Training a digital twin model

Abstract

According to an aspect, there is provided a computer-implemented method for determining values for a set of operational parameters for a part of a communication network. The method comprises performing (602) a first training process to train a digital twin model (200) to determine values of a plurality of adjustable operational parameters for the part of the communication network. The digital twin model (200) is configured to predict a performance of the part of the communication network based on input values for the set of operational parameters. A set of values for the plurality of adjustable operational parameters for the part of the communication network are defined in a first set of tensors (210), a set of values for a plurality of non-adjustable operational parameters for the part of the communication network are defined in a second set of tensors (212), a set of values for a plurality of performance parameters for the part of the communication network are defined in a third set of tensors (204), and a set of weights for the digital twin model (200) are defined in a fourth set of tensors. During the first training process (602), the first set of tensors is configured to be trainable, such that the values in the first set of tensors are varied by the first training process (602). In addition, during the first training process (602) the tensors in the second set of tensors, the third set of tensors, and the fourth set of tensors are configured to be non-trainable. Finally, during the first training process (602) the digital twin model (200) is trained to determine values of the adjustable operational parameters based on respective target values for one or more of the performance parameters.

Description

[0001] Training a digital twin model

[0002] Technical Field

[0003] This disclosure relates to a digital twin model of a communication network, and in particular to techniques for training a digital twin model.

[0004] Background

[0005] Digital twins (or digital twin models) have recently gained importance in the field of radio network optimisation based on Artificial Intelligence (Al), especially in cases where the optimal (or improved) values of parameters are obtained through an iterative approach where the agent that optimises the network needs to interact with the network several times. While the interaction with the network requires a certain level of exploration that can potentially degrade its performance during the intermediate stages of the process, the use of a digital twin (DT) that mimics the behaviour of the network prevents from this damage.

[0006] The article “Al: Enhancing customer experience in a complex 5G world” by Ericsson, A. B., published in Ericsson Mobility Report, 30 July 2021 describes how two different digital twins have successfully been used to assist Al-based algorithms to optimise a live network. One of the digital twins is based on a static network simulator that is used to pretrain a deep reinforcement learning (DRL) agent. The use of this digital twin is limited to the offline pre-training phase, and is not suitable for online training. The second digital twin is based on a big data database that stores pre-processed Call Trace Recording (CTR) data and uses behavioural models to predict the Key Performance Indicators (KPIs) associated with a cell transmission power change. This second type of digital twin can be very accurate but requires exhaustive computational and storage resources to operate.

[0007] Another approach to build a digital twin is the use of a neural network (NN). NNs require a long initialisation time, i.e. training time, but once they are trained, they are very efficient in computational time and in terms of the computational resources required to provide predictions. Designing a NN to behave as a digital twin is not always trivial, and expert engineering might be required depending on the needs of each case. To give an example, in WO 2023 / 208394 two different architectures based on NNs are proposed to build two digital twins that predict the performance of one or multiple cells simultaneously. One of these architectures is shown in Fig. 1 , and in particular Fig. 1 shows a digital twin architecture for KPI prediction based on a Graph NN (GNN) and an autoencoder. Summary

[0008] There currently exist certain challenge(s). Once a digital twin is available and ready to predict the performance of a network after a cell parameter is changed, it is possible to use any of the DRL techniques proposed in the article “Al: Enhancing customer experience in a complex 5G world” for network optimisation. However, such techniques are devised to achieve the best possible solution with the lowest number of interactions with the digital twin, which was expected to introduce a long delay to provide the network performance estimates in each iteration. However, since digital twins based on NNs provide performance estimates with very low delay, it is possible to explore a higher number of solutions when searching for the optimal (or satisfactory) one. On the other hand, DRL agents require many iterations to capture the necessary knowledge from the digital twin to build the optimisation policy.

[0009] Moreover, existing solutions for network optimisation based on a digital twin using DRL do not allow the user to define the network performance targets to achieve in a flexible and intuitive way, e.g. based on user-defined intents or objectives.

[0010] Certain aspects of the disclosure and their embodiments may provide solutions to these or other challenges. In particular, this disclosure proposes methods using a digital twin model (for example based on a neural network) to determine values of adjustable operational parameters that improve the network performance to a desired, or optimal, level. In some embodiments the network performance is improved / optimised at the cell level, and so the values of the adjustable operational parameters relate to operational parameters for the cell of interest. In other embodiments, the network performance is improved / optimised for a group or cluster of cells, or for the whole communication network.

[0011] The techniques described herein make use of tensors which are data structures used in ML frameworks to store and process data. Values for adjustable operational parameters for a part of the communication network are defined in a first set of tensors, values for non-adjustable operational parameters for the part of the communication network are defined in a second set of tensors, values for a plurality of performance parameters for the part of the communication network are defined in a third set of tensors, and a set of weights (e.g. coefficients) for the digital twin model are defined in a fourth set of tensors. Thus, the first and second sets of tensors comprise the inputs to the digital twin model, the third set of tensors comprise the outputs of the digital twin model, and the fourth set of tensors comprise the parameters of the digital twin model itself.

[0012] According to a first aspect, there is provided a computer-implemented method for determining values for a set of operational parameters for a part of a communication network. The method comprises performing a first training process to train a digital twin model to determine values of a plurality of adjustable operational parameters for the part of the communication network. The digital twin model is configured to predict a performance of the part of the communication network based on input values for the set of operational parameters. A set of values for the plurality of adjustable operational parameters for the part of the communication network are defined in a first set of tensors, a set of values for a plurality of non-adjustable operational parameters for the part of the communication network are defined in a second set of tensors, a set of values for a plurality of performance parameters for the part of the communication network are defined in a third set of tensors, and a set of weights for the digital twin model are defined in a fourth set of tensors. During the first training process, the first set of tensors is configured to be trainable, such that the values in the first set of tensors are varied by the first training process. In addition, during the first training process the tensors in the second set of tensors, the third set of tensors, and the fourth set of tensors are configured to be non-trainable. Finally, during the first training process the digital twin model is trained to determine values of the adjustable operational parameters based on respective target values for one or more of the performance parameters.

[0013] According to a second aspect, there is provided a computer program product comprising a computer readable medium having computer readable code embodied therein, the computer readable code being configured such that, on execution by a suitable computer or processor, the computer or processor is caused to perform the method according to the first aspect or any embodiments thereof.

[0014] According to a third aspect, there is provided an apparatus configured to perform the method according to the first aspect or any embodiments thereof.

[0015] According to a fourth aspect, there is provided an apparatus comprising a processor and a memory, said memory containing instructions executable by said processor whereby said apparatus is operative to perform the method according to the first aspect or any embodiments thereof.

[0016] Thus, the techniques described herein create an Al-based network optimiser by re-using the same digital twin model that mimics or models a communication network. In other words, the optimisation agent / recommender is the digital twin model itself. Embodiments of the techniques provide a loss function that permits the steering of the performance optimisation process towards a target defined by an intent of the user of the training process (e.g. an operator of the communication network).

[0017] Certain embodiments may provide one or more of the following technical advantage(s). The techniques provide that digital twin models, particularly neural network-based digital twin models, and ML optimisation algorithms are leveraged to find an improved or optimal network configuration in the most efficient way, in terms of computational time and quality of the obtained solutions.

[0018] The techniques provide faster optimisation executions compared to using external optimisation agents / recommenders that need exhaustive time to learn the network mechanisms underpinning the digital twin model by means of iterative interactions. With the proposed techniques, this knowledge is embedded in the optimiser from the first step(s).

[0019] The techniques enable an improved or optimal network configuration to be found that satisfies a predefined user intent. This introduces high flexibility and degrees of freedom to the user. For instance, it is possible for the user to define their intent for the training process at a cell level, at the level of a group or cluster of cells, or at a network level.

[0020] Brief Description of the Drawings

[0021] Some of the embodiments contemplated herein will now be described more fully with reference to the accompanying drawings, in which:

[0022] Fig. 1 shows a digital twin architecture for KPI prediction based on a Graph NN and an autoencoder;

[0023] Fig. 2 illustrates a digital twin model and various tensors in a model training process;

[0024] Fig. 3 illustrates a digital twin model and various tensors in an operational parameter training process;

[0025] Figs. 4(a) and 4(b) are graphs showing penalty values applied by an exemplary loss function;

[0026] Fig. 5 is a block diagram illustrating an open-radio access network (O-RAN) implementation of the techniques described herein;

[0027] Fig. 6 is a flow chart illustrating a computer-implemented method for determining values for a set of operational parameters for a part of a communication network in accordance with some embodiments;

[0028] Fig. 7 is a simplified block diagram of an apparatus according to some embodiments; and

[0029] Fig. 8 is a block diagram illustrating a virtualization environment in which functions implemented by some embodiments may be virtualized.

[0030] Detailed Description

[0031] Some of the embodiments contemplated herein will now be described more fully with reference to the accompanying drawings. Embodiments are provided by way of example to convey the scope of the subject matter to those skilled in the art.

[0032] As noted above, this disclosure proposes methods using a digital twin model (that is, for example, based on a neural network) to determine values of adjustable operational parameters that improve the performance of part of the communication network to a desired, or optimal, level. In some embodiments the part of the network for which the performance is improved / optimised is a cell, and so the values of the adjustable operational parameters relate to parameters for the cell of interest. In other embodiments, the part of the network is a group or cluster of cells, or the whole communication network.

[0033] The communication network can be any type of network architecture that enables communications with wireless devices, such as User Equipments (UEs). The communication network can have a large number of radio access nodes, such as base stations, access points, eNBs, gNBs, etc., with these radio access nodes defining one or more respective cells. The operational parameters can relate to the operation of the radio access nodes, and / or the cells defined by the radio access nodes. The communication network may operate according to one or more of the standards defined by the 3rdGeneration Partnership Project (3GPP), and for example could be a 4thGeneration (4G) network, a 5thGeneration (5G) network, a Long Term Evolution (LTE) network, a New Radio (NR) network, or a future 6thor subsequent generation network.

[0034] The design or structure of the digital twin model is outside of the scope of this disclosure, but those skilled in the art will be aware how such models can be constructed. For example, WO 2023 / 208394 describes how a digital twin model of a communication network can be constructed using neural network technology.

[0035] While a neural network is a preferred architecture for the digital twin model, other architectures can be used, provided that the model is differentiable. Differentiable means that, for any output of a model, a derivative can be computed with respect to the model inputs and model parameters. A neural network is differentiable, and another example of a differentiable model is a gradient-boosted decision tree.

[0036] The techniques described herein make use of tensors which are data structures used in ML frameworks to store and process data. Values for adjustable operational parameters for a part of the communication network are defined in a first set of tensors, values for non-adjustable operational parameters for the part of the communication network are defined in a second set of tensors, values for a plurality of performance parameters for the part of the communication network are defined in a third set of tensors, and a set of weights (e.g. coefficients) for the digital twin model are defined in a fourth set of tensors. Thus, the first and second sets of tensors comprise the inputs to the digital twin model, the third set of tensors comprise the outputs of the digital twin model, and the fourth set of tensors comprise the parameters of the digital twin model itself. A tensor can be configured as trainable or non-trainable. Values in a trainable tensor can be adjusted / tuned using model training processes. Values in a non-trainable tensor will not be modified during training (i.e. they are non-adjustable) and will keep its value constant.

[0037] The operational parameters can include parameters that are configurable / adjustable by the network operator in order to change the operation / performance of the network. Operational parameters can be set for individual cells, for a group or cluster of cells, or for the whole network.

[0038] The values of the various operational parameters for a part of the communication network define the operation of that part of the communication network (e.g. a cell or group of cells, etc.), and any of the operational parameters may have one of a number of different values for any given part.

[0039] Examples of operational parameters include: antenna tilt, antenna height, base station transmission power, uplink power control parameters such as Pzero and alpha, handover parameters for intra-frequency, inter-frequency, or inter-Radio Access Technology (RAT) mobility, cell selection and cell reselection parameters, admission control parameters, etc.

[0040] The values of various operational parameters are defined in two sets of tensors. The set of tensors includes values of the operational parameters whose values are to be configured / adjusted by the techniques described herein, and these parameters are referred to as “adjustable operational parameters”. The second set of tensors includes values of the (remaining) operational parameters whose values are not to be configured / adjusted by the operational parameter training process (these are referred to as “non-adjustable operational parameters”). In some cases a value of an operational parameter in the second set of tensors is not configurable or adjustable, for example in the case of antenna tilt where the tilt of the antenna is mechanically fixed, the antenna height, the antenna azimuth, or frequency bandwidth. In other cases a value of an operational parameter is not easily changed, for example where the transmission frequency / ies are defined by the network operator’s licence to operate the communication network. In other cases, it may be preferred for a value of an operational parameter to be left unchanged, in which case it is put into the second set of tensors. In this latter case, the network operator may prefer to use the same value (e.g. a default value) for an operational parameter in all parts (e.g. cells) of the communication network, in which case that operational parameter is to be considered as non-adjustable.

[0041] The performance parameters for the part of the communication network characterise different aspects of the performance of the part of the communication network. The performance parameters are also referred to herein as Key Performance Indicators (KPIs). The performance parameters can relate to the performance of an individual cell, a group or cluster of cells, or the whole network. Typically, but not exclusively, the performance parameters are parameters that can be affected by different values of the operational parameters. Examples of performance parameters include: data throughput, rate of packet loss, Signal to Noise Ratio (SNR), Signal to Interference plus Noise Ratio (SINR), spectral efficiency, utilisation, congestion rate, coverage rate, call setup failure rate, call drop rate, handover failure rate, etc.

[0042] Embodiments of the techniques described herein consist of two main training phases or processes, which are respectively referred to as a “model training process” and an “operational parameter training process”. Other embodiments of the techniques described herein comprise only the operational parameter training process outlined below.

[0043] The model training process aims to improve or optimise prediction weights for the digital twin model. Thus, the model training process can comprise training the digital twin model (e.g. a neural network) to model the communication network, so that the trained digital twin model is able to provide sufficiently accurate predictions of performance parameters representing the performance of part of the communication network based on input values of operational parameters for the part of the communication network. Optionally, the training of the digital twin model can also be based on input values of performance parameter values.

[0044] In the model training process, the first and second sets of tensors are configured to be “non- trainable”, and the fourth set of tensors, i.e. those that comprise the weights for the digital twin model, are configured to be “trainable”. In embodiments where values of performance parameters are also input to the digital twin model, these values can be comprised in a fifth set of tensors, and this fifth set of tensors is configured to be non-trainable during the model training process.

[0045] When a tensor is configured to be trainable, the values in the tensor can be varied by the training process, whereas when a tensor is configured to be non-trainable, the values are not varied by the training process.

[0046] Before the operational parameter training process starts, current values of the operational parameters are input to the trained digital twin model, and the output values of the performance parameters provide a “baseline configuration” for the communication network that the subsequent training process aims to improve or optimise. The current values of the operational parameters can be the values of the operational parameters that are being used in the (real) communication network (i.e. the communication network that is modelled by the digital twin). Therefore, the baseline configuration indicates the current performance of the communication network.

[0047] In the operational parameter training process, the digital twin model is trained to determine values of the adjustable operational parameters (i.e. those in the first set of tensors) for a part of the communication network based on target values for the performance parameters. This operational parameter training process is also described herein as optimising or improving the input operational parameters. As an initial step of the operational parameter training process, the tensors are configured such that the first set of tensors associated with the adjustable operational parameters are configured to be trainable by the training process, and the second set of tensors, the third set of tensors, and the fourth set of tensors, are configured to be non-trainable. If there is a fifth set of tensors comprising values of the performance parameters that are input to the digital twin model as described above, the fifth set of tensors is also configured to be non-trainable in the operational parameter training process.

[0048] In a further step of the operational parameter training process, a target loss for the training process can be defined. This loss can be defined based on a “user intent”, i.e. based on value(s) that the network operator would like to achieve. Thus, the user intent can be a list of one or more target values for one or more of the performance parameters. The values of the operational parameters will affect the values of these performance parameter(s). Target values can be defined in a relative manner with respect to the baseline values of the performance parameters, e.g. increase throughput by 10%, increase accessibility by 5%, etc. Alternatively, it is possible to define the target values in an absolute way, e.g. reach a 5thpercentile of downlink user throughput of 10 Mbps, i.e. guarantee that 95% of users have a downlink throughput higher than that value.

[0049] A loss function can be defined based on the values in the user intent. The loss can be defined in terms of a function (a “loss function”), which can penalise the training process when the value of the output performance parameters (e.g. KPIs) do not satisfy the user intent. In some embodiments, the penalty applied by the loss function may be higher when a value of an output performance parameter is worse than the original (baseline) performance parameter value.

[0050] Finally, the operational parameter training process comprises training or optimising the digital twin model using a ML training algorithm. This training results in modified / adjusted values for the operational parameters in the trainable tensors in the first set, and these operational parameter values can be implemented in, or applied to, the communication network.

[0051] In this operational parameter training process, a conventional training algorithm, for example backpropagation, can be used to find values of the input operational parameters that lead to performance parameter values that minimise, or otherwise satisfy, the loss function.

[0052] The following description of the techniques relates to a digital twin model that uses a neural network that is implemented using a tensor-based ML software library that carries out backwardforward propagation for neural network training, such as tensorflow (www.tensorflow.org). The digital twin model / neural network is for predicting values of performance parameters in a cell of a communication network.

[0053] As noted, embodiments of the techniques comprise two main processes, a model training process for determining or optimising weights for the digital twin model and an operational parameter training process for determining or optimising values of the adjustable operational parameters to meet target values for one or more performance parameters. The model training process is described further below with respect to Fig. 2, and the operational parameter training process is described further below with respect to Fig. 3.

[0054] Fig. 2 shows the model training process with the digital twin model in the form of a neural network 200 for a cell of a communication network. The inputs 202 to the neural network 200 comprise several sets of tensors, and the output of the neural network 200 comprise another set of tensors 204.

[0055] The inputs 202 comprise a first set of tensors 210 that have values of a plurality of adjustable operational parameters for the cell. The inputs 202 also comprise a second set of tensors 212 that have values of a plurality non-adjustable operational parameters for the cell. The output of the neural network 200 is the third set of tensors 204 that have values of a plurality of performance parameters (e.g. KPIs) for the cell. The weights (coefficients) of the neural network 200 are defined in a fourth set of tensors (not shown).

[0056] The optional fifth set of tensors 214 is also shown in Fig. 2, with the fifth set of tensors 214 comprising values of performance parameters that may be part of the input 202 to the neural network 200.

[0057] The values in the first set of tensors 210, the second set of tensors 212, the third set of tensors 204 and the fifth set of tensors 214 (if present) form a training dataset for the neural network 200.

[0058] As noted, the first set of tensors 210 comprise values of the adjustable operational parameter(s) for the cell that are to be optimised in the operational parameter training process. The tensors in the second set of tensors 212 comprise the values of the non-adjustable operational parameters for the cell. These non-adjustable operational parameters are labelled as CM (Configuration Management) parameters in Fig. 2. As noted above, examples of operational parameters can include antenna tilt, antenna height, base station transmission power, uplink power control parameters such as Pzero and alpha, handover parameters for intra-frequency, inter-frequency, or inter-RAT mobility, cell selection and cell reselection parameters, admission control parameters, etc. Any of these may be considered as an adjustable operational parameter or a non-adjustable operational parameter, depending on the situation, or preference of the network operator. For example, antenna height and antenna tilt might not be easily changed, and so these operational parameters are considered non-adjustable and are included in the second set of tensors 212. As another example the network operator may want to use the same admission control parameters in every cell of the network, and so this operational parameter can be considered to be non-adjustable and included in the second set of tensors 212. The model training process consists in the training of the neural network 200 that models the digital twin of the cell of the communication network. This training step can be carried out using conventional model training techniques. During the training all the tensors 210, 212, 214, and 204 associated with inputs and outputs of the neural network 200 are configured as non- trainable, meaning that they will not be modified during the training of the neural network 200. Only the fourth set of tensors that represent the weights of the neural network 200 are configured to be trainable, meaning that they will be modified during the training of the neural network 200.

[0059] The configuration of the various tensors as trainable and non-trainable can be applied to different types of neural network architecture. In some embodiments, the architecture of the neural network 200 can include or use an autoencoder, for example as described in WO 2023 / 208394. In this case the architecture attaches an input to the latent space of the autoencoder, and so the tensor associated with this input is configured as non-trainable.

[0060] The neural network 200 is trained using any suitable training algorithm, e.g. Adam (for example as described in the paper “Adam: A method for stochastic optimization” by Kingma, D. P. and Ba, J. L, ICLR 2015), and using the available training data to minimise a loss as defined by the designer of the digital twin model. This loss is typically a function of the target KPI prediction error.

[0061] Fig. 3 shows the operational parameter training process that makes use of the digital twin model (neural network) trained in the model training process. The trained neural network 300 is used for predicting values of performance parameters for a cell of the communication network. As noted above, the performance parameters can be KPIs, and can relate to the performance of an individual cell, a group or cluster of cells, or the whole network. Examples of the performance parameters include: data throughput, rate of packet loss, SNR, SINR, spectral efficiency, utilisation, congestion rate, coverage rate, call setup failure rate, call drop rate, handover failure rate, etc.

[0062] The trained neural network 300 has the same first set of tensors 210 as in the model training process, which comprise values of the plurality of adjustable operational parameters for the cell that are to be input to the trained neural network 300. The trained neural network 300 has the same second set of tensors 212 as in the model training process, which comprise values of the plurality of non-adjustable operational parameters for the cell that are to be input to the trained neural network 300. The values of the operational parameters in the first set of tensors 210 and second set of tensors 212 are the current values of the operational parameters for the cell (i.e. the values of the operational parameters that are being used in the (real) communication network). A third set of tensors 302 have values of the plurality of performance parameters (e.g. KPIs) for the cell that are output by the neural network 300. The values in the third set of tensors 302 include or represent a baseline configuration for the cell, which is predicted by the trained neural network 300 from the current values of the operational parameters in the cell. The weights (coefficients) of the trained neural network 300 are defined in the fourth set of tensors (not shown in Fig. 3). The optional fifth set of tensors 214 is also shown, with the fifth set of tensors 214 comprising values of performance parameters that may be input to the neural network 300.

[0063] As noted above, the operational parameter training process comprises three sub-steps, configuring the tensors, defining the loss based on the user intent, and optimisation of the operational parameters using a ML training algorithm.

[0064] Configuration of the tensors - Now that the neural network 300 is trained and is ready to predict the values of the operational parameters for the cell associated with any combination of values of one or more operational parameters for the cell, the trainable / non-trainable configuration of the tensors is modified to facilitate the adjustment / optimisation of the operational parameters.

[0065] In this process, the first set of tensors 210 that relate to values of the adjustable operational parameter(s) for the cell that are to be optimised in the operational parameter training process are set to be trainable, and the tensors in the second set of tensors 212 (relating to the non- adjustable operational parameters), the third set of tensors 302 and the fourth set of tensors (i.e. the weights of the neural network 300) are set to be non-trainable. Thus, the tensors that are trainable in the operational parameter training process are different to the tensors that are trainable in the model training process.

[0066] Definition of the loss based on the user intent - The user intent is the expression of declarative requirements that stem from the specific business needs of a Communication Service Provider (CSP) and its customers, as described in “Autonomous networks with multi-layer, intent-based operation” in Ericsson Technology Review, 31 August 2023, by Niembller, J et al. (available at: https: / / www.ericsson.com / en / reports-and-papers / ericsson-technology- review / articles / autonomous-networks-multi-layer-intent-based).

[0067] There are multiple ways to define an intent, e.g. using a utility function per performance parameter (KPI) that quantifies the degree of satisfaction as a function of the performance parameter (KPI) level. This disclosure provides a representative example of user intent, but it will be appreciated that other examples of user intent can be used.

[0068] User intent can be defined as a list of target values per performance parameter (per KPI). These performance parameter (KPI) targets can be defined at cell, cell cluster, cell group, or network level. The digital twin model 300 is able to predict performance parameter values (KPIs) per cell, but in other embodiments the digital twin model 300 can aggregate values to predict performance parameter values per group of cells, etc.

[0069] The user can define an intent that is usually in the form of a list of one or more performance parameters that are to be improved compared to the baseline configuration 302. The user can prioritise certain performance parameters (KPIs) and / or cells by setting more aggressive targets. It is also possible to include one or more performance parameter (KPI) targets that imply a degradation in the performance parameter, with the overall aim of getting much higher improvements for the rest of the performance parameters (KPIs).

[0070] Once all performance parameter (KPI) targets have been defined, they are used to define the loss function associated with the operational parameter training process. In embodiments of this disclosure, a loss function is proposed that only penalises for values of performance parameters (KPIs) when they do not reach the target values. In further embodiments, the loss function has two different slopes depending on whether the value of the performance parameter (KPI) is better or worse than the baseline configuration.

[0071] In the following, it is assumed that the neural network 300 is trained to predict performance parameter values for a single cell in the communication network. That is, the neural network 300 is a neural network-based digital twin that is for optimising the values of one or more operational parameters of a unique cell.

[0072] It is assumed that all performance parameter values (KPIs) are normalised in the range [0,1], where 1 is the best value for the performance parameter value, and 0 is the worst.

[0073] In this single cell case the loss is defined as: where KPItis the estimation of the i-th performance parameter (KPI) provided by the trained neural network 300, and KPI-urris the baseline value of the i-th performance parameter (KPI).

[0074] The function f implements the above scheme with two slopes and no penalty when a value of a performance parameter (KPI) exceeds the target, and can be defined as: where b, e [0,1] is used to define the penalty when the i-th performance parameter (KPI) keeps the baseline value, a, > 1 is the i-th performance parameter (KPI) target factor, and nt> 1 is the factor for an extra penalty due to degrading the i-th performance parameter (KPI), i.e. to be used when the i-th performance parameter (KPI) is below the baseline value. The i-th performance parameter (KPI) target factor aLdefines the desired gain for the i-th performance parameter (KPI) according to the intent.

[0075] The graphs in Figs. 4(a) and 4(b) show different sections of the functions f and f2respectively. In both graphs it can be seen that the penalty is higher when the performance parameter is below the baseline value (to the left of the 1 on the x-axis), and lower when the performance parameter is above the baseline value (1) and approaching the target value (cij).

[0076] The loss function in equation (1) can be extended to a multi-cell operational parameter optimisation case, by just replacing KPItand KPI-urrwith KPIt and KPI-rr, respectively, as shown in equation (3): where KPI^ is the estimation of the i-th performance parameter (KPI) of the j-th cell provided by the digital twin model 300, and KPI-rris the baseline value of the i-th performance parameter (KPI) of the j-th cell.

[0077] As noted above, it is possible to define an intent that implies target performance parameter (KPI) values at cell cluster level. This is possible by just considering “j” to be the cluster index in equation (3). It should be noted that the operational parameters could still be trained at cell level, but the loss function will use performance parameters (KPIs) aggregated at cluster level.

[0078] Optimisation using a ML training algorithm - Once the tensors are configured to be trainable or non-trainable as described above, and the loss function is defined as described above, the neural network 300 is trained again, this time based on the loss function. The training can use any suitable training algorithm, e.g. Adam algorithm, or Adaptive Moment Estimation. The tensors 210 relating to the adjustable operational parameter(s) are the only part that is trainable in this training process, and so the result of the training is adjusted or improved values of the operational parameters in the first set of tensors 210 that provide the improved performance parameter values.

[0079] The adjusted or improved operational parameter values can then be applied to the cell in the communication network to achieve the improved performance.

[0080] The techniques described herein can be implemented in a communication network that uses an Open Radio Access Network (O-RAN) architecture, as shown in Fig. 5. In particular, Fig. 5 shows an O-RAN based network 500, where the operational parameter-optimising digital twin model is implemented as a single “rApp” 502 (other rApps 504 are also shown) in a Non Real- Time Radio Access Network Intelligent Controller (Non-RT RIC) 506 that is located in a Service Management & Orchestrator (SMO) Framework 508. The rApp 502 can be used as a standalone app and / or the results can be used by other rApps.

[0081] Fig. 6 is a flow chart illustrating a computer-implemented method for determining a set of operational parameters for a part of a communication network according to various embodiments. The method can be performed by any suitable apparatus. The apparatus may perform the method in response to executing suitably formulated computer readable code. The computer readable code may be embodied or stored on a computer readable medium, such as a memory chip, optical disc, or other storage medium. The computer readable medium may be part of a computer program product.

[0082] The method in Fig. 6 shows two steps 601 , 602. Step 601 corresponds to the model training process described above, and is an optional step in the method of Fig. 6. The model training process is referred to as a “second training process” below. Step 602 corresponds to the operational parameter training process described above, and the operational parameter training process is referred to as a “first training process” below.

[0083] Thus, in step 602 a first training process (the operational parameter training process) is performed to train a digital twin model 200 to determine values of a plurality of adjustable operational parameters for the part of the communication network. The digital twin model 200 is configured to predict a performance of the part of the communication network based on input values for the set of operational parameters. A set of values for the plurality of adjustable operational parameters for the part of the communication network are defined in a first set of tensors 210, a set of values for a plurality of non-adjustable operational parameters for the part of the communication network are defined in a second set of tensors 212, a set of values for a plurality of performance parameters (e.g. KPIs) for the part of the communication network are defined in a third set of tensors 204, and a set of weights for the digital twin model 200 are defined in a fourth set of tensors. During the first training process 602, the first set of tensors is configured to be trainable, such that the values in the first set of tensors are varied by the first training process 602. In addition, during the first training process 602 the tensors in the second set of tensors, the third set of tensors, and the fourth set of tensors are configured to be non-trainable. Finally, during the first training process 602 the digital twin model 200 is trained to determine values of the adjustable operational parameters based on respective target values for one or more of the performance parameters.

[0084] The part of the communication network that the operational parameters and performance parameters relate to can be a cell, a group of cells, a cluster of cells, or all cells of the communication network.

[0085] The target values may have been defined by an operator or user of the communication network. The respective target values can be defined for a cell of the communication network, for a group of cells of the communication network, for a cluster of cells of the communication network, or for all cells of the communication network. The respective target values may be defined with respect to one or more of the values in the third set of tensors 204 prior to the first training process.

[0086] During step 602, the digital twin model 200 can be trained to determine values of the adjustable operational parameters based on a difference between the respective target values for the one or more performance parameters and respective predicted values for the one or more performance parameters.

[0087] A penalty can be defined for the training process that is to be applied in the first training process 602 if the respective predicted values for the one or more performance parameters do not meet or exceed the respective target values. In some cases, no penalty is to be applied in the first training process if the respective predicted values for the one or more performance parameters do meet or exceed the respective target values. In some cases, the value of the penalty depends on a difference between the respective predicted values for the one or more performance parameters and the respective target values. The value of the penalty may further depend on whether the respective predicted values are better or worse than one or more of the values in the second set of tensors prior to the first training process.

[0088] The first training process 602 can use a first loss function to train the digital twin model to determine values of the adjustable operational parameters.

[0089] As shown in Fig. 6, prior to the first training process in step 602, a second training process (the model training process) can be performed (step 601) to train the digital twin model to predict the performance of the part of the communication network based on input values of the set of operational parameters. During the second training process 601 , the first set of tensors, the second set of tensors and the third set of tensors are configured to be non-trainable, and the fourth set of tensors are configured to be varied by the second training process.

[0090] In the second training process the digital twin model can be trained to predict the performance of the part of the communication network to satisfy a second function based on a difference between a predicted performance of the part of the communication network and an actual performance of the part of the communication network. The second function may be a second loss function, and the digital twin model is trained to minimise the second loss function. The first set of tensors, the second set of tensors and the third set of tensors can provide a training dataset for training the digital twin model in the second training process 601.

[0091] The digital twin model may have an architecture that is differentiable, and in particular the digital twin model may be a neural network. Fig. 7 is a simplified block diagram of an apparatus 700 according to some embodiments that can be used to implement the techniques described herein. The apparatus 700 can be any type of network node, computer, server or other computing device.

[0092] The apparatus 700 comprises processing circuitry (or logic) 701. It will be appreciated that the apparatus 700 may comprise one or more virtual machines running different software and / or processes. The apparatus 700 may therefore comprise, or be implemented in or as one or more servers, switches and / or storage devices and / or may comprise cloud computing infrastructure that runs the software and / or processes.

[0093] The processing circuitry 701 controls the operation of the apparatus 700 to implement the methods described herein. The processing circuitry 701 can comprise one or more processors, processing units, multi-core processors or modules that are configured or programmed to control the apparatus 700 in the manner described herein. In particular implementations, the processing circuitry 701 can comprise a plurality of software and / or hardware modules that are each configured to perform, or are for performing, individual or multiple steps of the method described herein in relation to the apparatus 700.

[0094] The apparatus 700 may also comprise a communications interface 702. The communications interface 702 is for use in enabling communications with other nodes, apparatus, computers, servers, etc. For example, the communications interface 702 can be configured to transmit to and / or receive from other apparatus or nodes requests, acknowledgements, information, data, signals, or similar. The communication interface 702 can use any suitable communication technology.

[0095] The processing circuitry 701 may be configured to control the communications interface 702 to transmit to and / or receive from other apparatus or nodes, etc. requests, acknowledgements, information, data, signals, or similar, according to the methods described herein.

[0096] The apparatus 700 may comprise a memory 703. In some embodiments, the memory 703 can be configured to store program code that can be executed by the processing circuitry 701 to perform the method described herein in relation to the apparatus 700. Alternatively or in addition, the memory 703 can be configured to store any requests, acknowledgements, information, data, signals, or similar that are described herein. The processing circuitry 701 may be configured to control the memory 703 to store such information therein.

[0097] Fig. 8 is a block diagram illustrating a virtualization environment 800 in which functions implemented by some embodiments may be virtualized. In the present context, virtualizing means creating virtual versions of apparatuses or devices which may include virtualizing hardware platforms, storage devices and networking resources. As used herein, virtualization can be applied to any device described herein, or components thereof, and relates to an implementation in which at least a portion of the functionality is implemented as one or more virtual components. Some or all of the functions described herein may be implemented as virtual components executed by one or more virtual machines (VMs) implemented in one or more virtual environments 800 hosted by one or more of hardware nodes, such as a hardware computing device that operates as a network node, access network node, RAN node, UE, core network node, or host. Further, in embodiments in which the virtual node does not require radio connectivity (e.g. a core network node or host), then the node may be entirely virtualized. In some embodiments, the virtualization environment 800 includes components defined by the O-RAN Alliance, such as an O-Cloud environment orchestrated by a Service Management and Orchestration Framework via an 0-2 interface. Virtualization may facilitate distributed implementations of an access network node, network node, RAN node, UE, core network node, or host.

[0098] Applications 802 (which may alternatively be called software instances, virtual appliances, network functions, virtual nodes, virtual network functions, etc.) are run in the virtualization environment Q400 to implement some of the features, functions, and / or benefits of some of the embodiments disclosed herein.

[0099] Hardware 804 includes processing circuitry, memory that stores software and / or instructions executable by hardware processing circuitry, and / or other hardware devices as described herein, such as a network interface, input / output interface, and so forth. Software may be executed by the processing circuitry to instantiate one or more virtualization layers 806 (also referred to as hypervisors or virtual machine monitors (VMMs)), provide VMs 808a and 808b (one or more of which may be generally referred to as VMs 808), and / or perform any of the functions, features and / or benefits described in relation with some embodiments described herein. The virtualization layer 806 may present a virtual operating platform that appears like networking hardware to the VMs 808.

[0100] The VMs 808 comprise virtual processing, virtual memory, virtual networking or interface and virtual storage, and may be run by a corresponding virtualization layer 806. Different embodiments of the instance of a virtual appliance 802 may be implemented on one or more of VMs 808, and the implementations may be made in different ways. Virtualization of the hardware is in some contexts referred to as network function virtualization (NFV). NFV may be used to consolidate many network equipment types onto industry standard high volume server hardware, physical switches, and physical storage, which can be located in data centers, and customer premise equipment.

[0101] In the context of NFV, a VM 808 may be a software implementation of a physical machine that runs programs as if they were executing on a physical, non-virtualized machine. Each of the VMs 808, and that part of hardware 804 that executes that VM, be it hardware dedicated to that VM and / or hardware shared by that VM with others of the VMs, forms separate virtual network elements. Still in the context of NFV, a virtual network function is responsible for handling specific network functions that run in one or more VMs 808 on top of the hardware 804 and corresponds to the application 802.

[0102] Hardware 804 may be implemented in a standalone network node with generic or specific components. Hardware 804 may implement some functions via virtualization. Alternatively, hardware 804 may be part of a larger cluster of hardware (e.g. such as in a data center or CPE) where many hardware nodes work together and are managed via management and orchestration 810, which, among others, oversees lifecycle management of applications 802. In some embodiments, hardware 804 is coupled to one or more radio units that each include one or more transmitters and one or more receivers that may be coupled to one or more antennas. Radio units may communicate directly with other hardware nodes via one or more appropriate network interfaces and may be used in combination with the virtual components to provide a virtual node with radio capabilities, such as a radio access node or a base station. In some embodiments, some signalling can be provided with the use of a control system 812 which may alternatively be used for communication between hardware nodes and radio units.

[0103] Although the computing devices described herein may include the illustrated combination of hardware components, other embodiments may comprise computing devices with different combinations of components. It is to be understood that these computing devices may comprise any suitable combination of hardware and / or software needed to perform the tasks, features, functions and methods disclosed herein. Determining, calculating, obtaining or similar operations described herein may be performed by processing circuitry, which may process information by, for example, converting the obtained information into other information, comparing the obtained information or converted information to information stored in the apparatus, and / or performing one or more operations based on the obtained information or converted information, and as a result of said processing making a determination. Moreover, while components are depicted as single boxes located within a larger box, or nested within multiple boxes, in practice, computing devices may comprise multiple different physical components that make up a single illustrated component, and functionality may be partitioned between separate components. For example, a communication interface may be configured to include any of the components described herein, and / or the functionality of the components may be partitioned between the processing circuitry and the communication interface. In another example, non-computationally intensive functions of any of such components may be implemented in software or firmware and computationally intensive functions may be implemented in hardware. In certain embodiments, some or all of the functionality described herein may be provided by processing circuitry executing instructions stored on in memory, which in certain embodiments may be a computer program product in the form of a non-transitory computer-readable storage medium. In alternative embodiments, some or all of the functionality may be provided by the processing circuitry without executing instructions stored on a separate or discrete device- readable storage medium, such as in a hard-wired manner. In any of those particular embodiments, whether executing instructions stored on a non-transitory computer-readable storage medium or not, the processing circuitry can be configured to perform the described functionality. The benefits provided by such functionality are not limited to the processing circuitry alone or to other components of the computing device, but are enjoyed by the computing device as a whole, and / or by end users and a wireless network generally.

[0104] The foregoing merely illustrates the principles of the disclosure. Various modifications and alterations to the described embodiments will be apparent to those skilled in the art in view of the teachings herein. It will thus be appreciated that those skilled in the art will be able to devise numerous systems, arrangements, and procedures that, although not explicitly shown or described herein, embody the principles of the disclosure and can be thus within the scope of the disclosure. Various exemplary embodiments can be used together with one another, as well as interchangeably therewith, as should be understood by those having ordinary skill in the art.

Claims

Claims1 . A computer-implemented method for determining values for a set of operational parameters for a part of a communication network, the method comprising: performing a first training process to train a digital twin model to determine values of a plurality of adjustable operational parameters for the part of the communication network, wherein the digital twin model is configured to predict a performance of the part of the communication network based on input values for the set of operational parameters; wherein:- a set of values for the plurality of adjustable operational parameters for the part of the communication network are defined in a first set of tensors,- a set of values for a plurality of non-adjustable operational parameters for the part of the communication network are defined in a second set of tensors,- a set of values for a plurality of performance parameters for the part of the communication network are defined in a third set of tensors,- a set of weights for the digital twin model are defined in a fourth set of tensors,- during the first training process the first set of tensors is configured to be trainable, such that the values in the first set of tensors are varied by the first training process,- during the first training process the tensors in the second set of tensors, the third set of tensors, and the fourth set of tensors are configured to be non-trainable, and- during the first training process the digital twin model is trained to determine values of the adjustable operational parameters based on respective target values for one or more of the performance parameters.

2. The method as claimed in claim 1 , wherein the respective target values are defined by an operator or user of the communication network.

3. The method as claimed in claim 1 or 2, wherein the respective target values are defined for a cell of the communication network, for a group of cells of the communication network, for a cluster of cells of the communication network, or for all cells of the communication network.

4. The method as claimed in claim 3, wherein the respective target values are defined with respect to one or more of the values in the third set of tensors prior to the first training process.

5. The method as claimed in any of claims 1-4, wherein during the first training process the digital twin model is trained to determine values of the adjustable operational parameters based on a difference between the respective target values for the one or more performance parameters and respective predicted values for the one or more performance parameters.

6. The method as claimed in claim 5, wherein a penalty is defined that is to be applied in the first training process if the respective predicted values for the one or more performance parameters do not meet or exceed the respective target values.

7. The method as claimed in claim 5 or 6, wherein no penalty is to be applied in the first training process if the respective predicted values for the one or more performance parameters do meet or exceed the respective target values.

8. The method as claimed in claim 6 or 7, wherein the value of the penalty depends on a difference between the respective predicted values for the one or more performance parameters and the respective target values.

9. The method as claimed in claim 8, wherein the value of the penalty further depends on whether the respective predicted values are better or worse than one or more of the values in the second set of tensors prior to the first training process.

10. The method as claimed in any of claims 5-9, wherein a first loss function is used during the first training process to train the digital twin model to determine values of the adjustable operational parameters.

11. The method as claimed in any of claims 1-10, wherein the method further comprises: prior to the first training process, performing a second training process to train the digital twin model to predict the performance of the part of the communication network based on input values of the set of operational parameters.

12. The method as claimed in claim 11 , wherein, during the second training process the first set of tensors, the second set of tensors and the third set of tensors are configured to be non- trainable, and the fourth set of tensors are configured to be varied by the second training process.

13. The method as claimed in claim 11 or 12, wherein during the second training process the digital twin model is trained to predict the performance of the part of the communication network to satisfy a second function based on a difference between a predicted performance of the part of the communication network and an actual performance of the part of the communication network.

14. The method as claimed in claim 13, wherein the second function is a second loss function, and the digital twin model is trained to minimise the second loss function.

15. The method as claimed in claim 13 or 14, wherein during the second training process the first set of tensors, the second set of tensors and the third set of tensors provide a training dataset for the digital twin model.

16. The method as claimed in any of claims 1-15, wherein the digital twin model has an architecture that is differentiable.

17. The method as claimed in any of claims 1-16, wherein the digital twin model is a neural network.

18. The method as claimed in any of claims 1-17, wherein the part of the communication network is a cell, a group of cells, a cluster of cells, or all cells of the communication network.

19. The method as claimed in any of claims 1-18, wherein the plurality of performance parameters are a plurality of Key Performance Indicators, KPIs.

20. A computer program product comprising a computer readable medium having computer readable code embodied therein, the computer readable code being configured such that, on execution by a suitable computer or processor, the computer or processor is caused to perform the method of any of claims 1-19.

21. An apparatus configured to perform the method according to any of claims 1-19.

22. An apparatus comprising a processor and a memory, said memory containing instructions executable by said processor whereby said apparatus is operative to perform the method according to any of claims 1-19.

Images