FP_ABST
    Figure EP2025062870_19112026_PF_FP_ABST
Patent Text Reader

Abstract

Embodiments of the invention relate to a first network device and a second network device using evaluation of features in an Al air interface architecture. Therefore, the first network device (100) determines a set of quality indicators (Q) for a set of features (Z) of an Al air interface architecture comprising a set of Al models, wherein a quality indicator (Q ( i )) for each feature (Z( i )) in the set of features (Z) is determined based on inputting a set of data samples associated with the feature (Z ( i )) into an evaluation function (ε ( i )) for the feature (Z ( i )), wherein the set of data samples are obtained during an evaluation time period (Te ( i )) for the feature (Z ( i )), and wherein the evaluation function (ε ( i )) is a real function. The first network device (100) also determines a set of evaluations (V z ) for the set of features (Z) based on comparing, for each feature (Z( i )) in the set of features (Z), the quality indicator (Q( i )) for the feature (Z( i )) with a threshold value for the feature (Z( i )), wherein each evaluation in the set of evaluations (V z ) indicates whether the feature (Z ( i )) is valid or invalid as input in the Al air interface architecture. A first message (510) may be transmitted by the first network device (100) to the second network device (300), where the first message (510) indicates the set of quality indicators (Q) or the set of evaluations (V z ). Furthermore, embodiments of the invention also relate to corresponding methods.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] EVALUATION OF FEATURES IN Al AIR INTERFACE ARCHITECTURE

[0002] TECHNICAL FIELD

[0003] Embodiments of the invention relate to a first network device and a second network device using evaluation of features in an Al air interface architecture. Furthermore, embodiments of the invention also relate to corresponding methods.

[0004] BACKGROUND

[0005] In conventional solutions, two main approaches are considered to include artificial intelligence (Al) into the air interface of communication systems and networks i.e.:

[0006] Channel state information (CSI)-based air interface architectures use Al models for both CSI compression and decompression. The Al models provide a mechanism for efficient data transmission by compressing the CSI into a lower-dimensional representation, then decompressing it for use by conventional air interface algorithms. This process improves communication efficiency by reducing the overhead associated with transmitting raw CSI data.

[0007] Al feature-based air interface architectures, instead, process the CSI input by feeding it into feature extraction models, generating one or more feature vectors. Each feature extraction model can focus on different characteristics of the CSI, allowing for specific understanding and optimization of the air interface.

[0008] In both architectures, the deployed Al models have to be updated to guarantee optimal system performance in the long term.

[0009] 3GPP study item provides a set of general guidelines for the Al models Life-Cycle Management (LCM).

[0010] Accordingly, the following functional blocks may be defined:

[0011] Data collection gathers various types of data, including labeled training data, unlabeled inference data, and monitoring data. These formats support model training, real-time inference, and ongoing performance tracking, with different resolutions depending on the use case.

[0012] Model training uses collected data to create or update Al models. This process involves iterative learning, where the management function requests training based on performance feedback.

[0013] Management oversees the entire lifecycle, including model monitoring, and updates. It ensures that models perform optimally and are re-trained when necessary based on feedback from real- world data.

[0014] Model storage stores the Al models, making them available for inference or updates the models. These models may be transferred between network components to accommodate various use cases and ensure proper pre-training or fine-tuning.

[0015] Inference applies trained models to new, unlabeled data to generate outputs, guiding real-time decisions. It requires low latency and scalability to handle dynamic network conditions and ensure good performance.

[0016] SUMMARY

[0017] An objective of embodiments of the invention is to provide a solution which mitigates or solves the drawbacks and problems of conventional solutions.

[0018] Another objective of embodiments of the invention is to provide improved performance of Al air interface architecture compared to conventional solutions.

[0019] The above and further objectives are solved by the subject matter of the independent claims.

[0020] Further embodiments of the invention can be found in the dependent claims.

[0021] According to a first aspect of the invention, the above mentioned and other objectives are achieved with a first network device configured to:

[0022] determine a set of quality indicators for a set of features of an artificial intelligence, Al, air interface architecture comprising a set of Al models, wherein a quality indicator for each feature in the set of features is determined based on inputting a set of data samples associated with the feature into an evaluation function for the feature, wherein the set of data samples are obtained during an evaluation time period for the feature, and wherein the evaluation function is a real function; and

[0023] determine a set of evaluations for the set of features based on comparing, for each feature in the set of features, the quality indicator for the feature with a threshold value for the feature, wherein each evaluation in the set of evaluations indicates whether the feature is valid or invalid as input in the Al air interface architecture.

[0024] An advantage of the first network device according to the first aspect is that the set of evaluations ensures constant performance efficiency of the Al air interface architecture over time. For example, by changing the state of the Al air interface architecture and / or switching between different Al models so as to adapt to changing conditions in the communication network.

[0025] In an implementation form of a first network device according to the first aspect, the set of features are derived from a channel state information, CSI.

[0026] In an implementation form of a first network device according to the first aspect, each feature in the set of features is parametrized based on an identifier, a dimensionality, and a time-frequency span of feature validity.

[0027] These parameters are suitable for parametrizing the herein used features of the Al air interface architecture.

[0028] In an implementation form of a first network device according to the first aspect, the determining of the quality indicator for the feature comprises:

[0029] averaging the evaluation function over the set of data samples associated with the feature.

[0030] Thereby, temporary or instantaneous fluctuations can be mitigated.

[0031] In an implementation form of a first network device according to the first aspect, the set of data samples for the feature comprises a set of data associated with labels for evaluation of the feature.

[0032] In an implementation form of a first network device according to the first aspect, the first network device is configured to:

[0033] transmit a first message to a second network device, the first message indicating the set of quality indicators or the set of evaluations.

[0034] Thus, the set of quality indicators or the set of evaluations can be exchanged in the communication network so that they may be used in the Al air interface architecture at another location in the communication network.

[0035] In an implementation form of a first network device according to the first aspect, the first network device is configured to:

[0036] transmit the first message to the second network device upon receiving a second message from the second network device, the second message indicating a request for the set of quality indicators or the set of evaluations.

[0037] Thereby, a request / response signaling procedure is provided.

[0038] In an implementation form of a first network device according to the first aspect, the set of evaluations is indicated with a bitmap.

[0039] An advantage with this implementation form is the low overhead of the signaling needed to exchange feature evaluations-related information across different network devices.

[0040] In an implementation form of a first network device according to the first aspect, the first network device is configured to:

[0041] receive a third message from a second network device, the third message indicating a state of the Al air interface architecture; and

[0042] change a state of the Al air interface architecture based on the third message.

[0043] Thereby, a centralized solution is provided.

[0044] In an implementation form of a first network device according to the first aspect, the state of the Al air interface architecture is indicated with a bitmap.

[0045] An advantage with this implementation form is the low overhead of the signaling needed to exchange state-related information across different devices.

[0046] In an implementation form of a first network device according to the first aspect, the first network device is configured to:

[0047] change a state of the Al air interface architecture based on the set of evaluations.

[0048] Thereby, the performance of the Al air interface architecture can constantly be updated according to changes in the data samples.

[0049] In an implementation form of a first network device according to the first aspect, the state of the Al air interface architecture is:

[0050] a first state in which all features in the set of features are invalid as input into any Al model in the set of Al models; or a second state in which at least one feature in the set of features is valid as input into at least one Al model in the set of Al models; or

[0051] a third state in which all features in the set of features are valid as input into any Al model in the set of Al models.

[0052] In an implementation form of a first network device according to the first aspect, the first network device is configured to change the state of the Al air interface architecture:

[0053] from the first state to the second state when least one feature in the set of features is valid;

[0054] from the first state to the third state when all features in the set of features are valid;

[0055] from the second state to the first state when all features in the set of features are invalid;

[0056] from the second state to the third state when all features in the set of features are valid;

[0057] from the third state to the first state when all features in the set of features are invalid; and

[0058] from the third state to the second state when at least one feature in the set of features is valid.

[0059] Thereby, rules for change of state are provided for improved performance.

[0060] In an implementation form of a first network device according to the first aspect, changing the state of the Al air interface architecture comprises:

[0061] changing from a first Al model to a second Al model in the set of Al models.

[0062] An advantage with this implementation form is that the second Al model can be used in the Al air interface while waiting for a new training procedure of the first Al model.

[0063] In an implementation form of a first network device according to the first aspect, the first network device is a base station of a radio access network.

[0064] According to a second aspect of the invention, the above mentioned and other objectives are achieved with a second network device configured to:

[0065] receive a first message from a first network device, the first message indicating a set of quality indicators for a set of features or a set of evaluations for the set of features for an Al air interface architecture comprising a set of Al models, wherein a quality indicator for each feature in the set of features is determined based on inputting a set of data samples associated with the feature into an evaluation function for the feature, wherein the set of data samples are obtained during an evaluation time period for the feature, and wherein the evaluation function is a real function, and wherein each evaluation in the set of evaluations indicates whether a feature in the set of features is valid or invalid as input in the Al air interface architecture;

[0066] determine a state of the Al air interface architecture based on the first message; and

[0067] transmit a third message to the first network device, the third message indicating the determined state of the Al air interface architecture.

[0068] An advantage of the second network device according to the second aspect is that the set of quality indicators or the set of evaluations can be exchanged in the communication network so that they may be used in the Al air interface architecture at another location in the communication network. The set of evaluations ensures constant performance efficiency of the Al air interface architecture over time. For example, by changing the state ofthe Al air interface architecture and / or switching between different Al models so as to adapt to changing conditions in the communication network.

[0069] In an implementation form of a second network device according to the second aspect, the set of evaluations is indicated with a bitmap.

[0070] An advantage with this implementation form is the low overhead of the signaling needed to exchange feature evaluations-related information across different network devices.

[0071] In an implementation form of a second network device according to the second aspect, the state of the Al air interface architecture is indicated with a bitmap.

[0072] An advantage with this implementation form is the low overhead of the signaling needed to exchange feature evaluations-related information across different network devices.

[0073] In an implementation form of a second network device according to the second aspect, the first message indicates the set of quality indicators, and wherein the second network device is configured to:

[0074] determine the set of evaluations for the set of features based on comparing, for each feature in the set of features, the quality indicator for the feature with a threshold value for the feature.

[0075] In an implementation form of a second network device according to the second aspect, the second network device is configured to:

[0076] receive the first message in response to a transmission of a second message to the first network device, the second message indicating a request for the first message.

[0077] Thereby, a request / response signaling procedure is provided.

[0078] In an implementation form of a second network device according to the second aspect, the second network device is configured to:

[0079] transmit the second message to the first network device upon detecting that a quality-of-service, QoS, for the Al air interface architecture is below a QoS threshold value.

[0080] An advantage with this implementation form is the prompt reaction to performance degradation detected at the higher layers of the protocol stack. Such degradation can be due to sudden changes of the channel conditions, which could not be promptly compensated when waiting for periodic feature evaluation.

[0081] In an implementation form of a second network device according to the second aspect, the state of the Al air interface architecture is:

[0082] a first state in which all features in the set of features are invalid as input into any Al model in the set of Al models; or a second state in which at least one feature in the set of features is valid as input into at least one Al model in the set of Al models; or

[0083] a third state in which all features in the set of features are valid as input into any Al model in the set of Al models.

[0084] In an implementation form of a second network device according to the second aspect, the second network device is configured to:

[0085] determine a change of the state of the Al air interface architecture:

[0086] from the first state to the second state when least one feature in the set of features is valid;

[0087] from the first state to the third state when all features in the set of features are valid;

[0088] from the second state to the first state when all features in the set of features are invalid;

[0089] from the second state to the third state when all features in the set of features are valid;

[0090] from the third state to the first state when all features in the set of features are invalid; and

[0091] from the third state to the second state when at least one feature in the set of features is valid.

[0092] Thereby, rules for change of state are provided for improved performance.

[0093] In an implementation form of a second network device according to the second aspect, changing the state of the Al air interface architecture comprises:

[0094] changing from a first Al model to a second Al model in the set of Al models.

[0095] An advantage with this implementation form is that the second Al model can be used in the Al air interface while waiting for a new training procedure of the first Al model.

[0096] According to a third aspect of the invention, the above mentioned and other objectives are achieved with a method for a first network device, the method comprises:

[0097] determining a set of quality indicators for a set of features of an artificial intelligence, Al, air interface architecture comprising a set of Al models, wherein a quality indicator for each feature in the set of features is determined based on inputting a set of data samples associated with the feature into an evaluation function for the feature, wherein the set of data samples are obtained during an evaluation time period for the feature, and wherein the evaluation function is a real function; and

[0098] determining a set of evaluations for the set of features based on comparing, for each feature in the set of features, the quality indicator for the feature with a threshold value for the feature, wherein each evaluation in the set of evaluations indicates whether the feature is valid or invalid as input in the Al air interface architecture.

[0099] The method according to the third aspect can be extended into implementation forms corresponding to the implementation forms of the first network 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 network device.

[0100] The advantages of the methods according to the third aspect are the same as those for the corresponding implementation forms of the first network device according to the first aspect.

[0101] According to a fourth aspect of the invention, the above mentioned and other objectives are achieved with a method for a second network device, the method comprises:

[0102] receiving a first message from a first network device, the first message indicating a set of quality indicators for a set of features or a set of evaluations for the set of features for an Al air interface architecture comprising a set of Al models, wherein a quality indicator for each feature in the set of features is determined based on inputting a set of data samples associated with the feature into an evaluation function for the feature, wherein the set of data samples are obtained during an evaluation time period for the feature, and wherein the evaluation function is a real function, and wherein each evaluation in the set of evaluations indicates whether a feature in the set of features is valid or invalid as input in the Al air interface architecture;

[0103] determining a state of the Al air interface architecture based on the first message; and

[0104] transmitting a third message to the first network device, the third message indicating the determined state of the Al air interface architecture.

[0105] The method according to the fourth aspect can be extended into implementation forms corresponding to the implementation forms of the second network 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 network device.

[0106] The advantages of the methods according to the fourth aspect are the same as those for the corresponding implementation forms of the second network device according to the second aspect.

[0107] 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.

[0108] Further applications and advantages of embodiments of the invention will be apparent from the following detailed description.

[0109] BRIEF DESCRIPTION OF THE DRAWINGS

[0110] The appended drawings are intended to clarify and explain different embodiments of the invention, in which:

[0111] - Fig. 1 shows a first communication device according to embodiments of the invention;

[0112] - Fig. 2 shows a flow chart of a method for a first communication device according to embodiments of the invention;

[0113] - Fig. 3 shows a second communication device according to embodiments of the invention;

[0114] - Fig.4 shows a flow chart of a method for a second communication device according to embodiments of the invention;

[0115] - Fig. 5 shows a communication system according to embodiments of the invention;

[0116] - Fig. 6 shows a signaling diagram according to embodiments of the invention;

[0117] - Fig. 7 shows Al features evaluation scheme according to embodiments of the invention;

[0118] - Fig. 8 illustrates timeline of model initialization, training, inference and evaluation according to embodiments of the invention;

[0119] - Fig. 9 shows Al features local and centralized evaluation timeline according to embodiments of the invention; and

[0120] - Fig. 10 illustrates a switching procedure according to embodiments of the invention.

[0121] DETAILED DESCRIPTION

[0122] By design, the CSI-based air interface architecture can only operate in two distinct states:

[0123] Al-enabled state: in this state, the air interface uses Al models for both CSI compression and decompression. The Al models provide a mechanism for efficient data transmission by compressing the CSI into a lower-dimensional representation, then decompressing it for use by conventional air interface algorithms. This process improves communication efficiency by reducing the overhead associated with transmitting raw CSI data.

[0124] Al-disabled state: in this state, Al models are not used, and the system falls back on conventional air interface techniques, such as 3GPP CSI compression. These methods typically involve simple quantization techniques for compression, without leveraging any data driven approach for compression and reconstruction, thus being less efficient than the Al-based approach in terms of overhead and accuracy.

[0125] The Al feature-based air interface architectures, instead, allows for a more flexible state definition and usage, which is not exploited by the conventional approaches. The conventional techniques for CSI-based air interface and the 3GPP LCM framework can be adopted for the Al feature-based architectures, but their direct usage does not fully exploit the potential of such architectures. In Al feature-based air interface, several states can be defined, each one depending on the set of active feature extractors. This allows for flexible Al features usage, e.g., it prevents deactivation and re-training of the full Al-based air interface.

[0126] Al feature-based architectures can be decoupled into two main stages:

[0127] Feature extraction: Here, specialized models are used to compress raw channel data, such as CSI or sensor data, into low-dimensional features that are easier to process and transmit. These models are responsible for reducing the size of raw CSI while retaining the essential information for subsequent processing.

[0128] Al-based air interface algorithms: Once features are extracted, other Al models that implement specific wireless functionalities, take these features as inputs. So-called Al wireless functionality models are deployed across different nodes in the network. These nodes may include the base station (BS), user equipment (UE), or the central unit (CU).

[0129] Al-based air interface architectures need to guarantee quality-of-service (QoS) for the long-term system performance. While ensuring QoS at each node is crucial, it is impossible to monitor the CSI reconstruction accuracy, as the CSI is not reconstructed in Al feature-based air interface architectures. If the extracted features lose their representational quality, the Al functionality models will produce inaccurate outputs, degrading overall system performance. There exists therefore a need to introduce a metric to evaluate the feature extraction performance, as it indirectly impacts the QoS performance. Thus, Al model re-trainingand fine-tuning may be triggered periodically when the representations are not accurate enough to ensure QoS of the air interface. Additionally, feature extraction models may need to be switched or replaced entirely if they no longer meet performance standards.

[0130] In conclusion, the long-term maintenance of QoS and the related representation accuracy in Al-based air interface architectures requires continuous monitoring, adaptation, and management of both feature extraction models and Al wireless functionality models. A set of procedures performed at the CU has to be defined to manage these aspects and thus ensure reliable and high-performance communication across all nodes.

[0131] Fig. 1 shows a first network device 100 according to embodiments of the invention. In the embodiment shown in Fig. 1, the first network 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 network device 100 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 110 coupled to the transceiver 104, while the wired communication capability may be provided with a wired communication interface 112 e.g., coupled to the transceiver 104.

[0132] 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 104 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 network device 100 is configured to perform certain actions can in this disclosure be understood to mean that the first network device 100 comprises suitable means and devices, such as e.g., the processor 102 and the transceiver 104, configured to perform the actions.

[0133] According to embodiments of the invention the first network device 100 is configured to: determine a set of quality indicators Q for a set of features Z of an artificial intelligence, Al, air interface architecture comprising a set of Al models, wherein a quality indicator for each feature Z!iiin the set of features Z is determined based on inputting a set of data samples associated with the feature Z!iiinto an evaluation function 8 for the feature ZLI, wherein the set of data samples are obtained during an evaluation time period

[0134]

[0135] for the feature Z1'. and wherein the evaluation function 8LIis a real function; and determine a set of evaluations Vzfor the set of features Z based on comparing, for each feature Z!iiin the set of features Z. the quality indicator

[0136]

[0137] for the feature Z!iiwith a threshold value for the feature Z!ii. wherein each evaluation in the set of evaluations Vzindicates whether the feature Z!iiis valid or invalid as input in the Al air interface architecture.

[0138] Furthermore, in an embodiment of the invention, the first network device 100 comprises a processor configured to: determine a set of quality indicators Q for a set of features Z of an Al air interface architecture comprising a set of Al models, wherein a quality indicator

[0139]

[0140] for each feature Z!iiin the set of features Z is determined based on inputting a set of data samples associated with the feature Z!iiinto an evaluation function 8 for the feature ZLI, wherein the set of data samples are obtained during an evaluation time period

[0141]

[0142] for the feature Z1'. and wherein the evaluation function 8LIis a real function; and determine a set of evaluations Vzfor the set of features Z based on comparing, for each featureL>in the set of features Z. the quality indicator Q(l)for the feature ZLIwith a threshold value for the feature Z1'. wherein each evaluation in the set of evaluations Vzindicates whether the featureL>is valid or invalid as input in the Al air interface architecture.

[0143] Moreover, in yet another embodiment of the invention, the first network device 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 a set of quality indicators Q for a set of features Z of an Al air interface architecture comprising a set of Al models, wherein a quality indicator

[0144]

[0145] for each feature Z!iiin the set of features Z is determined based on inputting a set of data samples associated with the feature Z!iiinto an evaluation function 8!iifor the feature Z!ii. wherein the set of data samples are obtained during an evaluation time period

[0146]

[0147] for the feature Z1'. and wherein the evaluation function 8LIis a real function; and determine a set of evaluations Vzfor the set of features Z based on comparing, for each feature Z!iiin the set of features Z. the quality indicator

[0148]

[0149] for the feature Z!iiwith a threshold value for the feature Z!ii. wherein each evaluation in the set of evaluations Vzindicates whether the feature Z!iiis valid or invalid as input in the Al air interface architecture.

[0150] Fig. 2 shows a flow chart of a corresponding method 200 which may be executed in a first network device 100, such as the one shown in Fig. 1. The method 200 comprises: determining 202 a set of quality indicators Q for a set of features Z of an Al air interface architecture comprising a set of Al models, wherein a quality indicator

[0151]

[0152] for each feature Z!iiin the set of features Z is determined based on inputting a set of data samples associated with the feature Z!iiinto an evaluation function 8!iifor the featureL>, wherein the set of data samples are obtained during an evaluation time period

[0153]

[0154] for the feature Z1'. and wherein the evaluation function 8LIis a real function; and determining 204 a set of evaluations Vzfor the set of features Z based on comparing, for each featureL>in the set of features Z. the quality indicator Q(l)for the feature ZLIwith a threshold value for the feature Z1'. wherein each evaluation in the set of evaluations Vzindicates whether the featureL>is valid or invalid as input in the Al air interface architecture.

[0155] Fig. 3 shows a second network device 300 according to embodiments of the invention. In the embodiment shown in Fig. 3, the second network 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 network 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.

[0156] 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 304 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 network device 300 is configured to perform certain actions can in this disclosure be understood to mean that the second network device 300 comprises suitable means and devices, such as e.g., the processor 302 and the transceiver 304, configured to perform the actions.

[0157] According to embodiments of the invention the second network device 300 is configured to: receive a first message 510 from a first network device 100, the first message 510 indicating a set of quality indicators Q for a set of features Z or a set of evaluations Vzfor the set of features Z for an Al air interface architecture comprising a set of Al models, wherein a quality indicator Q(l)for each feature ZLIin the set of features Z is determined based on inputting a set of data samples associated with the feature ZLIinto an evaluation function 8LIfor the feature Z1'. wherein the set of data samples are obtained duringan evaluation time period for the feature Z ‘8 and wherein the evaluation function £LIis a real function, and wherein each evaluation in the set of evaluations Vzindicates whether a feature Z!iiin the set of features Z is valid or invalid as input in the Al air interface architecture; determine a state S of the Al air interface architecture based on the first message 510; and transmit a third message 530 to the first network device 100, the third message 530 indicating the determined state S of the Al air interface architecture.

[0158] Furthermore, in an embodiment of the invention, the second network device 300 comprises a transceiver configured to: receive a first message 510 from a first network device 100, the first message 510 indicating a set of quality indicators Q for a set of features Z or a set of evaluations Vzfor the set of features Z for an Al air interface architecture comprising a set of Al models, wherein a quality indicator Q(l)for each feature ZLIin the set of features Z is determined based on inputting a set of data samples associated with the feature ZLIinto an evaluation function

[0159]

[0160] for the feature Z!ii. wherein the set of data samples are obtained during an evaluation time period

[0161]

[0162] for the feature Z<‘£ and wherein the evaluation function £LIis a real function, and wherein each evaluation in the set of evaluations Vzindicates whether a featureL>in the set of features Z is valid or invalid as input in the Al air interface architecture. The second network device 300 comprises a processor configured to: determine a state S of the Al air interface architecture based on the first message 510. The transceiver is configured to transmit a third message 530 to the first network device 100, the third message 530 indicating the determined state S of the Al air interface architecture.

[0163] Moreover, in yet another embodiment of the invention, the second network 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 message 510 from a first network device 100, the first message 510 indicating a set of quality indicators Q for a set of features Z or a set of evaluations Vzfor the set of features Z for an Al air interface architecture comprising a set of Al models, wherein a quality indicator Q(l)for each feature ZLIin the set of features Z is determined based on inputting a set of data samples associated with the feature ZLIinto an evaluation function £LIfor the feature Z ‘£ wherein the set of data samples are obtained during an evaluation time period

[0164]

[0165] for the feature Z ‘£ and wherein the evaluation function £LIis a real function, and wherein each evaluation in the set of evaluations Vzindicates whether a featureL>in the set of features Z is valid or invalid as input in the Al air interface architecture; determine a state S of the Al air interface architecture based on the first message 510; and transmit a third message 530 to the first network device 100, the third message 530 indicating the determined state S of the Al air interface architecture.

[0166] Fig. 4 shows a flow chart of a corresponding method 400 which may be executed in a second network device 300, such as the one shown in Fig. 3. The method 400 comprises: receiving 402 a first message 510 from a first network device 100, the first message 510 indicating a set of quality indicators Q for a set of features Z or a set of evaluations Vzfor the set of features Z for an Al air interface architecture comprising a set of Al models, wherein a quality indicator Q(l)for each feature ZLIin the set of features Z is determined based on inputting a set of data samples associated with the feature ZLIinto an evaluation function £LIfor the feature Z ‘£ wherein the set of data samples are obtained during an evaluation time period

[0167]

[0168] for the feature Z ‘£ and wherein the evaluation function £LIis a real function, and wherein each evaluation in the set of evaluations Vzindicates whether a featureL>in the set of features Z is valid or invalid as input in the Al air interface architecture; determining 404 a state S of the Al air interface architecture based on the first message 510; and transmitting 406 a third message 530 to the first network device 100, the third message 530 indicating the determined state S of the Al air interface architecture.

[0169] Fig. 5 shows a communication system or network 500 according to an embodiment of the invention. The communication system 500 in the disclosed embodiment comprises a first network device 100 and a second network device 300 configured to communicate and operate in the communication system 500. In this particular example, the first network device 100 is a BS of a radio access network (RAN) while the second network device 300 is a central node of a communication network 500. The first network device 100 is configured to serve terminal / client devices such as UEs. The communication between the first network device 100 and the terminal devices may be performed via downlink / uplink (DL / UL) channels according to standardized protocols. The first network device 100 and the second network device 300 are also configured to communicate with each other e.g., by signaling of first 510, second 520 and third 530 messages using a wired, a wireless or a wired and wireless backhaul connection such as a fiber backhaul for wired communications and fixed- wireless access link for wireless communications.

[0170] Further details related to embodiments of the invention will fully or partially be described in a 3GPP context. Thus, 3GPP terminology, definitions, expressions and system architecture may be used. Especially, the first network device 100 according to the invention may in these embodiments be configured to perform any of the described functions of a 3GPP BS 100. Correspondingly, the second network device 300 according to the invention may in these embodiments be configured to perform any of the described functions of a 3GPP central unit (CU) 300, which may be configured to control and process data streams in the whole wireless access network. It may however be noted that embodiments of the invention are not limited thereto and other suitable devices, units and functions of the communication network may be employed.

[0171] Fig. 6 shows a signaling diagram illustrating signaling between the first network device 100 and the second network device 300 according to embodiments of the invention. In these embodiments, the second network device 300 triggers the first network device 100 to change a state S of the Al air interface architecture. It is however noted that the first network device 100 in other embodiments of the invention may autonomously change a state S of the Al air interface architecture without the involvement of the second network device 300.

[0172] In step I in Fig. 6, the CU 300 sends a second message 520 to the BS 100. The second message 520 indicates a request for a set of quality indicators Q or a set of evaluations Vz.

[0173] In embodiments of the invention, the CU 300 transmits the second message 520 to the BS 100 upon detecting that a QoS for the Al air interface architecture is below a QoS threshold value. The QoS of the air interface may periodically be reported to CU 300. If the QoS does not meet the requirements specified by the CU 300, the CU 300 sends a special request to trigger the evaluation of the features that are used as input to the air interface. The QoS may refer to an application in the upper layers of the protocol stack rather than the complete air interface. The special request is represented by the second message 520 in this example.

[0174] The requested features have to be evaluated despite their evaluation period T^', after the evaluation, the evaluation timer is reset for a next evaluation stage.

[0175] In step II in Fig. 6, the BS 100 receives the second message 520 from the CU 300. The BS 100 decodes the second message 520 to obtain the information therein, i.e., the request for the first message 510.

[0176] In step III in Fig. 6, the BS 100 thus transmits a first message 510 to the CU 300 in response to the reception of the second message 520 from the CU 300. As previously mentioned, the first message 510 indicates the set of quality indicators Q or the set of evaluations Vzdetermined by the BS 100. In embodiments of the invention, the set of evaluations Vzis indicated with a bitmap in the first message 510.

[0177] In step IV in Fig. 6, the CU 300 receives the first message 510 from the BS 100. By decoding the first message 510, the CU 300 obtains the information about the set of quality indicators Q or the set of evaluations Vz.

[0178]

[0179] In step V in Fig. 6, the CU 300 transmits a third message 530 to the BS 100. The third message 530 indicates the determined state S of the Al air interface architecture. The state S of the Al air interface architecture may also be indicated with a bitmap in the third message 530 according to embodiments of the invention.

[0180] In step VI in Fig. 6, the BS 100 receives the third message 530 from the CU 300. By decoding the third message 530, the BS 100 derives the information about the determined state S of the Al air interface architecture.

[0181] In step VII in Fig. 6, the BS 100 changes a state S of the Al air interface architecture based on the third message 530. That is, changing to the state determined by the CU 300. As aforementioned, the BS 100 may also in other embodiments ofthe invention change the state independently of the control signaling in the third message 530. In such cases, the BS 100 is configured to change a state S of the Al air interface architecture based on the set of evaluations Vz.

[0182] The first 510, second 520 and third 530 messages may be any suitable message types for communication systems and networks. They may also be denoted control messages and may or may not conform to standardized protocols. The bitmaps herein described may be carried by these messages.

[0183] The air interface may be defined as specification of technologies and protocols, usually involving algorithms from the physical layer (i.e., LI) and data link layer (i.e., L2) of the open system interconnection (OSI) model, enabling the wireless communication between a transmitter and a receiver in a wireless communication system using a set of radio resources elements in the time-frequency grid, denoted as transmission block. Examples of air interface algorithms are: precoder design, scheduler design, user localization, beam forming, etc.

[0184] The Al-based air interface may refer to an air interface in which the set of protocols or algorithms implementing the specific wireless functionalities are Al-based, i.e., they use pre-trained models based on data to take the corresponding actions / decisions.

[0185] The air interface algorithm or wireless functionality is defined as the process or set of rules taking decisions / actions enabling the transmission between a transmitter and a receiver using a transmission block (TB). It is noted that each wireless functionality can choose a transmission block of different sizes, i.e., different L and F, depending on the needed properties. Indeed, in modem communications systems and networks, the actions performed by a wireless functionality usually depend on the propagation channel over which the communication takes place, and the dynamic of the channel is dependent on the considered resource block (RB).

[0186] A concept of the disclosed invention is the previously mentioned concept of quality indicator (QI) metric or quality-of-representation (QoR) metric.

[0187]

[0188] For features evaluation, the features may be evaluated with the agreed function QI;, as shown in Fig. 7 of Al features evaluation scheme.

[0189] Generally, the set of features Z are derived by the BS 100 from a CSI. Also, each feature ZLIin the set of features Z may be parametrized based on an identifier, a dimensionality, and a time-frequency span of feature validity. As for the QI Q for a feature, the QI Q for the feature Z!iimay be determined by averaging the evaluation function 8!iiover the set of data samples associated with the feature Z!ii. The set of data samples for the feature Z!iimay comprise a set of data associated with labels for evaluation of the feature Z!ii.

[0190] These properties of the features may be defined in an initialization stage, with an agreement stage between the BS 100 and the CU 300.

[0191]

[0192] Another concept of the disclosed invention is an LCM solution based on a discrete number of states of the Al air interface architecture and the procedure to change the state of the Al air interface architecture. Each state enables a different set of features based on their QI metric. A length-0 codeword may be used as bitmap to codify the specific state.

[0193] A set of motoring data samples and associated labels is gathered at the BS 100. In some cases, labels can be derived from the channel data through state-of-the-art signal processing. The monitored data is then used for the feature extraction models inference, which output the feature set For each feature in the feature set, QI indicators then computed and compared to the respective thresholds to trigger a state change.

[0194] As for the states and the change of state of the Al air interface architecture, the following may be noted which is applicable to both the first network device 100 and the second network device 300.

[0195] The state S of the Al air interface architecture may be any of the following.

[0196] A first state 5, in which all features in the set of features Z are invalid as input into any Al model in the set of Al models. Also denoted an Al disabled state: in this state, QI;< QI;for all i, and none of the extracted features can be shared with BSs and associated UEs, and Al-based functionalities are completely disabled. The bitmap assigned to this state may be 0 = [0,...,0].

[0197] A second state S2in which at least one feature in the set of features Z is valid as input into at least one Al model in the set of Al models. Also denoted Al partially enabled states: in these states having 2D— 2 states in total:

[0198] Only a subset DEof features are enabled;

[0199] Features not belonging to DEcannot be shared with the CU 300, the BSs 100, and associated UEs 600;

[0200] Features not belonging to DEcannot be used as input of the Al-based air interface algorithm; and

[0201] The associated bitmap may be x = [x1;..., xD] with x;= 1 if QI;> QI;, and x;= 0 if QI;> QI;.

[0202] A third state S3in which all features in the set of features Z are valid as input into any Al model in the set of Al models. Also denoted an Al enabled state: in this state, QI;> QI;for all i. All the extracted features can be shared with the CU 300, BSs 100, and associated UEs 600, and Al-based functionalities are enabled. The bitmap assigned to this state may be 1 = [1,...,1],

[0203] A fourth state may also be defined when all Al functionalities are completely disabled. Also denoted an Al uninitialized state: in this state, all Al functionalities are completely disabled, The BS 100 needs to coordinate with the CU 300 and the UEs 600 to start using Al functionalities in this state. No bitmap may be assigned to this state.

[0204] The timeline of model initialization, training, inference and evaluation is illustrated in Fig. 8.

[0205] During the initialization stage, the agreement of both the raw channel data format and the feature list, including QI evaluation functions and periods, between all the nodes in the network takes place and the transition from “Al Uninitialized” to the “Al Disabled” state occurs. In Fig. 8, the nodes are UEs 600, BS 100 and CU 300.

[0206] During the training stage, training data is collected according to the agreed format and the Al models for features extraction are trained. The training process of feature i extraction model ends when QI;> QI;. i.e., when the quality indicator QI;is above the requested threshold. The completion of the training procedure for one feature extraction model triggers a change of the state of the Al-based air interface. Subsequently, the obtained training features can be used as input to train or fine-tune the Al models for the desired air interface functionalities.

[0207] During the inference stage the Al-based air interface is used for communication purposes. Periodically, the extracted features are evaluated: channel data for monitoring purpose from both the CU 300 and associated UEs 600 is acquired and given as input to the feature extractors.

[0208] During the evaluation stage, the extracted features are evaluated, then the CU 300 informs the BS 100 on the features that have eventually to be disabled i.e., possible state transition by transmitting the corresponding bitmap.

[0209] The bitmap may consist in a sequence of ones in the position indexed the active features and zeros in the positions indexed by inactive features. The bit sequence is then transmitted by the CU 300 to the BS 100 to trigger a state update at the BS 100, which activates the features in the positions of the ones and deactivates the rest of the features.

[0210] The evaluation timelines are illustrated in Fig. 9 which shows the Al features local (left) and the centralized (right) evaluation timeline.

[0211] Local feature evaluation: QI evaluation may in examples of the invention be performed locally at the BS 100, which periodically reports to the CU 300 the QIfindicators of each feature, each with the specific period T^1. The features may be agreed between the BS 100 and the CU 300 during the initialization procedure.

[0212] Centralized feature evaluation: QI evaluation may, in examples of the invention, be performed at the CU 300. The BS 100 may thus periodically, according to the feature- specific period T®, extract the features to be evaluated and send them to the CU 300, and the CU 300 computes the QIfmetrics and eventually triggers a state transition e.g., by signaling to the BS 100 as previously described. Note that with this approach, the BS 100 cannot lie about their QI metrics.

[0213] Model switching: the state transition is triggered by the result of the feature extraction evaluation, which is performed with period T®.

[0214] The condition to change a state of the Al air interface architecture may conform to the following conditions. The change of the state S of the Al air interface architecture may follow the rules of change:

[0215] From the first state 5, to the second state S2when least one feature in the set of features Z is valid;

[0216] From the first state 5, to the third state S3when all features in the set of features Z are valid;

[0217] From the second state S2to the first state 5, when all features in the set of features Z are invalid;

[0218] From the second state S2to the third state S3when all features in the set of features Z are valid;

[0219] From the third state S3to the first state 5, when all features in the set of features Z are invalid; and

[0220] From the third state S3to the second state S2when at least one feature in the set of features Z is valid.

[0221] The model switching procedure allows to prevent deactivation of the Al-based air interface and unnecessary re-training of feature extractors or at the BS 100 when entering in partially enabled state.

[0222] Fig. 10 hence illustrates a switching procedure according to embodiments of the invention. The meaning of changing the state of the Al air interface architecture may in embodiments of the invention be that changing the state S of the Al air interface architecture comprises changing from a first Al model All to a second Al model AI2 in the set of Al models. Thus, in the initial state, model Aj for air interface functionality <f> is used, which takes as input a subset of features. If any of these features is disabled, and if an algorithm Ak,p,k j, for air interface functionality <f> taking as input a subset of enabled features exists, then it is adopted. If the air interface is deployed at the CU 300 or the UEs 600 nodes, the new Al model may be delivered by the BS 100 and trained. If none of the features needed for air interface functionality <f> is enabled, the conventional state-of-the-art model for air interface functionality <f> based on CSI is adopted, while the feature extractors are re-trained. Further changes in the state may then re-enable an Al model for functionality <f>.

[0223] The Al-based air interface architecture using wireless environment state information (WESI) has been described in the art. In the WESI architecture, a set of predicted features at time T + P is obtained executing the following procedure:

[0224] WESI acquisition: WESI for time T is acquired at the transmitter.

[0225] WESI tokenization: WESI for time T is tokenized using a pre-trained model.

[0226] WESI sequence predictor: Given past WESI tokens until time T within a time window spanning the predictor memory, the sequence of WESI tokens characterizing the transmission resource block starting time T + P is predicted.

[0227] WESI structured representation: Given the sequence of predicted WESI tokens, the structured WESI representation at time T + P is obtained, which gathers the D individual features. Each individual feature characterizes a physical

[0228] quantity defined in the in the “WESI Property List” at a specific time and frequency granularity covering the transmission resource block starting at time T + P.

[0229] The state definitions and transitions herein disclosed can be directly applied to the WESI Al-based air interface. Moreover, for this specific architecture, a global QI metric indicator, which is a trigger for tokenization training, can be used to deactivate the

[0230] Al functionalities. The global QI indicator may be defined as:

[0231]

[0232] where SGis a function evaluating all the extracted features simultaneously. When QIG< QIG, i.e., less than a threshold value, a transition to the Al disabled state is triggered, and the full WESI Al-based air interface is re-trained, including both the WESI tokenizer and the WESI predictor.

[0233] A BS herein may also be denoted as a radio network access node, an access network access node, an access point (AP), 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 such as mobile networks for the future, as well as in IEEE related Wi-Fi, worldwide interoperability for microwave access (WiMAX) and their evolutions.

[0234] A terminal device or a client device herein may be denoted as a user device, a user equipment (UE), a mobile station, an internet of things (loT) 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 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 such as mobile networks for the future, as well as in IEEE related Wi-Fi, WiMAX and their evolutions.

[0235] 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.

[0236] Moreover, it should be realized that the first network device 100 and the second network device 300 comprise any 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.

[0237] Therefore, the processor(s) of the first network device 100 and the second network device 300 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.

[0238] 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

CLAIMS1. A first network device (100) configured to:determine a set of quality indicators (Q) for a set of features (Z) of an artificial intelligence, Al, air interface architecture comprising a set of Al models, wherein a quality indicatorfor each feature (Z^) in the set of features (Z) is determined based on inputting a set of data samples associated with the feature (Z^) into an evaluation function (S^) for the feature (Z(l>), wherein the set of data samples are obtained during an evaluation time period (T^ ) for the feature (Z1>) and wherein the evaluation function (S^) is a real function; anddetermine a set of evaluations (Vz) for the set of features (Z) based on comparing, for each feature (Z^) in the set of features (Z), the quality indicatorfor the feature (Z'1’) with a threshold value for the feature (Z^), wherein each evaluation in the set of evaluations (Vz) indicates whether the feature (Z^) is valid or invalid as input in the Al air interface architecture.

2. The first network device (100) according to claim 1, wherein the set of features (Z) are derived from a channel state information, CSI.

3. The first network device (100) according to any one of the preceding claims, wherein each feature (Z'1’) in the set of features (Z) is parametrized based on an identifier, a dimensionality, and a time-frequency span of feature validity.

4. The first network device (100) according to any one of the preceding claims, wherein the determining of the quality indicator (Q) for the feature (Z'1’) comprises:averaging the evaluation function (S^) over the set of data samples associated with the feature (Z'1’).

5. The first network device (100) according to any one of the preceding claims, wherein the set of data samples for the feature (Z'1’) comprises a set of data associated with labels for evaluation of the feature (Z'1’).

6. The first network device (100) according to any one of the preceding claims, configured to:transmit a first message (510) to a second network device (300), the first message (510) indicating the set of quality indicators (Q) or the set of evaluations (Vz).

7. The first network device (100) according to claim 6, configured to:transmit the first message (510) to the second network device (300) upon receiving a second message (520) from the second network device (300), the second message (520) indicating a request for the set of quality indicators (Q) or the set of evaluations (Vz).

8. The first network device (100) according to claim 6 or 7, wherein the set of evaluations (Vz) is indicated with a bitmap.

9. The first network device (100) according to any one of claims 1 to 7, configured to:receive a third message (530) from a second network device (300), the third message (530) indicating a state (S) of the Al air interface architecture; andchange a state (S) of the Al air interface architecture based on the third message (530).

10. The first network device (100) according to claim 9, wherein the state (S) of the Al air interface architecture is indicated with a bitmap.

11. The first network device (100) according to any one of claims 1 to 7, configured to:change a state (S) of the Al air interface architecture based on the set of evaluations (Vz).a second state (S2) in which at least one feature in the set of features (Z) is valid as input into at least one Al model in the set of Al models; ora third state (S3) in which all features in the set of features (Z) are valid as input into any Al model in the set of Al models.

13. The first network device (100) according to claim 12, configured to change the state (S) of the Al air interface architecture:

14. The first network device (100) according to claim 13, wherein changing the state (S) of the Al air interface architecture comprises:changing from a first Al model (All ) to a second Al model (AI2) in the set of Al models.

15. The first network device (100) according to any one of the preceding claims, wherein the first network device (100) is a base station (630) of a radio access network.determine a state (S) of the Al air interface architecture based on the first message (510); andtransmit a third message (530) to the first network device (100), the third message (530) indicating the determined state (S) of the Al air interface architecture.

17. The second network device (300) according to claim 16, wherein the set of evaluations (Vz) is indicated with a bitmap.

18. The second network device (300) according to claim 16 or 17, wherein the state (S) of the Al air interface architecture is indicated with a bitmap.

20. The second network device (300) according to any one of claims 16 to 19, configured to:receive the first message (510) in response to a transmission of a second message (520) to the first network device (100), the second message (520) indicating a request for the first message (510).

21. The second network device (300) according to claim 20, configured to:transmit the second message (520) to the first network device (100) upon detecting that a quality-of-service, QoS, for the Al air interface architecture is below a QoS threshold value.

22. The second network device (300) according to any one of claims 16 to 21, wherein the state (S) of the Al air interface architecture is:a second state (S2) in which at least one feature in the set of features (Z) is valid as input into at least one Al model in the set of Al models; ora third state (S3) in which all features in the set of features (Z) are valid as input into any Al model in the set of Al models.

23. The second network device (300) according to claim 22, configured to determine a change of the state (S) of the Al air interface architecture:

24. The second network device (300) according to claim 23, wherein changing the state (S) of the Al air interface architecture comprises:changing from a first Al model (All ) to a second Al model (AI2) in the set of Al models.wherein each evaluation in the set of evaluations (Vz) indicates whether the feature (Z^) is valid or invalid as input in the Al air interface architecture.

26. A method (400) for a second network device (300), the method (400) comprising:receiving (402) a first message (510) from a first network device (100), the first message (510) indicating a set of quality indicators (Q) for a set of features (Z) or a set of evaluations (Vz) for the set of features (Z) for an Al air interface architecture comprising a set of Al models, wherein a quality indicatorfor each feature (Z^) in the set of features (Z) is determined based on inputting a set of data samples associated with the feature (Z^) into an evaluation function (S^) for the feature (Z^), wherein the set of data samples are obtained during an evaluation time period (T^) for the feature (Z^), and wherein the evaluation function (£^) is a real function, and wherein each evaluation in the set of evaluations (Vz) indicates whether a feature (Z^) in the set of features (Z) is valid or invalid as input in the Al air interface architecture;determining (404) a state (S) of the Al air interface architecture based on the first message (510); and transmitting (406) a third message (530) to the first network device (100), the third message (530) indicating the determined state (S) of the Al air interface architecture.

27. A computer program with a program code for performing a method according to claim 25 or 26 when the computer program runs on a computer.