Reconfiguration of access network nodes in wireless communication networks

Abstract

There is provided techniques for reconfiguring an access network node in a wireless communication network. A method is performed by a control device. The method comprises identifying an access network node in the wireless communication network that requires reconfiguration. The method comprises obtaining one feature vector per each access network node in the wireless communication network. The method comprises providing the feature vectors as input to the ML model. The method comprises reconfiguring the identified access network node in accordance with the modifiable features of the feature vector of another access network node in the wireless communication network that has the digest feature similarity vector that is closest to the digest feature similarity vector of the identified access network node.

Description

[0001]RECONFIGURATION OF ACCESS NODES IN WIRELESS COMMUNICATION NETWORKS TECHNICAL FIELD Embodiments presented herein relate to a method, a control device, a computer program, and a computer program product for reconfiguring an access network node in a wireless communication network. BACKGROUND Access network nodes serve as the backbone of wireless communication networks and play a pivotal role in ensuring the quality and coverage of communication services in the wireless communication networks. However, the performance of access network nodes can be significantly affected by a myriad of factors, such as hardware malfunctions, software glitches, environmental constraints, and evolving user demands. This has necessitated the development of sophisticated technologies for identifying underperforming access network nodes and reconfiguring them to enhance their operational efficiency and service quality. The identification of access network nodes with poor performance is a complex process that could involve the analysis of various key performance indicators (KPIs). These indicators can include, but are not limited to, data throughput rates, signal strength, error rates, connection drop rates, and user traffic patterns. Traditional approaches have relied heavily on manual monitoring and threshold-based alerts to identify performance degradation. However, such approaches often fall short in providing the real-time, granular insights needed to proactively address issues before they impact the user experience. To overcome these limitations, there has been a shift towards leveraging advanced data analytics, machine learning (ML) algorithms, and artificial intelligence (AI) to intelligently analyze vast amounts of performance data. These technologies enable the dynamic identification of performance anomalies and trends that may indicate underlying issues with access network nodes. By processing and interpreting data from multiple sources, including network logs, sensor data, and user feedback, it is possible to construct a comprehensive picture of an access network node's performance landscape. However, the fact that there might be a large number of factors that could impact the performance, in combination with that different KPIs are important for different uses cases, still makes this a cumbersome task. Once poorly performing access network nodes have been identified, the next critical action is to reconfigure them to improve their performance. This reconfiguration process can encompass a wide range of actions, from adjusting transmission power levels and antenna patterns to updating software and In more complex scenarios, it may involve reengineering the network topology, optimizing resource allocation, or deploying additional infrastructure components such as small cells or relays to enhance coverage and capacity. The reconfiguration process is inherently challenging due to the intricate interdependencies within the wireless communication network and the potential for unintended consequences on overall network performance and user experience. Hence, there is still a need for how to reconfigure access network nodes with poor performance. SUMMARY An object of embodiments herein is to address the above issues. A particular object is to address the fact that there might be a large number of factors that could impact the performance of an access network node in a wireless communication network. According to a first aspect there is presented a method for reconfiguring an access network node in a wireless communication network. The method is performed by a control device. The control device implements an ML model. The method comprises identifying an access network node in the wireless communication network that, according to a performance metric, requires reconfiguration. The method comprises obtaining one feature vector per each access network node in the wireless communication network. Each feature vector comprises elements that represent non-modifiable features and elements that represent modifiable features of the access network nodes. The method comprises providing the feature vectors as input to the ML model. In the ML model, per each of the vectors, the elements that represent non-modifiable features are collected in a respective feature similarity vector and the elements that represent modifiable features are collected in a respective cell parameter vector. In the ML model, a respective digest feature similarity vector is computed for each feature similarity vector and a respective digest cell parameter vector is computed for each cell parameter vector. The method comprises reconfiguring the identified access network node in accordance with the modifiable features of the feature vector of another access network node in the wireless communication network that, according to a vector distance metric, has the digest feature similarity vector that is closest to the digest feature similarity vector of the identified access network node. According to a second aspect there is presented a control device for reconfiguring an access network node in a wireless communication network. The control device implements an ML model. The control device comprises processing circuitry. The processing circuitry is configured to cause the control device to an access network node in the wireless communication network that, according to a performance metric, requires reconfiguration. The processing circuitry is configured to cause the control device to obtain one feature vector per each access network node in the wireless communication network. Each feature vector comprises elements that represent non-modifiable features and elements that represent modifiable features of the access network nodes. The processing circuitry is configured to cause the control device to provide the feature vectors as input to the ML model. In the ML model, per each of the vectors, the elements that represent non-modifiable features are collected in a respective feature similarity vector and the elements that represent modifiable features are collected in a respective cell parameter vector. In the ML model, a respective digest feature similarity vector is computed for each feature similarity vector and a respective digest cell parameter vector is computed for each cell parameter vector. The processing circuitry is configured to cause the control device to reconfigure the identified access network node in accordance with the modifiable features of the feature vector of another access network node in the wireless communication network that, according to a vector distance metric, has the digest feature similarity vector that is closest to the digest feature similarity vector of the identified access network node. According to a third aspect there is presented a control device for reconfiguring an access network node in a wireless communication network. The control device implements an ML model. The control device comprises an identify module configured to identify an access network node in the wireless communication network that, according to a performance metric, requires reconfiguration. The control device comprises an obtain module configured to obtain one feature vector per each access network node in the wireless communication network. Each feature vector comprises elements that represent non-modifiable features and elements that represent modifiable features of the access network nodes. The control device comprises a provide module configured to provide the feature vectors as input to the ML model. In the ML model, per each of the vectors, the elements that represent non-modifiable features are collected in a respective feature similarity vector and the elements that represent modifiable features are collected in a respective cell parameter vector. In the ML model, a respective digest feature similarity vector is computed for each feature similarity vector and a respective digest cell parameter vector is computed for each cell parameter vector. The control device comprises a reconfigure module configured to reconfigure the identified access network node in accordance with the modifiable features of the feature vector of another access network node in the wireless communication network that, according to a vector distance metric, has the digest feature similarity vector that is closest to the digest feature similarity vector of the identified access network node. According to a fourth aspect there is a computer program for reconfiguring an access network node in a wireless communication network. The computer program comprises computer code which, when run on processing circuitry of a control device that implements an ML model, causes the control device to perform actions. One action comprises the control device to identify an access network node in the wireless communication network that, according to a performance metric, requires reconfiguration. One action comprises the control device to obtain one feature vector per each access network node in the wireless communication network. Each feature vector comprises elements that represent non- modifiable features and elements that represent modifiable features of the access network nodes. One action comprises the control device to provide the feature vectors as input to the ML model. In the ML model, per each of the vectors, the elements that represent non- modifiable features are collected in a respective feature similarity vector and the elements that represent modifiable features are collected in a respective cell parameter vector. In the ML model, a respective digest feature similarity vector is computed for each feature similarity vector and a respective digest cell parameter vector is computed for each cell parameter vector. One action comprises the control device to reconfigure the identified access network node in accordance with the modifiable features of the feature vector of another access network node in the wireless communication network that, according to a vector distance metric, has the digest feature similarity vector that is closest to the digest feature similarity vector of the identified access network node. According to a fifth aspect there is presented a computer program product comprising a computer program according to the fourth aspect and a computer readable storage medium on which the computer program is stored. The computer readable storage medium could be a non-transitory computer readable storage medium. Advantageously, these aspects enable identification of access network nodes with suboptimal performance and reconfiguration of such identified access network nodes to achieve enhanced operational efficiency and service quality. By enabling real-time, data-driven decision-making, these aspects ensure that the access network nodes are operating at their full potential, thereby improving the quality of service for end-users and optimizing network resources in the wireless communication network. Advantageously, these aspects provide efficient reconfiguration of low-performance access network nodes in wireless communication networks. Advantageously, by means of computing vectors, the comparisons between different vectors become computationally feasible, even if the original vectors are of considerable length. Advantageously, by means of the ML model, the digest vectors that best represent the original vectors can be found in an efficient manner, and without the use of an expert or genie. Advantageously, if the ML model finds any feature of an original vector to not be useful then this feature will be automatically discarded. Advantageously, if the ML model finds any feature of an original vector to be important, then this feature will automatically have more relevance in the digest vector. Other objectives, features and advantages of the enclosed embodiments will be apparent from the following detailed disclosure, from the attached dependent claims as well as from the drawings. Generally, all terms used in the claims are to be interpreted according to their ordinary meaning in the technical field, unless explicitly defined otherwise herein. All references to "a / an / the element, apparatus, component, means, module, action, etc." are to be interpreted openly as referring to at least one instance of the element, apparatus, component, means, module, action, etc., unless explicitly stated otherwise. The actions of any method disclosed herein do not have to be performed in the exact order disclosed, unless explicitly stated. BRIEF DESCRIPTION OF THE DRAWINGS The inventive concept is now described, by way of example, with reference to the accompanying drawings, in which: Fig. 1 is a schematic diagram illustrating a communications network according to embodiments; Fig.2 is a flowchart of methods according to embodiments; Fig.3 is a block diagram according to embodiments; Fig.4 shows a schematic graph representation according to an embodiment; Fig.5 schematically illustrates a neural network architecture of an ML model according to an embodiment; Fig.6 schematically illustrates a loss-function for an ML model according to an embodiment; Fig. 7 is a schematic diagram showing units of a control device according to an embodiment; Fig.8 is a schematic diagram showing functional modules of a control device according to an embodiment; Fig.9 is a schematic diagram showing an O-RAN architecture according to an embodiment; and Fig. 10 shows one example of a computer program product comprising computer readable storage medium according to an embodiment. DETAILED DESCRIPTION The inventive concept will now be described more fully hereinafter with reference to the accompanying drawings, in which certain embodiments of the inventive concept are shown. This inventive concept may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided by way of example so that this disclosure will be thorough and complete, and will fully convey the scope of the inventive concept to those skilled in the art. Like numbers refer to like elements throughout the description. Any action or feature illustrated by dashed lines should be regarded as optional. Fig. 1 is a schematic diagram illustrating a wireless communication network 100 where embodiments presented herein can be applied. The wireless communication network 100 comprises access network nodes 110a, 110b, 110c. The access network nodes 11a:110c are provided in clusters, denoted “Cluster A”, “Cluster B”, and “Cluster C”. Each access network node 110a:110c could be any of a radio access network node, a radio base station, a base transceiver station, a node B (NB), an evolved node (eNB), a gNB, an integrated access and backhaul (IAB) node, a transmission and reception point (TRP), or the like. As noted above, there is still a need for how to reconfigure access network nodes with poor performance. It is therefore assumed that some of the access network nodes 110a:110c might have low performance. How such access network nodes can be identified and how the low performance can be quantified will be disclosed below. For illustration, it is assumed that access network node 110a has low performance, and thus needs reconfiguration. Access network node 110a is therefore referred to as a target cell in Fig. 1. The reconfiguration is defined by set of parameter values, below referred to as a cell parameter vector. However, it could be challenging to identify which cell parameter vector the access network node 110a should be reconfigured with. According to at least some of the herein disclosed embodiments, the cell parameter vector is identified on which of the other access network nodes 110b:110c that are most similar to the access network node 110a. In Fig.1 these other access network nodes 110b are referred to as cells similar to the target cell. The embodiments disclosed herein in particular relate to techniques for reconfiguring an access network node 110a in a wireless communication network 100. In order to obtain such techniques, there is provided a control device, a method performed by the control device, a computer program product comprising code, for example in the form of a computer program, that when run on a control device, causes the control device to perform the method. Fig.2 is a flowchart illustrating embodiments of a method 200 for reconfiguring an access network node 110a in a wireless communication network 100. The method 200 is performed by a control device 700, 800 (see, Figs.7, 8). The control device 700, 800 implements and ML model 500 (see, Fig.5). The method 200 is advantageously provided as computer programs 1020 (see, Fig.10). In general terms, the method is based on (i) identifying a low-performing access network node, (ii) calculating digest vectors for this access network node as well as for at least some of the other access network nodes in the wireless communication network 100, (iii) finding the digest feature similarity vector that is closest to the digest feature similarity vector of the identified access network node 110a, and (iv) reconfiguring the identified access network node with the cell parameter vector of the access network node that has the digest feature similarity vector closest to the identified access network node 110a. Such a method is defined by actions S102- S108 below. S102: The control device 700, 800 identifies an access network node 110a in the wireless communication network 100 that, according to a performance metric, requires reconfiguration. S104: The control device 700, 800 obtains one feature vector per each access network node 110a:110c in the wireless communication network 100. Each feature vector comprises elements that represent non-modifiable features and elements that represent modifiable features of the access network nodes 110a:110c. S106: The control device 700, 800 provides the feature vectors as input to the ML model 500. In the ML model 500, per each of the feature vectors, the elements that represent non- modifiable features are collected in a respective feature similarity vector and the elements that represent modifiable features are collected in a respective cell parameter vector. Further, in the ML model 500, a respective digest feature similarity vector is computed for each feature similarity vector and a respective digest parameter vector is computed for each cell parameter vector. In this way, the digest feature similarity vectors and the digest cell parameter vectors can be derived automatically by means of the ML model. The ML model can take all the available features (as included in the feature vectors) as inputs and automatically generate one digest feature similarity vector and one digest cell parameter vector, representing embedded features of the original vectors, that can be used later for cell similarity evaluation and reconfiguration, as in action S108. Here, the digest feature similarity vectors can be used to find similar access network nodes in terms of morphology and / or environment, whereas the digest cell parameter can be used to identify access network nodes with similar (or different) parameter configurations. S108: The control device 700, 800 reconfigures the identified access network node 110a in accordance with the modifiable features of the feature vector of another access network node 110b in the wireless communication network 100. According to a vector distance metric, this other access network node 110b has the digest feature similarity vector that is closest to the digest feature similarity vector of the identified access network node 110a. Embodiments relating to further details of reconfiguring an access network node 110a in a wireless communication network 100 as performed by the control device 700, 800 will now be disclosed. Reference is here made to the block diagram 300 of Fig. 3. In this block diagram, access network nodes are represented by cells, where thus each cell represents one access network node. The block diagram can be implemented in the control device 700, 800 for feature vector generation and similarity factor calculation. Each access network node 110a:110c in the wireless communication network 100 is in the control device 700, 800 represented by a set of features, in a vector format, in this disclosure referred to as feature vectors. These feature vectors can be obtained from different types of sources, such as signals, maps, or other types of documentation. For this purpose, the block diagram 300 comprises an input block 310 configured to interface these sources. The block diagram 300 further comprises a feature calculation block 320 configured to generate the feature vectors from the input obtained by the input block and thus to perform action S104. The block diagram 300 further comprises a similarity factor calculation block 330 configured to compare the features of the different access network nodes 110a:110c to each other. For this purpose, the similarity factor calculation block is configured to implement the ML model 500 and thus to perform actions S106 and S108. In general terms, since the digest vectors a digest of the original vectors, the digest vectors can be represented by fewer elements than the original vectors. Hence, the digest vectors can be shorter than the original vectors. Particularly, in some embodiments, the digest feature similarity vectors have fewer elements than the feature similarity vectors, and the digest cell parameter vectors have fewer elements than the cell parameter vectors. In any case, as disclosed above, the digest vectors are computed from the original vectors by the ML model 500. Further aspects of this will be disclosed below. In some aspects it is ensured that the digest cell parameter vector is different for the other access network node 110b in comparison to the digest cell parameter vector of the access network node 110a that is to be reconfigured. That is, according to some embodiments, according to the vector distance metric, the other access network node 110b has a digest cell parameter vector that differs more than a threshold value from the digest cell parameter vector of the identified access network node 110a. In this way it can be guaranteed that the identified access network node 110a is not reconfigured with the same feature, or parameter, values that it is already configured with and is currently using. Further, just because the control device 700, 800 reconfigures the identified access network node 110a as in step S108, this does not exclude the possibility that the identified access network node 110a is also reconfigured according one or more further feature vectors of yet further access network nodes. In this way, the identified access network node 110a can be reconfigured for example based on the feature vectors of the three, or even five, access network nodes that have their digest feature similarity vectors closest to the digest feature similarity vector of the identified access network node 110a. As follows from the above, a vector distance metric is used to find the access network node 110b that has the digest feature similarity vector that is closest (or at least is one of the closest ones, such as one of the three closest, or even one of the five closest) to the digest feature similarity vector of the identified access network node 110a. In this respect there could be different examples of vector distance metrics. In some non-limiting examples, the vector distance metric is a Euclidean distance metric (such as the squared distance between the vector elements), a Minkowski distance metric, or a Cosine distance metric. As further follows from the above, the assessment that the identified access network node 110a requires reconfiguration is made based on a performance metric. In some aspects, the identified access network node 110a requires reconfiguration when the Figure of Merit (FoM) for the given use case is worse than some threshold value (that could be either absolute or relative). In particular, in some embodiments, that the identified access network node 110a requires reconfiguration is given by the performance metric having an FoM, that is worse than a threshold value. In this respect, the value could (i) be fixed or (ii) be defined by the FoM of another access network node. Hence according to option (ii), some other access network node, such as access network node 110b, has a better FoM than the identified access network node 110a. That is, for option (ii), the FoM of the other access network node defines the threshold to which the FoM of the target access network node (i.e., the identified access network node 110a) is compared. Here, the FoM is composed of KPIs. The threshold value could be either a fixed threshold value (and thus be absolute) or be defined by the FoM of another access network node in the wireless communication network 100 (and thus be relative). An example of the former is that the identified access network node 110a requires reconfiguration when its FoM is worse than some benchmark performance for a given use case. An example of the latter is that the identified access network node 110a requires reconfiguration when its FoM is worse than the FoM of another access network node in the wireless communication network 100 for the same use case. There could be different ways to calculate the FoM. As disclosed above, the FoM is composed of KPIs. Then, in some embodiments, the FoM can be computed as a weighted sum of the KPIs. Each of the KPIs in the sum can be weighted according to a specific weight. Here, either the weights of the weighted sum and / or which KPIs the FoM is composed could depend on the performance requirements of the identified access network node 110a (and thus be given by the use). That is, the FoM can be computed as: KPIwhere KPI are all the KPIs where each KPI is normalizedin the range [0,1], where 0 is the worst performance and 1 is the best performance, and where is the weight associated to KPI . The use case, or equally, the performance requirements,can thus be defined mathematically by means of a list of KPIs and their associated weights. For example, a use case focused on data performance will consider KPIs such as user throughput and latency, whereas a use case focused on voice performance will consider KPIs such as voice drop call rate or voice call setup success rate. The elements of the digest vectors will therefore, for the same access network node, be different depending on the use case. An example list of five KPIs and their associated weights for two given use cases is provided in Table 1. WEIGHT WEIGHT KPI (data use case) (voice use case) Downlink user 1.0 0.2 throughput Uplink user 0.8 0.1 throughput Latency 0.2 0.8Data drop call rate 0.5 0.3Voice drop call rate 0.2 1.0Voice call setup 0.2 1.0 success rate Table 1. Example KPIs and weights. In some examples, which use case, and thus the performance requirements of the identified access network node 110a, might be defined by an Operations, Administration and Maintenance (OAM) system of the wireless communication network 100. As disclosed above, the feature vectors comprise elements that represent non-modifiable features and elements that represent modifiable features of the access network nodes 110a:110c. In general terms, the feature vectors represent all available information not related to performance for the identified access network node 110a and for the access network nodes at least neighboring the identified access network node 110a. Non-limiting examples of such features (some of which are always non-modifiable, some of which are always modifiable, and some of which that could be either non-modifiable or modifiable, for example depending on the use case) are (in no particular order): antenna type, antenna height, antenna beamwidth, operation frequency, spectral bandwidth, mechanical tilt, electrical tilt, clutter, terrain elevation, timing advance, maximum transmission power, cell reference signal gain, mobility parameters (for idle and connected mode user equipment), transmission power parameters for uplink power control, physical distance between access network nodes, relative antenna orientation, carrier aggregation parameters, traffic load, traffic profile, user equipment capabilities, etc. All, or some of these features can, depending their availability, be included as input features for the ML model. In some non-limiting examples, the non-modifiable features pertain to any, or any combination of: antenna type, antenna height, antenna mechanical tilt, terrain elevation, inter-access network node distance. In some non-limiting examples, the modifiable features pertain to any, or any combination of: antenna electrical tilt, transmission power, reference signal gain, beam width, uplink power control parameters, carrier aggregation configuration. However, for other uses cases, the list of non-modifiable features and / or the list of modifiable features might include different features. For example, maximum transmission power could either be a non- modifiable feature or a modifiable feature, depending on whether it is allowed to be modified or not in the given use case, etc. That is, in some embodiments, at least some of the elements that represent non-modifiable features and at least some of the elements that represent modifiable features differ between different performance requirements of the identified access network node 110a. As follows from the above, the non-modifiable features are used for cell similarity, and the modifiable features are used for parameter configuration. The non-modifiable features and the modifiable features might be stored in graph structures, with a respective graph for all the non- modifiable features per access network node, and another respective graph for all the modifiable features per access network node. Two different graphs will thus be generated per access network node; one for the non-modifiable features and another one for the modifiable features. That is, there are two graph structures per each access nodes and thus in total twice as many graph structures as there are access network nodes. In Fig.4 is provided a schematic graph representation 400 that describes cell features and neighbor relation features for the identified access network node and its main neighboring access network nodes. In Fig.4 only the acquiring of neighbor relation features between the identified access network node and one of the neighbor access network nodes is shown and only the acquiring of cell features from two access network nodes is shown. As follows from the above, and as illustrated in the neural network architecture of the ML model 500 illustrated in Fig. 5, the elements that represent the non-modifiable features might be provided in a first graph 510a per access network node, and the elements that represent the modifiable features might be provided in a second graph 510c per access network node. In some embodiments, a Graph Neural Network (GNN) 520a, 520b is utilized in the ML model 500 for collecting the non-modifiable features from the first graphs and for collecting the modifiable features from the second graphs. In the example of Fig.5 there is one GNN 520a for the non-modifiable features and one GNN 520b for the non-modifiable features. In general terms, each GNN 520a, 520b takes as graph and converts it to a vector 530a, 530b. Here the vector 530a is the feature similarity vector and the vector 530b is the cell parameter vector. A concatenation operator (denoted “CONCAT” takes two vectors as input and generates a new vector as output by concatenating the two vectors taken as input. For example, the feature similarity vector 530a can be concatenated with non-modifiable cluster features. Further, as also illustrated in Fig.5, Multi-Layer Perceptron (MLP) 540a, 540b can be utilized in the ML model 500 for computing the digest feature similarity vectors 550a and the digest cell parameter vectors 550b. In general terms, each MLP 540a, 540b takes as an input a vector 530a, 530b and converts the vector 530a, 530b to a different vector 550a, 550b by means of a Deep Neural Network (DNN). Here, the vectors 550a, 550b that are output from the MLPs 540a, 540b have fewer elements than the vectors 530a, 530b that are input to the MLPs 540a, 540b.With further reference to Fig.5, at inference, the prediction of the KPIs is not of interest. Rather, only the hidden layers labeled representing the digest vectors are of interest. These digest vectors can then be used for the comparison in action S108 and for verifying that the other access network node 110b has a digest cell parameter vector that differs more than a threshold value from the digest cell parameter vector of the identified access network node 110a. How the neural network architecture of the ML model 500 illustrated in Fig.5 can be used for supervised training will be disclosed below. As disclosed above, the digest vectors have fewer elements than the original vectors. In this respect there could be different ways to select how many elements to include in the digest vectors. In one example, a predetermined, fixed, number of elements are included in the digest vectors. In another example, the number of elements to be included in the digest vectors is adaptively determined. In this respect, typically, if the digest vectors have enough elements, then the performance of the model (measured using some loss metric) will not be degraded compared to if all the elements of the original vectors are included in the digest vectors. However, as illustrated in the plot 600 of Fig. 6, at a certain point, a degradation can be observed. The minimum length (as represented by the vertical dashed line) of the digest vectors should thus be selected as where the loss degradation starts but still is low. How to find this minimum length of the digest vectors can be found through training, where the length of the digest vector is successively reduced, and where the performance is observed for each length of the digest vector. In particular, in some embodiments, how many elements to include per digest feature similarity vector and how many elements to include per digest cell parameter vector are determined to, during training of the ML model 500, not degrade a loss metric for the feature similarity vector and the cell parameter vector more than a threshold value. Further, a supervised learning training can to train the ML model 500. This supervised learning training can use the same neural network architecture as in Fig.5. For this purpose a concatenation of the digest feature similarity vector 550a and the digest cell parameter vector 550b per access network node is provided as input to an MLP 560 that produces one digest KPI vector 570 per combination of non-modifiable features and modifiable features. The model 500 can be trained using supervised learning techniques, in which all available cluster, cell and relationship features (modifiable or not) can be used as input, and a list of different KPIs will be used as output. In general terms, the goal of the supervised training is to tune the parameters of the neural network that makes up the ML model so that the ML model can predict the different KPIs provided in the output using all the features available in the input to the ML model 500. That is, in some embodiments, training the ML model 500 comprises tuning parameters of the GNN and / or the MLP for prediction of a respective set of key performance indicators, KPIs, for different sets of feature similarity vectors and cell parameter vectors. The digest KPI vector can then be compared to a target KPI vector per given use case, and the parameters of the ML model 500 can be tuned to yield the digest KPI vector that minimizes some loss function with respect to the target KPI vector per given use case. That is, in some embodiments, the ML model 500 is trained based on a loss function that depends on a weighted difference between the predicted KPIs and a set of target KPIs. Here, the loss function could be the same vector distance metric as disclosed above, with a FoM being a weighted sum of the KPIs, and where the weights of the weighted sum and / or which KPIs the FoM is composed of depends on the given use case. That is, in some embodiments, the weights of the loss function and / or which KPIs to include in the loss function depends on performance requirements in the wireless communication network 100. In this way, any modification of the use case will change the output KPIs and / or the KPI weights, so the training process will be affected, and the optimized neural network parameters will be different. That is, in some embodiments, the set of target KPIs is dependent on the performance requirements, wherein the ML model 500 is trained for different sets of performance, where each set of performance requirements yield a respective tuning of the parameters of the GNN and / or the MLP. Further, since the digest feature similarity vectors and the digest cell parameter vectors are nothing more than just hidden layers of the neural network model, those vectors will be different in case there is a modification in the use case definition (i.e., in the list of KPIs and / or the weights). In this way, two access network nodes could have similar such vectors for a given use case, but quite different such vectors for a different use case (where thus the KPIs and / or the weights are different). Fig.7 schematically illustrates, in terms of of structural units, the components of a control device 700 according to an embodiment. Processing circuitry 710 is provided using any combination of one or more of a suitable central processing unit (CPU), multiprocessor, microcontroller, digital signal processor (DSP), etc., capable of executing software instructions stored in a computer program product 1010 (as in Fig.10), e.g. in the form of a storage medium 730. The processing circuitry 710 may further be provided as at least one application specific integrated circuit (ASIC), or field programmable gate array (FPGA). Particularly, the processing circuitry 710 is configured to cause the control device 700 to perform a set of operations, or actions, as disclosed above. For example, the storage medium 730 may store the set of operations, and the processing circuitry 710 may be configured to retrieve the set of operations from the storage medium 730 to cause the control device 700 to perform the set of operations. The set of operations may be provided as a set of executable instructions. As disclosed above, the control device 700 implements an ML model 500. The ML model 500 might at least partly be implemented in the processing circuitry 710, possibly in combination with the storage medium 730. Thus the processing circuitry 710 is thereby arranged to execute methods as herein disclosed. The storage medium 730 may also comprise persistent storage, which, for example, can be any single one or combination of magnetic memory, optical memory, solid state memory or even remotely mounted memory. The control device 700 may further comprise a communications (comm.) interface 720 at least configured for communications with other entities, functions, nodes, and devices, such as the access network nodes 110a:110c. As such the communications interface 720 may comprise one or more transmitters and receivers, comprising analogue and digital components. The processing circuitry 710 controls the general operation of the control device 700 e.g. by sending data and control signals to the communications interface 720 and the storage medium 730, by receiving data and reports from the communications interface 720, and by retrieving data and instructions from the storage medium 730. Other components, as well as the related functionality, of the control device 700 are omitted in order not to obscure the concepts presented herein. Fig.8 schematically illustrates, in terms of a number of functional modules, the components of a control device 800 according to an embodiment. The control device 800 of Fig. 8 comprises a number of functional modules; an identify module 810 configured to perform action S102, an obtain module 820 configured to perform action S104, a provide module 830 configured to perform action S106, and a reconfigure module 840 configured to perform action S108. The control device 800 of Fig.8 may further comprise a number of optional functional modules, as represented by functional module 850. In general terms, each functional module 810:850 may in one embodiment be only in hardware and in another embodiment with the help of software, i.e., the latter embodiment having computer program instructions stored on the storage medium 730 which when run on the processing circuitry makes the control device 700 perform the corresponding actions mentioned above in conjunction with Fig 8. It should also be mentioned that even though the modules correspond to parts of a computer program, they do not need to be separate modules therein, but the way in which they are implemented in software is dependent on the programming language used. Preferably, one or more or all functional modules 810:850 may be implemented by the processing circuitry 710, possibly in cooperation with the communications interface 720 and / or the storage medium 730. The processing circuitry 710 may thus be configured to from the storage medium 730 fetch instructions as provided by a functional module 810:850 and to execute these instructions, thereby performing any actions as disclosed herein. The control device 700, 800 may be provided as a standalone device or as a part of at least one further device. For example, the control device 700, 800 may be provided in a node of the radio access network or in a node of the core network. Alternatively, functionality of the control device 700, 800 may be distributed between at least two devices, or nodes. These at least two nodes, or devices, may either be part of the same network part (such as the radio access network or the core network) or may be spread between at least two such network parts. For example, the control device 700, 800 might be part of, integrated with, or collocated with, an operations support system (OSS), or a radio application (rAPP) in a Non Real-Time Radio Access Network Intelligent Controller (Non-RT RIC). An example of the latter is provided in Fig. 9, In Fig. 9 is provided an illustration of an open radio access network (O-RAN) architecture 900 where the control device is provided as an rAPP in a Non-RTC RIC located in a Service Management and Orchestrator (SMO) framework entity that is operatively connected to an O-RAN based network. In this respect, the control device can be provided as a standalone rAPP where the functionality of the control device is accessible by other rAPPs in the Non-RTC RIC. Generally, the control device can be provided in any network node that has access to the different inputs and outputs of the ML model. In general terms, instructions that are required to be performed in real time may be performed in a device, or node, operatively closer to the cell than instructions that are not required to be performed in real time. Thus, a first portion of the instructions performed by the control device 700, 800 may be executed in a first device, and a second portion of the of the instructions performed by the control device 700, 800 may be executed in a second device; the herein disclosed embodiments are not limited to any particular number of devices on which the instructions performed by the control device 700, 800 may be executed. Hence, the methods according to the herein disclosed are suitable to be performed by a control device 700, 800 residing in a cloud computational environment. Therefore, although a single processing circuitry 710 is illustrated in Fig.7 the processing circuitry 710 may be distributed among a plurality of devices, or nodes. The same applies to the functional modules 810:850 of Fig.8 and the computer program 1020 of Fig.10. Fig. 10 shows one example of a computer program product 1010 comprising computer readable storage medium 1030. On this computer readable storage medium 1030, a computer program 1020 can be stored, which computer program 1020 can cause the processing circuitry 710 and thereto operatively coupled entities and devices, such as the communications interface 720 and the storage medium 730, to execute methods according to embodiments described herein. The computer program 1020 and / or computer program product 1010 may thus provide means for performing any actions as herein disclosed. In the example of Fig.10, the computer program product 1010 is illustrated as an optical disc, such as a CD (compact disc) or a DVD (digital versatile disc) or a Blu-Ray disc. The computer program product 1010 could also be embodied as a memory, such as a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM), or an electrically erasable programmable read-only memory (EEPROM) and more particularly as a non-volatile storage medium of a device in an external memory such as a USB (Universal Serial Bus) memory or a Flash memory, such as a compact Flash memory. Thus, while the computer program 1020 is here schematically shown as a track on the depicted optical disk, the computer program 1020 can be stored in any way which is suitable for the computer program product 1010. The inventive concept has mainly been described above with reference to a few embodiments. However, as is readily appreciated by a person skilled in the art, other embodiments than the ones disclosed above are equally possible within the scope of the inventive concept, as defined by the appended patent claims.

Claims

CLAIMS 1. A method (200) for reconfiguring an access network node (110a) in a wireless communication network (100), wherein the method is performed by a control device (700, 800), wherein the control device (700, 800) implements a machine learning, ML, model (500), and wherein the method (200) comprises: identifying (S102) an access network node (110a) in the wireless communication network (100) that, according to a performance metric, requires reconfiguration; obtaining (S104) one feature vector per each access network node (110a:110c) in the wireless communication network (100), wherein each feature vector comprises elements that represent non-modifiable features and elements that represent modifiable features of the access network nodes (110a:110c); providing (S106) the feature vectors as input to the ML model (500), wherein, in the ML model (500), per each of the feature vectors, the elements that represent non-modifiable features are collected in a respective feature similarity vector and the elements that represent modifiable features are collected in a respective cell parameter vector, and wherein, in the ML model (500), a respective digest feature similarity vector is computed for each feature similarity vector and a respective digest cell parameter vector is computed for each cell parameter vector; and reconfiguring (S108) the identified access network node (110a) in accordance with the modifiable features of the feature vector of another access network node (110b) in the wireless communication network (100) that, according to a vector distance metric, has the digest feature similarity vector that is closest to the digest feature similarity vector of the identified access network node (110a).

2. The method (200) according to claim 1, wherein said another access network node (110b), according to the vector distance metric, has a digest cell parameter vector that differs more than a threshold value from the digest cell parameter vector of the identified access network node (110a).

3. The method (200) according to claim 1 or 2, wherein the vector distance metric is a Euclidean distance metric, a Minkowski distance metric, or a Cosine distance metric.

4. The method (200) according to any preceding claim, wherein that the identified access network node (110a) requires reconfiguration is given by the performance metric having afigure of merit, FoM, that is worse than a value, where the FoM is composed of key performance indicators, KPIs, and wherein the threshold value is either a fixed threshold value or defined by a FoM of another access network node in the wireless communication network (100).

5. The method (200) according to claim 4, wherein the FoM is a weighted sum of the KPIs, and wherein weights of the weighted sum and / or which KPIs the FoM is composed of depends on performance requirements of the identified access network node (110a).

6. The method (200) according to any preceding claim, wherein at least some of the elements that represent non-modifiable features and at least some of the elements that represent modifiable features differ between different performance requirements of the identified access network node (110a).

7. The method (200) according to any preceding claim, wherein the non-modifiable features pertain to any, or any combination of: antenna type, antenna height, antenna mechanical tilt, terrain elevation, inter-access network node distance.

8. The method (200) according to any preceding claim, wherein the modifiable features pertain to any, or any combination of: antenna electrical tilt, transmission power, reference signal gain, beam width, uplink power control parameters, carrier aggregation configuration.

9. The method (200) according to any preceding claim, wherein the elements that represent the non-modifiable features are provided in a first graph per access network node, wherein the elements that represent the modifiable features are provided in a second graph per access network node, and wherein a Graph Neural Network, GNN, is utilized in the ML model (500) for collecting the non-modifiable features from the first graphs and for collecting the modifiable features from the second graphs.

10. The method (200) according to any preceding claim, wherein the digest feature similarity vectors have fewer elements than the feature similarity vectors, and the digest cell parameter vectors have fewer elements than the cell parameter vectors.

11. The method (200) according to claim 10, wherein how many elements to include per digest feature similarity vector and how many elements to include per digest cell parameter vector are determined to, during training of the ML model (500), not degrade a loss metric for the feature similarity vector and the cell parameter vector more than a threshold value.

12. The method (200) according to any claim, wherein Multi-Layer Perceptron, MLP, is utilized in the ML model (500) for computing the digest feature similarity vectors and the digest cell parameter vectors.

13. The method (200) according to any preceding claim, wherein the ML model (500) has been trained through supervised learning training.

14. The method (200) according to a combination of claims 9, 12, and 13, wherein training the ML model (500) comprises tuning parameters of the GNN and / or the MLP for prediction of a respective set of key performance indicators, KPIs, for different sets of feature similarity vectors and cell parameter vectors.

15. The method (200) according to claim 14, wherein the ML model (500) is trained based on a loss function that depends on a weighted difference between the predicted KPIs and a set of target KPIs.

16. The method (200) according to claim 15, wherein weights of the loss function and / or which KPIs to include in the loss function depends on performance requirements in the wireless communication network (100).

17. The method (200) according to claim 16, wherein the set of target KPIs is dependent on the performance requirements, wherein the ML model (500) is trained for different sets of performance, and wherein each set of performance requirements yield a respective tuning of the parameters of the GNN and / or the MLP.

18. The method (200) according to any preceding claim, wherein the control device (700, 800) is part of, integrated with, or collocated with, an operations support system, OSS, or a radio application, rAPP, in a Non Real-Time Radio Access Network Intelligent Controller, Non- RT RIC.

19. A control device (700, 800) for reconfiguring an access network node (110a) in a wireless communication network (100), the control device (700, 800) implementing a machine learning, ML, model (500), the control device (700, 800) comprising processing circuitry (710), the processing circuitry being configured to cause the control device (700, 800) to: identify an access network node (110a) in the wireless communication network (100) that, according to a performance metric, requires reconfiguration; obtain one feature vector per each access network node (110a:110c) in the wireless communication network (100), wherein each feature vector comprises elements that representnon-modifiable features and elements that modifiable features of the access network nodes (110a:110c); provide the feature vectors as input to the ML model (500), wherein, in the ML model (500), per each of the feature vectors, the elements that represent non-modifiable features are collected in a respective feature similarity vector and the elements that represent modifiable features are collected in a respective cell parameter vector, and wherein, in the ML model (500), a respective digest feature similarity vector is computed for each feature similarity vector and a respective digest cell parameter vector is computed for each cell parameter vector; and reconfigure the identified access network node (110a) in accordance with the modifiable features of the feature vector of another access network node (110b) in the wireless communication network (100) that, according to a vector distance metric, has the digest feature similarity vector that is closest to the digest feature similarity vector of the identified access network node (110a).

20. A control device (700, 800) for reconfiguring an access network node (110a) in a wireless communication network (100), the control device (700, 800) implementing a machine learning, ML, model (500), the control device (700, 800) comprising: an identify module (810) configured to identify an access network node (110a) in the wireless communication network (100) that, according to a performance metric, requires reconfiguration; an obtain module (820) configured to obtain one feature vector per each access network node (110a:110c) in the wireless communication network (100), wherein each feature vector comprises elements that represent non-modifiable features and elements that represent modifiable features of the access network nodes (110a:110c); a provide module (830) configured to provide the feature vectors as input to the ML model (500), wherein, in the ML model (500), per each of the feature vectors, the elements that represent non-modifiable features are collected in a respective feature similarity vector and the elements that represent modifiable features are collected in a respective cell parameter vector, and wherein, in the ML model (500), a respective digest feature similarity vector is computed for each feature similarity vector and a respective digest cell parameter vector is computed for each cell parameter vector; and a reconfigure module (840) configured to reconfigure the identified access network node (110a) in accordance with the modifiable features of the feature vector of another accessnetwork node (110b) in the wireless network (100) that, according to a vector distance metric, has the digest feature similarity vector that is closest to the digest feature similarity vector of the identified access network node (110a).

21. The control device (700, 800) according to claim 19 or 20, further being configured to perform the method according to any of claims 2 to 18.

22. A computer program (1020) for reconfiguring an access network node (110a) in a wireless communication network (100), the computer program comprising computer code which, when run on processing circuitry (710) of a control device (700, 800) implementing a machine learning, ML, model (500), causes the control device (700, 800) to: identify (S102) an access network node (110a) in the wireless communication network (100) that, according to a performance metric, requires reconfiguration; obtain (S104) one feature vector per each access network node (110a:110c) in the wireless communication network (100), wherein each feature vector comprises elements that represent non-modifiable features and elements that represent modifiable features of the access network nodes (110a:110c); provide (S106) the feature vectors as input to the ML model (500), wherein, in the ML model (500), per each of the feature vectors, the elements that represent non-modifiable features are collected in a respective feature similarity vector and the elements that represent modifiable features are collected in a respective cell parameter vector, and wherein, in the ML model (500), a respective digest feature similarity vector is computed for each feature similarity vector and a respective digest cell parameter vector is computed for each cell parameter vector; and reconfigure (S108) the identified access network node (110a) in accordance with the modifiable features of the feature vector of another access network node (110b) in the wireless communication network (100) that, according to a vector distance metric, has the digest feature similarity vector that is closest to the digest feature similarity vector of the identified access network node (110a).

23. A computer program product (1010) comprising a computer program (1020) according to claim 22, and a computer readable storage medium (1030) on which the computer program is stored.

Images