INFORMATION

Technical field

The present disclosure relates to wireless communication and, more particularly, to obtaining information across frequencies.

Background

In a wireless cellular network, a UE (user equipment) performs inter-frequency measurements, e.g. to gather information about cells using different frequencies than its serving cell (e.g., Reference Signal Received Power (RSRP)). Inter-frequency measurements are required for many tasks, such as inter-frequency hand over, carrier aggregation or mobility load balancing. To provide reasonable Reference Signal Received Power (RSRP) and Quality of Service (QoS), inter-frequency measurements are performed during scheduled measurement gaps during which no transmission and reception happen so that the user can switch to the other frequency.

Summary of invention

The present disclosure is directed at limiting inter-frequency measurements.

According to an aspect, provided is a communication device comprising an interface configured for acquiring a current measured value of a measurable parameter at a serving frequency and circuitry configured for determining a predicted value of the measurable parameter at a target frequency based on the current measured value, wherein the circuitry is configured for determining the predicted value as an output of a neural network upon input of the current measured value, the neural network comprising a serving frequency auto-encoder which maps the measurable parameter at the serving frequency to a latent representation and a target frequency auto-decoder which maps the latent representation to the measurable parameter at the target frequency, the neural network is trained using historical measured values of the measurable parameter at the serving frequency and at the target frequency and includes a proximity constraint enforcing proximity between a mapping of the measurable parameter at the serving frequency to the latent representation by the serving frequency autoencoder and a mapping of the measurable parameter at the target frequency to the latent representation by a target-frequency auto-encoder which maps the measurable parameter at the target frequency to the latent representation.

The communication device facilitates determining a measurable parameter at a target frequency without a need for performing measurements at the target frequency.

In some embodiments, the communication device is a base station comprising a receiver configured for receiving the measured value from a user equipment, the receiver comprising the interface.

Accordingly, if calculation of the estimated parameter is performed by the base station, a UE does not need to perform processing for the estimation of the predicted value.

For instance, the circuitry of the base station is configured for initiating at least one of a switch by the user equipment from the serving frequency to the target frequency and an extension of a serving frequency band by the target frequency.

In some embodiments, the communication device is a user equipment and comprises a receiver configured for performing a measurement to obtain the measured value, the receiver comprising the interface.

Accordingly, if calculation of the estimated parameter is performed by the user equipment, calculation processes at the base station can be designed in a less complex manner.

For instance, the circuitry of the user equipment is configured for performing at least one of a switch from the serving frequency to the target frequency and an extension of a serving frequency band by the target frequency.

For example, the measurable parameter is a signal power.

In some embodiments, the signal power is the signal power of a reference signal.

Also provided is a communication method comprising acquiring a current measured value of a measurable parameter at a serving frequency and determining a predicted value of the measurable parameter at a target frequency based on the current measured value, wherein the predicted value is determined as an output of a neural network upon input of the current measured value, the neural network comprising a serving frequency auto-encoder which maps the measurable parameter at the serving frequency to a latent representation and a target frequency auto-decoder which maps the latent representation to the measurable parameter at the target frequency, the neural network is trained using historical measured values of the measurable parameter at the serving frequency and at the target frequency and includes a proximity constraint enforcing proximity between a mapping of the measurable parameter at the serving frequency to the latent representation by the serving frequency auto-encoder and a mapping of the measurable parameter at the target frequency to the latent representation by a target-frequency auto-encoder which maps the measurable parameter at the target frequency to the latent representation.

The communication method enables determining a measurable parameter at a target frequency without a need for performing measurements at the target frequency.

In some embodiments, the communication method is to be performed by a base station and comprises receiving the measured value from a user equipment.

Accordingly, if calculation of the estimated parameter is performed by the base station, a UE does not need to perform processing for the estimation of the predicted value.

For instance, the method includes initiating at least one of a switch by the user equipment from the serving frequency to the target frequency and an extension of a serving frequency band by the target frequency.

In some embodiments, the communication method is to be performed by a user equipment and comprises performing a measurement to obtain the measured value.

Accordingly, if calculation of the estimated parameter is performed by the user equipment, calculation processes at the base station can be designed in a less complex manner.

For instance, the method includes performing at least one of a switch by the user equipment from the serving frequency to the target frequency and an extension of a serving frequency band by the target frequency.

For example, the measurable parameter is a signal power.

In some embodiments, the signal power is the signal power of a reference signal.

According to another aspect, provided is a method for training a neural network for outputting a predicted value of a measurable parameter at a target frequency based on an inputted value of the measurable parameter at a serving frequency, the method including training a serving frequency auto-encoder which maps the measurable parameter at the serving frequency to a latent representation, and training a target-frequency auto decoder which maps the latent representation to the measurable parameter at the target frequency, wherein the training of the serving frequency auto-encoder and decoder and of the target frequency auto-encoder and decoder is performed using measured values of the measurable parameter at the serving frequency and at the target frequency and is based on a proximity constraint enforcing proximity between a mapping of the measurable parameter at the serving frequency to the latent representation by the serving frequency auto-encoder and a mapping of the measurable parameter at the target frequency to the latent representation by a target-frequency autoencoder which maps the measurable parameter at the target frequency to the latent representation.

A neural network trained accordingly enables a communication device to determine a measurable parameter at a target frequency without a need for performing measurements at the target frequency.

For instance, the method includes training a serving frequency auto-decoder which maps the latent representation to the measurable parameter and training the target frequency autoencoder.

In some embodiments, the training includes minimizing a cost function including a sum of distances of first measured values of the measurable parameter at the serving frequency and outputs of a composition of the serving frequency auto-encoder and the serving-frequency auto-decoder applied respectively on the first measured values and a sum of distances of second measured values of the measurable parameter at the target frequency and outputs of a composition of the target-frequency auto-encoder and the target-frequency auto-decoder applied respectively on the second measured values.

For instance, the cost function includes, as a representation of the proximity constraint, a function proportional to a sum of differences between outputs of the serving frequency autoencoder applied on the first measured values and the target-frequency auto-encoder applied on the second measured values.

In some embodiments, the number of network nodes of the latent representation is the same for the serving frequency auto-encoder and the target-frequency auto-encoder.

For instance, the method includes receiving the measured values of the measurable parameter at the serving frequency and at the target frequency.

For example, the measurable parameter is a signal power.

In some embodiments, the signal power is the signal power of a reference signal.

Also provided is a device for training a neural network for outputting a predicted value of a measurable parameter at a target frequency based on an inputted value of the measurable parameter at a serving frequency including circuitry configured for training a serving frequency auto-encoder which maps the measurable parameter at the serving frequency to a latent representation and training a target frequency auto decoder which maps the latent representation to the measurable parameter at the target frequency, wherein the circuitry is configured for performing the training of the serving frequency auto-encoder and the target- frequency auto-encoder using measured values of the measurable parameter at the serving frequency and at the target frequency and based on a proximity constraint enforcing proximity between a mapping of the measurable parameter at the serving frequency to the latent representation by the serving frequency auto-encoder and a mapping of the measurable parameter at the target frequency to the latent representation by a target-frequency autoencoder which maps the measurable parameter at the target frequency to the latent representation.

A neural network trained accordingly enables a communication device to determine a measurable parameter at a target frequency without a need for performing measurements at the target frequency.

For instance, device for training is configured for training a serving frequency auto-decoder which maps the latent representation to the measurable parameter and for training the target frequency auto-encoder.

In some embodiments, the training includes minimizing a cost function including a sum of distances of first measured values of the measurable parameter at the serving frequency and outputs of a composition of the serving frequency auto-encoder and the serving-frequency auto-decoder applied respectively on the first measured values and a sum of distances of second measured values of the measurable parameter at the target frequency and outputs of a composition of the target-frequency auto-encoder and the target-frequency auto-decoder applied respectively on the second measured values.

For instance, the cost function includes, as a representation of the proximity constraint, a function proportional to a sum of differences between outputs of the serving frequency autoencoder applied on the first measured values and the target-frequency auto-encoder applied on the second measured values.

In some examples, the device for training includes at least one interface configured for receiving the measured values of the measurable parameter at the serving frequency and at the target frequency.

In some embodiments, the number of network nodes of the latent representation is the same for the serving frequency auto-encoder and the target-frequency auto-encoder. For example, the measurable parameter is a signal power.

In some embodiments, the signal power is the signal power of a reference signal.

Further provided is a computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the steps of the communication method or the method for training a neural network according to any of the aforementioned aspects, embodiments, and examples.

Also provided is a non-transitory computer-readable medium comprising instructions which, when executed by a computer, cause the computer to carry out the steps of the communication method or the method for training a neural network according to any of the aforementioned aspects, embodiments, and examples.

Brief description of drawings

In the following detailed description, exemplary embodiments are described in more detail with reference to the accompanying figures and drawings, wherein:

Fig. 1 is a schematic drawing RSRP determination in a scenario with two frequencies;

Fig. 2 is a flow chart showing an overview of training and inference techniques;

Fig. 3 is a block diagram showing a communication device;

Fig. 4 is a flow chart showing a communication method;

Fig. 5 is a flow chart showing a method for training a neural network;

Fig. 6 is a block diagram of a device for training a neural network;

Fig. 7 is a block diagram showing circuitry of a device for training a neural network;

Fig. 8 is a schematic drawing of a neural network of an auto-encoder, auto-decoder, and a latent representation;

Fig. 9 is a flow chart showing a method for training a neural network;

Fig. 10 is a schematic drawing of an inter-frequency common latent space for a plurality of auto-encoders; Fig. 11 is an illustration of simultaneous training of auto-encoders and constrained representations.

Fig. 12 is an illustration of the use of trained auto-encoders and decoders in the inference phase.

Detailed description

As mentioned in the background, a UE performs inter-frequency measurements. To illustrate, consider the following example shown in Fig. 1 . Some user is connected to a given frequency F ₀ , and therefore knows the RSRP of other cells for ₀ , e.g. through measurements. As shown in Fig. 1 , RSRPs are known only for one frequency, F ₀ . Various situations may require the user to have access to the RSRP of other frequencies. Such situations may include a low RSRP of F ₀ , or increase or in a required data rate or other parameters, e.g. due to requirements of a newly requested service, e.g. a phone call, access of a website, transmission or reception of video or image data. Accordingly, one needs access to RSRP of F to decide whether the user should switch frequency or not. To this end, inter-frequency measurements are typically performed. Since the RSRPs experienced by a UE are associated with a position of the UE in the coverage area of a network, e.g. a relative position with respect to one or more cells or antennas of network nodes, the RSRP determination may also be called “Virtual grid positioning”, where the term “grid” represents the network or the arrangement of cells or network nodes.

However, although performed for improving the overall user experience, inter-frequency measurements may degrade user experience in the short term. In addition to the increased battery consumption induced by these measurements, no downlink or uplink transmissions are performed with the serving cell during the measurement gaps, so the overall throughput may be degraded.

Some work has been concerned with the downside of inter-frequency measurements. For instance, the focus has been laid on ranking the frequencies on which measurements should be performed, wherein prioritizing of a subset of frequency from a given set of frequencies obtained from the network, based on information such as a list of prior frequencies used by the user, has been suggested. As another example, to reduce the measurement gap and, thus, the measurement delay between a user equipment and a base station or network node such as an evolved NodeB (eNodeB) of LTE- (Long Term Evolution) or similar communication systems, techniques for enhancing inter-frequency measurement have been detailed. These techniques aim at prioritizing the measures or limit the drawbacks, but still involve measuring other frequencies RSRP and therefore degrading the short-term user experience.

The present disclosure is directed at limiting inter-frequency measurements by using historical inter-frequency measurement data, and predicting inter-frequency measurements without the need of measurement gaps to improve user experience.

The disclosed techniques rely on the following elements. First, auto-encoder neural networks are used to grasp the inherent structure of measurements and build a coherent latent space shared by different frequencies. Then, a latent constraint or proximity constraint is motivated to keep similar measurements, in particular measurements performed in different frequencies, close to one another. The latent constraint may be represented by a regularization function such as a Tikhonov regularization. Finally, it is shown how both auto-encoder neural network and latent constraint can be used simultaneously for training by formulating a loss function. The combination or simultaneous use of auto-encoders and latent constraints results in semi- supervised training. An overview on the disclosed techniques is provided in Fig. 2.

Provided is a communication device 300 shown in Fig. 3, comprising an interface 310 configured for acquiring a current measured value of a measurable parameter at a serving frequency and circuitry 320 configured for determining a predicted value of the measurable parameter at a target frequency based on the current measured value.

For instance, the communication device 300 is a communication device adapted to perform communication in one or more wireless communication networks or a radio communication networks such as UMTS, LTE or New Radio, 4G, 5G, 3GPP New Radio (NR), WiMAX, Wi-Fi or any other wireless network. E.g. the communication includes communication between a base station or network node and a UE, user terminal, or mobile communication device. In case of a hierarchical cellular communication system. The base station serves a cell which may be a microcell, femtocell, macrocell, or any other instance of a layer in the hierarchy.

The term“interface” refers to an input and/or output processing structure which may include one or more protocol layers defining in which format the data are received or transmitted and how the data are to be interpreted. Such interface may be wired or wireless.

The term“circuitry” refers to processing circuitry such as one or more processors or CPU(s) (central processing unit(s)), and includes hardware components such as ASIC (application specific integrated circuit), FPGA (field programmable gate array), software implementations running on any hardware, or any combination of hardware and software. Acquiring of a measured value may include obtaining the measured value, e.g. by performing a measurement or by receiving the value or an indication of the same from another communication device which may have performed the measurement or may have received the value from yet another device.

A current measured value is a result of a measurement of the measurable parameter which has been performed at a measurement time, wherein an interval between the measurement time and a current point in time is sufficiently small for the measurement value to represent current network or channel conditions.

The serving frequency may be a frequency, frequency band, sub-band, carrier or subcarrier at which the communication device currently receives or transmits data and/or signaling. The target frequency is a candidate for performing communication at a time following the current point in time within a sufficiently small time interval for which it is expected or assumed that channel conditions or network conditions still match or are represented by the current measured value. Serving frequency and target frequency, i.e. the actual frequencies at and for which measurement and/or prediction is performed, may be frequencies representative of a wider range of serving frequency band and target frequency band.

The interface 310 may acquire one or more current measured values at the serving frequency, e.g. a plurality of respective measured values of a plurality of cells, antennas, or base stations. Further, the circuitry 320 may determine one or a plurality of predicted values, e.g. a plurality of values corresponding a plurality of cells or antennas. The number of measured values and the number of predicted values may be the same or different, wherein the number of predicted values may be greater or smaller than the number of measured values.

Also provided is a communication method corresponding to the above disclosed communication device 300. As shown in Fig. 4, the communication method includes a step of acquiring S410 a current measured value of a measurable parameter at a serving frequency, and determining S420 a predicted value of the measureable parameter at a target frequency based on the current measured value.

As applies for both the communication device 300 and the corresponding communication method, the predicted value is determined as an output of a neural network upon input of the current measured value. The neural network comprises a serving frequency auto-encoder which maps the measurable parameter at the serving frequency to a latent representation and a target frequency auto-decoder which maps the latent representation to the measurable parameter at the target frequency. Further provided is a method for training a neural network for outputting a predicted value of a measurable parameter at a target frequency based on an inputted value of the measurable parameter at a serving frequency, which is shown in Fig. 5. The method, which may also be referred to as a“training method”, includes a step of training S510 a serving frequency autoencoder which maps the measurable parameter at the serving frequency to a latent representation, and further includes a step of training S520 a target frequency auto-decoder which maps the latent representation to the measurable parameter at the target frequency.

The order of method steps suggested by Fig. 5 is merely exemplary and not limiting. For instance, a training operation of the target frequency auto-decoder may be performed before or after a training operation of the serving-frequency auto-encoder.

In correspondence with the above-disclosed training method, provided is a device for training a neural network, the neural network being a network for outputting a predicted value of a measurable parameter at a target frequency based on an inputted value of the measurable parameter at a serving frequency.

The device 600 for training a neural network, which is shown in Fig. 6, includes circuitry 620 configured for training a serving frequency auto-encoder and a target frequency auto decoder. The serving frequency auto-encoder maps the measurable parameter at the serving frequency to a latent representation, and the target-frequency auto-decoder maps the latent representation to the measurable parameter at the target frequency.

For instance, the circuitry 620 of the device for training a neural network, or“training circuitry”, includes serving frequency training circuitry 721 and target frequency training circuitry 722, as shown in Fig. 7.

In the aspects and embodiments of the present disclosure, the neural network is trained using measured values of the measurable parameter at the serving frequency and at the target frequency and includes a proximity constraint enforcing proximity between a mapping of the measurable parameter at the serving frequency to the latent representation by the serving frequency auto-encoder and a mapping of the measurable parameter at the target frequency to the latent representation by a target-frequency auto-encoder which maps the measurable parameter at the target frequency to the latent representation.

The measured values used in the training include“historical” measured values which have been obtained prior to a measurement of a current measurement of a current measured value in the operation phase or inference phase when frequency estimation and wireless communication is performed by a user equipment and a base station. As mentioned, the present disclosure relates to communication in a wireless network, including communication between a base station and a UE.

In some embodiments, the communication device 300 is a base station. The base station comprises a receiver configured for receiving, as an example of “acquiring”, the measured value from a user equipment. Therein the user equipment which transmits the measured value may have performed the measurement to obtain the measured value. In such embodiments, the receiver of the base station embodies or comprises interface 310 configured for acquiring the measured value.

In some other embodiments, the communication device 300 is a UE and comprises a receiver configured for performing a measurement to perform the measured value. In this case, the receiver of the UE embodies or comprises the interface 310 configured for acquiring the measured value.

If the communication device 300 is a base station, the UE does not need to perform the possibly resource consuming neural network processing. In this case, the receiver of the UE comprises or embodies the interface 310. However, a base station, may need to perform calculations of predicted values for measurable parameters for a plurality of currently connected UEs. Accordingly, to allow for less complex processing at the base station, it may be advantageous that the prediction of the predicted values is performed by the UEs which also perform the measurement, in particular if the UEs have sufficiently strong processing power.

As mentioned, the target frequency is a candidate frequency at which communication may be performed after measuring the measurable parameter at the serving frequency. For instance, to perform communication at the target frequency, the base station initiates, e.g. by Downlink Channel Information (DCI) or Radio Resource Control (RRC), the UE to activate the target frequency for one or more subsequent time intervals (e.g., subframes, slots, minislots, etc.), and the UE activates the target frequency in accordance with the initiation.

Activation of the target frequency may include a switch from the serving frequency to the target frequency. A switch from one frequency or frequency band to another frequency or frequency band may be needed in case of poor signal or channel conditions at the serving frequency or a change in requirements, e.g. when a phone call or a data transmission is being scheduled. Moreover, activation of the target frequency may include an extension of a serving frequency band by the target frequency or by a range of frequencies including the target frequency. Extension of the frequency band may be performed in case of a need for a greater data rate in subsequent transmissions. The measurable parameter may be a signal power of a data signal, control signal, or reference signal transmitted from a base station and received by the UE. The measurable parameter may further include a signal to noise ratio, a signal to interference plus noise ratio, or a similar parameter indicative of the signal strength or signal quality of a serving cell or a neighboring cell as experienced by the UE.

In some embodiments, the signal power is a signal power of a reference signal. E.g., the measurable parameter is a reference signal received power (RSRP). As an example of RSRP, in LTE, measurements are performed to provide a cell-specific signal strength metric used to rank different cells according to their signal strength as an input for handover and cell reselection decisions. In some LTE systems, the RSRP is defined a linear average over the power contributions (in Watts) of the Resource Elements (REs) which carry cell-specific RS within the considered measurement bandwidth. However, the present disclosure is not limited to a particular determination of RSRP. Other types of average, or non-averaged power values RS may be also be applied, depending on the particular system or configuration.

Moreover, alternatively or in addition to RSRP, the present disclosure may be applied to RSRQ (Reference Signal Received Quality), which additionally takes the interference level into account. An exemplary definition of RSRQ from some LTE systems is N ^c RSRP / (LTE carrier RSSI), with N being the number of Resource Blocks (RBs) of the LTE RSSI Received Signal Strength Indicator bandwidth. RSSI is a measurement of a total received wideband power observed by a UE including non-serving cells, adjacent channel interference, and thermal noise.

In the following description, details on the neural network as well as the training or learning of the neural network according to the present disclosure will be provided.

A schematic drawing of a neural network is provided in Fig. 8. As mentioned above, a neural network includes an auto-encoder which maps the measurable parameter or a value of the measurable parameter at a given frequency, which is received as input, to a latent representation of the measurable parameter in latent space. The neural network further includes an auto-decoder which maps the latent representation to a reconstructed value of the inputted value of the measurable parameter.

A plurality of such neural networks, as shown in Fig. 8, may be trained with measured values, possibly including historical measurements, performed respectively at a plurality of frequencies. In the inference phase when the neural network is applied to current values in the operation of the radio communication network, any of the plurality of frequencies may become a serving frequency and a target frequency for a particular, switching, handover, band widening, or other frequency activation operation. In this disclosure, when the training phase or learning phase and the device and method for training a neural network are describes, the expressions“current frequency” and“target frequency” may be considered illustrative names which refer to at least two frequencies from among a plurality of frequencies for which training is performed, and which reflect the purpose of the neural network when applied to wireless communication. Accordingly, a plurality of auto-encoders and corresponding auto-decoders include the serving-frequency auto-encoder serving frequency auto-decoder, target-frequency auto-encoder, and target-frequency auto-decoder.

As has been mentioned and will be described further, in the inference phase, the serving- frequency auto-encoder performs a mapping of an inputted measured value at the serving frequency to the latent space, and the target-frequency auto-decoder maps the latent representation to a predicted value of the measurable parameter at the target frequency. However, in the training phase or learning phase, both the auto-encoder and the auto-decoder are trained for each frequency. Accordingly, in some embodiments, the training method further includes the steps of training a serving frequency auto-decoder which maps the latent representation to the measurable parameter and training the target frequency auto-encoder which maps the measurable parameter to the latent representation.

Moreover, as mentioned, the auto-encoder performs the mapping to the latent representation on a value of the measurable parameter received as input. Accordingly, as is indicated in Fig. 6 by dashed lines, the device 600 for training a neural network may include an interface 610 configured for receiving the measured values of the measurable parameter at the serving frequency and at the target frequency. Correspondingly, the training method may include a step of receiving S905 the measured values of the measurable parameter at the serving frequency and at the target frequency, as is shown in Fig. 9 in addition to the steps already shown in Fig. 5.

Fig. 8 shows a standard representation of an auto-encoder and corresponding auto-decoder. As shown, the neural network includes input nodes symbolized by black circles corresponding to the values of the measurable parameter which are received as input, and output nodes corresponding to the reconstructed values of the measurable parameter, which are to be trained to be sufficiently close to the input values. The central layer between auto-encoder and auto-decoder corresponds to the latent representation associated with the input data, including network nodes of the latent representation symbolized by white circles. The number of input nodes and output nodes is the same as a number of cells from which the measurable parameter such as a RSRP is received. Input nodes and output nodes are both connected to the network nodes in the latent representation (or“latent nodes” in short) representing the received input values. As shown in the figure, the number of input and output nodes is four, and the number of latent nodes is three. However, in general, the number of latent nodes may be greater, equal to, or smaller than the number of input nodes / output nodes. Connections of the input nodes and output nodes with the latent nodes correspond to the mapping from the measurable parameter to the latent representation and vice versa by the auto-encoder and, respectively, auto decoder.

The use of neural network auto-encoders is motivated by the need to map values of the measurable parameter, e.g., F _t RSRP coordinates, to latent representations, in an unsupervised way. In particular, it can be assumed that the values which RSRPs may take are restricted, for example to a manifold, rather than being arbitrary. Neural networks autoencoders allow to learn the manifold RSRPs belong to, which will be useful for prediction as some RSRP configurations are not possible in practice.

These RSRP configurations or constellations may include RSRPs occurring on a plurality of frequencies. Thus, the latent space to be learned will be common to any frequency used in the measurements, allowing different frequency measurements to be compared. In order to have meaningful representations (/. e., two measurements from the same location/time but at different frequencies will be close in the latent space), as an additional step to the latent representation learning, a latent constraint (which is also called“proximity constraint” in this disclosure) may be introduced.

In particular, the training of the neural network in accordance with some embodiments of the present disclosure relies on constraining inter-frequencies latent representations to be close for different frequencies, by introducing an inter-frequency latent constraint in the latent space which is common to the different frequencies. Inter-frequency latent constraints are illustrated in Fig. 10.

Fig. 10 shows two the common latent space (L _å ) associated with two auto-encoders. In the latent space, representations values of the measurable parameter, such as RSRP, corresponding to measurements performed at the same position, e.g. location of a UE performing the measurements, are constrained to be at a small distance to each other. In the figure, the latent constraint is symbolized by dashed ovals surrounding associated latent representations of measurements from different frequencies Fo and Fi. Three latent nodes are shown in Fig. 10 in the common latent space. Accordingly, the latent space is a three- dimensional space, of which only two-dimensions are shown in the figure for the purpose of illustration on a two-dimensional sheet. In accordance with the above description, the training of the neural network includes learning of a latent space and of constrained representations of measurable parameters.

The learning of both latent space and constrained representations may be implemented minimizing the loss function given by the following equation:

(expression. (1 )).

In expression (1 ), eo and ei are auto-encoders of the measurable parameter at frequencies F ₀ and Fi, respectively, and do and di are the auto-decoders corresponding to eo and ei, and . l is a proportionality coefficient. Index i runs over measured values of the measureable parameter x _t measured at Fo and Fi _, which contribute as input to the learning process. If the number of cells from which RSRP are measured or measurable at Fo and/or Fi is greater than one, x ° or xf ¹ may be a vector, the number of components being equal to the number of cells and values measured at the respective frequency. The latent representations of the input values may be vectors, the number of components being the number of latent nodes.

Accordingly, in some embodiments, the training of the neural network includes minimizing a cost function (which may also be called a“loss function”). The cost function includes at least a sum of distances (e.g. ||. || ² ) of first measured values of the measurable parameter at the serving frequency and outputs of a composition d ₀ (e ₀ (x)), of the serving frequency autoencoder and the serving-frequency auto-decoder applied respectively on the first measured values, and a sum of distances of second measured values of the measurable parameter at the target frequency and outputs of a composition of the target-frequency auto-encoder and the target-frequency auto-decoder applied respectively on the second measured values. E.g., the two sums of distances mentioned in this paragraph may the first two terms on the right- hand side of expression (1 ). Thus, the first two sums in the cost function correspond to the training of the auto-encoders: the latent space should be constructed so as to reproduce the input with the highest possible fidelity. The first two-sums, i.e. the auto-encoder/auto-decoder terms, are unsupervised terms corresponding to unsupervised learning. Minimizing these terms corresponds to minimizing reconstruction errors of the auto-decoders.

Moreover, in some embodiments, the cost function includes, as a representation of the proximity constraint or latent constraint, a function proportional to a sum of differences between outputs of the serving frequency auto-encoder applied on the first measured values and the target-frequency auto-encoder applied on the second measured values. This function representing the proximity constraint may be implemented by the last term of expression (1 ). For instance, the function may be a Tikhonov regularization, which has been mentioned above, or some other suitable regularization function, such as LASSO (least absolute shrinkage and selection operator). Accordingly, the rightmost term of the loss function is related to the constraint on the latent representations: same measurements at different frequencies should be close to one another. The term representing the latent constraint is a supervised term in equation (1 ), representing the supervised part of the semi-supervised learning techniques of the present disclosure.

Using a loss function, as provided by this disclosure and exemplified by expression (1 ), allows for simultaneous training of auto-encoders and corresponding auto-decoders and constrained latent representation across frequencies. This is illustrated for two frequencies Fi and F _å by Fig. 1 1.

Fig. 1 1 shows respectively four input nodes for both frequencies and F. However, corresponding to measurement values of the measurable parameter such as RSRP from different cells. However, in general, the number of cells operating at the different frequencies for which the measurable parameter is input may be different. Correspondingly, the number of input nodes of the respective auto encoders may also be different for different frequencies (and consequently also the number of output nodes). This is indicated by the different indexes (cell _n , cell _m ) in the last lines of the different encoders shown in Fig. 1 1.

Moreover, the number of summands in the respective three terms of expression (1 ) may be the same or different, depending on the measurement values available for the respective frequencies. The greater the amount of summands is in the first or the second term, the greater may be the fidelity of auto-encoder and auto-decoder output to the original input for the respective frequency F ₀ or F . A greater number of summand in the last sum corresponding to pairs of measured values of frequencies F ₀ or F may achieve a closer proximity between latent representations corresponding to measurements across frequencies.

However, even when the number of cells or input/output nodes is different for different frequencies, the number of network nodes of the latent representation is the same for the serving frequency auto-encoder and the target-frequency auto-encoder.

Accordingly, each latent representation of one or more measured values measured at one frequency has a corresponding latent representation of one or more measured values at another frequency, and the proximity constraint can be applied on corresponding latent representations. Although the example shown in the figures of this disclosure show two frequencies F ₀ and Fi, the learning and training techniques of the disclosure may likewise be extended to the case of the number of frequencies greater than two, which, in the operation phase, may find application when there are more than one target frequencies as candidates for activation.

To this end, the loss function of expression (1 ) may be extended to an arbitrary number k of frequencies in a straight-forward manner. In this case, the loss function would contain k sums related to the training of k corresponding auto-encoders/auto-decoders corresponding to the first two sums of expression (1 ). Moreover, to include constraints for each pair of frequencies, there may be k(k-1 )/2 constraint terms corresponding to the rightmost term of expression (1 ).

As can be seen from the present description of learning and communication techniques, the embodiments of the present disclosure are based on the mapping of frequency-related data to a latent space. A mapping is learned for each considered frequency, and the resulting latent space is shared by all of these frequencies, allowing measurements from different frequencies to be compared to each other. Such comparison of two or more measurements at different frequencies is facilitated by the transformation or mapping into the latent space. Accordingly, the present disclosure facilitates using information from any frequency to predict a particular measurement, or to predict a result of a measurement instead of actually performing the measurement.

As mentioned, the neural networks of the present disclosure may be trained with ad applied to datasets which come from telecommunication networks. In the following, an example will be provided to illustrate the applicability of the neural networks of the present disclosure in the operation phase.

It may be assumed that a user or UE is connected to a given frequency F ₀ , and the RSRP at frequency F ₀ of all surrounding cells, such as a serving and neighboring cells, have been measured and collected. As has been mentioned, various situations may require the user to have access to the RSRP of other frequencies ( e.g ., RSRP of F ₀ is low). To this end, conventionally, one needs to scan other frequencies such as F to access RSRPs at F to decide whether the user should switch frequency or not.

However, it may now further be assumed that one or more scans have already been performed at F _{l t} wherein the RSRPs at F _{l t} have been collected. These scans may have been used to construct a dataset that may be used by the learning or training methods according to the present disclosure. When this learning has been performed, for a user connected to a cell at frequency F ₀ , a system capable of inferring the RSRP of the surrounding cells at F (which may be the same cells or different cells) have been trained. This system enables knowing or predicting the RSRPs of the surrounding cells at F without scanning frequency F during operation at F ₀ .

The use of an auto-encoders and auto-decoders in the inference phase is illustrated in Fig. 12. In this example, F ₀ may be assumed to be the serving frequency, and F may be assumed to be the target frequency. As shown in Fig. 12, the serving frequency auto-encoder,“Encoder Fo" in this case, encodes inputted current values, namely RSRP, measured at Fo, mapping them to the latent representation in the latent space which is common to the serving frequency and the target frequency, Fi in this case. Due to the above-described semi-supervised training including the proximity constraint of latent representations of RSRPs at different frequencies, the latent representation e ₀ (RSRP^ _rrent ) of the measured values at F ₀ is taken as an approximation of the latent representation e ₀ {RSRP^ _rrent ) of the corresponding RSRPs at Fi which have not been measured. This is symbolized by the equals sign in the figure. Accordingly, the latent representation e ₀ (RSRP^ _rrent ) is input into the target frequency autodecoder (“Decoder Fi”), which maps it to the estimated value of the RSRP at the target frequency. Thus, the estimated value of the measurable parameter at the target frequency is di (e ₀ (RSRP _c ^F ° _rrent )). As can be seen from Fig. 12, in the inference phase, the serving frequency auto-decoder and the target frequency auto-encoder are not applied. However, they are needed for training the corresponding serving frequency auto-encoder and target- frequency auto-decoder which are used in the inference phase.

Any details, definitions, embodiments and examples provided in this disclosure are meant to refer to each of the communication device, communication method, training device, training device, and training method, unless the context or explicit statement indicates otherwise.

Moreover, in the description and in some drawings, the reference to“RSRP” as a measurable parameter is exemplary. The present disclosure may be applied to measurable parameters the values of which are correlated across frequencies for a given location of measurement, allowing for the construction and learning of a common latent space where a proximity constraint ensures sufficient closeness of the latent representations of associated measurements across frequencies.

Further, the present disclosure refers to a learning phase (or“training” phase) in which the neural network is trained to learn auto encoders/decoders, latent representation, and latent constraint, and an inference phase (or“operation phase”) where the neural network is actually applied to estimation of a measurable parameter at frequencies where a user equipment currently does not operate. In order to allow for sufficiently precise prediction of measured values, the training phase shall precede the inference phase. However, raining and operation phases may be overlapping in time, and the neural network may be updated during operation to be enhanced by considering newly received measured values. Summarizing, provided are a communication device, a communication method, a training device and a training method. The communication device comprises an interface for acquiring a current measured value of a measurable parameter at a serving frequency, and circuitry for determining a predicted value of the measurable parameter at a target frequency based on the current measured value as an output of a neural network comprising a serving frequency auto- encoder mapping the measurable parameter at the serving frequency to a latent representation upon input of the current measured value and a target frequency auto-decoder mapping the latent representation to the measurable parameter at the target frequency. The neural network is trained using measured values of the measurable parameter at the serving frequency and at the target frequency and includes a proximity constraint enforcing proximity between mappings of the measurable parameter at the serving frequency and target frequency to the latent representation.