Distribution of measured data recording into plural phases of training of machine learning model

Abstract

To provide: a method of distributing a predetermined group of measured data recording for use in training of a machine learning model to a plurality of predetermined different phases of the training; a program; and a computer.SOLUTION: A method according to the present invention has the steps of: specifying a sequence of a reference point which covers a space of measured data recording and does not overlap with the measured data recording; specifying to which reference point one or more measured data recording among a predetermined group of the measured data recording are located closest by using a predetermined distance measure and associating the measured data recording with the reference point; distributing the reference point to a predetermined phase of training so that one or more reference points should be associated with each of phases of the training: and associating the measured data recording associated with the reference point with the phase associated with each of the reference points of the training.SELECTED DRAWING: None

Description

[Technical Field]

[0001] The present invention relates to training machine learning models, such as neural networks, for evaluating measurement data, for example within the framework of at least partially automated steering of a vehicle or robot. [Background technology]

[0002] Background technology In at least partially automated operation of a vehicle or robot on a company's premises or in public transport, it is assumed that the vehicle's or robot's surroundings are constantly monitored using measurement technology, and the measurement data thus obtained are evaluated with respect to planning the vehicle's or robot's further behavior. For this evaluation, machine learning models are used in particular. Such models are trained using a finite set of training examples with sufficient variability, and it is assumed based on the generalization power of the machine learning model that if the machine learning model provides accurate results for these training examples, it will also be able to accurately evaluate unknown data. Summary of the Invention [Problem to be solved by the invention]

[0003] A predetermined set of training examples can be used for training, for example, by optimizing a machine learning model based on a portion of these training examples, while another portion of the training examples can be suppressed during this optimization and used after the optimization to determine the performance of the machine learning model based on unknown measurement data. [Means for solving the problem]

[0004] Disclosure of the Invention The present invention provides a method for distributing a predetermined set of measurement data records for training a machine learning model into multiple different predetermined phases of the training. Each measurement data record contains values ​​of one or more measurements. In this case, the term "record" can be considered analogous to the English term "record," which refers to a collection of consecutive measurement and other data values, comparable to the contents of an index card. The intuitive term "dataset" is already used in the linguistic idiom of the machine learning field to refer to a collection of multiple records, comparable to a card box filled with multiple index cards. The measurement data records may include, in particular, sensor data recorded by any type of sensor, such as a measuring device, camera, radar sensor, lidar sensor, or ultrasonic sensor for a specific measurement.

[0005] The term "machine learning model" refers in particular to a model that embodies a parameterized function, for example with adjustable parameters, with a high generalization power. When training a machine learning model, the parameters can be adjusted in particular so that when training data are input to the model, a target output belonging to these training data is reproduced as well as possible. A machine learning model may in particular include and / or be an artificial neural network (KNN (English: ANN)).

[0006] Within the scope of this method, a sequence of reference points is identified that covers the space of the measurement data records and does not overlap with the measurement data records. This covering can be, in particular, a uniform covering. A uniform covering means, in particular, that, for example, there is neither a significant frequency of reference points nor a significant region without reference points in the space of the measurement data records. The space of the measurement data records can, in particular, have many more dimensions than the three dimensions of a Cartesian space. To this extent, this space can be interpreted as a hyperspace.

[0007] For one or more measurement data records from a predetermined set of measurement data records, it is determined using a predetermined distance measure to which reference point the measurement data record is located closest. The measurement data record is then associated with the reference point. In this case, therefore, in particular, one or more measurement data records can be associated with one reference point, in which case: One or more measurement data records are not associated with any reference point; and / or No measurement data records are associated with one or more reference points. is perfectly acceptable.

[0008] The distance measure may in particular be, for example, the Euclidean distance measure, but it is also possible to use any other distance measure, for example the 1-norm.

[0009] The reference points are distributed to predetermined phases of training so that one or more reference points are associated with each phase of training. The measurement data records associated with each reference point are also associated with the phase of training associated with that reference point. Figuratively speaking, each reference point receives a color when distributed to a phase of training, and all measurement data records associated with that reference point also receive this color.

[0010] It has been found that in this way measurement data records that are similar to one another and thus located adjacent to one another in the space of measurement data records are preferably associated with the same reference point, which results in these measurement data records being associated with the same phase of training.

[0011] Regardless of the specific configuration of the machine learning model, and regardless of whether the training is performed in a supervised or unsupervised manner, the training may include, among other things, e.g., at least one optimization phase in which parameters describing the behavior of the machine learning model to be trained are optimized; At least one test phase in which the results of the optimization phase are controlled; This may include, for example, "k-fold cross-validation", in which the set of available measurement data records is distributed into k equal partitions. Then, one of these partitions i=1,...k is always alternately used for the testing phase and the remaining part of the measurement data records for the optimization phase. In common linguistic usage in the field of machine learning, the optimization phase is generally referred to as "training". However, in the context of the present invention, it is expedient to understand training as a unit consisting of optimization and subsequent testing.

[0012] The division of training into an optimization phase and a test phase particularly aims to counteract the tendency toward so-called "overfitting." In particular, if only a relatively small number of measurement data records are provided, relative to the size and complexity of the machine learning model's architecture, the machine learning model may take the path of least resistance in the optimization phase, optimizing the model in a way that approximates "memorization" of the presented measurement data records rather than the general knowledge extraction embodied herein. Testing on measurement data records not seen in the optimization phase should, in particular, confirm whether the trained machine learning model is truly capable of generalization or has merely "memorized" without generalization capabilities. For the test to be convincing in this regard, the measurement data records presented in the optimization phase, on the one hand, and the measurement data records presented in the test phase, on the other hand, must be sufficiently different from each other. If substantially equivalent measurement data records are presented in both phases, the trained machine learning model may pass the test with good results even if it has merely "memorized" them.

[0013] Therefore, preferably, the same or similar measurement data records are provided for the same phase of training, which simultaneously promotes the existence of different measurement data records in different phases of training. This makes it impossible for a machine learning model that simply memorizes data to achieve good test results. That is, the error measured in testing is convincing for the generalization ability of the trained machine learning model.

[0014] The performance measured in this way is itself an essential criterion that the trained machine learning model can be used in a concrete application, for example for controlling a technical system such as a vehicle or a robot. A good test result roughly corresponds to a badge awarded after passing a major inspection, with which the vehicle is released for use in road traffic. Therefore, the testing, and thus the distribution of the entirely available measurement data records preceding the testing, into different training phases is not simply a mathematical distribution of elements into different sets as such, but is an integral part of the entire process that begins with the provision of a given set of measurement data and culminates in the use of a completed machine learning model in a concrete technical application.

[0015] The measurement data records can optionally be annotated with a target output ("label"), which is the output that should ideally be provided by the trained machine learning model when presented with each measurement data record. Training can then be performed as "supervised" training, in the sense that deviations between the machine learning model's output and each target output are evaluated as errors and potentially used as feedback for further optimization. Since "labeling" measurement data records is often an expensive manual process, "labeled" measurement data records for supervised training are a scarce resource in many applications. Labels can be associated with individual measurement data records, but may also be associated with, for example, trajectory segments of a time series of measurement data records.

[0016] Between the optimization phase and the testing phase, a validation phase, for example, can be optionally inserted. Thus, for example, the optimization phase can be repeated using different hyperparameters that define the topology and / or size of the machine learning model or the optimization strategy, and the validation phase can examine how each change in the hyperparameters affects the performance of the machine learning model. In this case, the finally trained machine learning model can be tested in a testing phase, again using other data, and released for use.

[0017] Particularly advantageously, each measurement data record from a predetermined set of measurement data records is associated with one reference point, and the available stock of measurement data records is then fully distributed over the different phases of training and thus optimally utilized.

[0018] A predetermined set of measurement data records can in particular form, for example, a time series or a sequence, in which case a clue as to which measurement data records are similar and therefore should be associated with the same phase of training already becomes apparent from the time series or sequence, respectively.

[0019] In particular, if the measurement data records form a time series, the measurement data records can be expanded by one or more further components indicative of the history of the time series, for example in a pre-processing step. A sequence of reference points in the space of the expanded measurement data records can then be identified. The further components indicative of the history of the time series can, for example, emerge from a low-pass filtered time series of one or more measured quantities. In this way, for example, to identify the output of a machine learning model, only measurement data records within a certain past time range can be taken into account, while further past measurement data records remain unaccounted for.

[0020] If the measurement data records form a sequence that, unlike a time series, only indicates the order in which the measurement data records are recorded and does not need to be annotated with time information, in a further advantageous embodiment, when associating the measurement data records with a reference point, it is facilitated to associate multiple measurement data records that are consecutive in the sequence of measurement data records with the same reference point. For example, if the measurement data is detected according to a test plan, this test plan is essentially tailored to exhaust a specific sequence for efficient measurement. Such test plans also include stationary phases in which consecutive measurement data records are in many cases more similar to each other than measurement data records located farther apart in the sequence. Determining the similarity in this way may be more efficient than directly calculating multiple distances to reference points in a multidimensional space.

[0021] In a further particularly advantageous embodiment, the section of the time series having the measurement data records located respectively closest to the reference point is associated with this reference point. In this way, the time series of measurement data records can be divided into different sections for use in different training phases, with the measurement data records still in the correct order within these sections. This makes it possible, in particular, to model transient system behavior, in which the value of a quantity of interest at time t to be predicted by the machine learning model depends on the evolution (history) of the measurement data records within a time range prior to time t.

[0022] In this case, advantageously, only sections of the time series having a predetermined minimum length, measured in terms of time and / or number of measurement data records, can be associated with the reference point and / or with the training phase. That is, sections that are too short can be excluded already when associating the measurement data records with the reference point, but can also only be excluded when associating the measurement data records with the training phase. The required minimum length can be determined, in particular, based on, for example, a time constant by means of which the respective technical system considered reacts to changes in its operating conditions embodied in the measurement data records.

[0023] In a further particularly advantageous embodiment, only measurement data records from a space-filling subset (SFS) of measurement data records of a predetermined set of measurement data records are associated with the reference point. In this way, the set of measurement data records can be reduced to a substantial set that characterizes the behavior of the considered technical system via the measurement data. In this case, in particular, an existing SFS can be distributed to several different training phases, for example. However, it is also possible, for example, to create a new SFS in the same work process and distribute it to several different training phases.

[0024] For this purpose, it is possible, in particular, for example, to provide that no measurement data record or the measurement data record located closest to the respective reference point is associated with it, i.e., that in this case there may be reference points that do not have any associated measurement data record, which is entirely acceptable.

[0025] In a particularly advantageous embodiment, a random or pseudo-random sequence of reference points in the space of the measurement data record is determined, so that the reference points are irregularly spaced from one another, so that the regularity of this grid does not induce artifacts, as would be the case if the space were covered by a regular grid.

[0026] Particularly advantageously, a Sobol sequence is identified as the sequence of reference points, characterized in that later reference points are located in the gaps between earlier reference points, so that as the sequence progresses, space is covered more and more densely.

[0027] In a further particularly advantageous embodiment, the sequence of reference points is partially distributed to predetermined phases of training. In particular, in connection with a Sobol sequence as the sequence of reference points, this has the advantage that each of the intervals also forms a space-filling sequence. That is, for example, if the sequence of reference points has M reference points, then for example, the first p reference points can be assigned to the optimization phase of training, and the remaining (1-p) reference points can be assigned to the test phase of training.

[0028] In a further particularly advantageous embodiment, in a preprocessing step, the measurement data records are scaled to a hypercube in which all coordinates take values ​​within the same value range (e.g., between 0 and 1). In this way, when calculating the distance measure determined by the correspondence between the measurement data records and the reference points, all measured quantities occurring in the measurement data records are treated identically, regardless of the number of digits of the values ​​of these measurement data records. That is, a measured quantity does not acquire a particularly large significance in the distance measure solely due to the fact that all measured values ​​of the measured quantity vary in absolute terms by a particularly large number of digits, nor does a measured quantity substantially deviate from the distance measure solely due to the fact that all measured values ​​of the measured quantity vary in absolute terms by a particularly small number of digits. Otherwise, the significance of the individual measured quantities would depend on the choice of measurement unit (e.g., ohms or kiloohms, volts or millivolts).

[0029] In a further particularly advantageous embodiment, the machine learning model is trained in a given training phase using the measurement data records respectively associated with that phase, such that an improved distribution of the measurement data records to the training phases is reflected in training success, which promises better generalization to data not seen during training.

[0030] In a further particularly advantageous embodiment, the machine learning model trained in this way is supplied with further measurement data records recorded by at least one sensor. A drive signal is determined from the output of the machine learning model. A vehicle, a driver assistance system, a robot, a quality control system, an area monitoring system, and / or a medical imaging system are driven by the drive signal. In this way, the probability that the reaction of the respective driven technical system to the drive signal for the situation embodied in the further measurement data records is appropriate is advantageously increased due to the improved generalization power of the machine learning model.

[0031] The method may in particular be fully or partly computer-implemented. The invention therefore also relates to a computer program comprising machine-readable instructions for causing one or more computers and / or computing instances to carry out the above method when executed on the computer program. In this sense, control units for vehicles and control units for embedded systems for technical devices that are likewise capable of executing machine-readable instructions should also be considered as computers. The computing instances may for example be virtual machines, containers or serverless execution environments that may be provided in particular in the cloud.

[0032] The invention likewise relates to a machine-readable data medium and / or a download product comprising a computer program, which is a digital product that can be transmitted over a data network, i.e. downloaded by a user of the data network, and which can be sold for immediate download, for example, in an online shop.

[0033] Additionally, a computer program, machine-readable data medium, or downloadable product may be implemented on one or more computers and / or computing instances.

[0034] Further improvements of the invention are set out in more detail below in conjunction with the description of preferred embodiments of the invention based on the drawings. [Brief explanation of the drawings]

[0035] [Figure 1] FIG. 1 illustrates an embodiment of a method 100 for distributing a predetermined set of measurement data records 2 into a number of different predetermined phases 1a, 1b, 1c of training a machine learning model 1. [Figure 2]FIG. 1 illustrates the complete distribution of a set of measurement data records 2 into an optimization phase 1a of training and a testing phase 1c. [Figure 3] FIG. 1 illustrates the allocation of only a space-filling subset SFS of the set of measurement data records 2 to the optimization phase 1a and the test phase 1c of training. DETAILED DESCRIPTION OF THE INVENTION

[0036] Example 1 is a schematic flow chart of an embodiment of a method 100 for distributing a predetermined set of measurement data records 2 for training a machine learning model 1 into a number of different predetermined phases 1a, 1b, 1c of said training. Each measurement data record 2 contains values ​​of one or more measurements. The measurement data records belong to a space 3.

[0037] A predetermined set of measurement data records 2 can form a time series and / or a sequence according to block 105. Then, for example according to block 105a, in a pre-processing step, the measurement data records 2 can be expanded by one or more further components that describe the history of the time series. The measurement data records 2' thus expanded belong to a new space 3'.

[0038] In a pre-processing step, according to block 106, the measurement data record 2 can be scaled into a hypercube in which all coordinates take values ​​within the same range of values. Again, the modified measurement data record 2' occurs in a new space 3', but this new space 3' has the same dimensions as the original space 3.

[0039] In step 110, a sequence of reference points 4 that covers the space 3 of the measurement data record 2 and that does not overlap with the measurement data record 2 is identified.

[0040] If the measurement data record 2 has been expanded into an expanded measurement data record 2' in a space 3' according to block 105, then the sequence of reference points 4 in this space 3' can be identified according to block 111.

[0041] According to block 112, a random or pseudo-random sequence of reference points 4 in the space 3 of the measurement data record 2 can be identified. In particular, for example, according to block 113, a Sobol sequence can be identified as the sequence of reference points 4.

[0042] In step 120, for one or more measurement data records 2 from a predetermined set of measurement data records 2, it is determined to which reference point 4 the measurement data record 2 is located closest using a predetermined distance measure 5. The distance 5a and, accordingly, the closest reference point 4 are obtained. In step 130, the measurement data records 2 are respectively associated with the closest reference point 4.

[0043] According to block 131 , each measurement data record 2 from a given set of measurement data records 2 can be associated with one reference point 4 .

[0044] When associating a measurement data record 2 with a reference point 4 according to block 132, it may be facilitated to associate multiple measurement data records 2 that are consecutive in the sequence of measurement data records 2 with the same reference point 4.

[0045] According to block 133, for example, only measurement data records 2 from a subset SFS of a given set of measurement data records 2 that covers a space 3 of measurement data records 2 can be associated with a reference point 4.

[0046] If a time series of measurement data records 2 exists according to block 105, then according to block 134, the section of the time series having the measurement data records 2 located respectively closest to the reference point 4 can be associated with this reference point 4. Then, in particular, only the sections of the time series having a predetermined minimum length, measured in terms of time and / or number of measurement data records 2, can be associated with the reference point 4, for example according to block 134a.

[0047] In step 140, the reference points 4 are distributed to the given phases 1a, 1b, 1c of the training such that one or more reference points 4 are associated with each phase 1a, 1b, 1c of the training.

[0048] If, according to block 134, the sections of the time series having the measurement data records 2 located closest to the reference point 4 are associated with that reference point 4, then, according to block 141, for example, only sections of the time series having a predetermined minimum length measured in terms of time and / or number of measurement data records 2 can be associated with phases 1a, 1b, 1c of the training.

[0049] According to block 142, the sequence of reference points 4 can be distributed in parts into predetermined phases 1a, 1b, 1c of training.

[0050] Follow step 143: at least one optimization phase 1a in which parameters characterizing the behavior of the machine learning model 1 to be trained are optimized; At least one test phase 1c in which the results of optimization phase 1a are controlled and 1c can be selected as training phases 1a, 1b, and 1c. Between these phases, for example, a validation phase 1b can be performed, which can be used to examine, for example, the dependence of the performance of the machine learning model 1 on hyperparameters, as described above.

[0051] In step 150, the phases 1a, 1b, 1c of the training associated with each reference point 4 are also associated with the measurement data records 2 associated with that reference point 4.

[0052] In the example shown in Figure 1, in step 160, a machine learning model 1 is trained in a given phase 1a, 1b, or 1c of training using measurement data records 2 associated with the respective phases 1a, 1b, or 1c. The completed state of the machine learning model is marked with the reference symbol 1*.

[0053] 1, further in step 170 the trained machine learning model 1* is supplied with further measurement data records 2 recorded by at least one sensor 6. In this case, the trained machine learning model 1* provides an output 1d. In step 180, a drive signal 180a is determined from this output 1d. In step 190, the vehicle 50, the driver assistance system 51, the robot 60, the quality control system 70, the area monitoring system 80 and / or the medical imaging system 90 are driven by the drive signal 180a.

[0054] FIG. 2 shows how a given set of measurement data records 2 can be perfectly distributed between the optimization phase 1a of training and the testing phase 1c.

[0055] In the example shown in Figure 2, the measurement data record 2 has been scaled to a (now two-dimensional) unit cube in which both coordinates can only take values ​​between 0 and 1. This results in a modified measurement data record 2' in a new space 3' with the same dimensions as the original space 3.

[0056] The reference points 4 in the space 3' are distributed among the two training phases 1a (optimization phase) and 1b (test phase). For this purpose, the interval of the sequence of reference points 4 in the optimization phase 1a and the remaining part of the sequence in the test phase 1b are written. Since the sequence of reference points 4 is a Sobol sequence, there are no frequent occurrences of reference points 4 belonging to the optimization phase 1a or the test phase 1b in the space 3'. Instead, a quasi-random mixture of these two categories of reference points 4 is performed, and the distinction is not shown in FIG. 2 for the sake of clarity.

[0057] A measurement data record 2 is assigned to either the optimization phase 1a or the test phase 1b depending on whether the measurement data record 2 is located closest to a reference point 4 belonging to the optimization phase 1a or to the test phase 1b. The sequence or time series of measurement data records 2 therefore alternates pseudo-randomly between the optimization phase 1a and the test phase 1b. However, in this case, there are always certain intervals of different lengths that belong exclusively to the optimization phase 1a or to the test phase 1b. In applications involving time series of measurement data records 2, it is necessary that the intervals of such a time series are at least long enough so that the phenomenon to be studied can appear in each interval.

[0058] FIG. 3 shows, using a similar set of measurement data records 2 as an example, how a space-filling subset SFS of this set of measurement data records 2 can be distributed between the optimization phase 1a and the test phase 1c of training.

[0059] 3, by scaling the measurement data record 2 to a unit cube, the modified measurement data record 2' also occurs within a unit cube in which both coordinates can only take values ​​between 0 and 1. The new space 3' in which this modified measurement data record 2' resides has the same dimensions as the original space 3.

[0060] Unlike in Fig. 2, only individual measurement data records 2' are associated with the optimization phase 1a or the test phase 1b. These measurement data records 2' belong to a space-filling subset SFS that was previously identified based on the same reference point 4, but in an otherwise arbitrary way. Like the sections of the sequence of measurement data records 2' identified in Fig. 2, these individual measurement data records 2' also alternate quasi-randomly between the optimization phase 1a and the test phase 1b.

Claims

1. A method (100) for distributing a predetermined set of measurement data records (2) for training a machine learning model (1) into a plurality of different predetermined phases (1a, 1b, 1c) of said training, each measurement data record (2) comprising values ​​of one or more measurements, the method comprising: a sequence of reference points (4) covering a space (3) of said measurement data record (2) and not overlapping said measurement data record (2) is identified (110); - for one or more measurement data records (2) from the predetermined set of measurement data records (2), it is determined (120) to which reference point (4) the measurement data record (2) is located closest using a predetermined distance measure (5) and the measurement data record (2) is associated (130) with this reference point (4); - distributing (140) reference points (4) to the predetermined phases (1a, 1b, 1c) of the training so that one or more reference points (4) are associated with each phase (1a, 1b, 1c) of the training; - associating (150) the measurement data records (2) associated with each reference point (4) with the phases (1a, 1b, 1c) of the training associated with that reference point (4); A method (100) comprising:

2. Each measurement data record (2) from the predetermined set of measurement data records (2) is associated (131) with one reference point (4); The method (100) of claim 1.

3. the predetermined set of measurement data records (2) forming a time series and / or sequence (105); 3. The method (100) of claim 1 or 2.

4. Associating the measurement data records (2) with the reference points (4) facilitates associating (132) successive measurement data records (2) in the sequence of the measurement data records (2) with the same reference point (4). The method (100) of claim 3.

5. In a pre-processing step, the measurement data record (2) is extended (105a) by one or more further components indicative of the history of the time series, A sequence of reference points (4) within the space (3') of the measurement data record (2') thus expanded is identified (111), The method (100) of claim 3 or 4.

6. A section of the time series having a measurement data record (2) located closest to each reference point (4) is associated with that reference point (4) (134). The method (100) of claim 3.

7. Only sections of the time series having a predetermined minimum length, measured in terms of time and / or number of the measurement data records (2), are associated (134a) with reference points (4) and / or with phases (1a, 1b, 1c) of the training (141), The method (100) of claim 6.

8. only the measurement data records (2) from a subset SFS of the predetermined set of measurement data records (2) that covers the space (3) of measurement data records (2) are associated (133) with a reference point (4); The method (100) of claim 1.

9. To each reference point (4), either no measurement data record (2) or the measurement data record (2) located closest to the reference point (4) is associated (133a); The method (100) of claim 8.

10. A random or pseudo-random sequence of reference points (4) within the space (3) of the measurement data record (2) is identified (112); The method (100) of any one of claims 1 to 9.

11. The Sobol sequence is identified (113) as the sequence of the reference point (4). The method (100) of any one of claims 1 to 10.

12. The sequence of reference points (4) is distributed (142) in parts into the predetermined phases (1a, 1b, 1c) of the training; The method (100) of any one of claims 1 to 11.

13. In a pre-processing step, the measurement data records (2) are scaled (106) into a hypercube in which all coordinates take values ​​within the same range of values. The method (100) of any one of claims 1 to 12.

14. at least one optimization phase (1a) in which parameters describing the behavior of the machine learning model (1) to be trained are optimized; at least one test phase (1c) in which the results of said optimization phase (1a) are controlled; is selected as the training phase (1a, 1b, 1c) (143), The method (100) of any one of claims 1 to 13.

15. In the predetermined phases (1a, 1b, 1c) of the training, the machine learning model (1) is trained (160) using measurement data records (2) associated with the respective phases (1a, 1b, 1c); The method (100) of any one of claims 1 to 14.

16. The trained machine learning model (1*) is fed (170) with further measurement data records (2) recorded by at least one sensor (6), A drive signal (180a) is identified (180) from the output (1d) of the machine learning model (1); a vehicle (50), a driving assistance system (51), a robot (60), a quality control system (70), an area monitoring system (80), and / or a medical imaging system (90) is driven (190) by the drive signal (180a); 16. The method (100) of claim 15.

17. 17. A computer program comprising machine-readable instructions that, when executed on one or more computers and / or computing instances, cause said one or more computers and / or computing instances to perform the method (100) of any one of claims 1 to 16.

18. A machine-readable data carrier and / or download product comprising a computer program according to claim 17.

19. One or more computers and / or computing instances comprising a computer program according to claim 17 and / or comprising a machine-readable data carrier and / or downloadable product according to claim 18.

Images