Method, device and program for predicting future quality of service of wireless communication link

Abstract

The present invention relates to a method, device and program for predicting the future quality of service of a wireless communication link, and more particularly to a device, method and computer program for predicting the future quality of service of a wireless communication link based on a predicted future environment model predicted using a time series prediction. The method includes determining multiple environment models of one or more active transceivers in an environment of a mobile transceiver at multiple time points. The method includes using a time series prediction of multiple environment models to determine a predicted future environment model of one or more active transceivers at a future time point. The method includes using a machine learning model to predict the future quality of service of a wireless communication link at a future time point. The machine learning model is trained to provide information about the predicted quality of service of a given environment model. The predicted future environment model is used as input to the machine learning model.

Description

Technical Field

[0001] The present invention relates to an apparatus, method and computer program for predicting the future quality of service of a wireless communication link based on a predicted future environment model predicted using time series forecasting. Background Art

[0002] Communication between mobile transceivers is an area of ​​research and development. For example, in vehicular applications, research is ongoing to improve both the performance and predictability of wireless communications between vehicles in a changing environment. For example, within the scope of collaborative driving, when the quality of service (QoS) conditions change, the prediction of the future quality of service (QoS) of the wireless communication link between two vehicles improves the functionality of vehicle applications. In fact, when predictive QoS (PQoS) is not provided, the application may only react to the changes and may therefore be limited to the lower bound performance of the communication system. The PQoS system can operate on the vehicle, the communication node using a radio access technology (RAT), such as LTE-V (Long Term Evolution of Vehicle Communications) or 5G-V2X (the fifth generation mobile communication system for vehicle-to-anything communications) in its independent mode or IEEE 802.11p (standard of the Institute of Electrical and Electronics Engineers of the United States). Combinations of these technologies can also be applied to multi-RAT systems. In such a PQoS system, vehicles can exchange information about the communication surroundings in order to provide PQoS.

[0003] In the literature, channel models (semi-stochastic, such as spatial channel models (SCM); and deterministic, such as ray tracing) provide estimates of path loss and interference from other communicating nodes. Statistical models can provide an idea of ​​some mapping between surrounding vehicles and quality of service. An example is shown in "Performance and Reliability of DSRC Vehicular Safety Communication: A Formal Analysis and IEEE 802.11pVANets: Experimental evaluation of packet inter-reception time provide such a model" by Ma, Chen, and Refai. The paper "Prediction of Packet Inter Reception Time for Platooning using Conditional Exponential Distribution" by Jornod, El Assad, Kwoczek, and Kürner provides a statistical connection between environmental density and inter-packet reception time. The paper also shows a method to divide the surrounding environment into circular areas to represent the traffic density around the transmitter. It uses the distance between transceivers to estimate the QoS of the link. Summary of the invention

[0004] An improved concept for predicting the quality of service of a wireless communication link between vehicles may be needed.

[0005] Embodiments are based on the discovery that previous methods for predicting quality of service focus on providing a prediction for a single point in time, rather than using a prediction method that tracks the gradual development of the environment of a mobile transceiver and, therefore, the gradual development of the quality of service of a wireless communication link between the mobile transceiver and another mobile transceiver. In an embodiment of the present disclosure, a future environment model of active transceivers around a mobile transceiver is predicted by performing a time series prediction based on a previously generated environment model. Based on the environment model predicted for a future time point (or multiple time points), a machine learning model is used to determine the quality of service at one or more future time points. Therefore, by using a time series prediction of the environment model, the development of the quality of service of the wireless communication link affected by changes in the environment is predicted.

[0006] An embodiment provides a method for predicting the future quality of service of a wireless communication link between a mobile transceiver and another mobile transceiver. The method includes determining multiple environmental models of one or more active transceivers in the environment of the mobile transceiver at multiple time points. The method includes using time series predictions of multiple environmental models to determine predicted future environmental models of one or more active transceivers at future time points. The method includes using a machine learning model to predict the future quality of service of the wireless communication link at future time points. The machine learning model is trained to provide information about the predicted quality of service of a given environmental model. The predicted future environmental model is used as an input to the machine learning model. By performing time series predictions, the radio environment of the mobile transceiver is modeled at future time points. The predicted future environmental model can in turn be used to predict the future quality of service of the wireless link via the machine learning model.

[0007] In various embodiments, time series prediction is performed based on a statistical fitting function or based on a temporal autocorrelation function. This statistical-based approach has a lower computational overhead.

[0008] Alternatively, another machine learning model can be used to perform time series forecasting.Time series forecasting based on machine learning may be useful in scenarios with a large number of related features and at a high computational cost.

[0009] For example, a time series forecast may be determined such that the progress of an environment model toward a predicted future environment model is predicted. In other words, data under multiple environment models may be extrapolated to a predicted future environment model.

[0010] For example, time series forecasting may generate a predicted future environment model. Future service quality is predicted based on the predicted future environment model. In other words, time series forecasting may not be applicable to service quality itself, but to the environment model that is the basis for the prediction of future service quality.

[0011] In various embodiments, the future quality of service of the wireless communication link is predicted for at least two future time points. For example, the predicted future environment model can be predicted for at least two future time points and then used to predict the future quality of service at at least two future time points.

[0012] For example, by determining a predicted future environment model of one or more active transceivers at at least two future time points, and using the predicted future environment model of one or more active transceivers at at least two future time points as an input to a machine learning model, the future quality of service of the wireless communication link at at least two future time points may be predicted. Thus, a prediction of the future quality of service may be provided on a timeline of future time points.

[0013] In an embodiment, a plurality of environmental models of one or more active transceivers are determined at a plurality of time points. The method may include determining a quality of service of a wireless communication link at a plurality of time points. The method may include training a machine learning model using the plurality of environmental models at a plurality of time points as training inputs and the quality of service of the wireless communication link at a corresponding plurality of time points as training outputs of the training of the machine learning model. Thus, the machine learning model may be trained on environmental models generated by and / or for the mobile transceiver.

[0014] For example, a machine learning model can be trained to implement a regression algorithm. A regression-based machine learning algorithm can be used to determine a value (within a certain range), such as future quality of service.

[0015] In various embodiments, a machine learning model can be trained to provide a probability distribution of predicted quality of service for a given environment model. Doing so can avoid situations where common quality of service prediction values ​​bias the prediction.

[0016] In some embodiments, one or more active transceivers are placed on a grid within the environment model. The grid may include multiple adjacent cells. One or more active transceivers may be clustered in each cell within the grid. For example, the grid may be used to facilitate maintenance of the environment model and limit the number of inputs to the machine learning model.

[0017] For example, the grid may be a circular grid. Transmissions by other active transceivers affect the wireless communication link based on their distance, which may be modeled by the circular grid.

[0018] Predicting the quality of service may involve at least one of inter-packet reception time, packet error rate, latency, and data rate. These are quality of service characteristics that can be predicted using the above-mentioned machine learning model.

[0019] The embodiments of the present disclosure also provide a computer program. When the computer program is executed on a computer, a processor or a programmable hardware component, the computer program has a program code for executing the above method.

[0020] Embodiments of the present disclosure also provide an apparatus for predicting the future quality of service of a wireless communication link between a mobile transceiver and another mobile transceiver. The apparatus includes one or more interfaces for communicating in a mobile communication system. The apparatus includes a control module configured to execute the above method. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Some other features or aspects will be described using the following non-limiting embodiments of an apparatus or method or computer program or computer program product, by way of example only, and with reference to the accompanying drawings, in which:

[0022] Figure 1a and Figure 1b A flow chart illustrating an embodiment of a method for predicting future quality of service of a wireless communication link between a mobile transceiver and another mobile transceiver;

[0023] Figure 1c A schematic diagram showing an apparatus for predicting future quality of service of a wireless communication link between a mobile transceiver and another mobile transceiver and a mobile transceiver (such as a vehicle) including the apparatus;

[0024] Figure 1d A flow chart illustrating an embodiment of a method for training a machine learning model;

[0025] Fig. 2 shows a table showing the development of service quality attributes over time in relation to an environment model;

[0026] Figure 3a to Figure 3d A schematic diagram related to the training of a machine learning model is shown. DETAILED DESCRIPTION

[0027] Various exemplary embodiments will now be described more fully with reference to the accompanying drawings, in which some exemplary embodiments are shown. In the accompanying drawings, the thickness of lines, layers or regions may be exaggerated for clarity. Optional components may be represented by broken lines, intermittent lines or dot-dash lines.

[0028] Therefore, although the exemplary embodiment is capable of various modifications and alternative forms, its embodiments are shown in the drawings by way of example and will be described in detail herein. However, it should be understood that the exemplary embodiments are not intended to be limited to the specific forms disclosed, but rather, the exemplary embodiments will cover all modifications, equivalents, and alternatives falling within the scope of the present invention. Throughout the description of the drawings, the same numbers refer to the same or similar elements.

[0029] As used herein, the term "or" refers to a non-exclusive "or" unless otherwise specified (e.g., "otherwise" or "or in the alternative"). In addition, as used herein, words used to describe the relationship between elements should be broadly interpreted to include a direct relationship or the presence of intermediate elements, unless otherwise specified. For example, when an element is referred to as being "connected" or "coupled" to another element, the element may be directly connected or coupled to the other element, or there may be intervening elements. In contrast, when an element is referred to as being "directly connected" or "directly coupled" to another element, there are no intervening elements. Similarly, words such as "between," "adjacent," etc. should be interpreted in a similar manner.

[0030] The terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments. As used herein, the singular forms "a", "an", and "the" are also intended to include plural forms, unless the context clearly indicates otherwise. It should also be understood that when used herein, the terms "include", "comprise", or "contain" specify the presence of stated features, integers, steps, operations, elements, or parts, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, parts, or combinations thereof.

[0031] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the exemplary embodiments belong. It should also be understood that terms (such as those defined in commonly used dictionaries) should be interpreted as having a meaning consistent with their meaning in the context of the relevant art, and will not be interpreted in an idealized or overly formal sense unless explicitly defined as such herein.

[0032] Figure 1a and Figure 1b A flow chart of an embodiment of a method for predicting a future quality of service of a wireless communication link between a mobile transceiver 100 and another mobile transceiver 102 is shown. The method includes determining 110 a plurality of environmental models of one or more active transceivers 104 in an environment of the mobile transceiver at a plurality of time points. The method includes determining 120 a predicted future environmental model of the one or more active transceivers at a future time point using a time series prediction 125 of the plurality of environmental models. The method includes predicting 130 a future quality of service of the wireless communication link at a future time point using a machine learning model. The machine learning model is trained to provide information about the predicted quality of service for a given environmental model. The predicted future environmental model is used 135 as an input to the machine learning model.

[0033] Figure 1c A schematic diagram of a corresponding apparatus 10 for predicting a future quality of service of a wireless communication link between a mobile transceiver 100 and another mobile transceiver 102 and a mobile transceiver 100 (such as a vehicle) comprising the apparatus 10 is shown. The apparatus comprises one or more interfaces 12 for communicating in a mobile communication system. The apparatus comprises a control module 14, which is configured to perform Figure 1a , Figure 1b and / or Figure 1d In general, the control module 14 may provide the functionality of the device 10, for example in conjunction with one or more interfaces 12.

[0034] The following description involves Figure 1a and / or Figure 1b Methods and Figure 1cThe device combines both. Figure 1a and / or Figure 1b The features described by the method can also be applied to Figure 1c device.

[0035] Embodiments of the present disclosure relate to methods, apparatus and computer programs for predicting the future quality of service of a wireless communication link between a mobile transceiver and another mobile transceiver. In order to predict future service availability and QoS, it may be beneficial to have a good understanding of the radio environment. Typical characteristics of the radio environment may be path loss, interference conditions, load of the system, number of carrier frequencies, multiple radio access technologies (RATs), etc. The more detailed the radio environment is simulated, the greater the amount of information required for its classification.

[0036] As described above, the future QoS of the wireless communication link is determined. In this case, the prefix "future" indicates a prediction of the QoS at a certain point in time in the future. To achieve this, the presence (and therefore the activity) of one or more active transceivers can be predicted for the future (e.g., using the trajectory of one or more active transceivers), and the prediction of future QoS can be based on the prediction of future activities of one or more active transceivers. The predicted QoS may include one or more aspects, such as the predicted minimum, average and / or maximum data transmission rate on the wireless link, the minimum, average and / or maximum packet or bit error rate, the minimum, average or maximum time between two successful transmissions of a packet (e.g., also expressed as the inter-packet reception time PIR), the minimum, average or maximum delay, etc. In other words, the predicted quality of service may involve at least one of the inter-packet reception time, packet error rate, delay and data rate of the wireless communication link. Typically, the predicted QoS may indicate the expected performance and / or expected reliability of the wireless link.

[0037] In an embodiment, the future QoS of the radio link may be predicted by the mobile transceiver, ie by a receiver of the radio transmission on the radio link. Thus, the method may be performed by the mobile transceiver, and / or the mobile transceiver may include Figure 1c device.

[0038] The method includes determining 110 a plurality of environmental models of one or more active transceivers 104 in an environment of a mobile transceiver at a plurality of time points. Typically, the environmental model of the one or more active transceivers may model the environment of the mobile transceiver for the presence and / or transmission activity of the one or more active transceivers. For example, the environmental model may include and / or represent the location of the one or more active transceivers in the environment of the mobile transceiver. In various embodiments, the environmental model may be limited to a predetermined range around the mobile transceiver, such as a predetermined circular distance, or according to a predetermined size of a grid. For example, one or more active transceivers may be placed on a grid within the environmental model. The grid may include a plurality of adjacent cells. One or more active transceivers may be clustered in each cell within the grid. In other words, the location and / or distance of the one or more mobile transceivers may be represented by the cell of the grid in which they are placed.

[0039] For example, in some cases, the grid may be a two-dimensional rectangular grid. In other words, the cells of the grid may have a rectangular shape. In addition, each cell of the grid may have (substantially) the same size / dimension.

[0040] Alternatively, the grid may be a circular grid, such as a one-dimensional circular grid. In a one-dimensional circular grid, the cells of the grid are formed along a dimension (i.e., the distance from the center) so that a plurality of concentric circles form a grid, and the space between the circles is a cell of the grid. In a two-dimensional circular grid, each space between two adjacent circles is further subdivided, such as into quadrants. In other words, the grid may be formed by circles, and the space between the two circles is (in a one-dimensional circular grid) or includes (in a two-dimensional circular grid) a cell of the grid. If a two-dimensional circular grid is used, each space between the two circles may be divided into, for example, quadrants, so that there are multiple cells in the space between the two circles. Each circle of the circular grid may represent a distance. For example, an active transceiver placed in a cell between the center point of the circular grid (i.e., the location of the mobile transceiver) and the first circle (from the center point) may have a distance of at most n meters from the mobile transceiver, an active transceiver placed in a cell between the first circle and the second circle may have a distance of more than n meters and at most 2n meters from the active transceiver, and so on. Such a grid is applied to the table of FIG. 2.

[0041] In an embodiment, determining 110 multiple environmental models may include obtaining information about one or more active transceivers. For example, the method may include collecting information about the location of one or more active transceivers via wireless messages (e.g., wireless vehicle-to-vehicle messages, if the active transceivers are vehicles) of one or more active transceivers. For example, periodic status messages of one or more active transceivers may be processed to collect information about the location of one or more active transceivers. Therefore, the method may include receiving wireless transmissions of one or more active transceivers. The mobile transceiver may receive wireless transmissions of one or more active transceivers (e.g., other vehicles / transceivers). Based on the received wireless transmissions, the location of one or more active transceivers may be determined. For example, the wireless transmission may include information about the location of the active transceiver that has sent the corresponding wireless message. The method may include determining the environmental model of one or more active transceivers based on the received wireless transmissions of one or more active transceivers. More generally, the method may include generating an environmental model based on the determined location of one or more active transceivers.

[0042] In various embodiments, the method may include receiving at least a portion of at least a subset of the plurality of environment models from another mobile transceiver (e.g., from a mobile transceiver of one or more active transceivers). For example, after determining their environment models, the mobile transceivers may share information with other mobile transceivers, such as by broadcasting the corresponding environment models. In other words, the method may include broadcasting the environment models to other mobile transceivers.

[0043] Based on the information collectively collected by the mobile transceivers, each mobile transceiver / vehicle can perform a prediction of the future QoS of the wireless links it maintains with other mobile transceivers.

[0044] In an embodiment, multiple environmental models of one or more active transceivers are determined at multiple time points. Each of the multiple environmental models may represent one or more active transceivers in the environment of a mobile transceiver at one of the multiple time points. Multiple environmental models may be generated over time (e.g., periodically). Figure 2 shows an example of multiple environmental models generated at multiple time points (e.g., every 0.5 seconds). In the context of the present disclosure, the term "at or for one or more time points (in the future)" merely means that an action related to the one or more time points is performed, and does not necessarily mean that the action is performed exactly at the same one or more future time points. However, in some cases, the corresponding actions may be performed at the same time point, such as when the quality of service is determined (rather than predicted) at a specific time point.

[0045] The method includes determining 120 a predicted future environmental model of one or more active transceivers at a future time point using a time series prediction 125 of a plurality of environmental models. Typically, the time series prediction predicts the development of one or more numerical values ​​over a time interval (including a plurality of time points) based on historical data about one or more numerical values. In other words, the trend of one or more numerical values ​​may be predicted based on historical data about the numerical values, and the time series of the development of the numerical values ​​may be predicted. In an embodiment, one or more numerical values ​​may represent (i.e., form) an environmental model. For example, each environmental model may be represented by a plurality of numerical values. For example, each environment may include digital information about the number of active transceivers of each cell of a grid, wherein the number of active transceivers of each cell of the grid is a value for which a time series prediction is performed. Thus, a plurality of environmental models of one or more active transceivers in an environment of a mobile transceiver may be historical data about one or more numerical values ​​(representing a predicted future environmental model).

[0046] There are different methods to perform time series prediction. In some embodiments, a statistical-based method can be used. For example, time series prediction can be performed 125 based on a statistical fitting function or based on a time autocorrelation function. In the statistical fitting function, the trend of the numerical value can be predicted by performing a fit on historical data about the numerical value. Time autocorrelation is an autocorrelation function performed on a time series, which is used to predict the development of the time series at a future time point or time period. Both functions can be applied to the above-mentioned objectives. For example, a statistical fitting function or a time autocorrelation function can be applied to multiple environmental models to determine a predicted future environmental model (at a future time point, or at at least two future time points).

[0047] Alternatively, machine learning can be used to perform time series prediction. In other words, time series prediction 125 is performed using another machine learning model. Typically, as will be described later, the prediction of numerical values ​​via a machine learning model can be performed using a regression-based machine learning algorithm. In order to train another machine learning model, multiple environmental models (represented by numerical values) can be divided into training inputs, training outputs, and unused environmental models. For example, if the future environmental model is to be predicted at a time point outside a predetermined time interval, for each training sample, a subset of multiple environmental models can be selected as a training input for training another machine learning model, and an environmental model determined for a time outside a predetermined time interval from a subset of environmental models (in the future) can be used as a training output. This subdivision can be repeated in a sliding window manner on multiple environmental models. Based on the training, another machine learning model can be configured to generate a predicted future environmental model at a certain point in time in the future, which is a predetermined time interval input into the other machine learning model. Again, the other machine learning model can be used on a subset of multiple environmental models in a sliding window manner to determine the predicted future environmental model of at least two / more future time points (or more precisely, at least two / more predicted future environmental models). The same concept can be applied to statistics-based functions.

[0048] Both approaches attempt to predict the development of the environment model in the future (i.e., until a future point in time) so that the predicted future environment model can then be used to predict future quality of service. In other words, a time series prediction may be determined 125 such that the progress of the environment model towards the predicted future environment model is predicted. In any case, the prediction of the future environment model may attempt to model the movement of one or more active transceivers and the corresponding movement of the mobile transceiver (and another mobile transceiver).

[0049] Contrary to intuition, time series forecasting may not be applied directly to the prediction of future service quality (as this may not be very reliable without some additional input), but rather to the environment model that underlies the prediction of future service quality. In other words, time series forecasting may produce a predicted future environment model, where the future service quality is predicted based on the predicted future environment model.

[0050] The method includes using a machine learning model to predict a future quality of service of a wireless communication link at a future time point 130. The predicted future environment model is used as an input to the machine learning model 135. The output of the machine learning model may be the future quality of service of the wireless communication link, or the future quality of service of the wireless communication link may be predicted based on the output of the machine learning model.

[0051] Machine learning refers to algorithms and statistical models that can be used by computer systems to perform specific tasks without explicit instructions, but relying on models and reasoning. For example, in machine learning, data transformations inferred from analysis of historical and / or training data can be used instead of rule-based data transformations. For example, a machine learning model or a machine learning algorithm can be used to analyze the content of an image. In order for a machine learning model to analyze the content of an image, a machine learning model can be trained using training images as input and training content information as output. By training a machine learning model with a large number of training images and associated training content information, the machine learning model "learns" to recognize the content of the image, so the machine learning model can be used to recognize the content of the image not included in the training image. The same principle can also be used for other types of sensor data: by training a machine learning model using training sensor data and expected outputs, the machine learning model "learns" the conversion between sensor data and output, which can be used to provide output based on non-training sensor data provided to the machine learning model. In an embodiment, the machine learning model is trained to provide a conversion between an environmental model and the quality of service of a wireless communication link predicted according to the environmental model. In other words, the machine learning model is trained to associate the environmental model with the predicted quality of service of the wireless communication link.

[0052] The machine learning model is trained using training input data. The examples specified above use a training method called "supervised learning". In supervised learning, a machine learning model is trained using multiple training samples, each of which may include multiple input data values ​​and multiple expected output values, i.e., each training sample is associated with an expected output value. By specifying both the training samples and the expected output values, the machine learning model "learns" which output value to provide based on input samples similar to the samples provided during training. Supervised learning can be based on supervised learning algorithms, such as classification algorithms, regression algorithms, or similarity learning algorithms. A classification algorithm can be used when the output is restricted to a finite set of values, i.e., when the input is classified as one of the finite set of values. A regression algorithm can be used when the output may have any numerical value (within a certain range). A similarity learning algorithm is similar to both the classification algorithm and the regression algorithm, but is based on learning from examples using a similarity function that measures the similarity or correlation between two objects.

[0053] Typically, a machine learning model is trained to provide information about a predicted quality of service for a given environment model. For example, the predicted quality of service of a wireless communication link may be represented by a numerical value, which may be obtained using a regression-based machine learning algorithm. For example, the predicted quality of service may involve at least one of an inter-packet reception time, a packet error rate, a delay, and a data rate, each of which may be represented by a numerical value. Thus, a machine learning model may be trained to implement (i.e., be based on) a regression algorithm, for example, to predict at least one of an inter-packet reception time, a packet error rate, a delay, and a data rate for a wireless communication link.

[0054] In some embodiments, the machine learning model is trained specifically for the mobile transceiver. Alternatively, the machine learning model may be a generic machine learning model applicable to different mobile transceivers. In either case, the method may include training the machine learning model, or the machine learning model may be trained by another entity (for which another method may be used, such as Figure 1d shown).

[0055] Typically, the machine learning model can be trained to associate (predict) the quality of service with the environmental model of one or more active transceivers in the environment of the mobile transceiver. Therefore, the environmental model and the corresponding quality of service can be used to train the machine learning model. Typically, a plurality of training samples are used to train the machine learning model. In an embodiment, a plurality of environmental models of one or more active transceivers are determined at a plurality of time points. Accordingly, the method may include determining 140 the quality of service of the wireless communication link at a plurality of time points. In other words, for a plurality of time points, the quality of service (based on one or more active transceivers in the environment of the mobile transceiver) can be determined. For example, the quality of service can be determined by determining a metric of the wireless communication link (such as the inter-packet reception time, packet error rate, latency and / or data rate of the wireless communication link). The method may include training 145 the machine learning model using a plurality of environmental models at a plurality of time points as training inputs and the quality of service of the wireless communication link at the corresponding plurality of time points as training outputs of the training of the machine learning model. In other words, the environmental models in the plurality of environmental models can be used together with the corresponding quality of service as training samples for training the machine learning model.

[0056] In some embodiments, the output of the machine learning model may be something other than the "raw" values ​​of various quality of service characteristics. For example, some important characteristics (such as the inter-packet reception time (PIR time) of a wireless link) are typically at their minimum value (i.e., the transmission time) because the corresponding transmission was successful the first time. Training the machine learning model in this case may produce a machine learning model that is biased towards this minimum value. This can be avoided by training the machine learning model to output proxy values, which can be used to determine the actual quality of service characteristics. One such proxy is a probability distribution that models how high the probability is for a given value of a quality of service characteristic (such as inter-packet reception time, packet error rate, latency, and data rate) for a given environmental model. In other words, the machine learning model can be trained to provide a probability distribution of predicted quality of service for a given environmental model. For example, as combined with Figure 3a to Figure 3d As shown, some characteristics of quality of service, such as packet inter-reception time (PIR time), can be modeled using an exponential distribution of probability. Such an exponential distribution can be modeled using the following formula, where γ is a quality of service characteristic and λ is a rate.

[0057]

[0058] This rate may be modeled as a function λ(x) of the environment model (ie, the distribution of one or more active transceivers in the mobile transceiver's environment) and the distance between the transmitter and the receiver (inter-antenna distance, IAD).

[0059]

[0060] The rate indicates an (exponential) probability distribution of a characteristic of the quality of service (e.g. PIR time). Thus, the machine learning model may be trained to provide a (predicted) rate of the (predicted) quality of service given the environment model, and thus provide a probability distribution. The method may comprise determining a predicted future quality of service of the wireless link based on the rate / probability distribution provided by the machine learning model. Further details may be found in conjunction with Figure 3a to Figure 3d The examples provided include training a suitable machine learning model based on simulated data.

[0061] This rate can then be used to determine the predicted future quality of service. For an exponential distribution, 1 / λ can be used to obtain the predicted future quality of service value. The quantile function of the exponential distribution (also called the inverse cumulative distribution function) can be used to obtain the quantile of the predicted future quality of service value:

[0062]

[0063] For example, the third quantile is obtained using p=0.75, and the ninth quantile is obtained using p=0.9.

[0064] In various embodiments, future quality of service is predicted for multiple time points in the future (i.e., at least two time points in the future). For example, the method may include predicting the development of future quality of service at multiple time points in the future. By determining a predicted future environmental model (or more precisely, multiple predicted future environmental models) at multiple time points in the future, and using the predicted future environmental model as an input to a machine learning model, embodiments of the present disclosure may be scaled at multiple time points in the future. In other words, by determining 120 a predicted future environmental model for one or more active transceivers at at least two future time points, and using 135 a predicted future environmental model for one or more active transceivers at at least two future time points as an input to a machine learning model, the future quality of service of a wireless communication link at at least two future time points may be predicted.

[0065] In the following, another method is shown which can be used to train a machine learning model independently of the method provided above. Figure 1d A flow chart of an embodiment of a method for training a machine learning model is shown. Figure 1d The method includes: determining 140 the quality of service of a wireless communication link at multiple time points; and training 145 a machine learning model using multiple environmental models at multiple time points as training inputs and the quality of service of the wireless communication link at the corresponding multiple time points as training outputs of the training of the machine learning model.

[0066] Machine learning algorithms are typically based on machine learning models. In other words, the term "machine learning algorithm" may refer to a set of instructions that can be used to create, train, or use a machine learning model. The term "machine learning model" may refer to a data structure and / or rule set that represents learned knowledge, for example, based on training performed by a machine learning algorithm. In an embodiment, the use of a machine learning algorithm may mean the use of an underlying machine learning model (or multiple underlying machine learning models). The use of a machine learning model may mean that the machine learning model and / or the data structure / rule set that is the machine learning model is trained by a machine learning algorithm.

[0067] For example, the machine learning model may be an artificial neural network (ANN). ANN is a system inspired by biological neural networks such as those found in the brain. ANN includes multiple interconnected nodes and multiple connections (so-called edges) between nodes. There are usually three types of nodes, namely, input nodes that receive input values, hidden nodes that are (only) connected to other nodes, and output nodes that provide output values. Each node may represent an artificial neuron. Each edge can transmit information from one node to another. The output of a node can be defined as a (non-linear) function of the sum of its inputs. The input of a node can be used in a function based on the "weight" of the edge or the node that provides the input. The weights of the nodes and / or edges can be adjusted during the learning process. In other words, the training of an artificial neural network may include adjusting the weights of the nodes and / or edges of the artificial neural network, i.e., achieving the desired output for a given input. In at least some embodiments, the machine learning model may be a deep neural network, such as a neural network including one or more layers of hidden nodes (i.e., hidden layers) (preferably multiple layers of hidden nodes).

[0068] Figure 3c Schematic diagram of an exemplary architecture of a machine learning model is shown. The machine learning model includes an input layer with input nodes for the distance d between a mobile transceiver and another mobile transceiver and for the active transceivers in cells of a circular grid of width g (denoted by A 1，g ...A n，g denoted by ), where n is the number of cells of the circular grid considered. The architecture also includes n h hidden layers, each with n n Nodes The architecture also includes an output layer that outputs a probability distribution of the rate λ and thus the quality of service.

[0069] In some embodiments, the machine learning model may be a support vector machine. A support vector machine (i.e., a support vector network) is a supervised learning model with an associated learning algorithm that can be used, for example, to analyze data in classification or regression analysis. A support vector machine can be trained by providing an input with multiple training input values ​​belonging to one of two categories. A support vector machine can be trained to assign new input values ​​to one of the two categories. Alternatively, the machine learning model may be a Bayesian network, which is a probabilistic directed acyclic graphical model. A Bayesian network can use a directed acyclic graph to represent a set of random variables and their conditional dependencies. Alternatively, the machine learning model may be based on a genetic algorithm, which is a search algorithm and heuristic technique that simulates the process of natural selection.

[0070] The device 10 and the mobile transceivers 100, 102, 104 (e.g., vehicles or entities) may communicate via a mobile communication system. The mobile communication system may, for example, correspond to one of the mobile communication networks standardized by the Third Generation Partnership Project (3GPP), wherein the term mobile communication system is used as a synonym for mobile communication network. Thus, messages (input data, measurement data, control information) may be communicated via a plurality of network nodes (e.g., the Internet, routers, switches, etc.) and the mobile communication system, which generates delays or latency considered in the embodiments.

[0071] The mobile or wireless communication system may correspond to a fifth generation mobile communication system (5G or new radio), and may use millimeter wave technology. The mobile communication system may correspond to or include, for example, long term evolution (LTE), advanced long term evolution (LTE-A), high speed packet access (HSPA), universal mobile telecommunications system (UMTS) or UMTS terrestrial radio access network (UTRAN), evolved UTRAN (e-UTRAN), global system for mobile communications (GSM) or enhanced data rates for GSM evolution (EDGE) network, GSM / EDGE radio access network (GERAN) or a mobile communication network with different standards (e.g., Worldwide Interoperability for Microwave Access (WIMAX) network IEEE802.16 or wireless local area network (WLAN) IEEE 802.11), typically an orthogonal frequency division multiple access (OFDMA) network, a time division multiple access (TDMA) network, a code division multiple access (CDMA) network, a wideband CDMA (WCDMA) network, a frequency division multiple access (FDMA) network, a space division multiple access (SDMA) network, etc.

[0072] Service provision may be performed by a network component such as a base transceiver, a relay station or a UE, which coordinates service provision, for example, in a cluster or group of multiple UE / vehicles. The base transceiver may be operable or configured to communicate with one or more active mobile transceivers / vehicles, and the base transceiver may be located in or near the coverage area of ​​another base transceiver (e.g., a macro cell base transceiver or a small cell base transceiver). Therefore, an embodiment may provide a mobile communication system including two or more mobile transceivers / vehicles 100, 102 and one or more base transceivers, wherein the base transceiver may establish a macro cell or a small cell, such as a micro cell, a metropolitan cell or a femto cell. A mobile transceiver or UE may correspond to a smart phone, a cellular phone, a laptop, a notebook computer, a personal computer, a personal digital assistant (PDA), a universal serial bus (USB) stick, a car, a vehicle, a road participant, a traffic entity, a traffic infrastructure, etc. According to 3GPP terminology, a mobile transceiver may also be referred to as a user equipment (UE) or a mobile phone. For example, the mobile transceiver, another mobile transceiver, and / or at least a subset of the one or more active transceivers may be a vehicle, such as a land vehicle, a road vehicle, a car, an automobile, an off-road vehicle, a motor vehicle, a truck, or a van.

[0073] A base transceiver station may be located in a fixed or stationary part of a network or system. A base transceiver station may be or correspond to a remote radio head, a transmission point, an access point, a macro cell, a small cell, a micro cell, a femto cell, a metro cell, etc. A base transceiver station may be a wireless interface of a wired network that enables the transmission of radio signals to a UE or a mobile transceiver. Such radio signals may conform to radio signals standardized, for example, by 3GPP, or generally conform to one or more of the systems listed above. Thus, a base transceiver station may correspond to a NodeB, an eNodeB, a gNodeB, a base transceiver station (BTS), an access point, a remote radio head, a relay station, a transmission point, etc., which may be further subdivided into remote units and central units.

[0074] The mobile transceiver or vehicle 100 may be associated with a base transceiver station or a cell. The term cell refers to the coverage area of ​​the radio service provided by a base transceiver station such as a NodeB (NB), an eNodeB (eNB), a gNodeB, a remote radio head, a transmission point, etc. The base transceiver station may operate one or more cells on one or more frequency layers, and in some embodiments, the cell may correspond to a sector. For example, the sector may be implemented using a sector antenna, which provides the characteristics of an angular segment covering the surrounding of the remote unit or base transceiver station. In some embodiments, the base transceiver station may, for example, operate three or six cells covering sectors of 120° (in the case of three cells), 60° (in the case of six cells), respectively. The base transceiver station may operate multiple sectorized antennas. In the following, a cell may represent the corresponding base transceiver station generating the cell, or similarly, a base transceiver station may represent a cell generated by a base transceiver station.

[0075] In an embodiment, the apparatus 10 may be comprised in a server, a base station, a NodeB, a UE, a mobile transceiver, a relay station or any service coordination network entity. It should be noted that the term network component may include multiple subcomponents, such as a base station, a server, and the like.

[0076] In an embodiment, one or more interfaces 12 may correspond to any device for obtaining, receiving, sending or providing analog or digital signals or information, such as any connector, contact, pin, register, input port, output port, conductor, channel, etc., which allows providing or obtaining signals or information. The interface may be wireless or wired, and it may be configured to communicate with other internal or external components, that is, to send or receive signals, information. One or more interfaces 12 may include other components to achieve corresponding communications in a mobile communication system, which may include transceiver (transmitter and / or receiver) components, such as one or more low noise amplifiers (LNA), one or more power amplifiers (PA), one or more duplexers, one or more duplexers, one or more filters or filter circuits, one or more converters, one or more mixers, correspondingly adapted radio frequency components, etc. One or more interfaces 12 may be connected to one or more antennas, which may correspond to any transmitting and / or receiving antennas, such as horn antennas, dipole antennas, patch antennas, sector antennas, etc. Antennas may be arranged in defined geometric settings, such as uniform arrays, linear arrays, circular arrays, triangular arrays, uniform field antennas, field arrays, combinations thereof, etc. In some examples, one or more interfaces 12 may be used for purposes of sending or receiving, or both sending and receiving information, such as information, input data, control information, further information messages, or the like.

[0077] like Figure 1cAs shown, the corresponding one or more interfaces 12 are connected to the corresponding control module 14 at the device 10. In an embodiment, the control module 14 can be implemented using one or more processing units, one or more processing devices, any means for processing (such as a processor, a computer, or a programmable hardware component that can be operated using correspondingly adapted software). In other words, the functions of the control module 14 can also be implemented in software, which is then executed on one or more programmable hardware components. Such hardware components may include general-purpose processors, digital signal processors (DSPs), microcontrollers, etc.

[0078] In an embodiment, communication (i.e., transmission, reception, or both) may occur directly between the mobile transceivers / vehicles 100, 102, such as forwarding input data or control information to / from a control center. Such communication may utilize a mobile communication system. Such communication may be performed directly, such as by means of device-to-device (D2D) communication. Such communication may be performed using the specifications of a mobile communication system. An example of D2D is direct communication between vehicles, also referred to as vehicle-to-vehicle communication (V2V), vehicle-to-vehicle communication, dedicated short range communication (DSRC), respectively. Technologies supporting such D2D communication include 802.11p, 3GPP systems (4G, 5G, NR and above), etc.

[0079] In an embodiment, one or more interfaces 12 may be configured to communicate wirelessly in a mobile communication system. To this end, radio resources, such as frequency, time, code and / or space resources, are used, which can be used for wireless communication with a base transceiver station and for direct communication. The allocation of radio resources may be controlled by the base transceiver station, i.e., determining which resources are used for D2D and which are not used. Here and below, the radio resources of the corresponding components may correspond to any conceivable radio resources on a radio carrier, and they may use the same or different granularity on the corresponding carrier. Radio resources may correspond to resource blocks (such as RBs in LTE / LTE-A / LTE-unlicensed (LTE-U)), one or more carriers, subcarriers, one or more radio frames, radio subframes, radio time slots, one or more code sequences potentially having corresponding spreading factors, one or more spatial resources (such as spatial subchannels, spatial precoding vectors), any combination thereof, etc. For example, in direct cellular vehicle-to-everything (C-V2X), where V2X includes at least V2V, V2-Infrastructure (V2I), etc., transmissions according to 3GPP Release 14 and above can be managed by the infrastructure (so-called Mode 3) or run in the UE.

[0080] Further details and aspects of the method, apparatus or mobile transceiver may be combined with the proposed concept or one or more examples described above or below (e.g., FIG. 2 to FIG. 3 ). Figure 3d) is mentioned. The method, device or mobile transceiver may include one or more additional optional features corresponding to one or more aspects of the proposed concept or one or more examples described above or below.

[0081] Various embodiments relate to a method for increasing the prediction time domain in QoS prediction using time series prediction. Within the scope of collaborative driving, when quality of service (QoS) conditions change, the prediction of future quality of service enables vehicle applications. In fact, when predicted QoS (PQoS) is not provided, applications can only react to changes and are therefore limited to the lower bound performance of the communication system. The PQoS system can operate on a vehicle (communication node) using a radio access technology (RAT) such as LTE-V or 5G-V2X or IEEE 802.11p in its independent mode. Combinations of these technologies can also be applied to multi-RAT systems. In such a PQoS system, vehicles can exchange information about the communication surroundings in order to provide PQoS. Therefore, an embodiment may relate to a method for predicting QoS in the future.

[0082] In other methods, such as the paper "Prediction of Packet Inter Reception Time for Platooning using Conditional Exponential Distribution" by Jornod, El Assad, Kwoczek and Kürner, a statistical connection between ambient density and inter-packet reception time is provided. However, in the method shown in the paper, the prediction is only instantaneous, while the embodiments are intended to predict future quality of service. More specifically, in a statistical sense, prediction involves modeling instantaneous data and forecasting predictions of future values. According to this definition, prediction is made using time series prediction, and the results of the prediction are used to predict future quality of service.

[0083] Embodiments may provide a method for forecasting QoS in the future using historical data and time series prediction.

[0084] The method may include one or more of the following features:

[0085] 1. Collect transmission data (i.e. wireless transmission) from the receiver (i.e. mobile transceiver) and the location of surrounding communication nodes (i.e. one or more active transceivers).

[0086] 2. Calculate the QoS metrics of interest (i.e., quality of service) from this receiver

[0087] 3. Modeling the surrounding communication environment (eg, by generating an environment model), such as obtaining indicators of the surrounding communication environment.

[0088] 4. Matching the calculated QoS indicators of the transmitted data with a model of the surrounding communication environment (e.g., associating the quality of service with the corresponding environment model) with respect to the time and location of the transmitter and the receiver, e.g., enabling estimation of the correlation between the indicators of the surrounding environment and the indicators of the QoS.

[0089] 5. Train a distribution model (e.g., a machine learning model) on the collected data (see table)

[0090] 6. Perform time series prediction in the future to predict the future QoS in the prediction time domain (e.g., use time series prediction to determine the predicted future environment model), for example, being able to predict the QoS based on the change in the correlation between the indicators of the surrounding environment and the indicators of the QoS

[0091] 7. Changing the prediction time domain to obtain a dynamic link QoS graph along time / direction (e.g., determining a predicted future environment model at at least two time points in the future), for example, to be able to create a QoS graph that enables decisions to be made in adjusting application settings.

[0092] 8. Repeat one or more previous features with accumulated data to improve learning (e.g., reinforcement learning)

[0093] The communication environment data collection can be performed by using a radio receiver as a sensor or by sharing the information with the communication. Modeling of the surrounding communication environment can be performed using a grid-like abstraction method for radio activity. For example, the surrounding communication environment can be modeled using a grid-like division of the environment, and radio activity numerical levels are assigned according to the number of communicating vehicles and the periodicity of their messages. For example, future time series prediction can be performed to predict the future QoS of the prediction time domain by using the change in the correlation between the previous environmental prediction factor values ​​and the previous QoS values. Time series prediction can be performed by using the data from the past t0-t p To simulate the sample of t0 (Figure 2 210) at t0+t f (Figure 2 230) is performed by distributing the QoS (where t0 is the current time (Figure 2 220), t p is the replay time, for example 10s, and t f is the prediction horizon). The change in the prediction horizon may result in a multi-link QoS graph, one for each future time step. It may change with time / displacement, in this sense it is dynamic. The QoS metrics calculated and predicted may be packet error rate (PER), packet inter-reception (PIR) time, latency, data rate. For example, time series prediction may use statistical prediction methods such as multi-layer perceptrons (e.g., Figure 3c shown).

[0094] FIG. 2 provides a method for combining Figure 3a to Figure 3d One of the examples provided is an example of data for training a prediction model. FIG2 shows a table of the development of a quality of service attribute associated with an environmental model over time. The table includes various columns, the first column showing a timestamp (time), the second and third columns showing the identification of the destination and source nodes (dst and src), the fourth column showing the size of the packet (size), the fifth column showing the PIR value (pir), the sixth column showing the distance between the transmitter and the receiver (dist, which remains fairly constant), and the seventh and subsequent columns showing the number of nodes d* in a geographic area (i.e., a cell or annulus of a circular grid) surrounding the receiver (src), where * represents the maximum distance (e.g., d30=0 until 30 meters, d60=greater than 30, until 60, etc.). Rectangle 210 defines the look-back data, rectangle 220 defines the current time step (defining a predefined time interval to the prediction horizon), and rectangle 230 defines the target prediction horizon, e.g., 8 seconds. This "sliding window" can be applied to all time steps in order to train a prediction model, which will effectively become a forecast model when we look at future data.

[0095] In various embodiments, the method includes monitoring an indicator of the surrounding communication environment (such as the number of surrounding communication vehicles in a grid cell) and correlating the change with a change in QoS. For example, the method may include predicting QoS (applicable to V2V communication) based on a correlation between a change in an indicator of surrounding communication vehicles and a change in QoS. The method may include using information about the surrounding communication environment for a V2V link. The method may include creating a QoS graph. The method may include performing a match between communication environment indicators and QoS changes. The method may include using past changes in the correlation between environmental predictors and QoS to predict future values ​​of QoS. Embodiments may use aggregated communication activity at a higher level (message level) where we wish to predict the impact of hidden nodes and channel overloads due to surrounding communication vehicles.

[0096] Figure 3a to Figure 3d A schematic diagram related to the training of a machine learning model is shown. In the following, two similar methods are described, which involve the training of a machine learning model suitable for providing a probability distribution over predicted quality of service for a given environment model. Although the machine learning model is trained using a simulated environment model, the same principles can be applied to data collected in a "real" mobile transceiver / vehicle.

[0097] In the following, the background of both approaches is presented. An interesting application of AQoSA (Adaptive Quality of Service Adaptation) is High Density Platooning (HDPL): This is a collaborative vehicle application where vehicles coordinate their control to achieve a small inter-vehicle distance (IVD) below 15m. Within the scope of AQoSA, a prediction system is needed that can predict the future QoS. This prediction system can use its knowledge of the surrounding environment to function. The following scenario features a five-truck HDPL driving at a speed of 25m / s on an oval multi-lane test track. The target of this platoon is an IVD between 5 and 25m. It is coordinated by broadcasting Platooning Control Messages (PCM), which are 700B messages transmitted at a rate of 20Hz using the IEEE 802.11p Radio Access Technology (RAT). To challenge the communication system, a vehicle traffic with increasing density in the lanes in the opposite direction is introduced. This surrounding traffic reaches more than 200 vehicles within 400m around the transmitter. The surrounding vehicles broadcast 400B CAMs (Collaborative Awareness Messages) at a rate of 10Hz using the same RAT. The mobility of the surrounding traffic is modeled in Simulation of Urban Mobility (SUMO), a simulation tool that takes full advantage of its ability to generate random vehicles with a specific density. This traffic simulator is combined with the network simulator ns-3, which uses its IEEE802.11p model to operate the wireless transmissions. This setup allows the simulation of realistic vehicle movements as well as a fairly accurate model of the communication system. A total of 106 transmission observations were collected using the tracking capabilities of ns-3. These observations include the observed PIR times and the environmental predictors described in the following. The PIR time, denoted γ, is defined as the duration between two consecutive messages in a pair of communication partners measured from the receiver. The PIR values ​​are the result of consecutive transmission failures, which produces a low representation of high PIR values. However, these higher values ​​are very important for applications as they may limit their performance. Direct prediction of γ most likely leads to a systematic prediction of a PCM transmission period of 0.05 seconds, as it is the minimum PIR and the most representative value. Instead, a PIR probability distribution can be modeled. The first set of features is the IAD d, the distance between the transmitter and the receiver, calculated between their antennas. The second group includes the number of communicating vehicles D in the annulus around the transmitter. n，g =card(a n，g ) or consisting of. n，g The difference in their radius and their indices to define. A further description of these characteristics is given in "Packet inter-reception time modeling for high-density platooningin varying surrounding traffic density" by G. Jornod, A. El Assaad, A. Kwoczek and T. Kürner, 2019 28th IEEE European Conference on Networking and Communications. In the simulations, a representation of the relationship between the PIR, the total number of vehicles within a 400m radius and the IAD was calculated. In this case, the PIR increases sharply when the number of surrounding communicating vehicles exceeds 150. The total number also has an impact on the target; this impact decreases sharply when the IAD becomes larger than 100m.

[0098] In the following, a prediction model is provided based on the learning of a nonlinear function that provides the parameters of an exponential distribution (which can be combined with Figure 1a to Figure 1d Machine learning models introduced).

[0099] The distribution of PIR can be viewed as a conditional exponential distribution. In the literature, it is shown that PIR is well modeled by exponential distribution, lognormal distribution, Weibull distribution, and gamma distribution, with the last two distributions being preferred, which pass the 90% significance test. Therefore, in the following, PIR can be modeled by exponential distribution. Exponential distribution can be viewed as a special case of gamma distribution, which has the advantage of having a single parameter, i.e., rate λ>0. Its probability density function (PDF) is expressed as:

[0100]

[0101] And in Figure 3a Several rates of interest λ are shown in Figure 3b Its cumulative density function (CDF) is shown in . λ is parameterized by the spatial predictor:

[0102]

[0103] where x is the number of interactions only (e.g. omitting d 2 ) predictors d and D n，g This modeling allows to calculate the PIR density for any combination of predictors within the range of the training data. The goal of modeling is to approximate the nonlinear function λ(x) using the observed data. For this purpose, an MLP regressor is used (see Figure 3c ). A Keras-based implementation was used, where the model was trained using the Adam optimization algorithm with negative log-likelihood as the loss function:

[0104]

[0105] where Ω is the set of indices containing the training data. A k-fold cross validation was performed for the choice of size and the number of hidden layers (n∈{1,2,3} and m∈{100,500,1000}, respectively) with k=8 using scikit-learn. One limitation of the model is that it does not exploit previous values ​​of the predictors. In fact, the PIR time represents the time between two successful events: the source of consecutive grouping failures may occur between two observations. However, considering that the predictors are moving vehicles, their motion is smooth enough that the model is robust to this limitation. The use of a combination of these two python libraries enabled rapid prototyping of the learning strategy. The MLP regressor was chosen for its ability to capture the nonlinearities highlighted during the data exploration phase.

[0106] Another approach is proposed below. In the following, the prediction of PIR distribution using the spatial distribution of communication nodes is shown. Previous studies have shown that PIR can be modeled using an exponential distribution. Furthermore, IAD and the surrounding communication traffic density change the performance of the communication system. In the following concept, these two premises are used to build a model that relates the distribution of the prediction target to the changes in the spatial distribution of the nodes. Subsequently, the rapid learning of this relationship in a full-scale system-level vehicle ad hoc network (VANET) simulation featuring a five-truck HDPL is studied.

[0107] Using the tracing capability of ns-3, a total of 4 10 7 transmission observations. Among them, 2.5·10 6 On the receiving side there are queue members, and 9.10 5 Describes intra-platoon communications. These observations include the observed PIR time, source and destination, and information about the positions of surrounding vehicles. Figure 3dAn exemplary distribution of PIR is shown. The PIR times show a heavy-tailed distribution, which motivates further modeling of their distribution rather than simply predicting the values. In reality, the PIR values ​​are the result of consecutive transmission failures, which produce a low representation of high PIR values. However, these higher values ​​are very important for applications because they may limit the performance of the application. As emphasized in Jornod, El Assaad, Kwoczek, and Kürner, “Prediction of Packet Inter-Reception Time for Platooning using Conditional Exponential Distribution,” in 16th Int. IEEE Symp. Wireless Commun. Sys. (ISWCS), 2019, pp. 265–270, a simple prediction most likely leads to a systematic prediction of the most representative value of 0.05 seconds. However, the distribution of the number of vehicles is different from previous studies. In reality, the vehicles are introduced into the scene in a random manner. The vehicles then drive along random routes in the scene and may cluster, which is an experimental design choice to study the robustness of the method to unknown or less frequent situations.

[0108] In a previous study, the importance of considering the interaction between IAD and the number of surrounding vehicles was shown. The average PIR for the combination of IAD and number spacing was simulated. A sharp increase in the average PIR values ​​above 125m was observed. In general, it also increases with both the number of surrounding vehicles and the IAD. In the region between 80 and 100 vehicles, higher PIR values ​​were observed. This region as well as the second gradient it induces were not seen in the previous results. Indeed, in this previous study, the surrounding vehicles were restricted to straight roads. In this new scenario, the spatial distribution of surrounding vehicles changes more sharply. This new pattern and the appearance of a second increasing path from the origin to the top of the ellipse are the results of this new distribution. This further motivates the modeling of this distribution and the use of nonlinear regression models.

[0109] In the following, the modeling strategy is presented. The first step is to calculate the target key performance indicator (KPI), PIR time (i.e., quality of service). The second step is the modeling of the surrounding communication environment (i.e., environment model), which includes location log preprocessing and abstraction of the environment. The third step is to formalize the relationship between these environmental characteristics and the distribution of our goal (i.e., generation of training data). The fourth and last step is to create a strategy for the learning of this relationship. In this example, the prediction target is the PIR time. As mentioned before, this metric is gaining more and more attention in VANET research. In the scope of HDPL, it measures the time that a vehicle cannot rely on communication to coordinate with other vehicles. In addition, modern control systems are able to cope with low update input rates due to prediction algorithms. However, they perform poorly for irregular inputs. This regularity is captured by the PIR distribution. PIR is the time difference between the reception of two messages measured from the receiver of each source. This KPI is studied for all received messages for queue members. Because these messages are periodic messages, the PIR time is a multiple of the transmission period, plus or minus the delay experienced. It reflects the number of consecutive dropped messages.

[0110] In the following, environmental features are introduced. One goal is to take into account the locations of all communicating vehicles (i.e., one or more active transceivers). The main challenge is that the number of inputs is variable (theoretically, this number can span from a few vehicles to infinity if the point of interest is not within a specific range). Moreover, most prediction methods require fixed-size and ordered inputs even when the range is narrowed to a specific range. As a result, appropriate aggregation methods for the surrounding communication environment information are used (to represent the environment model). In G. Jornod, A. El Assaad, A. Kwoczek, and T. Kürner, “Packet inter-reception time modeling for high-density platooning in varying surrounding trafic density,” in 2019 28th IEEE European Conference on Networking and Communications (EuCNC), pages 187–192, a torus-based environment model (i.e., based on a circular grid) is introduced, for example in Figure 3a to Figure 3d The core idea is to divide the space in concentric circles around the receiver and count the number of vehicles present in the formed annulus. The radius of these circles is a multiple of the granularity parameter r. As a result, A n，r is included in The number of vehicles in the receiver with radius difference The communication vehicles in the concentric annuli, is the index of the torus.

[0111] Using this model, the effect of the spatial distribution of surrounding vehicles on the channel load is captured. The annulus is defined as:

[0112]

[0113] A n，r =card(a n，r )

[0114] in is the set of surrounding vehicle position vectors, n is the torus index, r is its granularity value, and x r is the position vector of the receiver. The torus partitioning is designed for highway usage and for specific scenarios where vehicles enter in opposite lanes.

[0115] In this example, this modeling (i.e., the environment model) is improved by introducing a sector partitioning. This partitioning captures the location of the interfering nodes, especially relative to the transmitter location, when the IAD is relatively high, the transmitter may not detect these interferences. The space is divided into n s The sectors are centered on the receiver (i.e., the center point of the circular grid) and aligned with the receiver-transmitter segment. The division method can be defined as:

[0116]

[0117]

[0118]

[0119]

[0120]

[0121] in, is the function that provides the angle between the positive x-axis and the vector x, and β is the function that produces the angle between the Rx–Tx and Rx–interference vehicle vectors. The angle α is 2π / n s The mth sector Finally, we introduce an offset of π / 2 to represent the number of vehicles in the s = 2, instead of left / right. By combining the torus-based model and the sector-based model, the so-called torus-sector model is obtained. Its segments are defined by the intersection of the torus and the sector. In this example, Represents a torus With sectors Similarly, the group of vehicles contained in a segment and their cardinality are defined as:

[0122]

[0123]

[0124] It can be noted that and In this example, the sectors are always oriented relative to the transmitter and the annulus reflects the distance of the interferer. When combined with IAD through the step of polynomial combination of features, this environment model offers the possibility to solve the hidden node problem. This problem can be solved in the modeling of the relationship between the target and the features through a weighting process.

[0125] gather represents the node positions around it. This is distinguished into two cases: global knowledge and local knowledge. In the case of global knowledge, it includes all nodes in the simulation and is represented as This set can be obtained from the transmission log and the location log. In the case of local knowledge, this set is represented as And collect the nodes from which the receiver has received CAM in the last T seconds. Depending on the channel load, the number of nodes included in this set can be drastically reduced with lower T values. When the collective sensing system is not implemented, this reduced set also reflects the realistic knowledge of the surrounding communication environment. Similarly, the transmission log can be used to calculate the set. In the evaluation of the example, T is set to 10 seconds.

[0126] In this example, the prediction target is the PIR time. In the following, Г represents the PIR as a random variable. Figure 3d As shown, its distribution is heavy-tailed. This characteristic of the distribution may hinder the use of classical regression methods aimed at directly predicting the PIR. In fact, since lower values ​​are more representative than larger values, a simple prediction will lead to systematic predictions of more representative values, which happen to be the transmission period. Instead, the distribution of the target values ​​can be predicted, which adds a step of distribution modeling before solving the learning method. In the literature, it is shown that the PIR is well modeled by the exponential distribution, the lognormal distribution, the Weibull distribution, and the gamma distribution.

[0127] The probability density function (PDF) of the exponential distribution is given by:

[0128]

[0129] The advantage is that it has a single parameter, the rate λ. Therefore, the modeling task may be to find a suitable λ that fits the collected data. It can be noted that in the absence of relative delay, the PIR time is a discrete variable. In MErenda, G.Resta, P.Santi, F.Martelli, and A.Franchini, "IEEE 802.11p VANets: Experimental evaluationof packet inter-reception time," Comput. Commun., vol. 75, pp. 26–38, 2016, the PIR is modeled using a geometric distribution. The exponential distribution is a continuous analog of this distribution. It makes it possible to avoid the step of eliminating relative delays in the collected data.

[0130] The previous subsection described the environmental characteristics. Before that, the premise for studying the impact of the number of surrounding vehicles and IAD and their interaction was shown. It is shown that the number of vehicles and IAD jointly affect the average PIR of the CAM information. Conditional CDFs are calculated based on the number of vehicles around the receiver on the one hand and on the IAD on the other hand. The first observation is that all the provided CDFs resemble the CDF of an exponential distribution. The second observation is that the rate λ decreases with the number of vehicles n v and IADd. When the interval (n v，1 and n v，1 ), which shows the influence of the spatial distribution of surrounding nodes. This motivates the parameterization of λ with environmental characteristics. It is suggested to make λ a function of a polynomial combination of spatial distribution characteristics and IAD:

[0131]

[0132] where x is the number of interactions only (e.g. omitting d 2 ). Therefore, the remaining step is to approximate the function λ(x) such that Γ ≈ Exp(λ(x)). This process is called conditional density estimation (CDE).

[0133] In the following, a learning method (e.g., for training a machine learning model) is shown. The goal may be to provide a flexible method that enables fast learning of the PIR distribution. The nonlinear function λ(x) can be approximated by a multilayer perceptron (MLP). The Keras interface can be fully utilized and combined with scikit-learn for hyperparameter optimization. The number of hidden layers n is then automatically determined by a cross-validated grid search. h and the number of nodes in the layer n nThe model (i.e., machine learning model) is trained using the Adam optimization algorithm with negative log-likelihood as the loss function:

[0134]

[0135] Where Ω1 is the set of training data. The MLP is fed with a polynomial combination of features (i.e., a polynomial combination based on the environment model) and outputs a rate λ (which can be used to determine the probability distribution of the quality of service). Figure 3c This process is shown for a torus-based environment model. The learning model has three parameters: the learning rate LR, the number of hidden layers n h and the number of neurons in each layer n n For example, we can assume that there are the same number of neurons in each layer. The performance of the torus-sector model can be determined based on the parameters r and n. s This is done using an MLP regressor on data collected by the cohort members in the previous 30 minutes (i.e., multiple environment models and corresponding quality of service). A k-fold cross-validation strategy is used for the split between training and test sets. This strategy is applied to each model evaluated and is reported as the average performance. This model selection step may also include three learning parameters (learning rate, size and number of layers). r and n are reported. s The best performing model for each combination of is then used for the total simulation duration. The MLP regressor is iteratively trained with data collected by the training fleet members. In this training process, the weights of the regressor are updated at each iteration. The performance of the model is evaluated on the test data, which includes all observations collected by the remaining trucks for the simulation duration. The log-likelihood given above is used as the loss function for the training of the MLP model.

[0136] The evaluation found that the best performing torus-sector model parameters are r=30, n s =8,LR=0.0001,n n =1000 and n h =9, for the local range r=60, n s =4,LR=0.0001,n n =500 and n h = 10. The following values ​​were evaluated: r = n30m, where n∈{1, 2, 3, 4, 6, 13}, n s ∈{1, 2, 4, 8}, LR={0.1, 0.01, 0.001, 0.0001}, n n ={50, 100, 500, 1000} and n h ={2, 3, ..., 10}.

[0137] The selected model is trained in an online manner throughout the simulation using the results of the previous model comparison. The fitted model is updated every time a training node receives a transmission. The performance of the model is performed on two data sets: (i) data collected up to the reception time of the node of interest, which is a subset of the training set; and (ii) all data collected by the other queue members.

[0138] Evaluations show that the model can continue learning even after convergence to improve its robustness.

[0139] Further details and aspects of the concept may be combined with the proposed concept or one or more examples described above or below (e.g., Figure 1a The concept may include one or more additional optional features corresponding to one or more aspects of the proposed concept or one or more examples described above or below.

[0140] As already mentioned, in an embodiment, the corresponding method may be implemented as a computer program or code, which can be executed on corresponding hardware. Therefore, another embodiment is a computer program with a program code, which, when executed on a computer, a processor or a programmable hardware component, is used to perform at least one of the above methods. Another embodiment is a computer-readable storage medium storing instructions, which, when executed by a computer, a processor or a programmable hardware component, causes the computer to implement one of the methods described herein.

[0141] Those skilled in the art will readily recognize that the steps of the various methods described above can be performed by a programmed computer, for example, the position of a time slot can be determined or calculated. Here, some embodiments are also intended to cover program storage devices, such as digital data storage media, which are machine or computer readable and encode machine-executable or computer-executable instruction programs, wherein the instructions perform some or all of the steps of the methods described herein. The program storage device may be, for example, a digital memory, a magnetic storage medium such as a disk and a tape, a hard drive, or an optically readable digital data storage medium. Embodiments are also intended to cover computers programmed to perform the steps of the methods described herein or (field) programmable logic arrays ((F)PLA) or (field) programmable gate arrays ((F)PGA) programmed to perform the steps of the above methods.

[0142] The description and drawings merely illustrate the principles of the present invention. Therefore, it should be understood that those skilled in the art will be able to design various arrangements that, although not explicitly described or shown herein, embody the principles of the present invention and are included in its spirit and scope. In addition, all examples described herein are primarily intended to be explicitly used only for teaching purposes to help readers understand the principles of the present invention and the concepts contributed by the inventors to advance the art, and are interpreted as not being limited to these specifically described examples and conditions. In addition, all statements describing the principles, aspects, and embodiments of the present invention and their specific examples herein are intended to include their equivalents.

[0143] When provided by a processor, the functions may be provided by a single dedicated processor, a single shared processor, or multiple separate processors, some of which may be shared. In addition, the explicit use of the term "processor" or "controller" should not be interpreted as referring exclusively to hardware capable of executing software, and may implicitly include, but is not limited to, digital signal processor (DSP) hardware, network processors, application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), read-only memories (ROMs), random access memories (RAMs), and non-volatile storage for storing software. Other conventional or customized hardware may also be included. Their functions may be implemented by the operation of program logic, by dedicated logic, by the interaction of program control and dedicated logic, or even manually, and the specific technology can be selected by the implementer when more specifically understood according to the context.

[0144] Those skilled in the art will appreciate that any block diagram herein represents a conceptual diagram of an illustrative circuit that embodies the principles of the present invention. Similarly, it will be appreciated that any flow chart, flow chart, state transition diagram, pseudo code, etc., represent various processes that can be substantially represented in a computer-readable medium and thus executed by a computer or processor, whether or not such a computer or processor is explicitly shown.

[0145] In addition, the appended claims are hereby incorporated into the detailed description, wherein each claim may stand on its own as a separate embodiment. Although each claim may stand on its own as a separate embodiment, it should be noted that although a dependent claim may refer to a specific combination with one or more other claims in the claims, other embodiments may also include a combination of the dependent claim with the subject matter of each other dependent claim. Unless it is stated that a specific combination is not intended, such a combination is suggested herein. In addition, it is also intended to include features of a claim to any other independent claim, even if the claim is not directly subordinate to the independent claim.

[0146] It should also be noted that the methods disclosed in the specification or claims may be implemented by a device having means for performing each of the respective steps of these methods.

[0147] Reference numerals list

[0148] 10. Installation

[0149] 12 One or more interfaces

[0150] 14 Control Module

[0151] 100 Mobile Transceiver

[0152] 102 Another mobile transceiver

[0153] 104 One or more active transceivers

[0154] 110 Determine multiple environmental models

[0155] 120 Determine the predicted future environmental model

[0156] 120 Performing Time Series Forecasting

[0157] 130 Predicting future service quality

[0158] 135 Using predicted future environmental models as input to machine learning models

[0159] 140 Determine the quality of service

[0160] 145 Training Machine Learning Models

Claims

1. A method for predicting a future quality of service of a wireless communication link between a mobile transceiver (100) and another mobile transceiver (102), the method comprising: determining (110) a plurality of environment models of one or more active transceivers (104) in the surrounding environment of the mobile transceiver at a plurality of time points; determining (120) a predicted future environmental model of the one or more active transceivers at a future point in time using a time series prediction (125) of the plurality of environmental models, wherein the time series prediction (125) of the plurality of environmental models predicts development of the one or more values ​​over a time interval including the plurality of time points based on historical data regarding the one or more values; and predicting (130) the future quality of service of the wireless communication link at a future point in time using a machine learning model, wherein the machine learning model is trained to provide information about the predicted quality of service for a given environment model, Wherein, the predicted future environment model is used as an input to the machine learning model (135).

2. The method according to claim 1, characterized in that The time series prediction is performed (125) based on a statistical fitting function or based on a time autocorrelation function.

3. The method according to claim 1, characterized in that The time series forecasting (125) is performed using another machine learning model.

4. The method according to claim 1, characterized in that: The time series forecast is determined (125) such that a progression of the environment model toward the predicted future environment model is predicted.

5. The method according to claim 1, characterized in that The time series forecasting generates the predicted future environment model, wherein the future quality of service is predicted based on the predicted future environment model.

6. The method according to claim 1, characterized in that The future quality of service of the wireless communication link is predicted (130) for at least two future points in time.

7. The method according to claim 6, characterized in that The future quality of service of the wireless communication link is predicted (130) for the at least two future time points by determining (120) the predicted future environment model of the one or more active transceivers at the at least two future time points and using (135) the predicted future environment model of the one or more active transceivers at the at least two future time points as an input to the machine learning model.

8. The method according to any one of claims 1 to 7, characterized in that The method comprises: determining (140) the quality of service of the wireless communication link at the multiple time points; and training (145) the machine learning model using the multiple environmental models at the multiple time points as training inputs and the quality of service of the wireless communication link at the corresponding multiple time points as training outputs of the training of the machine learning model.

9. The method according to any one of claims 1 to 7, characterized in that: The machine learning model is trained (145) to implement a regression algorithm.

10. The method according to any one of claims 1 to 7, characterized in that The machine learning model is trained (145) to provide a probability distribution of the predicted quality of service given an environmental model.

11. The method according to any one of claims 1 to 7, characterized in that: The one or more active transceivers are placed on a grid within the environment model, the grid comprising a plurality of contiguous cells, wherein the one or more active transceivers are clustered in each cell within the grid.

12. The method according to claim 11, characterized in that The grid is a circular grid.

13. The method according to any one of claims 1 to 7, characterized in that The predicted quality of service relates to at least one of inter-packet reception time, packet error rate, delay and data rate.

14. A computer program product comprising a computer program for performing the method of any one of claims 1 to 13 when the computer program is executed on a computer, a processor or a programmable hardware component.

15. An apparatus (10) for predicting a future quality of service of a wireless communication link between a mobile transceiver and another mobile transceiver, the apparatus comprising: one or more interfaces (15) for communicating in a mobile communication system; and A control module (14) configured to execute the method according to any one of claims 1 to 13.

Images