Technique for controlling a wireless access network

Abstract

A technique for configuring parameters of a machine learning system, ML system (110) that controls (306; 406) a wireless access network (500) is described. As to a method aspect of the technique, at least two different sequences (1400) of measurement values (1300) indicative of wireless data traffic (800) between the wireless access network (500) and one or more wireless devices (504) are obtained (302). The different sequences (1400) are obtained (302) using different time domain processing of the wireless data traffic (800). Machine learning, ML, of the ML system (110) is performed (304) based on each of the at least two sequences (1400) resulting in at least two sets (1302) of parameter values, wherein the ML system (110) is configured (304-2) with one of the at least two sets (1302) based on a performance metric evaluated (304-1) for the ML system (110).

Description

[0001] Technique for controlling a wireless access network

[0002] Technical Field

[0003] The present disclosure relates to a technique for controlling a wireless access network. More specifically, and without limitation, a method and a device are provided for controlling a wireless access network.

[0004] Background

[0005] In today's rapidly evolving telecommunication, one of the cardinal techniques that has significantly enhanced data transmission rates and seamless connectivity across networks is Carrier Aggregation (CA). CA facilitates mobile networks to combine multiple carrier frequencies, enhancing bandwidth and boosting network throughput. This is crucial in areas of dense data consumption and demand for faster and / or more reliable mobile internet, driven by advanced applications like streaming services, augmented reality, and the Internet of Things (loT).

[0006] However, while CA effectively multiplies the data-carrying capacity, the underlying technology's management, especially the dynamic activation and deactivation of secondary carriers, poses significant operational challenges. Conventional systems use static or rule-based algorithms that often fail to cope with rapidly shifting network conditions and varying user demands. This results in either underutilization of allocated wireless resources or unwanted delays. Particularly, power efficiency and user experience are conflicting goals exacerbated by insensitive handling of secondary cells, either activating them too early or deactivating them too late.

[0007] The advent of machine learning (ML) system, also referred to as artificial intelligence (Al), has brought improvement in predictive capabilities concerning data demands and network adjustments. For example, M. Elsayed, R. Joda, H. Abou-zeid, R. Atawia, A. Bin Sediq, G. Boudreau and M. Erol-Kantarci, "Reinforcement Learning Based Energy-Efficient Component Carrier Activation- Deactivation in 5G," in IEEE Global Communications Conference (GLOBECOM), Madrid, Spain, 2021, consider CA as a key enabler technology for delivering higher rates to users of LTE and 5G networks. However, the increased transmission rate comes with the price of higher energy consumption which stems from users continuously monitoring the control channel of the active component carriers (CCs) whether data transmission is ongoing or not. In order to reduce energy consumption, the activation-deactivation procedure at the medium access control (MAC) layer of LTE / 5G network uses a reinforcement learning-based algorithm to improve energy-efficiency by dynamically activatingdeactivating secondary component carriers (SCCs) with awareness of the user traffic profiles. The proposed algorithm aims to predict the arrival of data and identify SCCs to activate for each user.

[0008] However, Al-based systems explored in the existing literature for controlling wireless access networks, such as multi-agent reinforcement learning approaches, are predominantly confined to macro-level control and operate at a timescale removed from real-time flux of network traffic. These Al models, while progressive, are often not optimally aligned with the operational intricacy of modern wireless access networks which require granular and swift adjustments.

[0009] Summary

[0010] Accordingly, there is a need for a technique that more accurately predicts and dynamically controls the operation of a wireless access network, such as activation and deactivation of secondary carriers in carrier aggregation, considering real-time data traffic and operational nuances.

[0011] As to a first method aspect, a method of configuring parameters of a machine learning system (ML system) for controlling a wireless access network is provided. The method comprises a step of obtaining at least two different sequences of measurement values indicative of wireless data traffic between the wireless access network and one or more wireless devices, wherein the different sequences are obtained using different time domain processing of the wireless data traffic.

[0012] The method further comprises a step of performing machine learning (ML) of the ML system based on each of the at least two sequences resulting in at least two sets of parameter values. The ML system is configured with one of the at least two sets based on a performance metric evaluated for the ML system.

[0013] Using one of the at least two sets of parameter values based on the performance metric may be referred to as selecting the set of parameters based on the performance metric. Based on the selected set of parameters, embodiments of the method may control the wireless access network considering real-time data traffic and / or operational nuances of the wireless access network. For example, by capturing fluctuating network conditions and / or user demands, as representable by the measurement values of the wireless data traffic, with the time domain processing that is superior for predicting the wireless data traffic of the specific combination of wireless access network and one or more wireless devices, the ML system can control the wireless access network with enhanced spectral efficiency and reduced energy consumption, thereby addressing the inefficiencies of the prior art.

[0014] The first method aspect may be implemented alone or in combination with any one of the embodiments described herein below. Alternatively or in addition, embodiments of the first method aspect may comprise features or steps disclosed below in the context of the second method aspect (e.g., in a system method).

[0015] By configuring the parameters of the ML system using different time domain processing when measuring the wireless data traffic, method embodiments can improve the control of the wireless access network by configuring the ML system to be based on the most relevant time domain characteristics of the wireless data traffic according to the performance metric of the ML system. Same or further embodiments of the method can adapt the ML system by changing the time domain processing of the wireless data traffic at an input of the ML system and changing the parameter values of the ML system in accordance with the changed time domain processing.

[0016] For example, the time domain processing and the configuration of the ML system may be simultaneously and / or consistently changed according to the performance metric. Alternatively or in addition, the performance metric may be (e.g., dynamically) evaluated for each of the at least two different sequences (i.e., for each of the at least two different ways of time domain processing). As the wireless data traffic occurs, the at least two sequences of measurement values may be collected in real-time in the obtaining step. These measurements may reflect a current state and / or dynamics and / or interactions of the wireless access network such as bandwidth usage, connection density, latency, signal strength, etc. The set of parameters may also be referred to as a ML model. Alternatively or in addition, the first method aspect may be performed in a training phase.

[0017] The parameters of the ML system may be configured (e.g., provided) by performing the ML of the ML system. Alternatively or in addition, the method may further comprise a step of configuring (e.g., providing) the ML system. The step of configuring (e.g., providing) the ML system may comprise sending (e.g., from a centralized unit, CU, to a distributed unit, DU) the set of parameters for controlling (e.g., at the DU) the wireless access network. Alternatively or in addition, the method may comprise a step of controlling the wireless access network using the ML system configured with the one of the at least two sets of parameters.

[0018] The method may further comprise steps performed in an inference phase, e.g. the obtaining and / or the controlling. For example, the method may comprise obtaining, in the inference phase, a sequence of measurement values indicative of wireless data traffic between the wireless access network and one or more wireless devices, wherein the sequence is obtained using the time domain processing selected according to the evaluated performance metric. Alternatively or in addition, the method may comprise controlling the wireless access network using the ML system configured with the selected one set of parameters (i.e., in the inference phase), wherein the obtained sequence (e.g., obtained in the inference phase) is applied to the ML system.

[0019] Herein, the different time domain processing may relate to different modes of sampling and / or analyzing the wireless data traffic. Alternatively or in addition, the sequence of measurement values may be a (e.g. periodic or aperiodic) time series of measurement values.

[0020] The different sequences may be obtained using different time domain processing of the same wireless data traffic. For example, the same wireless data traffic between the same pair of network node and wireless device in the same time window may be the basis for the at least two different sequences of measurement values.

[0021] Herein, measuring may encompass sampling. For example, sampling may refer to an observation of the data traffic at a certain point in time and / or over a period of time that is the shortest period technically possible for the hardware used for the obtaining of the measurement values indicative of the wireless data traffic. Each of the measurement values in a sequence may correspond to a different sample value or may be based on multiple sample values. The point in time and / or the shortest period technically possible for sampling a sample value may be referred to as sampling occasion (e.g., independent of whether the sampling occasion contributed to one of the measurement values). Each sample value may correspond to one of sampling occasions (e.g., wherein not necessarily all sampling occasions are used for sampling a sample value). The sequence of measurement values may be sub-sampling the wireless data traffic at a rate less frequently than the sampling occasions depending on the time domain processing of the wireless data traffic for the respective one of the different sequences.

[0022] The different time domain processing of the wireless data traffic may relate to different time scales on which the wireless data traffic is measured and / or processed for obtaining the different sequences of measurement values. The different time domain processing may provide (e.g., filter) different features (e.g., temporal patterns) of the wireless data traffic in the different sequences of the measurement values. The different time domain processing may output the measurement values at different temporal resolutions or different temporal scales for the ML. For example, a pertinent time scale for predicting the wireless data traffic and / or for controlling the wireless access network may (e.g., dynamically) change. Thus, the ML system can adapt to the pertinent time scale without requiring the resources and the computational complexity of sampling and processing the wireless data traffic permanently at the smallest time scale (e.g., the finest time scale or highest sampling rate among the at least two different time domain processing). In other words, by identifying the correct time scale, which is achieved by selecting the set of parameter values according to the performance metric, the ML system can detect features (e.g., temporal patterns) that are the basis (e.g., an indicator or a trigger) for effectively (e.g., preemptively) controlling the wireless access network and can be shielded from details in the wireless data traffic that are irrelevant for controlling the wireless access network. Hence, the wireless access network can be controlled more accurately and / or with more lead time and / or with less resources required for the ML system.

[0023] Herein, a time scale of a sequence of measurement values may relate to the temporal resolution at which the measurement values are indicative of wireless data traffic or an inverse frequency of a highest-frequency peak in a Fourier transformation of the sequence of measurement values.

[0024] The lead time of the ML system may relate to a timing advance (i.e., a negative time offset) between prediction of a change in the wireless data traffic predicted by the ML system and the occurrence of said change in the wireless data traffic.

[0025] Furthermore, the number of parameters can be reduced and / or the lead time of the ML system can be improved (e.g., compared to a method that always uses the sequence of measurement values with the finest time scale or highest sampling rate as input of the ML system) by selecting the sequence of measurement values that represents the pertinent time scale. For example, the method can select the least complex sequence of measurement values and / or the least complex set of parameter values for the ML system, which are necessary to monitor a broad range of different time scales in the sequence of measurement values for feature detection.

[0026] The feature detection by means of the ML system may comprise detecting a feature in the wireless data traffic that indicates (e.g., requires or triggers) a change in controlling the wireless access network.

[0027] Alternatively or in addition, an architecture of the ML system can be simplified (e.g., compared to a ML system architecture for multi-scale feature extraction) by using the wireless data traffic-specific time domain processing. For example, a hierarchical or pyramid structures in the architecture of the ML system can be avoided. Alternatively or in addition, a complexity of the ML system in terms of time scale decomposition or temporal scale decomposition can be reduced, whereas conventional implementations would require the ML system to determine the most relevant time scale in the sequence at inference. For example, the ML system does not have to monitor a broad range of different time scales in the sequence of measurement values for feature detecting of a feature that triggers the controlling of the wireless data traffic.

[0028] Each of the different time domain processing may be based on sample values of the wireless data traffic. A result of the different time domain processing may comprise the different sequences of measurement values. For example, the different time domain processing may comprise at least one of different sampling rates of the sample values, different time intervals (e.g., for averaging one or more sample values), different triggers defining the beginning of the time interval, different wavelets for wavelet analysis of the sample values (e.g., which decomposes a time series of the sample values into components of different frequencies), different Fourier components (e.g., in the frequency domain) of a Fourier transformation of the sample values, and / or a different sizes of a (e.g., sliding) time window within which sample values are recorded and / or processed.

[0029] An output of the ML system may directly (e.g., as a control indicator, optionally a control command) or indirectly (e.g., as a prediction of wireless data traffic) control the wireless access network. The output of the ML system may comprise a prediction of the wireless data traffic and / or a change of use in the wireless data traffic and / or an indicator of an anomaly and / or a control indicator (e.g., a control command) for controlling the wireless access network. The output may be a regression or a classification.

[0030] Performing the ML of the ML system may also be referred to as training the ML system. Configuring parameters of the ML system may encompass determining the parameter values for the parameters of the ML system (e.g., in a training phase) and setting the parameters of the ML system according to the determined values (e.g. for an inference phase).

[0031] A set of parameter values for the ML system (optionally in combination with an architecture of the ML system) may be referred to as Al model. The Al model may refer to a specific instantiation of the ML system, which may comprise both the architecture and the set of trained parameter values. The Al model may result from the ML (i.e., the training phase) in which the respective one of the at least two different sequences of measurement values is used to adjust the parameters.

[0032] The parameters of the ML system may comprise weights or biases, e.g. in case the ML system comprises at least one neural network.

[0033] A structure of the ML system independent of the concrete parameter values may be referred to as Al architecture. In other words, the Al architecture may relate to a structural design of the ML system, independent of the specific parameter values. The Al architecture may outline components of the ML system, e.g. one or more layers in a neural network (e.g., convolutional, recurrent, and / or fully connected layers), optionally their connectivity (i.e., how layers or nodes are linked), or other characteristics like activation functions and / or dropout layers. The Al architecture may encompass at least one of: a number of layers of a neural network, an input layer, one or more intermediate layers, an output layer, one or more convolutional layers, a feedforward across different layers, a feedback loop across different layers, an encoder, and a decoder. The Al architecture may imply a certain number of the parameters of the ML system (e.g., as a weight at each node of the neural network). Preferably, the Al architecture provides a structural feature of the ML system, which when combined with the selected set of parameter values (i.e., trained parameters), constitutes the ML system with an Al model operative for inference.

[0034] The same ML technique may be used for the performing of the ML based on each of the at least two sequences. Herein, the ML technique may encompass one of reinforcement learning, supervised learning and unsupervised learning. Alternatively or in addition, the same hyperparameters may be used for the performing of the ML based on each of the at least two sequences. Herein, hyperparameters may comprise any condition for the performing of the ML and / or at least one of learning rate, number of epochs, and batch size. Alternatively or in addition, the different sequences may imply or correspond to (or may require) different numbers of parameters of the ML system and / or different architectures of the ML system. For example, a number of input neurons in an input layer of the ML system may correspond to a number of measurement values (optionally, and associated time stamps) in the respective one of the at least two different sequences.

[0035] The ML performed based on the at least two sequences may result in the at least two sets of the parameter values, respectively. For example, the ML of the ML system may be independently (e.g., in parallel) performed based on each of the at least two sequences. The at least two sets may be the result of the at least two ML operations, respectively.

[0036] The wireless access network may be a radio access network (RAN). The RAN may operate according to a radio access technology (RAT). The wireless device may be a radio device. The RAN may serve the radio devices and / or the radio devices may be in a (e.g., radio resource control, RRC) connected state relative to the RAN. Alternatively or in addition, the wireless access network may be an optical access network (OAN). The wireless data traffic may use free-space optical (FSO) communications.

[0037] The (e.g., different) time domain processing may comprise (e.g., different) preprocessing of sample values sampled from the wireless data traffic for mapping the sampled values to the measurement values at an input of the ML system. The ML system may comprise a neural network system. The input of the ML system may comprise an input layer of neurons of the neural network system.

[0038] The (e.g., different) time domain processing of the wireless data traffic (e.g., used to obtain the different sequences) may relate to different methods of processing (e.g., aggregating) values sampled from the wireless data traffic and / or different sampling rates of values sampled from the wireless data traffic for obtaining the measurement values of the at least two different sequences. Herein, the values sampled from the wireless data traffic and the sample values of the wireless data traffic may be synonymous.

[0039] In an embodiment, the measurement values may be indicative of one or more traffic characteristics of the wireless data traffic. Alternatively or in addition, the measurement values may be indicative of one or more network behaviors of the wireless access network. Alternatively or in addition, the measurement values may be indicative of a network bandwidth usage by the wireless data traffic. Alternatively or in addition, the measurement values may be indicative of a relative bandwidth usage of one or more active component carriers by the wireless data traffic. Alternatively or in addition, the measurement values may be indicative of a relative bandwidth usage, by the wireless data traffic, of a master cell group (MCG) and / or one or more secondary cell groups (SCGs) of the wireless access network. Alternatively or in addition, the measurement values may be indicative of a data rate of the wireless data traffic or a data volume per time of the wireless data traffic. Alternatively or in addition, the measurement values may be indicative of a number of data packets in the wireless data traffic. Alternatively or in addition, the measurement values may be indicative of a size of data packets in the wireless data traffic. Alternatively or in addition, the measurement values may be indicative of a distribution of time intervals between subsequent data packets in the wireless data traffic. Alternatively or in addition, the measurement values may be indicative of source internet protocol addresses and / or target internet protocol addresses in headers of the wireless data traffic. Alternatively or in addition, the measurement values may be indicative of applications underlying the wireless data traffic. Alternatively or in addition, the measurement values may be indicative of classifications based on deep packet inspection of the wireless data traffic.

[0040] Packets that fulfill a size requirement (e.g., being greater than a predefined threshold value) may indicate video streaming as the application underlying the wireless data traffic. Alternatively or in addition, source and target internet protocol (IP) addresses in headers of the wireless data traffic may enable the ML system to track session specifics and / or discern network traffic routes and / or destinations. The classifications based on deep packet inspection (DPI) of the wireless data traffic may include determining underlying applications (e.g., social media, video streaming, or online gaming) and / or categorizing types of traffic (e.g., Hypertext transfer protocol (HTTP or HTTPS) or File Transfer Protocol (FTP or FTPS)), e.g. which influence resource allocation decisions as an example of controlling the wireless access network. The distribution of time intervals between subsequent data packets in the wireless data traffic may include a minimum time interval, an average time interval, and / or a maximum time interval, e.g. relevant for diagnosing network congestion and jitter.

[0041] Each of the measurement values in the sequence may be indicative of a measurement quantity (e.g., a numerical quantity or a classification quantity (i.e., a classifier)) of the wireless data traffic.

[0042] The different time domain processing (e.g., different processes) used for the different sequences may be implemented by varying a sampling rate underlying the measurement values of the different sequences, by varying a measurement time slot (e.g. start and / or duration T of the measurement time slot) underlying each measurement value of the different sequences, and / or by varying a history time window (e.g. a number of measurement values and / or a temporal length W) underlying each of the different sequences. The measurement time slot may be briefly referred to as time slot. The history time window may be briefly referred to as history window. The (temporal) length of the history time window may also be referred to as history window size.

[0043] In an embodiment, the obtaining of the at least two different sequences of measurement values using the different time domain processing of the wireless data traffic may comprise sampling the wireless data traffic at different sampling rates for the at least two different sequences of measurement values. Alternatively or in addition, the obtaining of the at least two different sequences of measurement values using the different time domain processing of the wireless data traffic may comprise sampling the wireless data traffic for at least one of the at least two different sequences of measurement values according to an irregular (e.g., aperiodic) pattern of sample values of the wireless data traffic.

[0044] The least two different sequences of measurement values may be based on different sampling rates. Alternatively or in addition, the at least two different sequences may be obtained by sampling the wireless data traffic at different sampling rates.

[0045] The different time domain processing of the wireless data traffic may comprise different sampling (e.g., different periodic sampling rates and / or different periodic and aperiodic sampling patterns) for the sampling of the wireless data traffic.

[0046] Alternatively or in addition, measuring may comprise the time domain processing after the sampling and / or the measurement values may result from the time domain processing of multiple sample values.

[0047] In a first variant of any embodiment, the same sample values from the wireless data traffic may be subject to the different time domain processing, which results in the different sequences of measurement values. The measurement values of the different sequences may be different values (e.g. as a result of the different time domain processing) of the same measurement quantity of the wireless data traffic (e.g., data volume per time). In the first variant, the at least two different sequences may be obtained by effectively sampling the wireless data traffic at different sampling rates by selectively down-sampling (as an example of the time domain processing) the sample values of the data traffic.

[0048] For instance, after sampling the wireless data traffic at a constant sampling rate, the sample values can be down-sampled, yielding the measurement values, by selecting only every n-th sample value, where n represents a down-sampling factor. This will effectively reduce the sampling rate by a factor of n and generate a different sequence of measurement values, which may enable reducing the complexity of the ML system. Alternatively or in addition, the sample values may be processed in the time domain by apply a running average filter, which smooths the sample values by averaging a set number of consecutive sample values together. A running average filter may be applied by replacing each sample value with the average of the sample value and its k-1 neighbors, wherein k is the size of a time slot T over which the average is computed. Alternatively or in addition, a time-domain filter such as a low-pass Butterworth filter may be employed. Another example of the time domain processing may comprise decimation, which combines down-sampling with a low-pass filtering step before reducing the sampling rate. This can help to prevent aliasing, a phenomenon where higher frequency components in the data traffic are misinterpreted as lower frequencies due to insufficient sampling rates. An implementation of decimation may involve applying a Finite Impulse Response (FIR) filter followed by taking every m-th sample value from the filtered output. Independent of the implemented filter, averaging can further lower high-frequency components of the data traffic and configure the ML system to focus on the relevant time scale for predicting the data traffic and / or controlling the wireless access network.

[0049] In a second variant of any embodiment, the measurement values are sample values sampled from the wireless data traffic at different sampling rates corresponding to the different sequences. In other words, the time domain processing may be implemented at the sampling stage. This variant can reduce the complexity of monitoring the data traffic as a reduction in a time resolution of the ML system can translate directly in a reduction of data acquisition for sampling the data traffic.

[0050] In any variant, each of the measurement values in the respective sequence may be indicative of at least one of a data rate of the wireless data traffic, a subsampled data rate of the wireless data traffic (e.g., according to a sampling rate), a data volume of the wireless data traffic at one or more sampling occasions, and a sum of data volumes of the wireless data traffic associated with sampling occasions.

[0051] The data rate resulting from sampling of (e.g., quantitative) features of data packets may be a subsampled data rate. The data packet features are also referred to as packet data, e.g. metadata of packet-oriented data traffic. For example, the data rate may be a result of sampling the packet data (e.g., a data volume or a number of data packets) of the wireless data traffic at a sampling rate that is less (e.g., multiple times less) than a data packet rate of the wireless data traffic. The sample values sampled from the wireless data traffic may be indicative of a measurement quantity of data packets of the wireless data traffic. The sample values sampled from the wireless data traffic may comprise values of a measurement quantity (e.g., data volume per time) of the wireless data traffic and / or comprise no content (e.g., payload) of the wireless data traffic.

[0052] At least one of the at least two different sequences of measurement values may be based on aperiodic samples of the wireless data traffic.

[0053] By including at least one aperiodically sampled sequence of measurement values, the ML system can detect aperiodic patterns in the wireless data traffic and / or can control the wireless access network to handle (e.g., burst-like) transmissions that are correlated in the time domain. More specifically, by including at least one aperiodically sampled sequence of measurement values in the at least two different sequences used for the ML, the method can selectively configure the parameters of the ML system to detect an aperiodic pattern in the wireless data traffic if (e.g., and only if) such detection is advantageous according to the performance metric. For example, if the aperiodic pattern is absent in the wireless data traffic at the time of performing the ML, there may be no improvement in terms of the performance metric, so that an additional energy consumption for aperiodic sampling can be avoided without additional delay in predicting the wireless data traffic and / or controlling the wireless access network.

[0054] In an embodiment, each of the measurement values may be associated with a measurement time slot. Alternatively or in addition, each of the measurement values may be indicative of an aggregation of one or more sample values of the wireless data traffic within a measurement time slot. Alternatively or in addition, the different time domain processing of the wireless data traffic may correspond to different durations (T) of the measurement time slot.

[0055] Each measurement value may correspond to one measurement time slot. The measurement time slot (e.g., within which sample values are aggregated according to the aggregation) may be referred to as aggregation time slot or aggregation time window. The aggregation may be referred to as grouping of the sample values and / or measuring of the measurement values. Herein, the measurement time slot may encompass a (e.g., continuous) time window during which the wireless data traffic is sampled (e.g., discretely at individual sampling occasions). For example, the wireless data traffic may be sampled during the measurement time slot. The measurement value may be (e.g., exclusively) based on the sample values in the measurement time slot.

[0056] The aggregation may comprise a sum and / or an average of individual sample values in the measurement time slot.

[0057] In an embodiment, measurement time slots of different measurement values may be consecutive in the time domain. Alternatively or in addition, the measurement time slots of different measurement values may be regular, e.g. repeating periodically, in the time domain. Herein, regular (e.g., regular measurement time slots) may encompass a pattern of one or more periodic events (e.g., measurement time slots repeating according to one or more time periods).

[0058] Above-mentioned consecutive and / or disjoint measurement time slots may be the afore-mentioned measurement time slots (which different durations T are related to or define the different time domain processing).

[0059] The obtaining of the sequences using consecutive measurement time slots may be responsive to detecting a continuous wireless data traffic. Alternatively or in addition, the consecutive measurement time slots may enable monitoring seamlessly the wireless data traffic. This can reduce a latency (e.g., a lead time or time delay) in predicting the wireless data traffic and / or in controlling the wireless access network.

[0060] The measurement time slots of different measurement values may be (e.g., pairwise) disjoint in the time domain.

[0061] For example, the wireless traffic may be sampled at a constant sampling rate (e.g., at least within the measurement time slot or the history time window).

[0062] Alternatively or in addition, the measurement values may comprise the aggregated or estimated data volume of those data packets that occur in the associated measurement time slot at or closest after each sampling occasion. The measurement values may be sub-sampled (i.e., under-sampled), e.g. in the sense that the measurement values do not comprise the aggregated data volume of all data packets in the associated measurement time slot, e.g. only those that occur at or closest after a sampling occasion. Herein, occurrence of a data packet may refer to the availability (e.g., the reception or the processing after the reception, or the transmission or the pendency for the transmission) of the data packet at the network node, e.g., a centralized unit (CU) of the network node.

[0063] In an embodiment, measurement time slots of different measurement values may be non-consecutive in the time domain. Alternatively or in addition, the measurement time slots of different measurement values may be event-triggered, optionally triggered by the availability for transmission or by the transmission of a data packet in the wireless data traffic.

[0064] Also in this alternative, the measurement time slots of different measurement values may be disjoint in the time domain.

[0065] Event-triggered starts of the measurement time slots may reduce a power consumption for obtaining of the measurement values of the wireless data traffic (e.g., in the training phase and / or in the inference phase) and / or for performing the ML of the ML system (i.e., in the training phase) and / or for controlling the wireless access network based on the ML system (i.e., in the inference phase).

[0066] In an embodiment, a length of each sequence of measurement values may correspond to a history time window. Alternatively or in addition, the different time domain processing of the wireless data traffic may correspond to different durations (W) of the history time window.

[0067] Each sequence of measurement values may correspond to a history time window whose length may vary for the different sequences. The different time domain processing of the wireless data traffic may be specifically tied to these different durations of the history time window. As a result, embodiments of the method can determine the parameters of the ML system according to the length of the history time window, e.g. allowing for controlling of the wireless access network based on varying time-based contexts in an analysis of the wireless data traffic.

[0068] Herein, the time duration of the sequence of measurement values applied as the input of the ML system may be referred to as history time window. The symbol "W" may refer to the history time window or its duration (e.g., if the history time window has a constant duration for more than one of (or all) predictions or control indications of the ML system. For example, a number of input neurons of the ML system may be proportional to the ratio of W / T, wherein W is the duration of the history time window and T is the duration of the measurement time slots within the history time window.

[0069] The sequence of measurement values used for the ML (i.e., in the training phase) may comprise at least one further measurement value measured subsequently to the measurement values applied as the input to the ML system. The at least one further measurement value measured subsequently to the measurement values applied as the input may be at least one labeled value (e.g., at least one "ground truth" value). A deviation between an output of the ML system (e.g., in the training phase) and the at least one labeled value may be back-propagated into a neural network of the ML system (e.g., in the training phase) to configure the parameters of the ML system (e.g., for the inference phase).

[0070] In an embodiment, the ML system may output a prediction for the wireless data traffic, the wireless access network being controlled based on the prediction for the wireless data traffic. Alternatively or in addition, the ML system may output a control indicator, the wireless access network being controlled based on the control indicator.

[0071] The prediction for the wireless data traffic may comprise one or more predicted measurement values, e.g. a predicted continuation of the respective one of the sequences of measurement values. The prediction may comprise a regression by the ML system (e.g., for the one or more predicted measurement values) or a classification by the ML system (e.g., for outputting the control indicator).

[0072] Herein, the control indicator may be a control command.

[0073] In an embodiment, the controlling of the wireless access network by the ML system may comprise radio resource management (RRM) of the wireless access network. Alternatively or in addition, the controlling of the wireless access network by the ML system may comprise selectively activating or deactivating a wireless carrier of the wireless access network. Alternatively or in addition, the controlling of the wireless access network by the ML system may comprise selectively activating or deactivating a component carrier of the wireless access network for carrier aggregation (CA) of the one or more wireless devices. Alternatively or in addition, the controlling of the wireless access network by the ML system may comprise selectively activating or deactivating a wireless connection to a secondary cell for dual connectivity of the one or more wireless devices. Alternatively or in addition, the controlling of the wireless access network by the ML system may comprise changing a modulation and coding scheme (MCS) for the one or more wireless devices. Alternatively or in addition, the controlling of the wireless access network by the ML system may comprise load balancing the wireless data traffic and / or the one or more wireless devices across available network resources of the wireless access network. Alternatively or in addition, the controlling of the wireless access network by the ML system may comprise mobility management of the one or more wireless devices.

[0074] Whenever referring to controlling the wireless access network (e.g., performing RRM of the wireless access network), an embodiment may perform any one of the above actions. Furthermore, controlling the wireless access network by the ML system may comprise that the ML system initiates the action or that the action is performed if two or more conditions are fulfilled, wherein the output of the ML system to perform the action is one of the two or more conditions. In other words, the ML system may cause the action, optionally subject to one or more further conditions being fulfilled.

[0075] In an embodiment, the at least two different sequences of measurement values may be obtained per wireless device. Alternatively or in addition, the ML of the ML system may be performed per wireless device. Alternatively or in addition, the controlling of the wireless access network may comprise performing radio resource management (RRM) per wireless device or may comprise (e.g., selectively) activating or deactivating a component carrier per wireless device.

[0076] The ML system may output a prediction for the wireless data traffic per wireless device. The wireless access network may be controlled based on the prediction of the wireless data traffic for the wireless device. Alternatively or in addition, the ML system may output a control indicator (e.g., a control command) for controlling the wireless access network for the wireless device. The wireless access network may be controlled based on the control indicator for the wireless device. In an embodiment, the measurement values indicative of the wireless data traffic may be received from a centralized unit (CU) of a network node of the wireless access network. Alternatively or in addition, an output of the ML system for the controlling of the wireless access network may be sent to a distributed unit (DU) of a network node of the wireless access network.

[0077] In a first variant of any embodiment, a time duration (e.g., the afore-mentioned W) of a sequence currently selected out of the at least two sequences of measurement values (i.e., currently applied to the ML system for controlling the wireless access network) and / or a time duration (e.g., the afore-mentioned W) of each of the at least two sequences may be less than a time lag between real-time and a time-stamp of the last measurement value in the sequence of measurement values applied to the ML system for controlling the wireless access network.

[0078] In a second variant of any embodiment, which may be combined with the first variant, a time duration (W) of the sequence of measurement values used for performing the ML may be less than a time lag between real-time and a timestamp of the last measurement value in the sequence of measurement values used for performing the ML.

[0079] The ML may be performed in real-time and / or based on a current sequence of measurement values. The ML system may process the at least two sequences as they are obtained, e.g. continuously updating its configuration (e.g., model) parameters based on the influx of measurement values. For example, a value (for the wireless data traffic) predicted by the ML system at time t may be used for both controlling the wireless access network and for determining a deviation [e.g., between the predicted value and the (e.g., subsequently available) measurement value for the time t] for performing the ML of the ML system. This continuous adaptation may allow the ML system to remain consistently sensitive for the most relevant time scale and / or current network conditions as represented by the time domain processing selected based on the performance metric.

[0080] As to a second method aspect, a method of controlling a wireless access network is provided. The method comprises receiving a configuration message indicative of parameters of a machine learning system (ML system) and a time domain processing of wireless data traffic. The method further comprises obtaining a sequence of measurement values indicative of wireless data traffic between the wireless access network and one or more wireless devices. The sequence is obtained using the received time domain processing of the wireless data traffic. The method further comprises controlling the wireless access network using the ML system configured with the received parameters. The obtained sequence is applied to the ML system.

[0081] Herein, the expressions "received parameters" and "received time domain processing" may be shorthand for the parameters and the time domain processing, respectively, as indicated in the "received" configuration message.

[0082] The received parameters may be a result of machine learning (ML) for the ML system based on each of at least two sequences resulting in at least two sets of parameter values, wherein the ML system is configured with the received set of parameters being one of the at least two sets selected based on a performance metric evaluated for the ML system.

[0083] The second method aspect may be implemented alone or in combination with any one of the embodiments disclosed herein.

[0084] The second method aspect may further comprise any feature and / or any step disclosed in the context of the first method aspect, or a feature and / or step corresponding thereto, e.g., a receiver counterpart to a transmitter feature or step.

[0085] In any aspect, the performance metric may depend on a predefined quality of service (QoS) level, e.g., a QoS of the wireless data traffic and / or the wireless device and / or a service underlying the wireless data traffic. Whenever referring to QoS, the QoS may be indicated by a QoS class identifier (QCI).

[0086] Without limitation, for example in a 3GPP implementation, any wireless device may be a "radio device" and / or a user equipment (UE).

[0087] Alternatively or in addition, the method may further comprise receiving a control message (e.g., from a source or target of the wireless data traffic, e.g. a wireless device or a server) indicative of the QoS. The wireless access network may provide wireless access to the wireless device according to a radio access technology (RAT) such as 3GPP New Radio (NR), which is an example of fifth generation (5G) RAT, or beyond 5G.

[0088] The measurement values for the wireless data traffic may be based on, or may be an extension of, performance measurements according to the 3GPP document TS 32.425, version 18.0.0 and / or key performance indicators according to 3GPP document TS 32.451, version 18.0.0.

[0089] Any wireless device (e.g., radio device) may be a user equipment (UE), e.g., according to a 3GPP specification. The wireless device and the wireless access network (e.g., a radio access network, RAN) may be wirelessly connected in an uplink (UL) and / or a downlink (DL), optionally through a Uu interface. Alternatively or in addition, a sidelink (SL) may enable a direct radio communication between proximal wireless devices, e.g., a remote wireless device that is out of coverage of the wireless access network and a relay wireless device that is in coverage of the wireless access network, optionally using a PC5 interface. Services underlying the wireless data traffic provided using the SL or the PC5 interface may be referred to as proximity services (ProSe). Any wireless device (e.g., the remote wireless device and / or the relay wireless device and / or a further wireless device) supporting a SL may be referred to as ProSe-enabled wireless device. The relay wireless device may also be referred to as ProSe UE-to-Network Relay.

[0090] The one or more wireless devices and / or the wireless access network may form, or may be part of, a wireless network (e.g., a radio network), e.g., according to the Third Generation Partnership Project (3GPP) or according to the standard family IEEE 802.11 (Wi-Fi). Each of the first method aspect and the second method aspect and third method aspect may be performed by one or more embodiments of a network node of the wireless access network (e.g., a base station) or component of the network node (e.g., the CU or DU). Alternatively or in addition, the first method aspect may be performed by a training server serving the wireless access network and / or the network node of the wireless access network.

[0091] The wireless access network (e.g., RAN) may comprise one or more base stations, e.g., performing the first and / or second method aspects. Alternatively or in addition, the wireless network may be a vehicular, ad hoc and / or mesh network comprising two or more wireless devices, e.g., acting as the remote wireless device and / or the relay wireless device. The relay wireless device may perform the first and / or second method aspects for controlling the SL (as an example of an ad hoc wireless access network) between the remote wireless device (as an example of the wireless device) and the wireless access network.

[0092] Any of the one or more wireless devices may be a 3GPP user equipment (UE) or a Wi-Fi station (STA). The wireless device may be a mobile station or a portable station, a device for machine-type communication (MTC), a device for narrowband Internet of Things (NB-loT) or a combination thereof. Examples for the UE and the mobile station include a mobile phone, a tablet computer and a self-driving vehicle. Examples for the portable station include a laptop computer and a television set. Examples for the MTC device or the NB-loT device include robots, sensors and / or actuators, e.g., in manufacturing, automotive communication and home automation. The MTC device or the NB-loT device may be implemented in a manufacturing plant, household appliances and consumer electronics.

[0093] Whenever referring to the wireless access network (e.g., RAN), the wireless access network may be implemented by one or more network nodes (e.g., base stations).

[0094] Each of the one or more wireless devices may be wirelessly connected or connectable (e.g., according to a radio resource control, RRC, state or active mode) with the wireless access network, e.g. at least one base station of the wireless access network.

[0095] The network node (e.g., a base station) may encompass any station that is configured to provide wireless access (e.g., radio access) to any of the wireless devices (e.g., radio devices). The base station may be a cell, a transmission and reception point (TRP), a central unit (CU), a distributed unit (DU), a radio access node or an access point (AP). The wireless access network (and / or the relay wireless device) may provide a data link to a host computer providing user data to the (e.g., remote) wireless device or gathering user data from the (e.g., remote) wireless device. Examples for the base stations may include a 3G base station or Node B (NB), 4G base station or eNodeB (eNB), a 5G base station or gNodeB (gNB), a Wi-Fi AP and a network controller (e.g., according to Bluetooth, ZigBee or Z- Wave). The RAN may be implemented according to the Global System for Mobile Communications (GSM), the Universal Mobile Telecommunications System (UMTS), 3GPP Long Term Evolution (LTE) and / or 3GPP New Radio (NR).

[0096] Any aspect of the technique may be implemented on a Physical Layer (PHY), a Medium Access Control (MAC) layer, a Radio Link Control (RLC) layer, a packet data convergence protocol (PDCP) layer, and / or a Radio Resource Control (RRC) layer of a protocol stack for the radio communication.

[0097] Herein, referring to a protocol of a layer may also refer to the corresponding layer in the protocol stack. Vice versa, referring to a layer of the protocol stack may also refer to the corresponding protocol of the layer. Any protocol may be implemented by a corresponding method.

[0098] As to another aspect, a computer program product is provided. The computer program product comprises program code portions for performing any one of the steps of the first and / or second method aspects disclosed herein when the computer program product is executed by one or more computing devices. The computer program product may be stored on a computer-readable recording medium. The computer program product may also be provided for download, e.g., via the radio network, the RAN, the Internet and / or the host computer.

[0099] Alternatively, or in addition, the method may be encoded in a Field-Programmable Gate Array (FPGA) and / or an Application-Specific Integrated Circuit (ASIC), or the functionality may be provided for download by means of a hardware description language.

[0100] As to a first device aspect, a device (e.g. network entity) for configuring parameters of a machine learning system (ML system) controlling a wireless access network is provided. The device may be configured to perform any one of the steps of the first method aspect. Alternatively or in addition, the device comprises processing circuitry (e.g., at least one processor and a memory). Said memory comprises instructions executable by said at least one processor whereby the device is operative to perform any one of the steps of the first method aspect.

[0101] The device may be a network node (e.g., a base station such as a gNB or access point), particularly a centralized unit (CU) of a base station (e.g., a gNB-CU) or a training server of the wireless access network. As to a second device aspect, a device (e.g. network entity) for controlling a wireless access network is provided. The device may be configured to perform any one of the steps of the second method aspect. Alternatively or in addition, the device comprises processing circuitry (e.g., at least one processor and a memory). Said memory comprises instructions executable by said at least one processor whereby the device is operative to perform any one of the steps of the second method aspect.

[0102] The device may be a network node (e.g., a base station such as a gNB or access point), particularly a distributed unit (DU) of a base station (e.g., a gNB-DU).

[0103] As to a further aspect, a wireless access network is provided. The wireless access network comprises at least one first network entity according to the first device aspect and at least one second network entity according to the second device aspect.

[0104] As to a still further aspect a communication system including a host computer is provided. The host computer comprises a processing circuitry configured to provide user data, e.g., included in the wireless data traffic. The host computer further comprises a communication interface configured to forward the user data to a wireless access network (e.g., a cellular network, a RAN and / or a base station) for transmission to a wireless device (e.g., a UE). A processing circuitry of the wireless access network is configured to execute any one of the steps of the first and / or second method aspects.

[0105] The communication system may further include the UE. Alternatively, or in addition, the cellular network may further include one or more base stations configured for radio communication with the UE and / or to provide a data link between the UE and the host computer using the first and / or second method aspects.

[0106] The processing circuitry of the host computer may be configured to execute a host application, thereby providing the user data and / or any host computer functionality described herein. Alternatively, or in addition, the processing circuitry of the UE may be configured to execute a client application associated with the host application. Any one of the devices, network entities, base stations, wireless access network, and communication system, or any node or station for embodying the technique, may further include any feature disclosed in the context of the first and / or second method aspects, and vice versa. Particularly, any one of the units and modules disclosed herein may be configured to perform or initiate one or more of the steps of the method aspect.

[0107] Brief Description of the Drawings

[0108] Further details of embodiments of the technique are described with reference to the enclosed drawings, wherein:

[0109] Fig. 1 shows a schematic block diagram of an embodiment of a device for configuring parameters of a ML system controlling a wireless access network;

[0110] Fig. 2 shows a schematic block diagram of an embodiment of a device for controlling a wireless access network;

[0111] Fig. 3 shows a flowchart for a method of configuring parameters of a ML system controlling a wireless access network, which method may be implementable by the device of Fig. 1;

[0112] Fig. 4 shows a flowchart for a method of controlling a wireless access network, which method may be implementable by the device of Fig. 2;

[0113] Fig. 5 exemplifies a wireless access network scenario depicting embodiments of the devices of Figs. 1 and 2 performing the methods of Figs. 3 and 4, respectively;

[0114] Fig. 6 and Fig. 7 illustrate examples of controlling the wireless access network related to carrier aggregation and dual connectivity, respectively, which can be controlled by embodiments of the methods of Figs. 3 and 4; Fig. 8 schematically represents an example of the wireless data traffic including data packets transmitted over different data radio bearers within specified time intervals;

[0115] Fig. 9 shows a schematic block diagram of an embodiment of the device of Fig. 1;

[0116] Fig. 10 shows a flowchart of an embodiment of the method of Fig. 3;

[0117] Fig. 11 shows a flowchart of an embodiment of the method of Fig. 4;

[0118] Fig. 12A schematically illustrates a wireless data traffic and exemplary points in time for subsampling the wireless data traffic;

[0119] Fig. 12B schematically illustrates an example of obtaining measurement values over regular time intervals;

[0120] Fig. 12C schematically illustrates an example of obtaining measurement values over irregular time intervals;

[0121] Fig. 13 schematically illustrates an example of performing ML from a sequence of measurement values to forecast future wireless data traffic.

[0122] Fig. 14 displays an example of predicting traffic (e.g., generating output as prediction based on the input) and / or performing ML (e.g., how the input and a desired output as the label is generated) based on a sliding history time window defining the sequence of measurement values at the input of the ML system;

[0123] Fig. 15 showcases a signaling diagram of an exemplary communication process within a wireless access network when performing an embodiment of the method of Fig. 4;

[0124] Fig. 16 shows a schematic block diagram of a centralized unit embodying the device of Fig. 1; and Fig. 17 shows a schematic block diagram of a distributed unit embodying the device of Fig. 2.

[0125] Detailed Description

[0126] In the following description, for purposes of explanation and not limitation, specific details are set forth, such as a specific network environment in order to provide a thorough understanding of the technique disclosed herein. It will be apparent to one skilled in the art that the technique may be practiced in other embodiments that depart from these specific details. Moreover, while the following embodiments are primarily described for a New Radio (NR) or 5G implementation, it is readily apparent that the technique described herein may also be implemented for any other radio communication technique, including a Wireless Local Area Network (WLAN) implementation according to the standard family IEEE 802.11, 3GPP LTE (e.g., LTE-Advanced or a related radio access technique such as MulteFire), for Bluetooth according to the Bluetooth Special Interest Group (SIG), particularly Bluetooth Low Energy, Bluetooth Mesh Networking and Bluetooth broadcasting, for Z-Wave according to the Z-Wave Alliance or for ZigBee based on IEEE 802.15.4.

[0127] Moreover, those skilled in the art will appreciate that the functions, steps, units and modules explained herein may be implemented using software functioning in conjunction with a programmed microprocessor, an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA), a Digital Signal Processor (DSP) or a general purpose computer, e.g., including an Advanced RISC Machine (ARM). It will also be appreciated that, while the following embodiments are primarily described in context with methods and devices, the invention may also be embodied in a computer program product as well as in a system comprising at least one computer processor and memory coupled to the at least one processor, wherein the memory is encoded with one or more programs that may perform the functions and steps or implement the units and modules disclosed herein.

[0128] The diagram illustrated in Fig. 1 depicts a schematic block diagram of an embodiment of a device for configuring parameters of a Machine Learning (ML) system that controls a wireless access network. The device is generically referred to by reference sign 100. The modules illustrated in Fig. 1 perform the corresponding steps of the first method aspect.

[0129] In an embodiment, the device 100 comprises a sequences obtaining module 102 that handles the acquisition of measurement values that feed into a ML system. The sequences obtaining module 102 is responsible for gathering two or more sequences of measurement values indicative of wireless data traffic. These sequences are collected using different time domain processing techniques of the wireless data traffic. For example, the module 102 is not only collecting data but also applying varied analytical techniques and / or controlling the data acquisition to provide diverse temporal resolutions of the wireless data traffic. The module 102 serves as the data foundation for the subsequent ML processes.

[0130] The device 100 further comprises a machine learning module 104 that configures the ML system by performing ML based on each sequence of measurement values obtained from the sequences obtaining module 102. The machine learning module 104 uses the diverse sequences (i.e., sequences resulting from the different time domain processing) to derive multiple sets of parameter values. Each sequence or time domain processing potentially yields different insights into the wireless data traffic, thus the varied sets of parameters are crucial for optimizing the system response to fluctuating network conditions and traffic patterns.

[0131] Following the determination of the parameter values for each sequence, the ML module 104 integrates these parameters into the ML system. The module 104 decides which set of parameters from the machine learning best meets the performance metric, thus enabling the most effective configuration of the ML system. This choice influences how the wireless access network is controlled, optimizing its performance based on the adapted time domain processing.

[0132] An optional controlling module 106 of the device 100 may play a crucial role after the ML system 110 has been properly configured. It takes charge of controlling the wireless access network using the set of parameter values selected by the machine learning module 104 in combination with the associated time domain processing.

[0133] Any of the modules of the device 100 may be implemented by units configured to provide the corresponding functionality. The device 100 may also be referred to as, or may be embodied by, a first network entity 100, e.g., a network node of the wireless access network or a centralized unit of such a network node. The first network entity 100 and a second network entity (e.g., a distributed unit), which controls the wireless access network taking the output of the configured ML system into account, may be in signaling communication, e.g., at least for updating the parameters of the ML system when the time domain processing is changed, e.g., by sending a configuration message. The distributed unit may be embodied by the following device 200.

[0134] Fig. 2 schematically illustrates a block diagram of an embodiment of a device for controlling a wireless access network. The device is generically referred to by reference sign 200.

[0135] Embodiments of the device 200 may leverage a machine learning (ML) system for controlling the wireless access network. Same or further embodiments of the device 200 may receive configuration parameters and time domain processing directives for handling wireless data traffic, obtaining measurement values of said traffic accordingly, and applying these configurations to control the wireless access network.

[0136] Referencing Fig. 2, the device 200 comprises a configuration receiving module 201 that receives a configuration message indicative of parameters of a machine learning system and a time domain processing of wireless data traffic. The module 201 may effectively capture and process a configuration which encompasses operational parameters set for the machine learning (ML) system, as well as the methodologies to be employed in analyzing the time domain aspects of monitored wireless data traffic.

[0137] The device 200 further comprises a sequence obtaining module 202 that obtains a sequence of measurement values indicative of wireless data traffic, which is guided by the received time domain processing settings. The module 202 may actively gather and organize data from network interactions ensuring that the measurement values align with the specified analytical procedures dictated by the configuration received at module 201.

[0138] A controlling module 206 of the device 200 controls the wireless access network using the configured machine learning (ML) system where the obtained sequence is applied to the ML system. The module 206 may interpret the outputs from the ML system, which are informed by the structured data provided by the sequence obtaining module 202, to make real-time decisions that optimize and control various network functions and resources based on learned and predicted traffic patterns.

[0139] The device 200 may be embodied as a second network entity capable of applying the result of ML to dynamically control a wireless access network. The modules may transform raw data of the wireless data traffic into actionable control commands, which are used to maintain efficient operations of the wireless access network. This can ensure the network's performance is aligned with the everchanging conditions and demands of network traffic.

[0140] Any of the modules of the device 200 may be implemented by units configured to provide the corresponding functionality.

[0141] The device 200 may also be referred to as, or may be embodied by, a second network entity of the wireless access network (e.g., a network node of the wireless access network or a distributed unit of the network node). The second network entity 200 and a centralized unit that performs the ML and the selection of the tie domain processing may be in signaling communication, e.g., at least for the receiving of the configuration message. The centralized unit may be embodied by the above-mentioned device 100 (e.g., the first network entity 100).

[0142] Fig. 3 shows an example flowchart for a method 300 of configuring parameters of a ML system that controls a wireless access network.

[0143] In the embodiment of Fig. 3, the method 300 begins with step 302 where at least two different sequences of measurement values indicative of wireless data traffic are obtained. These sequences are obtained using different time domain processing of the wireless data traffic. In practical terms, this step may involve collecting data from network traffic and applying various time domain processing techniques, to the same history time window of the same wireless data traffic, to generate distinct sets of measurement value, e.g. allowing for an analysis of network behavior and / or wireless device behavior over different time scales. Following this, machine learning (ML) of a ML system is performed based on each of the sequences as per step 304 of the method 300, resulting in at least two sets of parameter values. This step may apply a ML algorithm to the different sequences obtained, optimizing and tuning the ML system to derive actionable output that guides network management decisions.

[0144] Within step 304, an optional substep involves evaluating a performance metric for the ML system, which guides a selection of one set of the parameters for controlling the wireless access network. This evaluation may compare different sets of the parameter values based on their ability to predict network conditions accurately or optimize network performance.

[0145] Another optional substep of the step 304 includes configuring the ML system (e.g., an instance of the ML system that is collocated with the second network entity that controls the wireless access network) with one of the at least two sets of parameter values. The one set may be selected based on the performance metric evaluated in above substep. For example, the selected parameter set may be integrated into the ML system to make it operational, e.g. for real-time network traffic control and adjustments.

[0146] Alternatively or in addition, the first network entity may itself control the wireless access network in an optional step 306 (as indicated by dotted a line box).

[0147] As schematically illustrated in Fig. 3, embodiments of the method 300 cover where network traffic data is processed and analyzed using ML to derive optimized control strategies, refining the wireless access network's 500 operations to align with current and anticipated conditions of the wireless access network and / or current and anticipated network requirements of one specific wireless device. The multi-time domain processing enables flexibility in the ML of the ML system targeting specific performance metrics crucial to network efficiency and user satisfaction.

[0148] The method 300 may be performed by the device 100. For example, the modules 102, 104 and 106 may perform the steps 302, 304 and 306, respectively.

[0149] Fig. 4 shows an example flowchart for a method 400 of controlling a wireless access network, which may begin with a step 401 where a configuration message indicative of parameters of a machine learning system (ML system) and a time domain processing of wireless data traffic is received. The step 301 may involve receiving information that sets the operational conditions for the machine learning system (e.g., the ML model as represented by the set of parameter values) that will subsequently be applied to control the wireless access network (e.g., to manage the wireless data traffic). Furthermore, the configuration may include time domain parameters that dictate how the ML system is fed and / or should interpret and analyze the wireless data traffic, e.g. specifications on how to segment and process the wireless data traffic time-wise.

[0150] A step 402 of the method 400 involves obtaining a sequence of measurement values indicative of the wireless data traffic between the wireless access network (e.g., one specific network node of the wireless access network) and one or more wireless devices (e.g., one specific wireless device), where the sequence is obtained using the received time domain processing of the wireless data traffic. The step 402 may collect wireless data traffic in a format that is analyzable by the ML system using the earlier received configuration specifying time domain processing methods. This can entail gathering real-time or near-real-time data about the wireless data traffic, which may then be formatted or filtered according to the specified time domain methodologies to ensure compatibility and effectiveness in subsequent analysis steps.

[0151] The sequence obtained in step 402 is applied in a step 406 of the method 400 to control the wireless access network using the ML system configured with the received parameters. This step may yield the actionable output of the ML system, configured per the previously received parameters, which is utilized to make realtime adjustments or optimizations in the operation of the wireless access network. This may involve adjusting bandwidth allocations, activating or deactivating network channels, secondary cells, component carrier, or other control actions aimed at enhancing network performance and efficiency based on the predictive analytics provided by the ML system.

[0152] The method 400 may be performed by the device 200. For example, the modules 201, 202 and 206 may perform the steps 401, 402 and 406, respectively.

[0153] In any aspect, embodiments of the technique may use a performance metric that evaluates the effectiveness of the machine learning (ML) models (i.e., the different sets of parameter values) developed from different sequences of measurement values to control the wireless access network. A concrete example of the performance metric within this scenario may be the Root Mean Square Error (RMSE) or another metric to measure an error of a ML model in predicting the measurement values. It assesses the average magnitude of the errors in a set of predictions, without considering their direction. Lower RMSE values may indicate a better fit of the ML model to the wireless data traffic, making it a valuable performance metric for selecting the one set out of the parameter sets for configuring the ML system that controls the wireless access network.

[0154] Alternatively or in addition, the performance metric may be indicative, for each model (i.e., for each of the sets of parameter values), of a similarity of the performance of the model (i.e., the ML system configured with the respective set of a parameters) when applying a training dataset and a validation or test dataset. Lacking similarity should be avoided and / or may indicate an overfitting and eventually low quality performance when such model would be used with future sequences.

[0155] This RMSE may be specifically used to compare the predictive accuracy of different ML models (developed from different time domain sequences of the wireless data traffic) in forecasting network traffic conditions or wireless device demand, thereby guiding the selection of the most appropriate model (i.e., set of parameters) to deploy for dynamic network control tasks like carrier aggregation management, resource allocation, and dynamic cell activation / deactivation.

[0156] In any aspect, for concreteness, and without limitation of the technique, below description refers to radio access network (RAN) as an example of the wireless access network, a user equipment (UE) as an example of the wireless device, and a fifth generation network node (gNB) as an example of the network entities.

[0157] Furthermore, while the technique is primarily described for wireless data traffic in the uplink (UL) and / or downlink (DL), the technique is also applicable to direct communications between wireless devices, e.g., device-to-device (D2D) communications or sidelink (SL) communications. In that case, the role of the wireless access network may be replaced one or more relay wireless devices. Moreover, each of the first network entity 100 and the second network entity 200 may be a relay radio device or a base station. Herein, any radio device may be a mobile or portable station and / or any radio device wirelessly connectable to a base station or RAN, or to another radio device. For example, the radio device may be a user equipment (UE), a device for machine-type communication (MTC) or a device for (e.g., narrowband) Internet of Things (loT). Two or more radio devices may be configured to wirelessly connect to each other, e.g., in an ad hoc radio network or via a 3GPP SL connection. Furthermore, any base station may be a station providing radio access, may be part of a radio access network (RAN) and / or may be a node connected to the RAN for controlling the radio access. For example, the base station may be an access point, for example a Wi-Fi access point.

[0158] The different time domain processing used by the first network entity for the different sequences and / or the selected time domain processing indicated from the first network entity to the second network entity may be out of a predefined codebook of time domain processing. Furthermore, "predefined" may encompass stored in memory of the network entity, or hard-coded or hard-wired in the network entity, or preconfigured or configured by a network node or radio access network (RAN) for the network entity.

[0159] Herein, a list of the form A, B, and / or C (also written as A, B and / or C) may correspond to at least one or each of A, B, and C, i.e., A and / or B and / or C.

[0160] Any embodiment of any aspect may be implemented as a ML-assisted method for carrier activation and / or deactivation. Alternatively or in addition, the controlling of the wireless access network may relate to carrier aggregation (CA), e.g. activation and / or deactivation of one or more secondary carriers.

[0161] Fig. 5 provides a schematic overview of an exemplary wireless access network scenario showing interactions and application of various embodiments of the first network entity 100 and the second network entity 200. In one variant of any embodiment, one or more distributed units (DU) 200 embody the second network entity, and a centralized unit (CU) 100 embodies the first network entity 100. The ML system (MLS) 110 resides within the CU 100 for both training phase (e.g., including the step 304) and inference phase (e.g., including the step 306). Each DU may provide radio access to one or more radio device 504. For example, each DU 200 covers a cell 502 of the radio access network (RAN) 500. Optionally, these cells overlap for dual connectivity as an example of controlling 306 or 406 the RAN.

[0162] In the first variant depicted by solid lines, the CU 100 houses the MLS 110 for both the training phase 304 and the inference phase 306, e.g. contributing to the adaptability and intelligence of the network's operational protocols.

[0163] In a second variant, depicted by dashed lines, the inference functionality 406 of the MLS 110 is integrated within one or more distributed units 200 placed at different locations in the RAN 500. Such a decentralized approach may benefit from a reduced complexity of the MLS 110 as a result of selecting the time domain processing and data handling which enhances responsiveness and reliability.

[0164] In a third variant, the training according to the step 304 or the entire method 300 is performed by a training server (e.g., in a core network serving the RAN 500), as illustrated in the right-hand side of Fig. 5. In the third variant, the CU and / or the one or more DUs may act as the second network entity 200 performing the method 400.

[0165] UEs as wireless devices 504 are in a connected state with the distributed units 200 of the RAN 500. The controlling 306 or 406 may comprise decision-making concerning network control, potentially based on predictions of the ML system 110, which can be performed close to the network edge, reducing latency and increasing the efficiency of the network's responses to dynamic conditions.

[0166] A mobility of wireless devices 504, encompassing possibly mobile phones or other mobile connected devices, and their interaction with multiple distributed units 200 may be an example of the network control 306 or 406 for a hand-over between DUs 200 or cells 502 of the RAN 500. Based on the output of the ML system 110, the DUs 200 are capable of controlling the RAN 500 under changing network environments and user locations, e.g., essential for maintaining seamless connectivity and service quality. The network control 306 or 406 may comprise assigning different radio resources to the wireless device 504, e.g. in the time domain (e.g., scheduling grants), the frequency domain (e.g. component carriers), and / or the spatial domain (e.g., number of multiple-input multipleoutput layers, MIMO layers).

[0167] In a variant of any embodiment, the controlling 306 or 406 of the RAN 500 comprises Carrier Aggregation (CA). Fig. 6 schematically illustrates CA.

[0168] Carrier aggregation is a radio telecommunication technology that combines one or more component carriers (CCs, also: secondary CCs or SCCs) along with a primary carrier (also: primary component carrier, or PCC) to jointly provide data transmission at a higher capacity. Fig. 6 illustrates an example of CA in the case of Long Term Evolution (LTE) and New Radio (NR).

[0169] The controlling 306 or 406 of the RAN 500 may comprise selectively activating and deactivating the one or more SCCs depending on a volume of the data traffic predicted by the ML system 110 for the particular UE 504. When activated, UL and / or DL data traffic between a network node 200 (e.g., a gNB or the DU 200) and the UE 504 is split at a medium access control (MAC) layer of a radio communication protocol stack, which may be split between the CU 100 and the DU 200.

[0170] In other words, the method 300 or 400 may control 306 or 406 the wireless access network by controlling a medium access control (MAC) layer, e.g. for activating or deactivating carriers for dual or multi connectivity between a network node of a wireless access network 500 and a wireless device 504. The MAC layer may be a layer of a radio protocol stack, e.g. according to the 3GPP document TS 38.321, version 18.1.0.

[0171] Alternatively or in addition to CA, the controlling 306 or 406 of the RAN 500 may relate to dual connectivity (DC), which is schematically illustrated in Fig. 7. User traffic is split between carriers in a packet data convergence protocol (PDCP) layer.

[0172] As described in the 3GPP standard, for CA or DC to be working for a traffic session of a UE 504, CA or DC would first require a CA or DC configuration and then a secondary cell activation process. Similarly, to stop using CA or DC for a UE traffic session, an explicit secondary cell deactivation process is also required. Note that activation and deactivation may imply different delays depending on technology and standards. For example, currently on average around 90 ms and 20 ms delays for activation and deactivation, respectively.

[0173] In a variant of any embodiment, the wireless data traffic 800 may be structured in data radio bearer (DRB) sessions and Internet Protocol (IP) flows. Fig. 8 schematically illustrates DRB sessions comprising one or more IP flows.

[0174] When obtaining 302 the different sequences in the method 300 and / or when obtaining 402 the sequence in the method 400, the wireless data traffic (e.g., an incoming UE traffic session) of each of the one or more wireless devices (e.g. UEs) may be categorized in two or more types, e.g. on two or more different levels. Firstly, each sequence may distinguish the wireless data traffic at a flow-level, i.e., incoming packets within the same IP flow, which may correspond to a single application or destination server. Secondly, e.g. in combination with the firstly mentioned IP flow, each sequence may distinguish the wireless data traffic at a DRB-level (or a DRB session). E.g., incoming IP packets of all IP flows (which may correspond to all applications) within a DRB session may be included in the measurement values. Fig. 8 visualizes an example of a change of the wireless data traffic between two levels. A time for the change (tc), e.g. due to different usage of the wireless data traffic by different applications, may be a configuration parameter of the wireless device or its radio configuration. A value of tcmay be 10 seconds or greater and / or may depend on a data preparation at a server (for downlink wireless data traffic) and / or a data preparation at the wireless device (for uplink wireless data traffic).

[0175] The ML system configured for the selected one of the at least two sets (based on the performance metric) can detect a precursor of the change in the usage of the wireless data traffic for preemptively controlling the wireless access network based on a prediction of the change of the wireless data traffic.

[0176] Embodiments of the network entities 100 and 200 and the methods 300 and 400 can control the RAN 500 by recommending when to appropriately activate and / or deactivate a secondary cell for a traffic session of a UE 504, e.g., a Data Radio Bearer (DRB) session. The controlling 306 or 406 may be specific for a pair of a UE 504 and a base station 100 and / or 200. In this technical context, "appropriate" may refer to a situation when there exists enough UE data volume to be transmitted (e.g., pending data packets) that would require carrier aggregation (CA) for the UE 504.

[0177] Thus, embodiments address an important problem, since activating 306 or 406 a secondary cell (as an umbrella term for any expansion of radio resources, e.g. in CA or DC) in case of wireless data traffic 800 with low data volume causes an efficiency problem in terms of power consumption of the UE 504 and wasted spectrum utilization for the RAN 500. In a similar fashion, if deactivation 306 or 406 occurs when the expanded radio resources (e.g., CA and / or DC) would be required, the deactivation would damage user quality of experience (QoE), since the UE 504 would receive less capacity for the wireless data traffic 800 than an expansion of radio resources (e.g., CA or DC) could offer.

[0178] Embodiments of the technique can combine the advantages of a fast (e.g., realtime) rule-based approach, which cannot adapt to changing conditions and demands, and the afore-mentioned technique of Elsayed et al. using a multi-agent Reinforcement Learning (RL)-based approach that predicts the arrival of data and identifies Secondary Component Carriers (SCCs) to activate for each user, i.e. one agent per user. Furthermore, embodiments can look further ahead in time due to the selected time domain processing, for which reason embodiments of the method 300 or 400 do not have to operate at the MAC level in contrast to the afore-mentioned existing ML-based technique.

[0179] Alternatively or in combination with any aspect and embodiment, a machine learning-assisted methods 300 and 400 for proactively activating and / or deactivating a secondary cell (e.g., for CA or DC) is disclosed. The method 300 applies a predictive model using machine learning to estimate per-UE traffic in a future fixed time horizon (e.g., 200 ms), optimizing parameters, e.g., a duration of a time slot (or time interval) and / or a size (e.g., duration or number of measurement values) of a history time window. Same method 300 or a further method 400 provides a recommendation to a secondary network entity 200 (e.g., a scheduler of the RAN 500) to decide activation and / or deactivation of the CA or DC. The methods 300 and 400 enables efficient operation of the RAN 500 in terms of actual throughput gain from the activation of the CA or DC and avoids or reduces power consumption of the UE 504 caused by delayed or late deactivation of the CA or DC. Fig. 9 shows a schematic block diagram of an embodiment of the first network entity 100 performing an embodiment of the method 300. In order to control the RAN 500 based on the ML system 110, the device 100 provides the configuration of parameters of the ML system 110, which plays a central role in optimizing network management.

[0180] Fig. 9 particularly illustrates an embodiment that provides a detailed view of the first network entity 100 involved in configuring the ML system 110 for network operation. In the diagram, critical components such as the sequence obtaining module 102 and the ML module 104 may be implemented in a separate apparatus or in a network node (e.g., a gNB) serving the UE 504, for example in a CU or a DU of the network node. The wireless data traffic relates to the pair of network node and UE 504.

[0181] The sequence obtaining module 102 is tasked with collecting measurement values indicative of the wireless data traffic, which are essential inputs for the ML system 104. The module 102 adapts multiple time domain processing strategies to capture detailed traffic sequences under various network conditions. This enables the system to accommodate diverse network scenarios and user behavior, providing a robust dataset for subsequent analysis.

[0182] Machine learning module 104 stands central in this configuration, receiving the diverse sequences from the sequence obtaining module 102. It performs the different time domain processing to sample values of the wireless data traffic and / or controls different sampling modes for acquiring the sample values of the wireless data traffic according to different time domain processing, and performs a ML process to derive multiple sets of parameter values, e.g. one set for each of the sequences. These parameter values are pivotal; they represent the potential configurations for the ML system 110 to effectively control the network. The ML system 110 (e.g. in the DU or CU) uses these parameters to predict and adapt to network and / or UE demands dynamically.

[0183] Moreover, ML 304 based on sequence 1, sequence 2, and sequence 3 illustrates the iterative processing capability of the module 104, whereby each sequence undergoes a distinct evaluative process. This analytical depth ensures that the parameters selected are tuned to the most effective operational strategy as judged by performance metrics. The selected set of parameters, shown as the inference model in Fig. 9, culminates the ML process. These parameters are determined to be the most effective through rigorous testing and are chosen for actual implementation in the ML system 110 to control the live network environment.

[0184] The device setup, including modules 102, 104 affiliated with the first network entity 100 (an optionally the second network entity node 200), creates a robust framework for real-time, data-driven decision-making in network control. By processing time-domain data variably, a system comprising the network entities 100 and 200 can optimize parameter settings continuously, ensuring network performance is maximized and aligned with current traffic conditions and predictions.

[0185] Alternatively or in combination with any aspect and embodiment, methods 300 and 400 may comprise a ML-assisted method for activation and / or deactivation of radio resources, e.g. as illustrated in Fig. 10 or Fig. 11. Furthermore, the methods 300 and 400 may at least one of: o use one or more sequences of measurement values according to a "regular" and / or "irregular" time series; o handle various sampling rates (e.g., as illustrated in Fig. 12A); o select the best time slot (also referred to as time window) for aggregating the wireless data traffic, e.g. as an example of a data driven technique (e.g., as illustrated in Fig. 10); and o provide recommendations (e.g., control commands) to control the RAN 500, e.g. to activation and / or deactivation network resources based on a prediction of the wireless data traffic for a specific time horizon (e.g., as illustrated in Fig. 11).

[0186] Moreover, any embodiment may be embedded in the RAN 500 by signaling between entities and deployment scenario for when this technique is realized in practice (e.g., as illustrated Fig. 15). The entities may comprise a dedicated apparatus where the method 300 or 400 is implemented or a separate apparatus comprising the ML system 110. Alternatively or in addition, the entities may comprise a gNB-CU 100 and a gNB-DU 200. Any of these entities may serve a UE 504 transmitting and / or receiving a data over a user plane (UP) in the RAN 500, i.e., in the wireless data traffic.

[0187] The following detailed embodiments may be implemented as such or as an extension of one of the afore-mentioned embodiments.

[0188] Fig. 10 shows a flowchart of an embodiment of the method 300 and / or a data collection and model (e.g., in a training phase of the technique).

[0189] Fig. 11 shows a flowchart of an embodiment of the method 400 and / or inference and recommendation for activation and / or deactivation a network resource (e.g., a radio resource), which is also referred to as consumer (e.g., in an inference phase of the technique).

[0190] The detailed embodiments may use at least one of the following prerequisites. It is assumed that UE flow level data (e.g. packets 802) can be collected in the user plane of the centralized unit (CU-UP) 100. In cases that the complexity of opening every packet and collecting data is expensive, a sampling strategy (e.g., subsampling) may be used, e.g. as shown in Fig. 12A.

[0191] For the simplicity of explanation, it is assumed that the goal of the ML model is to predict the time slot next time window volume. However, the process for predicting multiple time windows is similar except the model training phase needs to be adjusted.

[0192] Fig. 10 and Fig. 11 show the training and inference phases for the proposed solution. In training phase, first the UE level flow (packet level) is collected and then an iterative approach is used to train model which also selects the best time slot (details in the sequel) for grouping the packet data. Once the model is trained and the best partitioning data parameter is determined, the inference phase takes the results from the training phase and estimate the future flow. Then, a recommendation block gets the traffic estimate and provides recommendation to activation / deactivation consumer. The details of training and inference plus the recommendation phases are detailed below.

[0193] Data collection (AO & BO) The embodiments of the methods 300 and / or 400 collect the following related to incoming packet as pictured in Fig. 12A, with the differently hashed boxes indicating a quantitative feature of data packets of the wireless data traffic over time, e.g., time of arrival or availability of the respective data packet. For each incoming (e.g., received) or outgoing (e.g., transmitted) data packet, at least one of the following features may be collected:

[0194] - a packet volume of the data packet;

[0195] - a packet duration of the data packet;

[0196] - a number and / or rate of data packet;

[0197] - a protocol (e.g., according to an application layer or another layer of a communication protocol stack) associated with or used by the data packet;

[0198] - a direction of the data packet in the wireless data traffic (e.g., uplink or downlink);

[0199] - an IP address and / or a port of the respective one of the wireless devices (e.g., internet socket or network socket of the wireless device) in the data packet;

[0200] - an IP and a port of a server associated with the data packet;

[0201] - time of arrival (e.g., a time of availability), e.g. for transmission or reception, of the data packet; and

[0202] - a unique and / or temporary identifier of the respective wireless device, e.g. UE identifier (called UE ID in the following)

[0203] - If there are other data, e.g., cell ID, or geographical location, they will be used along with the available data.

[0204] Alternatively or in addition, the at least one of the following data fields may be included in the sequence per UE 504:

[0205] - time stamp

[0206] - persistent UE ID

[0207] - packet volume, duration, protocol, and / or direction; and

[0208] - sample packet.

[0209] Fig. 12A to 12C illustrate data preparation and / or wireless data traffic with regular (e.g., periodic) and / or irregular (e.g., aperiodic) time intervals in the steps 302 and In the one variant, if all packet information collection would require a to huge data storage, which cannot be satisfied, a data collection sampling method is then applied to overcome this difficulty. The steps 302 and 402 comprise of collecting the incoming packets at a slower rate as shown in Fig. 12A with black arrow above the packet flow, the other packets are ignored. In the example drawn in Fig. 12A, instead of collecting features of 9 packets after the sampling we collect features of 6 packets. The sampling rate is determined by technical limitation, like buffer size etc.

[0210] The step 302 or 402 further comprises a data preparation (A1&B1), i.e. the time domain processing. The time domain processing may be defined for a given time slot TS of duration T (e.g., as illustrated at reference sign 1200 in Fig. 12), optionally according to the configuration message received in the step 401 (which may be sent in the step 304), or any other a configuration parameter or as a result of cross validation (CV).

[0211] For example, the time domain processing may comprise at least one of:

[0212] - aggregate the traffic per UE per time slot T;

[0213] - select a window (via a configuration or a learning process);

[0214] - define the label as the most recent traffic in a time slot T; and

[0215] - for regression, create data frame for regression.

[0216] Once the data are collected in the step 302 and 402, a data preparation step may be required to make the collected data usable for a next step including model training 304.

[0217] For the data preparation step or the usage of the wireless data traffic, two implementations are disclosed. The data preparation or usage of the wireless data traffic may be in a regular (e.g., periodic) or in an irregular (e.g. aperiodic) time window, which are also referred to as "regular time interval" (e.g., as illustrated in Fig. 12B) and "irregular time interval" (e.g., as illustrated in Fig. 12C), respectively.

[0218] Each approach has its own advantages and use cases. In regular time series-based approach, the start of the next time frame 1200 immediately begins after the end of the current time slot 1200, making the synchronization easier. But the drawbacks are that it should continuously process even there is no arrival packet. While in irregular time series, the processing or aggregation only starts when the first packet arrives. But the drawback is that the time gaps between traffics in time slots are not considered.

[0219] In both implementations, we start by defining a time slot T to aggregate the traffic (e.g., volume) over T. The time slot could be defined empirically, heuristically based on the Proof of Concept or determined by technical limitations; further details will be provided later. Once time slot T is determined, the difference between the two implementations is the triggering moment.

[0220] In the "regular time interval"-based approach, the data are grouped every time slot T since the first incoming packet. According to the Fig. 12B, the collected data are grouped into 5 subsets starting from the first dotted bar, corresponding later to 5 inputs for a specific UE ID.

[0221] While in the "irregular time interval" implementation, the data are gathered over T according to the next ungrouped and incoming packets. In the example drawn in Fig. 12C, there are 4 inputs.

[0222] Model training, i.e., the step 304, may be implemented according to the sub-steps Al to A6 or A2 to A6 in Fig. 10.

[0223] If the time slot TS has a duration T as a default value, the data prepared in the last section will be used for training a model. Otherwise, the best value of T will be selected through training phase according to the performance metric in the step 304 and / or as explained below.

[0224] First define a range of T values, for example 100 ms, 200 ms, 300 ms, etc. A vector T = [Ti, T2,..., Tn] may define possible values for the duration of the time interval (i.e., the time slot 1200).

[0225] The process from Al to A6 contains at least one of the following steps: o Prepare time series data based on Tko Train a ML model for given Tk (a regression or a time series-based model) o Compare the performance of the current model (model k) with best model trained so far. If the model outperforms the best model developed so far, the best model will be replaced by model k and the best time slot by Tk o Change the time slot window to T(k+i) and continue the loop. o The process will be continued for all time window candidates.

[0226] Note that to have a fair comparison, the normalized volume prediction will be compared. That is, the volume will be normalized by T . The reason is that if the time interval gets smaller, the aggregated traffic (volume) in the interval also becomes smaller.

[0227] Fig. 13 schematically illustrates an implementation of the ML step 304. While the example may use supervised learning with controlled input and output values, any embodiment may alternatively use reinforcement learning.

[0228] Performing the ML 304 may comprise receiving the sequence of measurement values in a training phase of the ML system (i.e., as an input to the ML system, e.g. at an input layer of a neural network of the ML system).

[0229] The ML 304 may comprise supervised learning, e.g. wherein an output of the ML system 110 comprises one or more subsequent values (e.g., one or more predicted subsequent values that are compared with subsequent measurement values as the label in the training phase or predicted future values as the basis for the controlling of the wireless access network in the inference phase). Alternatively, the ML may comprise reinforcement learning (RL), e.g. if the output of the ML system 110 comprises an action for controlling the wireless access network. A state of the RL may comprise the sequence of measurement values and a current mode of operation of the wireless access network. The action may comprise a future mode of operation of the wireless access network, e.g. equal or different from the current mode of operation.

[0230] The ML may use reinforcement learning or supervised learning and / or back propagation based on one or more measurement values measured subsequently (e.g., as ground truth) after the sequence of measurement values received by the ML system in the training phase.

[0231] The ML problem can be cast as a time series or a regression problem, e.g. as illustrated in Fig. 13. The regression approach may be particularly suitable, since other features can be easily inserted to the data, e.g., cell ID or geographical location data. To train a regression model, first it may be required to prepare the data as a tabular format. To do that, a history window (W) 1402 should be defined. W could be a default parameter or it can be obtained through cross validation (a data driven approach).

[0232] The data driven approach is a preferred implementation of the technique. First, define a range of W as W=[Wi,W2,..,Wj]. For W=Wj, the step 304 may comprise at least one of: o Get the flow history and transform them to vector (sliding window), e.g., as illustrated in Fig. 14; o Split the data into portions for training and / or validation and / or testing; o Train a regression model on train and / or validation data to learn the relation between input and output. Using cross validation, the best parameters are selected. For model design, a relevant metric, e.g., RMSE (root-mean-square error), MAE (mean absolute error), will be chosen when training the model; o Evaluate 304-1 the model on test data and compare the performance with best model achieved for W=Wi to Wi. If the performance is improved replace the best model by the model developed for Wj otherwise keep the previous best model. o The best model along with the best W (say Wopt) will be kept for inference phase.

[0233] It is noted that when doing the transformation, some "not a number" (NaN) values could appear in the transformed data. The missing values will be handled via imputation. By imputation, the size of the data for model development will be kept the same for all Wj, i=l,..., J, resulting in a fairer comparison.

[0234] In one variant, the technique may use a fixed time interval-based model.

[0235] The input to the ML system 110 may comprise measurement values 1300 of time slots 1200 in the history time window 1402. The time slots 1200 may be either regular or irregular time intervals.

[0236] The output to the ML system 110 may predict the traffic for the next time slot.

[0237] Alternatively, for multi-time slots predictions, if T is small, the prediction may be at least two time slots predictions ahead. This can enable more reliable recommendation for the controlling 406 based on consistent predictions.

[0238] Fig. 14 schematically illustrates a row to vector transformation (i.e., reshaping) of the measurement values 1300.

[0239] A detailed embodiment of the method 400 for the inference phase is described.

[0240] Once the model is trained and the best T (Topt) is selected in the step 304, optionally if the default is not available, the inference phase will use the outcome of the training phase (e.g., as configured in the step 401) to provide an estimate of volume for future for the controlling 406.

[0241] Alternatively or in addition, the method 400 may comprise at least one of: o Data collection 402 is similar to the training phase (i.e., 302) with the same sampling strategy. o Data preparation in the step 402 uses the best time window T, from training and / or validation phase. o Use the size Woptfor the history time window to transform the time series data into data frame, similar to training phase. o Prediction: The ML system 110 predicts the volume of the wireless data traffic for the next time window (i.e., time slot 1200), e.g. regular or irregular, optionally in combination with a (e.g., 95%) confidence interval. o In the last step, the prediction will be scaled based on a sampling ratio. That is, if, instead of sampling frequency f, a smaller frequency fs, fs<f, is selected, the ratio of f / fs will be used to scale the prediction.

[0242] The controlling step 406 (and optionally the controlling step 306) may be implemented as a recommendation.

[0243] The recommender block gets the prediction plus the accuracy metric (in terms of confidence interval) and evaluate the future predictions and provides suggestions for activation.

[0244] The control module 206 may be referred to as a recommender block or briefly recommender. The step 406 may comprise at least one of: o For T is small, e.g., 20 ms or less, the recommender 206 request multi time slot estimates and the associated confidence intervals. o For large T, e.g., 200 ms or more, the recommender 206 may only require single estimate. o Evaluate the incoming traffic, in terms of volume: if the volume is larger than a threshold, suggest activation otherwise deactivation (if it was activated previously). o Assign the reliability of the decision in terms of probability, i.e., how confident the recommendation is. The reliability is linked to the confidence interval and also the previous decision.

[0245] Fig. 15 schematically illustrates a signaling diagram of how embodiments of the network entities 100 and 200 can be realized in practice. The signaling diagram of Fig. 15 relates the entities performing the methods 300 and 400 in the context of a 3GPP RAN 500.

[0246] Fig. 15 presents a signaling diagram visualizing the signals between entities in this invention. While the sub-steps B0 - B4 of the method 400 may be implemented in a dedicated apparatus comprising the ML system 110 (as indicated in Fig. 15), the method may be performed by the CU 100 or the DU 200 in a variant of any embodiment. Furthermore, the steps S0-S3 represent signals between RAN entities.

[0247] For example, in one variant, the dashed line represent that the apparatus implementing method 300 and / or 400 can be optionally geographically collocated together with the gNB-CU 100 and / or the gNB-DU 200 for various RAN deployment scenarios.

[0248] As an embodiment, the apparatus can be deployed in a distributed RAN (DRAN) 500, e.g., whereby gNB-CU 100 and gNB-DU 200 are geographically collocated, while in another embodiment, which may include a centralized RAN (CRAN), whereby a CU 100 and DUs 200 are geographically distributed.

[0249] In yet another embodiment, for example, in open RAN (O-RAN) according to the O- RAN Alliance, where an apparatus performing the method 300 and / or 400 may be implemented as in a Near-real-time Radio Intelligent Controller (Near-RT RIC), which is not geographically collocated with gNB CU and gNB DU. In any of the above embodiments, the "best" time slot may be the time slot that has the best predictive performance in terms of accuracy, e.g. as indicated by evaluating 304-1 the performance metric.

[0250] The iterative process of identifying the best duration of the time slot 1200 (e.g., in the training phase) may be based on Fig. 10 or method 300: For time slot T belong to a predefined sequence 1400 of measurement values 1300:

[0251] ■ Prepare the data using the current time slot T

[0252] ■ Train the model

[0253] ■ Evaluate the performance of the model

[0254] ■ If the performance is improved, assign T as the best time slot and ignore the previous one.

[0255] The inference phase may be implemented according to Fig. 11 or the method 400, which may comprise at least one of: o Use the selected best time slot from the training phase (Step 2 above) to prepare the data for inference. o Use the trained model (Step 2) to predict the traffic.

[0256] For regular and irregular time interval data preparation (e.g. as illustrated in Figs. 12B and 12C, the above approach may be used.

[0257] The controlling 306 or 406 may trigger Carrier Aggregation, Load Balancing, Mobility Prediction, and / or Channel Prediction.

[0258] The deployment scenario can have significant advantages. Firstly, the CU 100 generally has extra computational resources (available to execute ML inference) compared to DU 200. Secondly, the DU 200 is operating at a real-time scale (at the order of several microseconds). Really few ML models can achieve that latency in implementation. However, embodiments of the method 300 or 400 may optionally perform the step 306 (i.e., the inference) at the CU 100 at the slower loop (at few milliseconds) latency and giving the signal to real-time scheduler in the DU 200.

[0259] Fig. 16 shows a schematic block diagram for an embodiment of the device 100. The device 100 comprises processing circuitry, e.g., one or more processors 1604 for performing the method 300 and memory 1606 coupled to the processors 1604.

[0260] For example, the memory 1606 may be encoded with instructions that implement at least one of the modules 102, 104 and 106.

[0261] The one or more processors 1604 may be a combination of one or more of a microprocessor, controller, microcontroller, central processing unit, digital signal processor, application specific integrated circuit, field programmable gate array, or any other suitable computing device, resource, or combination of hardware, microcode and / or encoded logic operable to provide, either alone or in conjunction with other components of the device 100, such as the memory 1606, training functionality and optional inference functionality. For example, the one or more processors 1604 may execute instructions stored in the memory 1606. Such functionality may include providing various features and steps discussed herein, including any of the benefits disclosed herein. The expression "the device being operative to perform an action" may denote the device 100 being configured to perform the action.

[0262] As schematically illustrated in Fig. 16, the device 100 may be embodied by a centralized unit 1600, e.g., functioning as a training network node of the RAN 500. The CU 1600 comprises an interface 1602 (e.g., an Fl interface) coupled to the device 100 for communication with one or more distributed units (DUs), e.g., functioning as radio heads serving UEs.

[0263] Fig. 17 shows a schematic block diagram for an embodiment of the device 200. The device 200 comprises processing circuitry, e.g., one or more processors 1704 for performing the method 400 and memory 1706 coupled to the processors 1704.

[0264] For example, the memory 1706 may be encoded with instructions that implement at least one of the modules 201, 202 and 206.

[0265] The one or more processors 1704 may be a combination of one or more of a microprocessor, controller, microcontroller, central processing unit, digital signal processor, application specific integrated circuit, field programmable gate array, or any other suitable computing device, resource, or combination of hardware, microcode and / or encoded logic operable to provide, either alone or in conjunction with other components of the device 200, such as the memory 1706, network controlling functionality. For example, the one or more processors 1704 may execute instructions stored in the memory 1706. Such functionality may include providing various features and steps discussed herein, including any of the benefits disclosed herein. The expression "the device being operative to perform an action" may denote the device 200 being configured to perform the action.

[0266] As schematically illustrated in Fig. 17, the device 200 may be embodied by a distributed unit (DU) 1700, e.g., functioning as a scheduler of the RAN 500. The DU 1700 comprises an interface 1702 (e.g., the Fl interface) coupled to the device 200 for communication with an associated CU, e.g., functioning as a training network node.

[0267] As has become apparent from above description, at least some embodiments of the technique adapt to different scenarios with different sampling rates. It can estimate the best sliding window from the data and configure the parameters (i.e., adjust the ML model) accordingly, i.e. in a self-configurable way. Based on the configured parameters according to the method, the ML system can provide short as well as long time horizon predictions.

[0268] Same of further embodiment, can adopt a proactive approach by using traffic prediction from the configured ML system (i.e., the configured parameters, which represent trained ML models), and can make early decisions (e.g. compared to reactive approach). This results in actual throughput gain from controlling the wireless access network, e.g. carrier aggregation (CA) activation and reduces UE power consumption from late CA deactivation.

[0269] Same or further embodiment do not require that the ML system (which is performing the ML model according to the configured parameters) can operate at an execution frequency of MAC schedulers like in prior art, since it runs on the automatically identified time slots (e.g., 300 milliseconds). It is feasible to deploy in real RAN deployment using today's technology. Reference signs refer to features and steps as described in their context or may refer to the following features and steps.

[0270] 100 First network entity, e.g. a network node such as a gNB

[0271] 102 Sequences obtaining module

[0272] 104 Machine learning module

[0273] 106 Controlling Module

[0274] 110 Machine learning system (ML system)

[0275] 200 Second network entity

[0276] 201 Configuration receiving module

[0277] 202 Sequence obtaining module

[0278] 206 Controlling module

[0279] 300 Method of configuring a ML system controlling a wireless access network

[0280] 302 Step of obtaining measurement values

[0281] 304 Step of performing ML

[0282] 304-1 Substep of evaluating performance

[0283] 304-2 Substep of configuring the ML system

[0284] 306 Step of controlling the wireless access network

[0285] 400 Method of controlling a wireless access network

[0286] 401 Step of receiving configuration

[0287] 402 Step of obtaining measurement values

[0288] 406 Step of controlling the wireless access network

[0289] 500 Wireless access network

[0290] 504 Wireless device

[0291] 800 Wireless data traffic

[0292] 802 Data packet or metadata about packet-oriented data traffic (packet data)

[0293] 1200 Measurement time slot (TS)

[0294] 1300 Measurement values (MV)

[0295] 1302 Set of parameters, e.g., ML model

[0296] 1400 Sequence of measurements values

[0297] 1402 History time window (W)

[0298] 1600 Centralized Unit (CU) embodiment of the first network entity, e.g., gNB-CU

[0299] 1602 Interface of the CU, e.g. an Fl interface

[0300] 1604 Processing circuitry of the CU

[0301] 1606 Memory of the CU

[0302] 1700 Distributed Unit, e.g., gNB-DU

[0303] 1702 Interface of the DU, e.g. an Fl interface 1704 Processing circuitry of the DU

[0304] 1706 Memory of the DU

[0305] Furthermore, abbreviations have the technical meaning defined where the abbreviation is initially introduced or may have the following explanation.

[0306] Abbreviation Explanation

[0307] NR New radio

[0308] LTE Long Term Evolution gNB-CU gNB Centralized Unit gNB-DU gNB Distributed Unit

[0309] UE User equipment

[0310] CA Carrier aggregation

[0311] RS Reference signal

[0312] CSI Channel State Information

[0313] SSB Synchronization Signal Block

[0314] RSRP Reference Signal Received Power

[0315] RSRQ Reference Signal Received Quality RSSI Received Signal Strength Indicator SRS Sounding Reference Signal RLC Radio Link Control

[0316] UE User Equipment

[0317] SCC Secondary Component Carrier

[0318] SCell Secondary Cell

[0319] 3GPP Third Generation Partnership Project

[0320] RL Reinforcement Learning ML Machine Learning

[0321] CRAN Centralized Radio Access Network

[0322] DRAN Distributed Radio Access Network

[0323] Many advantages of the present invention will be fully understood from the foregoing description, and it will be apparent that various changes may be made in the form, construction and arrangement of the units and devices without departing from the scope of the invention and / or without sacrificing all of its advantages. Since the invention can be varied in many ways, it will be recognized that the invention should be limited only by the scope of the following claims.

Claims

Claims1. A method (300) of configuring parameters of a machine learning system, ML system (110), controlling (306; 406) a wireless access network (500), the method (300) comprising: obtaining (302) at least two different sequences (1400) of measurement values (1300) indicative of wireless data traffic (800) between the wireless access network (500) and one or more wireless devices (504), wherein the different sequences (1400) are obtained (302) using different time domain processing of the wireless data traffic (800); and performing (304) machine learning, ML, of the ML system (110) based on each of the at least two sequences (1400) resulting in at least two sets (1302) of parameter values, wherein the ML system (110) is configured (304-2) with one of the at least two sets (1302) based on a performance metric evaluated (304-1) for the ML system (110).

2. The method (300) of claim 1, wherein the measurement values (1300) are indicative of at least one of: one or more traffic characteristics of the wireless data traffic (800); one or more network behaviors of the wireless access network (500); a network bandwidth usage by the wireless data traffic (800); a relative bandwidth usage of one or more active component carriers by the wireless data traffic (800); a relative bandwidth usage, by the wireless data traffic (800), of a master cell group, MCG (502), and / or one or more secondary cell groups, SCGs (502), of the wireless access network (500); a data rate of the wireless data traffic (800) or a data volume per time of the wireless data traffic (800); a number of data packets (802) in the wireless data traffic (800); a size of data packets (802) in the wireless data traffic (800); a distribution of time intervals between subsequent data packets (802) in the wireless data traffic (800); source internet protocol addresses and / or target internet protocol addresses in headers of the wireless data traffic (800); applications underlying the wireless data traffic (800); andclassifications based on deep packet inspection of the wireless data traffic(800).

3. The method (300) of claim 1 or 2, wherein the obtaining (302) of the at least two different sequences (1400) of measurement values (1300) using the different time domain processing of the wireless data traffic (800) comprises at least one of: sampling the wireless data traffic (800) at different sampling rates for the at least two different sequences (1400) of measurement values (1300); and sampling the wireless data traffic (800) for at least one of the at least two different sequences (1400) of measurement values (1300) according to an irregular, optionally aperiodic, pattern of sample values of the wireless data traffic (800).

4. The method (300) of any one of claims 1 to 3, wherein each of the measurement values (1300) is associated with a measurement time slot (1200) and / or wherein each of the measurement values (1300) is indicative of an aggregation of one or more sample values of the wireless data traffic (800) within a measurement time slot (1200), and wherein the different time domain processing of the wireless data traffic (800) corresponds to different durations, T, of the measurement time slot (1200).

5. The method (300) of any one of claims 1 to 4, wherein measurement time slots (1200) of different measurement values (1300) are consecutive in the time domain; and / or wherein the measurement time slots (1200) of different measurement values (1300) are regular, optionally repeating periodically, in the time domain.

6. The method (300) of any one of claims 1 to 4, wherein measurement time slots (1200) of different measurement values (1300) are non-consecutive in the time domain; and / or wherein the measurement time slots (1200) of different measurement values (1300) are event-triggered, optionally triggered by the availability for transmission or by the transmission of a data packet (802) in the wireless data traffic (800).

7. The method (300) of any one of claims 1 to 6, wherein a length of each sequence (1400) of measurement values (1300) corresponds to a history time window (1402), and wherein the different time domain processing of the wireless data traffic (800) corresponds to different durations, W, of the history time window (1402) and / or wherein a duration, W, of the history time window (1402) is determined through cross-validation of the measurement values.

8. The method (300) of any one of claims 1 to 7, wherein the ML system (110) outputs a prediction for the wireless data traffic (800), the wireless access network (500) being controlled based on the prediction for the wireless data traffic (800), and / or wherein the ML system (110) outputs a control indicator, the wireless access network (500) being controlled based on the control indicator.

9. The method (300) of any one of claims 1 to 8, wherein the controlling (306; 406) of the wireless access network (500) by the ML system (110) comprises at least one of: radio resource management, RRM, of the wireless access network (500); selectively activating or deactivating a wireless carrier of the wireless access network (500); selectively activating or deactivating a component carrier of the wireless access network (500) for carrier aggregation, CA, of the one or more wireless devices (504); selectively activating or deactivating a wireless connection to a secondary cell for dual connectivity of the one or more wireless devices (504); changing a modulation and coding scheme, MCS, for the one or more wireless devices (504); load balancing the wireless data traffic (800) and / or the one or more wireless devices (504) across available network resources of the wireless access network (500); and mobility management of the one or more wireless devices (504).

10. The method (300) of any one of claims 1 to 9, wherein the at least two different sequences (1400) of measurement values (1300) are obtained per wireless device (504), and the ML of the ML system is performed per wireless device (504), and the controlling (306; 406) of the wireless access network (500)comprises performing radio resource management, RRM, per wireless device (504) or activating or deactivating a component carrier for the wireless device (504).

11. The method (300) of any one of claims 1 to 10, wherein the measurement values (1300) indicative of the wireless data traffic (800), or sample values underlying the measurement values (1300), are received from or obtained (302) at a centralized unit, CU, of a network node of the wireless access network (500); and / or wherein an output of the ML system (110) for the controlling (306; 406) of the wireless access network (500) is sent to a distributed unit, DU, of a network node of the wireless access network (500).

12. The method (300) of any one of claims 1 to 11, wherein a time duration, W, of each sequence (1400) of measurement values (1300) used for performing (304) the ML is less than a time lag between real-time and a time-stamp of the last measurement value (1300) in the sequence (1400) of measurement values (1300) used for performing (304) the ML.

13. A method (400) of controlling (406) a wireless access network (500), the method (400) comprising: receiving (401) a configuration message indicative of parameters of a machine learning system, ML system (110), and a time domain processing of wireless data traffic (800); obtaining (402) a sequence (1400) of measurement values (1300) indicative of wireless data traffic (800) between the wireless access network (500) and one or more wireless devices (504), wherein the sequence (1400) is obtained (402) using the received (401) time domain processing of the wireless data traffic (800); and controlling (406) the wireless access network (500) using the ML system (110) configured with the received (401) parameters, wherein the obtained (402) sequence (1400) is applied to the ML system (110).

14. The method (400) of claim 13, further comprising the feature or step of any one of claims 2 to 12, or further comprising a feature or step corresponding to any one of claims 2 to 12 according to a transmitter-receiver relationship.

15. A network entity (100; 1600) for configuring parameters of a machine learning system, ML system (110), controlling a wireless access network (500), thenetwork entity (100; 1600) comprising memory (1606) operable to store instructions and processing circuitry (1604) operable to execute the instructions, such that the network entity (100; 1600) is operable to: obtain at least two different sequences (1400) of measurement values (1300) indicative of wireless data traffic (800) between the wireless access network (500) and one or more wireless devices (504), wherein the different sequences (1400) are obtained using different time domain processing of the wireless data traffic (800); and perform machine learning, ML, of the ML system (110) based on each of the at least two sequences (1400) resulting in at least two sets (1302) of parameter values, wherein the ML system (110) is configured with one of the at least two sets (1302) based on a performance metric evaluated for the ML system (110).

16. The network entity (100; 1600) of claim 15, further operable to perform any one of the steps of any one of claims 2 to 12.

17. A network entity (200; 1700) for controlling a wireless access network (500), the network entity (200; 1700) comprising memory (1706) operable to store instructions and processing circuitry (1704) operable to execute the instructions, such that the network entity (200; 1700) is operable to: receive a configuration message indicative of parameters of a machine learning system, ML system (110), and a time domain processing of wireless data traffic (800); obtain a sequence (1400) of measurement values (1300) indicative of wireless data traffic (800) between the wireless access network (500) and one or more wireless devices (504), wherein the sequence (1400) is obtained using the received time domain processing of the wireless data traffic (800); and control the wireless access network (500) using the ML system (110) configured with the received parameters, wherein the obtained sequence (1400) is applied to the ML system (110).

18. The network entity (200; 1700) of claim 17, further operable to perform any one of the steps of any one of claims 13 to 14.

19. A network entity (100; 1600) for configuring parameters of a machine learning system, ML system (110), controlling a wireless access network (500), the network entity (100; 1600) being configured to:obtain at least two different sequences (1400) of measurement values (1300) indicative of wireless data traffic (800) between the wireless access network (500) and one or more wireless devices (504), wherein the different sequences (1400) are obtained using different time domain processing of the wireless data traffic (800); and perform machine learning, ML, of the ML system (110) based on each of the at least two sequences (1400) resulting in at least two sets (1302) of parameter values, wherein the ML system (110) is configured with one of the at least two sets (1302) based on a performance metric evaluated for the ML system (110).

20. The network entity (100; 1600) of claim 19, further configured to perform any one of the steps of any one of claims 2 to 12.

21. A network entity (200; 1700) for controlling a wireless access network (500), the network entity (200; 1700) being configured to: receive a configuration message indicative of parameters of a machine learning system, ML system (110), and a time domain processing of wireless data traffic (800); obtain a sequence (1400) of measurement values (1300) indicative of wireless data traffic (800) between the wireless access network (500) and one or more wireless devices (504), wherein the sequence (1400) is obtained using the received time domain processing of the wireless data traffic (800); and control the wireless access network (500) using the ML system (110) configured with the received parameters, wherein the obtained sequence (1400) is applied to the ML system (110).

22. The network entity (200; 1700) of claim 21, further configured to perform any one of the steps of any one of claims 13 to 14.

Images