Low latency machine learning

Abstract

It is provided a method comprising: preparing a set of plural input values based on a received first actual value of a quantity; calculating, for each of the plural input values, a respective inference by a respective machine learning instance, wherein the calculating for at least two of the plural input values is performed fully or partly in parallel; selecting one of the plural input values, wherein the selected one of the plural input values is closest to a second actual value of the quantity, and the second actual value is received after the first actual value; determining an estimated inference based on the inference calculated for the selected one of the plural input values.

Description

[0001] LOW LATENCY MACHINE LEARNING

[0002] Field of the invention

[0003] The present disclosure relates to machine learning.

[0004] Abbreviations

[0005] 3GPP 3rdGeneration Partnership Project

[0006] 4G / 5G / 6G 4th / 5th / 6thGeneration

[0007] 5GC 5G Core

[0008] Al Artificial Intelligence

[0009] BS Base Station

[0010] CN Core Network

[0011] CP Control Plane

[0012] CSI Channel State Information

[0013] CTF Channel Transfer Function

[0014] CU Central Unit

[0015] DNN Deep Neural Network

[0016] DU Distributed Unit gNB 5G NR Node B

[0017] GPU Graphical Processing Unit

[0018] HW Hardware

[0019] ID Identifier

[0020] JT CoMP Joint Transmission CoMP

[0021] LSTM Long-short Term Memory

[0022] MIMO Multiple Input - Multiple Output

[0023] ML Machine Learning mMIMO massive MIMO

[0024] MPC Multi Path Components

[0025] MU Multi-User

[0026] NLOS Non-Line of Sight

[0027] NR New Radio

[0028] O-DU Open Distributed unit

[0029] ORAN Open RAN

[0030] PHY Physical (layer) PUCCH Physical Uplink Control Channel

[0031] RAN Radio Access Network

[0032] RIC RAN Intelligent Controller

[0033] RNN Recurrent Neural Network

[0034] RRM Radio Resource Management

[0035] RT Real-time

[0036] RU Radio Unit

[0037] UCI Uplink Control Information

[0038] UE User Equipment

[0039] UP User Plane

[0040] URLLC Ultra-reliable low-latency communications

[0041] Background

[0042] Mobile radio networks are evolving towards increased usage of ML for radio interface procedures. In parallel, in some nationally funded projects the best integration of ML into PHY layer applications is being discussed.

[0043] ORAN - closely related to 3GPP - defines already a whole framework including interfaces for an AI / ML architecture. In this framework, the base station is split into CU (for UP and CP, respectively), one or more DUs and respective RUs. A first control loop runs on and between DU and RU. It is particularly to the PHY layer. The latency should be less than 10 ms. Plural base stations may be connected (from their O-CU-CP) via E2 interface with a Near Realtime RIC. The Near Realtime RIC control loop is near realtime and may have a latency equal to or larger than 10 ms but less than 1 second. On the service management and orchestration framework, a non-realtime RIC control loop may run, having a latency equal to or larger than 1 second.

[0044] Especially in combination with PHY layer applications, the processing time for ML instances becomes much more important compared to higher layer RRM applications, which run on a much longer time scales. Even the 10ms delay of the O-DU control loop may be too long for many PHY layer algorithms. See as an exemplary application channel estimation and prediction, which becomes outdated / useless after one to few ms, even for nomadic users. URLLC is another example with an obvious need of fast decision makings. Industry applications for industry 4.0 is another area, where fastest reaction times might be needed in some use cases. ML processing is often complex with high number of layers and weights, which results in a certain minimum latency even under the assumption of powerful HW implementations in the cloud and / or based on latest GPU technologies.

[0045] In detail, ML relies on various types of neural network architectures, like DNN, RNN, LSTM, which might be even combined (e.g. chained) to realize a more powerful overall network. The processing might be parallelized to some extent and one might use specific HW accelerators like GPUs, but often the inference in one network layer can only be calculated when the previous network layer outputs are available, i.e., there is often a serial processing included. This is even particularly important for classical iterative inference processes as they are quite usual in the context of PHY layer applications.

[0046] In addition, some applications can become extremely complex, for example, in case of accurate multi path component (MPC) parameter estimation for NLOS radio channels comprising potentially several tens to hundreds of MPC. In that case, iterative ML inferences - optimized step by step - might be needed so that for large number of iterations even low latencies per iteration become an issue. Or, contrary a relatively large processing time per iteration will limit the number of iterations with correspondingly reduced performance of the final inference. For example, in the estimation of MPC parameters used for a CSI channel estimation and prediction, a limited CSI estimation accuracy (large normalized mean square error of the CTF) will limit the channel prediction horizon and eventually will limit the performance of a MU MIMO and / or JT CoMP precoder. So, a low latency ML inference can become fundamental for advanced system level performance with highest spectral efficiency or capacity.

[0047] Note that channel estimation and prediction is only one example where calculating ML inference with low latency is important.

[0048] Ensemble learning comprises calculating multiple inferences on noisy input signals in parallel, and in the end selecting one of the inferences, e.g. by a majority voting. Ensemble learning is typically used to increase the reliability of ML inferences. It does not affect the latency of the inference process as such.

[0049] Summary It is an object of the present invention to improve the prior art.

[0050] According to a first aspect of the invention, there is provided an apparatus comprising: one or more processors, and memory storing instructions that, when executed by the one or more processors, cause the apparatus to perform: preparing a set of plural input values based on a received first actual value of a quantity; calculating, for each of the plural input values, a respective inference by a respective machine learning instance, wherein the calculating for at least two of the plural input values is performed fully or partly in parallel; selecting one of the plural input values, wherein the selected one of the plural input values is closest to a second actual value of the quantity, and the second actual value is received after the first actual value; determining an estimated inference based on the inference calculated for the selected one of the plural input values.

[0051] According to a second aspect of the invention, there is provided a method comprising: preparing a set of plural input values based on a received first actual value of a quantity; calculating, for each of the plural input values, a respective inference by a respective machine learning instance, wherein the calculating for at least two of the plural input values is performed fully or partly in parallel; selecting one of the plural input values, wherein the selected one of the plural input values is closest to a second actual value of the quantity, and the second actual value is received after the first actual value; determining an estimated inference based on the inference calculated for the selected one of the plural input values.

[0052] The method of the second aspect may be a method of machine learning.

[0053] According to a third aspect of the invention, there is provided a computer program product comprising a set of instructions which, when executed on an apparatus, is configured to cause the apparatus to carry out the method according to the second aspect. The computer program product may be embodied as a computer-readable medium or directly loadable into a computer.

[0054] According to some embodiments of the invention, at least one of the following advantages may be achieved:

[0055] • The negative effects of the relatively long latency of the inference calculation are reduced;

[0056] • Precision of the inference process and computational effort may be outweighed depending on the needs; • May enable ultra fast ML applications, despite possibly extreme complexity or implementation in a more distant cloud environment;

[0057] • May enable PHY layer applications including complex iterative ML inferences

[0058] • May support complex ML inference tasks in combination with URLLC applications.

[0059] It is to be understood that any of the above modifications can be applied singly or in combination to the respective aspects to which they refer, unless they are explicitly stated as excluding alternatives.

[0060] Brief description of the drawings

[0061] Further details, features, objects, and advantages are apparent from the following detailed description of the preferred embodiments of the present invention which is to be taken in conjunction with the appended drawings, wherein:

[0062] Fig. 1 illustrates a method according to some example embodiments of the invention;

[0063] Fig. 2 illustrates a method according to some example embodiments of the invention;

[0064] Fig. 3 shows how the range of input signals evolves in time based on the reported data value (assuming a smooth evolution of the input signals), according to some example embodiments of the invention;

[0065] Fig. 4 shows an example embodiment of the invention where multiple ML ensembles run in parallel. In addition, it shows that the inferences calculated from the multiple ML instances may be input into another ML instance (e.g. for precoding);

[0066] Fig. 5 shows an example embodiment of the invention applied to JT CoMP precoding;

[0067] Fig. 6 shows a message flow according to some example embodiments of the invention;

[0068] Fig. 7 shows an apparatus according to an example embodiment of the invention;

[0069] Fig. 8 shows a method according to an example embodiment of the invention; and Fig. 9 shows an apparatus according to an example embodiment of the invention;

[0070] Detailed description of certain embodiments

[0071] Herein below, certain embodiments of the present invention are described in detail with reference to the accompanying drawings, wherein the features of the embodiments can be freely combined with each other unless otherwise described. However, it is to be expressly understood that the description of certain embodiments is given by way of example only, and that it is by no way intended to be understood as limiting the invention to the disclosed details. Moreover, it is to be understood that the apparatus is configured to perform the corresponding method, although in some cases only the apparatus or only the method are described.

[0072] Some example embodiments of the invention enable very low latency ML inference, in particular for time critical Al applications.

[0073] Fig. 1 shows an example embodiment of the invention. There are K (K > 2) input values Apk, i.e., the set {... ,Apk, ... }K. The K input values are determined such that they are close to an actual (real) value received at time t0(details will be described below).K ML model instances run in parallel, and each of the K ML model instances pre-calculates a respective inference for one of the input values Apkfully or partly in parallel. These inference calculations start at time to (or shortly thereafter) and will be finished before time ti = to + 1 (note that Ί ’ indicates here a time step). At time ti = to + 1 , an actual (real) value is received for the time ti. Then, a selector selects of the K input values the one which is closest to the actual (real) value. The ML inference calculated for the selected input value is used to estimate the actual inference at the time ti. In the example of Fig. 1 , the selector selects input value Dr2 because it is closest to the actual value received at t1 and, thus, determines the actual inference at the time ti based on inference 2. For example, the estimated actual inference may be equal to the inference precalculated for the selected input value (equal to inference 2 in the example of Fig. 1).

[0074] The estimated inference may then be applied to a parameter which presumably influences the future actual values.

[0075] For example, the actual (real) values may be measured values of a physical quantity. As another option, they may be values of a quantity calculated by some algorithm. As a still other option, they may be estimated values of a physical quantity, wherein the estimation is based on one or more measured values of another physical quantity. The actual values may be single values or combinations of plural values (such as vectors, matrices, etc.).

[0076] Pre-calculating two inferences fully in parallel means that the pre-calculations start simultaneously, or at least with the minimum delay possible for the system architecture. Precalculating two inferences partly in parallel means that the calculation of the second inference starts some time later than the calculation of the first inference but still before the calculation of the first inference is finished. The rationale for this procedure is to trade-off latency for complexity, i.e., precompute several possible conditional inferences and select the correct one only when the measurement is available.

[0077] By selecting the best fitting ML instance at time ti from the already calculated inferences one can save the time duration for ML inference calculation from to to the end of the calculation time which is not later than ti. Instead one adds a time tseiect to identify the best fitting ML instance, but this time is typically extremely short compared to the ML inference processing time.

[0078] Fig. 2 shows another example embodiment of the invention. In this case we have a first inference by the set of parallel ML instances - the ML ensemble - and a selector, which based on the most recent data value and identifies the best output of the ML ensemble from the most recent ML inference. In the next time step the ML ensemble will also generate inferences relative to the most recent data value i.e., with quantized predefined value range around the latest known data value. That way one can track over time varying input data values.

[0079] For improved reliability there is a concept to run an ensemble of ML instances in parallel, where at each input signal a different Gaussian noise signal is added. The related different output signals are illustrated at the right side of Fig. 2. It can be seen that the output signals have a different range for the same input signal (different columns in the figure). In case of similar inference values for all instances of the ML ensemble one may assume an accurate inference.

[0080] According to some example embodiments of the invention, this type of ensemble learning is extended. Namely, a set of predefined input values may be used covering the whole possible signal range for the input signal or a substantial part thereof. Then the ML selector will use the inference fitting to the estimated input signal as the related output value (one value out of one of the columns).

[0081] The concept of pre-calculating inferences for several input values before the next actual value is received trades processing delay against processing overhead. In addition, it is a tradeoff between overhead and accuracy with an error similar to a quantization error. In case the set {... ,Apk, ... }K is large (comprises many input values) then the difference signal between the optimum Apkand the actual value at time ti may be small, but a large set of input values requires a high number of parallel ML instances. As a first alternative one might consider one single ML model, but considering in parallel many input options. At the time when the actual (really measured) value is available then the inference ‘collapses’ in the end into one decision based on the later arriving actual value. For that purpose, one might benefit from the typically sparse representations of DNNs, i.e., the high level of unused capacity in DNNs. That is, in the previous example embodiment, N parallel ML instances run in parallel with according processing complexity. In this alternative example embodiment, there is only a single neural network, but this single NN calculates inferences for multiple input values in parallel. One may then in later NN layers input a later received actual value so that the NN downselects the inference fitting best to the correct actual value. This alternative option is a more integrated solution and might be more efficient, as parts of the NN or weights are the same for all the input values

[0082] In some example embodiments, the pre-calculated inference corresponding to the input value being closest to the received actual value is taken as the estimated inference. In some example embodiments, some interpolation between the pre-calculated inferences is used to estimate the actual inference. The interpolation might be e.g. a linear or spline interpolation of the ML inferences for the quantized input values closest to the actual value at time ti.

[0083] In some example embodiments, some other side information may be additionally used to estimate the actual inference based on the pre-calculated inference. These example embodiments may or may not use interpolation such as linear or spline interpolation.

[0084] In some example embodiments, classical ensemble ML inferences for increased reliability are combined with the pre-calculation of the inferences and selection, as described above. For example, if close by input values generate close by inference results one may assume a reliable inference. Simultaneously, selecting the best fitting ML output achieves the low latency inference. The combination of ensemble and low latency ML instances helps to limit overall processing power. That is, if the inferences outputted for relatively close input values are close to each other (or in other ways look reasonable for the given problem) then one may assume that the calculated inferences are reliable. Otherwise, if these inferences are spread (randomly or over a large range), this will be an indication that the ML model is quite unreliable in this situation.

[0085] In some example embodiments, it may happen that some of the K inferences are precalculated sufficiently early before the next actual value will be received (or is expected to be received in case the actual values are received with some regularity (e.g. periodically)). In these embodiments, the respective ML model instances may use the time to pre-calculate further inference based on other input values. Thus, either the input values might be more dense, and / or the range of input values may cover a larger range of values.

[0086] The pre-calculation of the inferences and the selection may be performed for every n,hactual value, wherein n is a natural number which may be equal to 1 or larger than 1 .

[0087] The number of parallel inferred signal values of the input set { ... ,Apk, ... }K, defines the overall processing overhead. For that reason, in some example embodiments, these parameters may be tuned assuming a smooth signal variation over time:

[0088] • For example, one may attempt to cover by the input values the whole range expected for the actual value at the time ti. In the same gist, but less strict, one may attempt to cover a predetermined fraction (e.g. 50%, or 75%, or 90%) of the value range expected for the actual value. This obviously requires some knowledge about the underlying process, i.e., how fast the signal might evolve over time in the worst case.

[0089] • Fig. 3 illustrates an adaptation of the input signal range over multiple time slots according to some example embodiments of the invention. The dots indicate actual values being received at a certain time instance. Then a data range around this actual value may be used as input values for the next iteration of pre-calculating inferences and selecting. For example, the data range may be symmetric (or may have a predefined asymmetry) around the actual value for the next loop of multiple parallel inferences. This is then redone at each new time instance where an actual value is received, so that a floating of the signal range around the mid points is being achieved and the quantization error remains mainly constant. In case the variations of the actual values are not linear with known distributions, then uneven value ranges might be defined for the input values based on the respective actual value. For example, there might be a limited and known parameter space for MPC parameters. Assuming an iterative optimization loop of the MPC parameters then the value range is even more decreasing from iteration to iteration and therefore the signal range around the latest actual value should be reduced as well. In case of a distributed inference at UE and gNB (at O-DU interface), there might be multiple UE feedbacks over a varying parameter range so that these varying parameter ranges should be reported as side condition from UE to gNB (or vice versa). • Another option would be for the gNB to infer the range of input values, for example, from an outer loop ML. This would facilitate to employ example embodiments of the invention for legacy devices.

[0090] • In some example embodiments, the range of input values for the inference precalculation may be modified based on gNB compute power, power efficiency requirements and the UE mobility characteristics. E.g., stationary UEs are only subject to slow fading variations. Hence the number of inferences to be pre-calculated may be dependent on the time elapsed since the last parameter set has been inferred. If the elapsed time is relatively short, most of the previously calculated inferences might still be valid and need not to be recalculated.

[0091] • For mobile UEs with several serving cells, the compute load could be scaled so that the carrier with the best radio conditions reported in the past may have more input values (a more dense distribution of input values and / or a wider range of input values) compared to a carrier with worse radio conditions.

[0092] • It may happen that the actual value is outside the range of the input values. In this case, one might use an extrapolation.

[0093] • In some example embodiments, one may determine a difference between the closest input value and the actual value used for the selection. In particular if the ML inference is quite sensitive to variations of the input value, it may be beneficial to re-calculate the ML inference based on the actual value if the absolute value of the difference is too large. This leads to a trade off between additional latency, e.g., only for certain stages of the DNN, versus reliability of the estimated inference.

[0094] • Fig. 4 shows another example embodiment, which is particularly useful when the actual values have a faster repetition rate than the minimum required inference calculation time per ML instance. In that case, one may provide multiple parallel ML ensembles each operating according to the above described principle of pre-calculating plural inferences and selecting. In operation, one switches for each received actual value from one ML ensemble to another ML ensemble such that each ML ensemble pre-calculates inferences for every nth (n>1) actual value only. Each ML ensemble may use the most recent actual value to generate its input values for pre-calculation. The multiple parallel ML ensembles may start their inference calculation at different time slots based on the latest input values for this time slot. That way also the output inferences will be available with the same rate as the input values.

[0095] Fig. 4 (right part) shows in addition that the inferences calculated by the ML ensembles may be input into another ML instance, such as an ML instance for precoding, as shown in Fig. 4. In some example embodiments, only the selected inference from each of the parallel ML ensembles is input into the other ML instance (e.g. ML instance for precoding). In some example embodiments, all of the inferences calculated by the parallel ML ensembles are input into the other ML instance, wherein the selected inference is marked accordingly. The other ML instance may then use the additional information of the non-selected inferences to improve its own inferences.

[0096] The pre-calculated inference based on which the actual inference is estimated preferably should be selected as soon as the real measured signal value is reported. The selection might be very fast, e.g. in the ns to me range. The delay for the inference of a single ML instance might be in the range of 10, 20, or even 100 ms or even more depending on the underlying application. Correspondingly, the latency reductions might be quite large and given a sufficient processing power might be in the range of a factor of 100, 1000 or even more.

[0097] In some example embodiments, the factor of latency reduction may depend on other side conditions, like cell load, number of active users, and others. In such scenarios the number of ML instances and the achievable latency might be floating over time. In case of a distributed ML inference over UE and gNBs / cloud then a corresponding signalling may be defined for controlling the currently required maximum latency. Preferably, such new UCI messages may be standardized. Similarly, floating latency requirements may help to minimize the UE processing power, but new ‘maxjatency’ messages might be required. Such a new ‘maxjatency’ message may define the maximum latency allowed for the UE to provide the inference of interest. This can be, for example, a CSI estimation of the radio channel. Based on the time a conventional ML instance needs for one inference, the UE can decide about the required number of parallel ML inferences in a time staggered way so that the maximum latency requirement is always met and the UE can report its inference (e.g., estimated CSI) before the maxjatency time limit is reached. Then, for a larger value of maxjatency, the UE might reduce the number of parallel ML instances and, thus, saves processing power.

[0098] In some example embodiments, UEs may inform gNBs about their low latency capabilities including maximum number of ML instances, maximum time for inference selection, etc.

[0099] Some example embodiments of the invention may be applied to applications combining high inference complexity and very low latency. Channel prediction based on accurate MPC parameter estimation is a candidate, as (state of today) the MPC parameter estimation for 50 or more MPCs leads to complex iterative optimization loops. At the same time the inference of the parameters for channel prediction becomes obsolete within milliseconds due to channel aging.

[0100] 1) Another example is JT CoMP or MU MIMO precoding. A multi-site JT CoMP system may benefit from close to ideal CSI in combination with channel prediction. Spectral efficiency gains of 100% or more might be achievable for such systems. On high level, such a scenario could be implemented in the following way, as shown in Fig. 5. In Fig. 5, W is the common JT CoMP precoder for multiple cells, V is a (fixed) grid of beamformers per cell. Note that Each UE estimates continuously its CSI (and parameter range, i.e. the range of values of the MPC parameters)

[0101] 2) Each UE reports continuously its CSI estimation and “parameter range new” to the multiple beamformers (marked as “V”). The CSI report may be transmitted on PUCCH. The PDCCH message ‘parameter range new’ is to adapt the CSI reporting to the varying channel conditions and is based, for example, on the gNB channel predictor.

[0102] 3) Based on N different CSI reports, the common JT CoMP precoder (marked as “W”) creates a precoding matrix. The precoder infers possible (future) CSI reports (as input values) and pre-computes K possible precoding matrices

[0103] 4) Based on the actual CSI received from UE, the BS selects the most appropriate precoding matrix.

[0104] Fig. 6 shows a message flow according to some example embodiments of the invention. gNB (may be implemented in the cloud) instructs UE to perform a UE specific subtask. For example, in some example embodiments, UE infers all MPC parameters. In some example embodiments, gNB infers the MPC parameters and the UE infers only the delta of these parameters over time. UE may use one or more ML instances and use a corresponding (or lower) number of input values. As a result, UE reports an updated value range of input parameters for the parallel ML instances.

[0105] Upon receipt, gNB updates the input parameter value range. In addition, it calculates parallel inferences for the updated value range. The input parameter may be CSI.

[0106] Meanwhile, UE measures and / or estimates most recent values, such as CSI, and reports them to gNB. gNB compares the reported CSI with the input parameters (CSIs) and selects an inference corresponding to the input parameter closest to the reported CSI. This inference is used for precoding. Thus, an improved precoding matrix is obtained with a minimum delay after the CSI report is received from UE. As a result, UE may receive an improved signal with a very low mismatch due to channel aging.

[0107] Fig. 7 shows an apparatus according to an example embodiment of the invention. The apparatus may be a ML model ensemble or an element thereof. Fig. 8 shows a method according to an example embodiment of the invention. The apparatus according to Fig. 7 may perform the method of Fig. 8 but is not limited to this method. The method of Fig. 8 may be performed by the apparatus of Fig. 7 but is not limited to being performed by this apparatus.

[0108] The apparatus comprises means for preparing 210, means for calculating 220, means for selecting 230, and means for determining 240. It may optionally comprisemeans for applying 250. The means for preparing 210, means for calculating 220, means for selecting 230, means for determining 240 and means for applying 250 may be a preparing means, calculating means, selecting means, determining means, and applying means, respectively. The means for preparing 210, means for calculating 220, means for selecting 230, means for determining 240 and means for applying 250 may be a preparer, calculator, selector, determiner, and applicant, respectively. The means for preparing 210, means for calculating 220, means for selecting 230, means for determining 240, and means for applying 250 may be a preparing processor, calculating processor, selecting processor, determining processor, and applying processor, respectively.

[0109] The means for preparing 210 prepares a set of plural input values based on a received first actual value of a quantity (S210). For each of the plural input values prepared in S210, the means for calculating 220 calculates a respective inference by a respective ML instance (S220). The calculating of S220 for at least two of the plural input values is performed fully or partly in parallel.

[0110] When a second actual value of the quantity is received, the means for selecting 230 selects one of the plural input values (S230). The selected one of the plural input values is closest to the second actual value of the quantity. The second actual value is received after the first actual value.

[0111] Based on the inference calculated for the selected one of the plural input values in S230, the means for determining 240 determines an estimated inference (S240). Optionally (marked by dashed lines in Figs. 7 and 8), the means for applying 250 applies the estimated inference to a parameter for influencing the quantity (S250). Fig. 9 shows an apparatus according to an embodiment of the invention. The apparatus comprises at least one processor 810, at least one memory 820 including computer program code, and the at least one processor 810, with the at least one memory 820 and the computer program code, being arranged to cause the apparatus to at least perform at least the method according to Fig. 8 and related description.

[0112] Some example embodiments are explained with respect to a 5G or 6G network. However, the invention is not limited to 5G or 6G. It may be used in other networks, too, e.g. in previous of forthcoming generations of 3GPP networks such as 4G or 7G, etc. The invention is not even limited to mobile communication networks but may be applied anywhere where latency of the inference process is considered to be critical. For example, it may be applied to non- 3GPP mobile communication networks, wired communication networks, industrial applications (such as automation), military applications (such as enemy detection) etc.

[0113] One piece of information may be transmitted in one or plural messages from one entity to another entity. Each of these messages may comprise further (different) pieces of information.

[0114] Names of network elements, network functions, protocols, and methods are based on current standards. In other versions or other technologies, the names of these network elements and / or network functions and / or protocols and / or methods may be different, as long as they provide a corresponding functionality.

[0115] If not otherwise stated or otherwise made clear from the context, the statement that two entities are different means that they perform different functions. It does not necessarily mean that they are based on different hardware. That is, each of the entities described in the present description may be based on a different hardware, or some or all of the entities may be based on the same hardware. It does not necessarily mean that they are based on different software. That is, each of the entities described in the present description may be based on different software, or some or all of the entities may be based on the same software. Each of the entities described in the present description may be deployed in the cloud.

[0116] According to the above description, it should thus be apparent that example embodiments of the present invention provide, for example, a base station (such as a gNodeB) or a component thereof, an apparatus embodying the same, a method for controlling and / or operating the same, and computer program(s) controlling and / or operating the same as well as mediums carrying such computer program(s) and forming computer program product(s). According to the above description, it should thus be apparent that example embodiments of the present invention provide, for example, a terminal (such as a UE) or a component thereof, an apparatus embodying the same, a method for controlling and / or operating the same, and computer program(s) controlling and / or operating the same as well as mediums carrying such computer program(s) and forming computer program product(s).

[0117] Implementations of any of the above described blocks, apparatuses, systems, techniques or methods include, as non-limiting examples, implementations as hardware, software, firmware, special purpose circuits or logic, general purpose hardware or controller or other computing devices, or some combination thereof. Each of the entities described in the present description may be embodied in the cloud.

[0118] It is to be understood that what is described above is what is presently considered the preferred example embodiments of the present invention. However, it should be noted that the description of the preferred example embodiments is given by way of example only and that various modifications may be made without departing from the scope of the invention as defined by the appended claims.

Claims

Claims:1 . Apparatus comprising: one or more processors, and memory storing instructions that, when executed by the one or more processors, cause the apparatus to perform: preparing a set of plural input values based on a received first actual value of a quantity; calculating, for each of the plural input values, a respective inference by a respective machine learning instance, wherein the calculating for at least two of the plural input values is performed fully or partly in parallel; selecting one of the plural input values, wherein the selected one of the plural input values is closest to a second actual value of the quantity, and the second actual value is received after the first actual value; determining an estimated inference based on the inference calculated for the selected one of the plural input values.

2. The apparatus according to claim 1 , wherein the instructions, when executed by the one or more processors, further cause the apparatus to perform: repeating a loop comprising the preparing, calculating, selecting, determining, and applying such that the second actual value of one loop becomes the first actual value of a subsequent loop.

3. The apparatus according to any of claims 1 and 2, wherein the one or more processors and the memory are configured such that the calculating for all of the plural input values is performed fully or partly in parallel.

4. The apparatus according to any of claims 1 to 3, wherein the one or more processors and the memory are configured such that the calculating for all of the plural input values is finished before a time when a reception of the second actual value is expected.

5. The apparatus according to any of claims 1 to 4, wherein the one or more processors and the memory are configured such that the estimated inference is determined based on the inference calculated for the selected one of the plural input values and the inference calculated for at least one of the plural input values different from the selected one of the plural input values.

6. The apparatus according to any of claims 1 to 4, wherein the one or more processors and the memory are configured such that the estimated inference is determined to be equal to the inference calculated for the selected one of the plural input values.

7. The apparatus according to any of claims 1 to 6, wherein the one or more processors and the memory are configured such that the set of plural input values is prepared such that it covers a predetermined fraction of a range of values expected for the second actual value.

8. The apparatus according to any of claims 1 to 7, wherein the one or more processors and the memory are configured such that, if it is expected to receive a respective actual value of the quantity regularly in time, the second actual value is the n,hreceived actual value after the first actual value is received, and n is a natural number equal to or larger than 1.

9. The apparatus according to any of claims 1 to 8, wherein the instructions, when executed by the one or more processors, further cause the apparatus to perform: checking if an absolute value of a difference between the selected one of the plural input values and the second actual value exceeds a predetermined threshold; calculating, for the second actual value, a respective inference by one of the machine learning instances if the absolute value of the difference exceeds the predetermined threshold; wherein the estimated inference is determined as the inference calculated for the second actual value.

10. The apparatus according to any of claims 1 to 9, wherein the instructions, when executed by the one or more processors, further cause the apparatus to perform: adapting a number of the plural input values based on a condition, wherein the condition comprises at least one of an installed processing power, a processing power available for the preparing, calculating, selecting, and applying, and a number of estimated inferences to be applied in parallel.

11. The apparatus according to any of claims 1 to 10, wherein the instructions, when executed by the one or more processors, further cause the apparatus to perform: applying the estimated inference to a parameter for influencing the quantity.

12. The apparatus according to any of claims 1 to 11 , wherein the one or more processors and the memory are configured such that the first and second actual values are received from aproviding device, and the instructions, when executed by the one or more processors, further cause the apparatus to perform: informing the providing device on at least one of• a number of input values for which the apparatus may calculate the respective inferences fully or partly in parallel;• a maximum time required for selecting the one of the plural input values;• a maximum time required for determining the estimated inference after the one of the plural input values is selected; and• a maximum overall time for performing the preparing, calculating, selecting, and determining depending on a number of input values.

13. Method comprising: preparing a set of plural input values based on a received first actual value of a quantity; calculating, for each of the plural input values, a respective inference by a respective machine learning instance, wherein the calculating for at least two of the plural input values is performed fully or partly in parallel; selecting one of the plural input values, wherein the selected one of the plural input values is closest to a second actual value of the quantity, and the second actual value is received after the first actual value; determining an estimated inference based on the inference calculated for the selected one of the plural input values.

14. The method according to claim 13, further comprising: repeating a loop comprising the preparing, calculating, selecting, determining, and applying such that the second actual value of one loop becomes the first actual value of a subsequent loop.

15. The method according to any of claims 13 and 14, wherein the calculating for all of the plural input values is performed fully or partly in parallel.

16. The method according to any of claims 13 to 15, wherein the calculating for all of the plural input values is finished before a time when a reception of the second actual value is expected.

17. The method according to any of claims 13 to 16, whereinthe estimated inference is determined based on the inference calculated for the selected one of the plural input values and the inference calculated for at least one of the plural input values different from the selected one of the plural input values.

18. The method according to any of claims 13 to 16, wherein the estimated inference is determined to be equal to the inference calculated for the selected one of the plural input values.

19. The method according to any of claims 13 to 18, wherein the set of plural input values is prepared such that it covers a predetermined fraction of a range of values expected for the second actual value.

20. The method according to any of claims 13 to 19, wherein, if it is expected to receive a respective actual value of the quantity regularly in time, the second actual value is the n,hreceived actual value after the first actual value is received, and n is a natural number equal to or larger than 1 .

21. The method according to any of claims 13 to 20, further comprising: checking if an absolute value of a difference between the selected one of the plural input values and the second actual value exceeds a predetermined threshold; calculating, for the second actual value, a respective inference by one of the machine learning instances if the absolute value of the difference exceeds the predetermined threshold; wherein the estimated inference is determined as the inference calculated for the second actual value.

22. The method according to any of claims 13 to 21 , further comprising: adapting a number of the plural input values based on a condition, wherein the condition comprises at least one of an installed processing power, a processing power available for the preparing, calculating, selecting, and applying, and a number of estimated inferences to be applied in parallel.

23. The method according to any of claims 13 to 22, further comprising: applying the estimated inference to a parameter for influencing the quantity.

24. The method according to any of claims 13 to 23, wherein the first and second actual values are received from a providing device, and the method further comprises: informing the providing device on at least one of• a number of input values for which the method may calculate the respective inferences fully or partly in parallel;• a maximum time required for selecting the one of the plural input values;• a maximum time required for determining the estimated inference after the one of the plural input values is selected; and• a maximum overall time for performing the preparing, calculating, selecting, and determining depending on a number of input values.

25. A computer program product comprising a set of instructions which, when executed on an apparatus, is configured to cause the apparatus to carry out the method according to any of claims 13 to 24.

26. The computer program product according to claim 25, embodied as a computer-readable medium or directly loadable into a computer.

Images