Configuring network nodes in a wireless network

Abstract

There is provided a method for determining a configuration of a first network node (NN). The method comprises obtaining first topology data which indicates a first topology of a plurality of interconnected NNs and determining a first sub-topology that includes the first NN. The method further comprises selecting, from a first set of sub-topologies included in the first topology, one or more sub-topologies that are most similar to the first sub-topology and obtaining configuration information which indicates configurations of one or more NNs included in said one or more selected sub-topologies. The method further comprises determining the configuration of the first NN based on the configuration information.

Description

CONFIGURING NETWORK NODES IN A WIRELESS NETWORK TECHNICAL FIELD

[0001] This disclosure relates to configuring network nodes in a wireless network (e.g., a mobile wireless network). In this disclosure, “a wireless network,” “a telecommunication network,” “a telecom network,” and “a mobile wireless network” are used interchangeably. BACKGROUND

[0002] Mobile wireless network operators around the globe are striving to provide seamless user experiences to users. As mobile wireless networks (i.e., telecom networks) continuously evolve, optimizing the networks is becoming an important process to ensure reliable services, thereby allowing to provide such seamless user experiences.

[0003] However, the diversity and complexity of technologies that power the mobile wireless networks bring a challenge to manage the networks effectively and efficiently. Such challenge is exacerbated in 5G networks and beyond, where several new software (SW) features and configuration parameters are being introduced at a higher pace. The challenge in managing the mobile wireless networks is described below in detail.

[0004] A radio access network (RAN) consists of a set of interconnected sites located in distributed geographical regions across a city where each cell in a site provides a coverage in a geographic area, often referred to as a service area or a coverage area. In each site, a user equipment (UE) or any other wireless device can communicate via RAN to one or more core networks (CN). Each service area is a geographical area where radio coverage is served by radio network node(s) such as eNB for Long Term Evolution (LTE), gNB for 5G New Radio (NR), or radio network node(s) operating in a mixed mode of LTE and 5G NR

[0005] Since a service area or coverage area can have distinctive characteristics such as built-in infrastructures and socio-demographic factors, the choice of hardware (HW) equipment and SW configurations in each radio network node is dependent on service area characteristics. Moreover, the hierarchical and spatial nature of radio network nodes in RAN is important for some of RAN functionalities (e.g., allowing a handover). Therefore, it is important to manage RAN configuration and RAN node parameters while taking in consideration various deployment scenarios and service area characteristics.

[0006] Managing a configuration of a telecom network is an intricate and time-consuming process because there are thousands of configurable attributes that are associated with HW components and SW applications, and configuring a great number of these attributes is mandatory.

[0007] Even though most of the mandatory attributes can be set to default values that are provided either by a HW / SW vendor or defined by communication service providers (CSPs) themselves, there are also a significant number of attributes that should be set based on deployment settings of 4G / 5G sites and telecommunication cells. Configurations of these attributes is largely set by network engineers with domain-knowledge and can be set in multiple stages. First, during node provisioning, configuration files needed for the configurations are initially prepared, and values of some attribute parameters may be set using the configuration files. Second, during post-integration, for a newly deployed site / cell and / or existing site(s) / cell(s) in the vicinity of the newly deployed site / cell, values of other remaining attribute parameters may be set if needed. SUMMARY

[0008] Certain challenges presently exist in managing the configuration of a mobile wireless network. For example, manually configuring a mobile wireless network is a costly and time-consuming process. Also, such manual configuration may not be optimal as it is often based on assumptions, experiences, and / or expectations of network designers.

[0009] In another example, some telecom SW vendors typically provide golden parameters that are often generated in a simulation of a controlled environment, and this simulation includes mimicking RAN configuration with the same HW, SW, and traffic replication of RAN. However, the simulation doesn’t always capture changes in network usage.

[0010] In further example, in some existing art, a trial-error process is used for finding optimal configurations of network nodes in a telecommunication network. For instance, a reinforcement learning (RL) may be repeatedly performed to find such optimal configurations. This trial-error process may provide good results, but it may take a long time to find such optimal configurations (e.g., in case of an RL, the RL must be performed until the neural network involved in the RL stabilizes), thereby increasing the OPEX required for the configuration process.

[0011] Accordingly, there is a need for a way to configure network nodes while takingnetwork topology in consideration.

[0012] Accordingly, in one aspect of some embodiments of this disclosure, there is provided a method for determining a configuration of a first network node, NN. The method comprises obtaining first topology data which indicates a first topology of a plurality of interconnected NNs; and determining a first sub-topology that includes the first NN. The method further comprises selecting, from a first set of sub-topologies included in the first topology, one or more sub-topologies that are most similar to the first sub-topology; obtaining configuration information which indicates configurations of one or more NNs included in said one or more selected sub-topologies; and determining the configuration of the first NN based on the configuration information.

[0013] In another aspect, there is provided a computer program comprising instructions which when executed by processing circuitry cause the processing circuitry to perform the method of the embodiment described above.

[0014] In another aspect, there is provided an apparatus comprising: a memory; and processing circuitry coupled to the memory, wherein the apparatus is configured to perform the method of the above embodiment.

[0015] Some embodiments of this disclosure allow reducing manual intervention in configuring context-dependent attributes of a telecommunication network. More specifically, according to some embodiments, values of context-dependent attributes of a telecommunication network are configured using a machine learning (ML) model (e.g., neural networks, auto- encoders, etc.) with a minimal manual intervention, thereby reducing lead time and operating expenditures (OPEX) required for the configuration process. Here, the ML model may be a one- shot learner, and thus it may not require retraining using a gradient descent when new training data is introduced. Here, one-shot learning may imply that the new training data may be learned by generating its vector representations (a.k.a., “embeddings” or “vectors in an embedding space”).

[0016] In some embodiments, each network node in a telecommunication network may be represented as a vector representation, which corresponds to an aggregation of itself with other neighboring network nodes where the semantics in the embedding space are consistent with the natural contexts of the network nodes in the telecommunication network. As compared to othertraditional methods (e.g., a clustering techniques) of configuring network nodes, the configuring methods according to these embodiments can capture more complex underlying spatial and temporal distributions (e.g., by performing the configuring iteratively in a time domain) of network nodes in the telecommunication network, thereby allowing to generate a more meaningful configuration that is consistent with the rest of the telecommunication network.

[0017] Some embodiments of this disclosure also allow generating configuration values of context-dependent attributes at scale, meaning that a multitude of attributes can be configured with a single generalized model.

[0018] Lastly, as briefly mentioned above, in the existing art, an RL may be used to find optimal configurations of the network nodes in a telecommunication network. But performing the RL to find the optimal configurations may take a long time, which may lead to a higher OPEX. On the contrary, in some embodiments of this disclosure, the optimal configurations of the network nodes may be transferred from the configurations of already operating network nodes. This reduces the need of domain-knowledge in configuring context-dependent attributes in the telecommunication network. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The accompanying drawings, which are incorporated herein and form part of the specification, illustrate various embodiments.

[0020] FIG. 1 shows an exemplary RAN.

[0021] FIG. 2 shows a process according to some embodiments.

[0022] FIG. 3 shows an exemplary topology graph.

[0023] FIG. 4 shows exemplary subgraphs.

[0024] FIG. 5 shows an exemplary subgraph.

[0025] FIG. 6 shows exemplary subgraphs.

[0026] FIG. 7A shows a process of training an ML model according to some embodiments.

[0027] FIG. 7B shows a process of training an ML model according to some embodiments.

[0028] FIG. 8 shows a process according to some embodiments.

[0029] FIG. 9 shows an apparatus according to some embodiments. DETAILED DESCRIPTION

[0030] FIG. 1 shows an exemplary RAN 100. In the RAN 100, managed objects (MOs) 104-122 are configured to provide wireless services to user equipment (UEs) 102. Even though FIG.1 shows that each of the MOs 104-122 is a radio unit (RU), in some embodiments, the MOs may be any other network node (NN). For example, in some embodiments, each of the MOs 104-122 may be one or more logical cells. In this disclosure, MOs and NNs are used interchangeably. Note that the numbers of the entities shown in FIG. 1 (e.g., the MOs 104-122 and / or the UEs 102) are provided for simple explanation purposes, and do not limit the embodiments of this disclosure in any way. The RAN 100 may be operated in Long Term Evolution (LTE), 5G New Radio (NR), a mixed mode of LTE and 5G NR, or as components of Open RAN (O-RAN) (e.g., intelligent controller, O-RU (Radio Unit), O-DU (Distributed Unit), O-CU (Centralized Unit), etc.).

[0031] Let’s assume that a network operator wants to add a new NN 124 to the RAN 100. In order to add the new NN 124 to the RAN 100, the new NN 124 needs to be configured. FIG. 2 shows a process 200 for configuring the new NN 124 according to some embodiments. In the process 200, the new NN 124 may be configured by transferring configuration information from one or more NNs included in the already operational RAN 100. In some embodiments, the process 200 may be performed by a single entity (e.g., a 5G network data analytics function (NWDAF)). However, in other embodiments, the process 200 may be performed by multiple entities distributed in a cloud.

[0032] The process 200 may begin with step s202. The step s202 comprises obtaining RAN data from the RAN 100. In some embodiments, the RAN data may be collected continuously as a part of SW monitoring capabilities. The RAN data is data related to the NNs 104-122. More specifically, in some embodiments, the RAN data may include configuration management (CM) data which indicates one or more configurations of the NNs 104-122 and performance management (PM) data which indicates performance of the NNs 104-122.

[0033] The RAN data (e.g., the CM data included in the RAN data) may include attribute data indicating values of attribute parameters of each of the NNs 104-122. The attribute data may include a first type of attribute data (a.k.a., “X-set of attribute data”) and a second type ofattribute data (a.k.a., “Y-set of attribute data”).

[0034] As mentioned above, configuring the attribute data (i.e., setting the values of the attribute parameters) may be performed in two stages -- the cell and frequency planning stage and the node provisioning stage. The X-set of attribute data may be manually configured during the cell and frequency planning stage while the Y-set of attribute data corresponds to the remaining context-dependent attributes. The X-set of attribute data may be needed and / or used to make inference for the corresponding Y-set of attribute data.

[0035] Examples of the X-set of attribute data includes a channel bandwidth (BW), evolved universal terrestrial radio access (E-UTRA) absolute radio frequency channel number (EARFCNDL), a base station channel bandwidth (bSChannelBwDL), and an absolute radio- frequency channel number (ARFCN). Examples of the Y-set of attribute data includes pZeroNominalPusch, crsGain, and preambleInitialReceivedTargetPower.

[0036] After obtaining the RAN data in the step s202, the process 200 may proceed to step s204. The step s204 comprises extracting relations between the NNs 104-122 using the RAN data obtained in the step s202. Here each of the extracted relations may be any one or a combination of a mobility relation, a spatial relation, and / or a transport relation between two NNs.

[0037] The mobility relation between two NNs may indicate whether a handover of a wireless communication session of a UE can be performed from one of the two NNs to another one of the two NNs and / or the other way. In some embodiments, two NNs may be considered to have the mobility relation in case a handover can be performed between the two NNs.

[0038] The spatial relation between two NNs may indicate a distance between the two NNs. The distance between the two NNs may be calculated based on locations of the NNs (e.g., using GPS coordinates). In some embodiments, two NNs may be considered to have the spatial relation in case the distance between the two NNs is less than a threshold distance value.

[0039] The transport relation between two NNs may indicate a relation corresponding to a physical link between two radio nodes / sites in a network. The link may be a part of the transport layer. X2 is a logical representation of this relation.

[0040] After extracting the relations between the NNs 104-122, the process 200 may proceed to step s206. The step s206 comprises building a topology graph using the RAN data and the extracted relations. More specifically, in the step s206, the RAN 100 shown in FIG. 1(which consists of the interconnected NNs 104-122) is modeled as a graph. An example of the topology graph 300 corresponding to the RAN 100 is shown in FIG. 3.

[0041] The graph 300 modeling the RAN 100 may be represented as a graph ^ = (^, ℰ) with n nodes and d dimensional attributes ∈ ℝ^∗^. The graph may consist of vertices ^ (e.g., NNs) that are connected to edges ℰ each of which represents a relation.

[0042] Each vertex in the graph 300 may represent a configurable NN (i.e., MO) such as an eNB, a gNB, and / or a cell. For example, a vertex 304 may represent the NN 104, a vertex 306 may represent the NN 106, a vertex 308 may represent the NN 108, and so on and so forth. Each vertex may have the two sets of associated attribute data -- the X-set of attribute data and the Y-set of attribute data discussed above.

[0043] The lines (a.k.a., “edges”) between the vertices may represent the relations extracted in the step s204 -- the relations between the NNs in the RAN. As discussed above, each of the relations may be any one or a combination of the mobility relation, the spatial relation, and the transport relation.

[0044] Examples of the relations include cell relations, X2-relations, and / or relations between cells associated with the same eNB / gNB. As shown in FIG. 3, the relations between the vertices indicated in the topology graph 300 are similar to the spatial relations between the NNs 104-122 in the RAN 100 shown in FIG. 1.

[0045] After building the topology graph 300, the process 200 may proceed to step s208. The step s208 comprises generating subgraphs based on the topology graph 300. A subgraph may be represented by a configurable NN as the center vertex and the center vertex’s neighbor NNs as defined by the relations in the RAN. In the step s208, given the topology graph 300 (^ = (^, ℰ)), a set of subgraphs around each vertex may be extracted using the graphic sampling method described in https: / / arxiv.org / abs / 1706.02216, which is hereby incorporated by reference.

[0046] FIG. 4 shows subgraphs 402 and 404 that are generated based on the topology graph 300. The subgraph 402 is associated with a sub-topology of the vertices 304-312 corresponding to the NNs 104-112 and the subgraph 404 is associated with a sub-topology of the vertices 314-322 corresponding to the NNs 114-122.

[0047] In the subgraph 402, there is the center vertex 306 and its neighbor vertices 304, 308, and 312. Similarly, in the subgraph 404, there is the center vertex 320 and its neighbor vertices 314, 316, 318, and 322. Each of the subgraphs 402 and 404 may have a hyperparameterK = 1. The value of the hyperparameter may indicate a tier that the farthest neighbor vertices belong to with respect to the center vertex. Here, a tier may be defined by the number of edges between the center vertex and a non-center vertex. For example, in FIG. 4, the number of edges between the center vertex 306 and each of the remaining vertices 304, 308, and 312 is 1, and thus the non-center vertices 304, 308, and 312 are in tier 1. Thus, the subgraph 402 has the hyperparameter K=1. On the contrary, a subgraph 500 shown in FIG. 5 consists of a center vertex 502, tier 1 neighbor vertices 504, 506, and 510, and tier 2 neighbor vertices 508 and 512. Thus, the subgraph 500 has the hyperparameter K=2.

[0048] After generating the subgraphs 402 and 404 of the topology graph 300, the process 200 may proceed to step s210. The step s210 comprises, using the generated subgraphs, training an ML model for generating a vector representation of each subgraph and for determining a similarity between the subgraphs of the topology. The detailed steps of training the ML model are described below in detail. As a result of training the ML model in the step s210, the trained ML model would have the knowledge of similarities between the subgraphs.

[0049] After training the ML model, the process 200 may proceed to step s212. The step s212 comprises obtaining a new NN (e.g., 124 in FIG. 1) to be added to the RAN 100. Note that even though FIG. 2 shows that the step s212 is performed after the step s210, the step s212 can be performed after or before any of the steps shown in FIG. 2.

[0050] After obtaining the new NN, the process 200 may proceed to step s214. The step s214 comprises adding the new NN to the topology graph, thereby generating an updated topology graph. The way that the topology graph is updated is similar to the way that the topology graph is built in the step s206 (e.g., obtaining the RAN data for the new NN, determining relations between the new NN and other NNs in the topology, and rebuilding a topology graph based on the relations).

[0051] After updating the topology graph, thereby obtaining the updated topology graph, the process 200 may proceed to step s216. The step s216 comprises generating subgraphs based on the updated topology graph. This process of generating the subgraphs is similar to the step s208, and thus detailed explanation of the step s216 is omitted. In this disclosure, the subgraph that includes the new NN is called a new subgraph. In the new subgraph, the new NN corresponds to the center vertex.

[0052] Note that the steps s214 and s216 are performed in case the new NN is added tothe existing topology graph. But in some scenarios, the new NN may not be added to an existing topology graph but rather may be deployed in a new area. In such scenarios, instead of performing the steps s214 and s216, a new subgraph including a group of new NNs may be formed.

[0053] Also, in some scenarios, the existing topology graph may not need to be updated entirely in the step s214. For example, in case the new NN is added to an edge of the existing topology, the new NN may only need to be related to the NN(s) located at the edge of the existing topology. In such case, there is no need to update the entire topology graph. Only some subgraphs located at the edge of the topology graph may need to be updated. Therefore, according to some embodiments, in the step s214, the updated topology graph is generated only by updating a portion of the existing topology graph, and, in the step s216, only subgraphs corresponding to and / or related to the updated portion of the existing topology graph are generated (or re-generated).

[0054] After obtaining the new subgraph, the process 200 may proceed to step s218. The step s218 comprises selecting, from among the subgraphs generated in the step s216, ^ subgraphs that are most similar to the new subgraph. In case, only a portion of the existing topology graph is updated in the step s214, in the step s218, the ^ most similar subgraphs are selected from the updated and / or new subgraphs generated in the step s216 and other non- updated subgraphs generated in the step s208. Here, ^ may be greater than or equal to 1.

[0055] In selecting the ^ most similar subgraphs, the ML model trained in the step s210 may be used. More specifically, in some embodiments, the X-set attribute data of the new subgraph (i.e., a combination of the X-set of attribute data of the NNs included in the new subgraph) and the X-set attribute data of each of other subgraphs generated in the step s216 may be provided to the ML model, and the ML model may calculate a vector representation of the new subgraph and a vector representation of each of other subgraphs based on the X-set attribute data and the topology of the sub-graphs. Based on the difference between the vector representation of the new subgraph and the vector representation of each of other subgraphs, ^ subgraphs of which the vector representations are most similar to the vector representation of the new subgraph are identified.

[0056] As explained below, in the next step of the process 200, i.e., step s220, the new NN is configured based on configuration(s) of NNs included in the ^ most similar subgraphs.However, one or more NNs included in the ^ most similar subgraphs may not be stable. In other words, there may be NN(s) that are from unstable networks. In such case, it is desirable not to use the configuration(s) of those NN(S) to configure the new NN.

[0057] Thus, in some embodiments, the ^ most similar subgraphs may be filtered according to PM data which indicates performance of one or more NNs included in the ^ most similar subgraphs. Note that, as mentioned above, the PM data may be included in the RAN data obtained in the step s202.

[0058] More specifically, in one example, the PM data and / or the CM data included in the RAN data may be used to calculate key performance indicators (KPIs) of one or more NNs belonging to a certain sub-topology. The KPIs are listed in 3GPP TS 28.554 V.18.0.0. In this example, in case at least one of (or at least a certain number of or all of) the KPIs of said one or more NNs is not within an associated acceptable range, the subgraph corresponding to the certain sub-topology is excluded from the ^ most similar subgraphs that are used in the step s218.

[0059] In other words, as explained above, the next step of the process 200, i.e., step s220, is configuring the new NN based on configuration(s) of NN(s) included in the ^ most similar subgraphs, and in case a certain subgraph corresponding to a certain sub-topology is excluded from the ^ most similar subgraphs as explained above, configuration(s) of NN(s) included in the certain sub-graph will not be used in configuring the new NN. The step s220 is explained in detail below.

[0060] After identifying the ^ most similar subgraphs, the process 200 may proceed to the step s220. The step s220 comprises configuring the new NN based on configuration(s) of NN(s) included in the ^ most similar subgraphs.

[0061] For example, in case ^ = 1 (i.e., in case one most similar subgraph is found in the step s218), the configuration of the new NN may be set to be same as the configuration of the NN corresponding to the center vertex in the most similar subgraph. For example, in FIG. 6, let’s assume that the subgraph 606 is most similar to the new subgraph 602 that includes the new NN corresponding to the center vertex 324. In this case, the configuration of the NN corresponding to the vertex 324 may be set to be same as the configuration of the NN corresponding to the center vertex 320 in the subgraph 606 that is most similar to the new subgraph 602.

[0062] In another example, in case ^ = ^, the configuration of the new NN in the newsubgraph may be set to be an aggregation (e.g., an average, a weighted average, etc.) of the configurations of the NNs corresponding to the center vertices in the ^ closest subgraphs. For example, in FIG. 6, let’s assume that ^ = 2. Then, the configuration of the NN corresponding to the center vertex 324 in the new subgraph 602 may be set to be an aggregation of the configurations of the NNs corresponding to the center vertices 306 and 320 in the subgraphs 604 and 606 (assuming that the subgraphs 604 and 606 are most similar to the new subgraph 602).

[0063] Referring back to the step s210, according to some embodiments, a similarity between two subgraphs – first and second subgraphs – may be determined based on a difference between a vector representation of the first subgraph and a vector representation of the second subgraph. In generating a vector representation of a subgraph, a main model and a reference model may be used. The ML model in the step s210 may correspond only to the main model or may correspond to both the main model and the reference model.

[0064] In some embodiments, the main model is a self-supervised model leveraging a contrastive learning algorithm which uses information in the X-set attribute data (i.e., a combination of the X-set attribute data of the NNs included in the subgraph), the Y-set attribute data (i.e., a combination of the Y-set attribute data of the NNs included in the subgraph), and the structure of the subgraph. The main model is configured to (1) identify a similarity between vector representations of identical subgraphs while (2) identifying a dissimilarity between a vector representation of a subgraph and a vector representation of an unrelated subgraph. In some embodiments, Siamese Graph Neural Network (S-GNN) may be used as the main model.

[0065] An S-GNN may comprise two parallel neural network components and a function to calculate the similarity of the outputs of the two neural network components. In order to train the S-GNN to serve as the main model for finding a similarity between two subgraphs, training data comprising pairs of similar subgraphs and pairs of dissimilar subgraphs are needed. Therefore, according to some embodiments, the training data may be obtained as follows.

[0066] First, a subgraph is selected. Then, the selected subgraph is paired with another subgraph in the network. There are different ways of pairing the selected subgraph with another subgraph. In one example, subspace learning and random-pairing (i.e., randomly pairing the selected subgraph with another subgraph) may be used. After the pairing, the Y-set attribute data of the selected subgraph and the Y-set attribute data of said another subgraph is compared and based on the comparison of the two Y-set attribute data, a similarity score indicating whetherthe two subgraphs are similar or not is generated. These steps of selecting a subgraph, performing a subgraph pairing, and comparing the Y-set attribute data, and generating a similarity score indicating a similarity between the paired subgraphs may be performed repeatedly until all subgraphs in the topology are selected. The subgraphs that are determined to be similar may serve as positive training data while the subgraphs that are determined to be dissimilar may serve as negative training data.

[0067] In some embodiments, instead of performing the above steps to obtain the training data, subspace learning may be used to find more informative training data. The subspace learning is described in https: / / arxiv.org / abs / 2001.06543, which is hereby incorporated by reference. The subspace learning may be summarized as follows:

[0068] First, subgraphs are generated. Then, each of the generated subgraphs is provided to an encoder model, thereby generating a vector representation (a.k.a., “embeddings”) of the subgraph. After finding all embeddings of the generated subgraphs, a clustering is performed among the similar embeddings. Then for embeddings included in the same cluster, embedding pairs are randomly sampled, the randomly sampled embedding pairs’ corresponding subgraphs are identified, and the identified subgraphs are paired (e.g., in a tuple).

[0069] The training data obtained via the processes described above may be used for training the S-GNN. More specifically, the similar and dissimilar pairs in the training data may be provided to the S-GNN, and then the S-GNN may be trained to learn such similarity and dissimilarity. In other words, during the training phase, the S-GNN learns how to distinguish similar pairs and dissimilar pairs from each other.

[0070] The following is an exemplary conceptual algorithm for training the S-GNN according to some embodiments. 1. For Epoch in range of 1 through E; 1) Create a dataset comprising pairs of subgraphs. 2) Split the dataset into a training set and a validation set. 3) Generate training batches. 4) For a batch (that includes a plurality of subgraphs) in the training batches, a. Generate embeddings (a.k.a., vector representations) of two subgraphs (by providing the subgraphs to the upper and lower encoders (which are neuralnetworks) shown in FIG. 7A). b. Calculate a contrastive loss using a contrastive loss function as shown in FIG. 7A. c. Update weights of the encoders with stochastic gradient descent (SGD) and backpropagation. 5) After updating the weights of the encoders, for subgraphs in the validation set a. Generate embeddings of two subgraphs. b. Calculate a validation contrastive loss using the contrastive loss function. c. Perform the validation of the encoders. 6) Save the model (i.e., the weights of the encoders in the S-GNN – Note that since the weights of the two encoders are identical, only the weights of one of the encoders need to be saved) 2. Save the embeddings in data store.

[0071] As explained above, in training the S-GNN, a contrastive loss function may be employed. An example of the contrastive loss function with respect to FIG. 7A is provided below.

[0072] In some may compared against the reference model. In some embodiments, the reference model may be an unsupervised model that leverages the information in the X-set attribute data and the graph structure to promote a similarity between the vector representations of the identical subgraphs while (2) promoting a dissimilarity between the vector representation (i.e., the embedding) of a subgraph with the vector representation (i.e., the embedding) of an unrelated subgraph. In some embodiments, Graph Autoencoder (GAE) may be used as the reference model.

[0073] The following is an exemplary conceptual algorithm for training the GAE according to some embodiments.1. Load a dataset comprising subgraphs. 2. Split the loaded dataset into a training set and a validation set. 3. Perform neighbor sampling to create training batches of subgraphs. 4. For each Epoch in range of 1 through E, 1) For each of the training batches, a. Generate reconstructed graphs (the output of the decoder in FIG. 7B). b. Calculate a mean squared error (MSE) loss. c. Update weights of the encoder / decoder (e.g., shown in FIG. 7B) with SGD and backpropagation. 5. Save the model (i.e., save the updated weights of the encoder / decoder).

[0074] FIG. 8 shows a process 800 for determining a configuration of a first NN (e.g., 124) according to some embodiments. The process 800 may begin with step s802. The step s802 comprises obtaining first topology data which indicates a first topology (e.g., shown in FIG. 6) of a plurality of interconnected NNs. The step s806 comprises determining a first sub-topology (e.g., 602) that includes the first NN (e.g., the node 324 corresponds to the first NN 124). The step s808 comprises selecting, from a first set of sub-topologies included in the first topology, one or more sub-topologies (e.g., 606) that are most similar to the first sub-topology. The step s810 comprises obtaining configuration information which indicates configurations of one or more NNs (e.g., 320) included in said one or more selected sub-topologies. The step s812 comprises determining the configuration of the first NN based on the configuration information.

[0075] In some embodiments, the process comprises based on the first topology data, determining the first set of sub-topologies of interconnected NNs in step s804.

[0076] In some embodiments, selecting, from the first set of sub-topologies, said one or more sub-topologies that are most similar to the first sub-topology comprises: using a machine learning, ML, model, generating a vector representation of each sub-topology included in the first set; comparing the vector representation of the first sub-topology to each of the vector representations of other sub-topologies included in the first set; and based on the comparisons, selecting said one or more sub-topologies that are most similar to the first sub-topology.

[0077] In some embodiments, each sub-topology included in the first set includes a plurality of NNs, each of the plurality of NNs is configured with values of attribute parameters, and thevector representation of each sub-topology is generated using the ML model based on the values of the attribute parameters associated with each sub-topology.

[0078] In some embodiments, selecting, from the first set of sub-topologies, said one or more sub-topologies that are most similar to the first sub-topology comprises: determining a difference between the vector representation of the first sub-topology and each of the vector representations of other sub-topologies included in the first set; from the determined differences, selecting ^ lowest differences; and selecting said one or more sub-topologies associated with the ^ lowest differences.

[0079] In some embodiments, the first topology data indicates relations among the plurality of interconnected NNs included in the first topology, and the first set of sub-topologies is determined based on the relations among the plurality of interconnected NNs included in the first topology.

[0080] In some embodiments, each of the relations is any one or a combination of a mobility relation, a spatial relation, and / or a transport relation.

[0081] In some embodiments, the mobility relation indicates whether a handover of a wireless communication session of a user equipment, UE, can be performed between NNs.

[0082] In some embodiments, the spatial relation indicates how close NNs are from each other.

[0083] In some embodiments, each sub-topology in the first set comprises a center NN and one or more groups of interconnected NNs, said one or more groups of interconnected NNs comprise a first group of interconnected NNs, and in each sub-topology, a relation between the center NN and any other NN included in the first group is provided.

[0084] In some embodiments, said one or more groups of interconnected NNs comprise a second group of interconnected NNs, in each sub-topology, a relation between at least one NN included in the first group and at least one NN included in the second group is provided, and in each sub-topology, no relation between the center NN and any NN included in the second group is provided.

[0085] In some embodiments, in case said one or more selected sub-topologies includes only one sub-topology, the configuration of the first NN is set to be same as a configuration of a center NN included in the only sub-topology that is selected, and in case said one or more selected sub- topologies includes two or more sub-topologies, the configuration of the first NN is set based ona combination of configurations of center NNs included in said two or more sub-topologies that are selected.

[0086] In some embodiments, the process comprises (i) obtaining second topology data which indicates a second topology of a plurality of interconnected NNs, wherein the second topology does not include the first NN; (ii) adding the first NN to the second topology; (iii) determining relations between the first NN and one or more NNs included in the second topology, thereby generating the first topology; (iv) based on the relations determined in step (iii) above, identifying one or more NNs to be grouped with the first NN; and (v) grouping the identified one or more NNs with the first NN, thereby determining the first sub-topology.

[0087] In some embodiments, the ML model comprises Siamese graph neural network, GNN, the GNN comprises identical graph deep neutral networks, and the ML is trained to determine a similarity between sub-topologies based on a difference between vector representations of the sub-topologies.

[0088] In some embodiments, the ML model is trained by training data, the training data comprises a first type of attribute data (X-set) associated with NNs included in a set of sub- topologies, a second type of attribute data (Y-set) associated with the NNs included in the set of sub-topologies, and similarity data which indicates similarities between the sub-topologies included in the set, the first and second types are different, and the similarity data is obtained by: comparing the second type of attribute data of one sub-topology included in the set of sub- topologies to the second type of attribute data of another sub-topology included in the set of sub- topologies; and based on the comparison, determining a similarity between said one sub-topology and said another sub-topology.

[0089] In some embodiments, the first and second types of attribute data are extracted from configuration management, CM, data which indicates one or more configurations of the NNs included in the set of sub-topologies, the first type of attribute data is manually configured, and the second type of attribute data is context-dependent.

[0090] In some embodiments, the first topology indicated by the obtained first topology data comprises an initial set of sub-topologies, and the process further comprises: obtaining performance management, PM, data which indicates performance of one or more NNs included in the initial set; and filtering the initial set of sub-topologies based on the PM data, thereby obtaining the first set of sub-topologies.

[0091] In some embodiments, said one or more NNs included in the initial set belong to a certain sub-topology, the PM data comprises key performance indicators (KPIs) of said one or more NNs belonging to the certain sub-topology, and filtering the initial set of sub-topologies based on the PM data comprises: determining that each of the KPIs of said one or more NNs belonging to the certain topology is not within an associated acceptable range; and determining not to include the certain sub-topology in the first set of sub-topologies.

[0092] In some embodiments, the plurality of interconnected NNs is included in an open- radio access network, O-RAN.

[0093] FIG. 9 is a block diagram of an entity (e.g., a controller, a network node, etc.) 900, according to some embodiments. The entity 900 may be capable of performing the process 800 shown in FIG. 8. As shown in FIG. 9, the network node 900 may comprise: processing circuitry (PC) 902, which may include one or more processors (P) 955 (e.g., one or more general purpose microprocessors and / or one or more other processors, such as an application specific integrated circuit (ASIC), field-programmable gate arrays (FPGAs), and the like), which processors may be co-located in a single housing or in a single data center or may be geographically distributed (e.g., network node 900 may be a distributed computing apparatus comprising two or more computers or a monolithic computing apparatus consisting of a single computer); at least one network interface 948 (e.g., a physical interface or air interface) comprising a transmitter (Tx) 945 and a receiver (Rx) 947 for enabling network node 900 to transmit data to and receive data from other nodes connected to network 110 (e.g., an Internet Protocol (IP) network) to which network interface 948 is connected (physically or wirelessly) (e.g., network interface 948 may be coupled to an antenna arrangement comprising one or more antennas for enabling network node 900 to wirelessly transmit / receive data); and a storage unit (a.k.a., “data storage system”) 908, which may include one or more non-volatile storage devices and / or one or more volatile storage devices. In embodiments where PC 902 includes a programmable processor, a computer readable storage medium (CRSM) 942 may be provided. CRSM 942 may store a computer program (CP) 943 comprising computer readable instructions (CRI) 944. CRSM 942 may be a non-transitory computer readable medium, such as, magnetic media (e.g., a hard disk), optical media, memory devices (e.g., random access memory, flash memory), and the like. In some embodiments, the CRI 944 of computer program 943 is configured such that when executed by PC 902, the CRI causes network node 900 to perform steps described herein (e.g., steps described herein with reference tothe flow charts). In other embodiments, network node 900 may be configured to perform steps described herein without the need for code. That is, for example, PC 902 may consist merely of one or more ASICs. Hence, the features of the embodiments described herein may be implemented in hardware and / or software.

[0094] Summary of Embodiments A1. A method (800) for determining a configuration of a first network node, NN (e.g., 124), the method comprising: obtaining (s802) first topology data which indicates a first topology of a plurality of interconnected NNs; determining (s806) a first sub-topology that includes the first NN; selecting (s808), from a first set of sub-topologies included in the first topology, one or more sub-topologies that are most similar to the first sub-topology; obtaining (s810) configuration information which indicates configurations of one or more NNs included in said one or more selected sub-topologies; and determining (812) the configuration of the first NN based on the configuration information. A1a. The method of embodiment A1, the method comprising: based on the first topology data, determining (s804) the first set of sub-topologies of interconnected NNs. A2. The method of embodiment A1 or A1a, wherein selecting, from the first set of sub- topologies, said one or more sub-topologies that are most similar to the first sub-topology comprises: using a machine learning, ML, model, generating a vector representation of each sub- topology included in the first set; comparing the vector representation of the first sub-topology to each of the vector representations of other sub-topologies included in the first set; and based on the comparisons, selecting said one or more sub-topologies that are most similar to the first sub-topology.A2a. The method of embodiment A2, wherein each sub-topology included in the first set includes a plurality of NNs, each of the plurality of NNs is configured with values of attribute parameters, and the vector representation of each sub-topology is generated using the ML model based on the values of the attribute parameters associated with each sub-topology. A2b. The method of embodiment A2 or A2a, wherein selecting, from the first set of sub- topologies, said one or more sub-topologies that are most similar to the first sub-topology comprises: determining a difference between the vector representation of the first sub-topology and each of the vector representations of other sub-topologies included in the first set; from the determined differences, selecting ^ lowest differences; and selecting said one or more sub-topologies associated with the ^ lowest differences. A3. The method of embodiment A1 or A2 wherein the first topology data indicates relations among the plurality of interconnected NNs included in the first topology, and the first set of sub-topologies is determined based on the relations among the plurality of interconnected NNs included in the first topology. A4. The method of embodiment A3, wherein each of the relations is any one or a combination of a mobility relation, a spatial relation, and / or a transport relation. A5. The method of embodiment A4, wherein the mobility relation indicates whether a handover of a wireless communication session of a user equipment, UE, can be performed between NNs. A6. The method of embodiment A4 or A5, wherein the spatial relation indicates how close NNs are from each other.A7. The method of any one of embodiments A1-A6, wherein each sub-topology in the first set comprises a center NN and one or more groups of interconnected NNs, said one or more groups of interconnected NNs comprise a first group of interconnected NNs, and in each sub-topology, a relation between the center NN and any other NN included in the first group is provided. A8. The method of embodiment A7, wherein said one or more groups of interconnected NNs comprise a second group of interconnected NNs, in each sub-topology, a relation between at least one NN included in the first group and at least one NN included in the second group is provided, and in each sub-topology, no relation between the center NN and any NN included in the second group is provided. A9. The method of any one of embodiments A1-A8, wherein in case said one or more selected sub-topologies includes only one sub-topology, the configuration of the first NN is set to be same as a configuration of a center NN included in the only sub-topology that is selected, and in case said one or more selected sub-topologies includes two or more sub-topologies, the configuration of the first NN is set based on a combination of configurations of center NNs included in said two or more sub-topologies that are selected. A10. The method of any one of embodiments A1-A9, the method comprising: (i) obtaining second topology data which indicates a second topology of a plurality of interconnected NNs, wherein the second topology does not include the first NN; (ii) adding the first NN to the second topology; (iii) determining relations between the first NN and one or more NNs included in the second topology, thereby generating the first topology;(iv) based on the relations determined in step (iii) above, identifying one or more NNs to be grouped with the first NN; and (v) grouping the identified one or more NNs with the first NN, thereby determining the first sub-topology. A11. The method of any one of embodiments A2-A10, wherein the ML model comprises Siamese graph neural network, GNN, the GNN comprises identical graph deep neutral networks, and the ML is trained to determine a similarity between sub-topologies based on a difference between vector representations of the sub-topologies. A12. The method of embodiment A11, wherein the ML model is trained by training data, the training data comprises a first type of attribute data (X-set) associated with NNs included in a set of sub-topologies, a second type of attribute data (Y-set) associated with the NNs included in the set of sub-topologies, and similarity data which indicates similarities between the sub-topologies included in the set, the first and second types are different, and the similarity data is obtained by: comparing the second type of attribute data of one sub-topology included in the set of sub-topologies to the second type of attribute data of another sub-topology included in the set of sub-topologies; and based on the comparison, determining a similarity between said one sub-topology and said another sub-topology. B1. A computer program (943) comprising instructions (944) which when executed by processing circuitry (902) cause the processing circuitry to perform the method of any one of embodiments A1-A12.B2. A carrier containing the computer program of embodiment B1, wherein the carrier is one of an electronic signal, an optical signal, a radio signal, and a computer readable storage medium. C1. An apparatus (900) for determining a configuration of a first network node, NN (e.g., 124), the apparatus being configured to: obtain (s802) first topology data which indicates a first topology of a plurality of interconnected NNs; based on the first topology data, determine (s804) a first set of sub-topologies of interconnected NNs; determine (s806) a first sub-topology that includes the first NN; select (s808), from the first set of sub-topologies, one or more sub-topologies that are most similar to the first sub-topology; obtain (s810) configuration information which indicates configurations of one or more NNs included in said one or more selected sub-topologies; and determine (s812) the configuration of the first NN based on the configuration information. C2. The apparatus of embodiment C1, wherein the apparatus is configured to perform the method of any one of embodiments A2-A12. D1. An apparatus (900), the apparatus comprising: a memory (942); and processing circuitry (902) coupled to the memory, wherein the apparatus is configured to perform the method of any one of embodiments A1-A12.

[0095] While various embodiments are described herein, it should be understood that they have been presented by way of example only, and not limitation. Thus, the breadth and scope of this disclosure should not be limited by any of the above described exemplary embodiments. Moreover, any combination of the above-described elements in all possible variations thereof is encompassed by the disclosure unless otherwise indicated herein or otherwise clearly contradictedby context.

[0096] As used herein transmitting a message “to” or “toward” an intended recipient encompasses transmitting the message directly to the intended recipient or transmitting the message indirectly to the intended recipient (i.e., one or more other nodes are used to relay the message from the source node to the intended recipient). Likewise, as used herein receiving a message “from” a sender encompasses receiving the message directly from the sender or indirectly from the sender (i.e., one or more nodes are used to relay the message from the sender to the receiving node). Further, as used herein “a” means “at least one” or “one or more.”

[0097] Additionally, while the processes described above and illustrated in the drawings are shown as a sequence of steps, this was done solely for the sake of illustration. Accordingly, it is contemplated that some steps may be added, some steps may be omitted, the order of the steps may be re-arranged, and some steps may be performed in parallel.

[0098] In some scenarios, one or more nodes (hereinafter “misconfigured node”) may behave anomalously. To identify and correct the misconfigured node, according to some embodiments, the following steps may be performed. First, an anomaly detection is performed on embeddings associated with all (if not some) subgraphs included in the topology. More specifically, in some embodiments, an anomaly score may be generated for the embeddings associated with each subgraph and a determination is made as to whether the anomaly score associated with each subgraph satisfies a condition (e.g., whether the anomaly score is greater than a certain value). In case the anomaly score satisfies the condition, the subgraph may be selected for modification. Then, the X-values of the nodes (i.e., the attributes) and the relations between the nodes in the selected subgraph may be modified (e.g., by the operator). Then an updated topology may be generated based on some subgraphs (if not all) including the selected subgraph in the same manner as described in the above embodiments. More information about the subgraph updating based on the anomaly detection is explained more in the Appendix provided below.

[0099] Exemplary Implementation

[0100] In the paragraphs below, the model architectures and the method to generate the configuration information, e.g., such as the information obtained in the step s810, are described. The configuration information may comprise one or more configuration values. A framework for configuration management is proposed which aims to learn optimal configuration parameters fromstable operational network and recommend configuration parameters in other parts of the network. The proposed approach comprises three stages.

[0101] Stage I: represent multi-dimensional RAN data as a graph.

[0102] Stage II: Generate subgraphs using the graph-sampling method. Trains Siamese Graph Neural Network (S-GNN) using RAN data from one or multiple stable operational networks. S-GNN that consists of two components, each component is an encoder that produce embeddings from subgraphs which are then used to compare similarities between different subgraphs. The training process leverages a contrastive learning algorithm that uses the subgraph structure and RAN data to (1) promote a similarity between the vector representations of identical subgraphs while (2) promoting a dissimilarity between the vector representation of a subgraph with the vector representation of an unrelated subgraph in the network. To assess the method, S-GNN Model has been evaluated against a Graph Autoencoder (GAE).

[0103] Stage III: Recommending Configuration Values. Recommend configuration parameters in other parts of the network by finding the most similar subgraph from stable operational network.

[0104] - Expansion – new cell(s) are installed on already existing Radio Nodes.

[0105] - Greenfield Deployment – new Radio Node(s) are installed in a completely new region or a region with no preexisting Radio Node(s).

[0106] - Modification – Already existing cell(s) and Radio Node(s) in the network are reconfigured to eliminate misconfigurations.

[0107] The method aims at developing a machine assisted system to recommend a set of network configuration parameters and transfer knowledge from operational networks. The method can also help in detecting misconfiguration in telecom networks.

[0108] The method obtains RAN data from an operational mobile operator network. To this end, a graph neural network generative model is trained which is able to capture the underlying hierarchical distribution of RAN data. As the data modality of interest are graph in nature, the model takes a subgraph as input and generates an embedding of the subgraph. The model is trained such that subgraphs that are similar in terms of context and configuration-setup are also close in the corresponding embedding space. Such a model can extrapolate meaningful configuration parameters in similar scenarios in the acquired data to recommend parameters.

[0109] The proposed method allows to:

[0110] 1. Process RAN topology and configuration data which is then leveraged to transfer knowledge from existing operational network and generate configuration parameters for new network Radio Nodes. The method can generate consistent network configuration values, by leveraging generative graph neural network model.

[0111] 2. Configure a multitude of node attributes for managed objects in telecom networks at the same time.

[0112] The RAN data in this disclosure may be collected from multiple eNBs and / or gNBs. The following paragraphs detail the RAN data and how the RAN data is processed.

[0113] A. Feature Values and Pre-processing

[0114] For this end, each cell in the telecom network is represented by two feature vectors in the dataset:

[0115] - x – predictor values that are defined during cell planning,

[0116] - y – configuration values that display variation across the network.

[0117] Both vectors are composed of a union of Long-Term Evolution (LTE) - and New Radio (NR) related attributes. Some of the radio node configuration attributes used in this study are described in the Table provided below Table Attributes LTE NR x Channelbandwidth bSChannelBwDL … … Earfcndl arfcnDL y pZeroNominalPusch endcUlNrLowQua lThresh … … preambleInitialReceivedTargetPo rachPreambleRec wer TargetPower

[0118] Channelbandwidth relates to the amount of data that can be transmitted over a particular communications channel in each amount of time. Earfcndl specifies the channel number for the central downlink frequency. pZeroNominalPusch specifies the target power level that the eNB receives per resource block. preambleInitialReceivedTargetPower is the initial preamble power value.

[0119] LTE and NR node configuration parameters are collected individually for each cell in a network. Then intuition comes from the fact that some Radio Nodes can be functioning on one technology (e.g., 4G, 5G) or on mixed mode. Representing LTE & NR parameters mitigatethe challenges of generalizing the model for different deployments scenarios. These values are composed in 4 datasets: where N and M denoterespectively. The datasets are then preprocessed as described in the following steps:

[0120] 1. Min-max scale values in each set to the range [0,1]

[0121] 2. Concatenate the following sets:

[0122] The set of features for LTE and NR can be different so concatenation of the sets introduce missing values such that each ^ ∈ ^ and ^ ∈ ^ have the same dimension, respectively. The missing values appear where a feature is available for LTE but not for NR, and vice versa.

[0123] 3. Impute missing values introduced by concatenation with value 0.

[0124] B. RAN Graph

[0125] RAN is represented by an undirected graph as:is the set of vertices in the graph a the telecom network andis the set of relations between cells in the network. These are composed of cell relations and inter-Radio Node relations.

[0126] A subgraph is defined as:

[0127] ^(^^)and ^^^is the set of neighbors that are randomly sampled uniformly without replacement from ^(^^). Subgraphs are leveraged as it is assumed that only the local neighborhood is relevant when configuring Radio Nodes or cells. The cardinality of ^(^^) is unbounded which can result in larger subgraphs than needed to make accurate predictions.

[0128] C. Datasets for Inductive Graph Neural Networks

[0129] GNNs are a class of machine learning models designed for learning and reasoning over graph-structured data. Unlike transductive GNNs, inductive GNNs can generalize to unseen nodes and graphs. They achieve this by leveraging shared weights and aggregating information from node neighborhoods to learn node representations. Each subgraph ^^has an associated feature set which is defined as:

[0130] Also, ^^isfor the center cell ^^for which the subgraph isaround. The dataset is then defined as:

[0131] Model

[0132] The models implemented in this study are one-shot metalearners. The main model is a S-GNN based model that is trained using self-supervision through minimization of a contrastive loss. The S-GNN is compared to a GAE model that is trained through unsupervised learning.

[0133] The S-GNN and the GAE are functions represented by GNNs that are parameterized by ^ and are defined as:is the Encoder part of the GAE and representation.is the decoder that

[0134] ^ is aof the latent space. ^^is the set of all embeddings of vertices belonging to subgraph (^^, ^^^).

[0135] A. Training S-GNN

[0136] The parameters ^ in ^!"##$are trained with objective to minimize the loss: where M is aAre the labels,on their corresponding configuration setups.

[0137] B. Training GAE

[0138] The parameters ^ in ^"'($are trained with objective to minimize the loss:

[0139] values.

[0140] C. Inference

[0141] The function that yields the configuration values for the k:th cell that is introduced to the network after training is defined as: which produce agiven ^1and 2#. / .0!3. ^1is either, ^!"##$or ^(^)*^+,$, 2#. / .0!3is the dataset that includes the N+Mtraining and the k-1 samples that were introduced after the initial training.

[0142] new cell ^#. / .0to all other existing cells. The configuration values are generated using one of two methods.

[0143] (1) Using closest semantic subgraph neighbor:

[0144] Copy the configurations from its closest neighbor in the embedding space.

[0145] ^4^5^#. / .0in the network that yields the smallest 78− 9:;; with the embedding of the new cell. Where:

[0146] ^ ^5^< is thethe set of all non-negative integers, including 0.

[0147] (2) Using Majority voting:

[0148] Aggregate the configurations from its n closest neighbors in the latent space, where K is a pre-defined parameter of neighbors. ^^^ 9#. / .0is the n:thvalues, excluding the set

[0149] This yields,and ^- ^^^ #. / .0the n:th mostThus,where aggregate is to be specified per attribute included in y.

[0150] D. Subspace Learning

[0151] The S-GNN is trained by feeding pairs of subgraphs ^^%, ^&^ to ^!"##$. In a training set composed of N+M samples, the number of possible subgraph pairs amounts to:

[0152] pairs by identifying the most informative samples. Informative pairs are composed of ambiguous samples, that is, samples that are close in the embedding space but belong to different classes. Which class the samples belong to is unknown, however, the variable >%,&indicates how likely two samples a andb belong to the same class. Given (12) above, the corresponding embeddings generated by each (^^, ^^^) is included in the set: subgraph ^^. Only theembedding associated with the center vertex ^^is included in the set ^, thus, ?^is sampled from ^^. The set ^ is then partitioned in P-sets: using the K-means set of samples, definedas:

[0154] Theis:

[0155] A&istimes a A&is sampled and paired with AB%&+B+^,%.

[0156] E. Model Architecture

[0157] The internal mechanics of ^!"##$, ^(^)*^+,$, ^D+)*^+,$, each compose multi- headed Graph Attention (GATv2Conv) and aggregates the output of each head using a FeedForward Network (FFN). The forward propagation through the networks is detailed in equations (28) to (31).where E5and F5are vectors containing learnable parameters. ℎ^His the output of thelayer MLP ^ / IJto combine the outputs from the individual heads, yielding:where O denotes the

[0158] F. Hyperparameter Tuning

[0159] The Model hyperparameters are tuned to minimize the reconstruction error of the GAE model. This process is automated using Optuna. It automates the search for optimal hyperparameters by using various optimization algorithms, such as Bayesian optimization and Treestructured Parzen Estimators (TPE). It efficiently explores the hyperparameter space of machine learning models, helping to find parameter settings that maximize performance metrics while reducing manual tuning efforts.

[0160] G. Configuration Confidence Evaluation

[0161] Anomaly detection is used to evaluate the novelty of an input-sample A^to provide the operator with a confidence value for the generated configuration values. The confidence values correspond to an anomaly score that is generated by a learnable anomaly detection model that determines the novelty based on the distribution of the embeddings in the latent space. Given the set of embeddings (25), an anomaly detector model is defined as: and is trained to detectA^, the anomaly score is yielded by:

[0162] If thethe operators have the option to reevaluate A^until the anomaly score is in the acceptable range.

Claims

CLAIMS 1. A method (800) for determining a configuration of a first network node, NN, the method comprising: obtaining (s802) first topology data which indicates a first topology of a plurality of interconnected NNs; determining (s806) a first sub-topology that includes the first NN; selecting (s808), from a first set of sub-topologies included in the first topology, one or more sub-topologies that are most similar to the first sub-topology; obtaining (s810) configuration information which indicates configurations of one or more NNs included in said one or more selected sub-topologies; and determining (812) the configuration of the first NN based on the configuration information.

2. The method of claim 1, the method comprising: based on the first topology data, determining (s804) the first set of sub-topologies of interconnected NNs.

3. The method of claim 1 or 2, wherein selecting, from the first set of sub-topologies, said one or more sub-topologies that are most similar to the first sub-topology comprises: using a machine learning, ML, model, generating a vector representation of each sub- topology included in the first set; comparing the vector representation of the first sub-topology to each of the vector representations of other sub-topologies included in the first set; and based on the comparisons, selecting said one or more sub-topologies that are most similar to the first sub-topology.

4. The method of claim 3, wherein each sub-topology included in the first set includes a plurality of NNs, each of the plurality of NNs is configured with values of attribute parameters, andthe vector representation of each sub-topology is generated using the ML model based on the values of the attribute parameters associated with each sub-topology.

5. The method of claim 3 or 4, wherein selecting, from the first set of sub-topologies, said one or more sub-topologies that are most similar to the first sub-topology comprises: determining a difference between the vector representation of the first sub-topology and each of the vector representations of other sub-topologies included in the first set; from the determined differences, selecting ^ lowest differences; and selecting said one or more sub-topologies associated with the ^ lowest differences.

6. The method of any one of claims 1-5 wherein the first topology data indicates relations among the plurality of interconnected NNs included in the first topology, and the first set of sub-topologies is determined based on the relations among the plurality of interconnected NNs included in the first topology.

7. The method of claim 6, wherein each of the relations is any one or a combination of a mobility relation, a spatial relation, and / or a transport relation.

8. The method of claim 7, wherein the mobility relation indicates whether a handover of a wireless communication session of a user equipment, UE, can be performed between NNs.

9. The method of claim 7 or 8, wherein the spatial relation indicates how close NNs are from each other.

10. The method of any one of claims 1-9, wherein each sub-topology in the first set comprises a center NN and one or more groups of interconnected NNs, said one or more groups of interconnected NNs comprise a first group of interconnected NNs, andin each sub-topology, a relation between the center NN and any other NN included in the first group is provided.

11. The method of claim 10, wherein said one or more groups of interconnected NNs comprise a second group of interconnected NNs, in each sub-topology, a relation between at least one NN included in the first group and at least one NN included in the second group is provided, and in each sub-topology, no relation between the center NN and any NN included in the second group is provided.

12. The method of any one of claims 1-11, wherein in case said one or more selected sub-topologies includes only one sub-topology, the configuration of the first NN is set to be same as a configuration of a center NN included in the only sub-topology that is selected, and in case said one or more selected sub-topologies includes two or more sub-topologies, the configuration of the first NN is set based on a combination of configurations of center NNs included in said two or more sub-topologies that are selected.

13. The method of any one of claims 1-12, the method comprising: (i) obtaining second topology data which indicates a second topology of a plurality of interconnected NNs, wherein the second topology does not include the first NN; (ii) adding the first NN to the second topology; (iii) determining relations between the first NN and one or more NNs included in the second topology, thereby generating the first topology; (iv) based on the relations determined in step (iii) above, identifying one or more NNs to be grouped with the first NN; and (v) grouping the identified one or more NNs with the first NN, thereby determining the first sub-topology.

14. The method of any one of claims 3-13, whereinthe ML model comprises Siamese graph neural network, GNN, the GNN comprises identical graph deep neutral networks, and the ML is trained to determine a similarity between sub-topologies based on a difference between vector representations of the sub-topologies.

15. The method of claim 14, wherein the ML model is trained by training data, the training data comprises a first type of attribute data (X-set) associated with NNs included in a set of sub-topologies, a second type of attribute data (Y-set) associated with the NNs included in the set of sub-topologies, and similarity data which indicates similarities between the sub-topologies included in the set, the first and second types are different, and the similarity data is obtained by: comparing the second type of attribute data of one sub-topology included in the set of sub-topologies to the second type of attribute data of another sub-topology included in the set of sub-topologies; and based on the comparison, determining a similarity between said one sub-topology and said another sub-topology.

16. The method of claim 15, wherein the first and second types of attribute data are extracted from configuration management, CM, data which indicates one or more configurations of the NNs included in the set of sub- topologies, the first type of attribute data is manually configured, and the second type of attribute data is context-dependent.

17. The method of any one of claims 1-16, wherein the first topology indicated by the obtained first topology data comprises an initial set of sub-topologies, and the method further comprises:obtaining performance management, PM, data which indicates performance of one or more NNs included in the initial set; and filtering the initial set of sub-topologies based on the PM data, thereby obtaining the first set of sub-topologies.

18. The method of claim 17, wherein said one or more NNs included in the initial set belong to a certain sub-topology, the PM data comprises key performance indicators (KPIs) of said one or more NNs belonging to the certain sub-topology, and filtering the initial set of sub-topologies based on the PM data comprises: determining that each of the KPIs of said one or more NNs belonging to the certain topology is not within an associated acceptable range; and determining not to include the certain sub-topology in the first set of sub- topologies.

19. The method of any one of claims 1-18, wherein the plurality of interconnected NNs is included in an open-radio access network, O-RAN.

20. A computer program (943) comprising instructions (944) which when executed by processing circuitry (902) cause the processing circuitry to perform the method of any one of claims 1-19.

21. An apparatus (900), the apparatus comprising: a memory (942); and processing circuitry (902) coupled to the memory, wherein the apparatus is configured to perform the method of any one of claims 1-19.

Images