Method for training machine learning algorithm

Abstract

To provide a method for training a machine learning algorithm including uncertainties.SOLUTION: A method (1) includes the steps of: for each point in time of a plurality of points in time, determining in each case an influence that data detected at a corresponding point in time have on uncertainties instantaneously contained in an initial model (4); for each point in time of the plurality of points in time, determining a resolution of the corresponding detected data based on an established data level and on the respective influence that the corresponding detected data have on uncertainties instantaneously contained in the initial model (5); for each point in time of the plurality of points in time, transferring the detected data to a control unit based on the corresponding determined resolution (6); and retraining the initial model by the control unit based on the data transferred to the control unit (7).SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a method for training a machine learning algorithm that controls at least one controllable system, wherein the at least one controllable system is trained based on a machine learning algorithm with uncertainty, the machine learning algorithm can be retrained during operation of the at least one controllable system, and transmission of training data to a control device that retrains the machine learning algorithm is optimized. [Background technology]

[0002] Digital controllers are used in many applications for open-loop and closed-loop control of technical systems, also generally referred to as controllable systems below, by processing sensor signals or other input values ​​according to predetermined control algorithms to determine one or more output values ​​for functions, e.g., manipulated variables for controlling actuators.

[0003] The control algorithm here can be, for example, a machine learning algorithm. In this case, each individual controllable system collects data about its use, and these data are then used to train the machine learning algorithm accordingly. Typically, this type of machine learning algorithm is based on using statistical methods to train a data processing system to perform a specific task without the data processing system being explicitly programmed for this purpose from the beginning. In this case, the goal of machine learning is to build an algorithm that can learn from data and make predictions. These algorithms create mathematical models that can be used, for example, to classify data.

[0004] In this case, a method is known in which an algorithm is first pre-trained on initially collected data to obtain an initial model with uncertainties. This has the advantage that the machine learning algorithm can be used relatively quickly for the control of at least one controllable system, without the need to first laboriously and accurately train all possible assignments, especially those that occur rarely. Then, during operation of the at least one controllable system, the machine learning algorithm or the initial model can be retrained based on corresponding data detected during operation of the at least one controllable system to eliminate uncertainties.

[0005] However, it is known that a problem arises here when a controllable system has an increasing number of functions to be controlled or corresponding actuators, but if during operation of the controllable system data relating to all these functions to be controlled are detected here and thus a large number or amount of data is transmitted to the corresponding control device for retraining the machine learning algorithm, this may lead to problems in the data transmission, such as interaction effects, for example latency and interference.

[0006] German Patent Application No. 10 2016 216 945 discloses a method for executing a function based on model values ​​of a functional model of a database, in which the model values ​​of the functional model of the database are determined at a query time, a model accuracy indication or a model plausibility indication is determined which indicates the accuracy or plausibility of the model values ​​of the functional model of the database at the query time, and the function is executed depending on the model accuracy indication or the model plausibility indication. [Prior art documents] [Patent documents]

[0007] [Patent Document 1] German Patent Application Publication No. 102016216945 Summary of the Invention [Problem to be solved by the invention]

[0008] The present invention is therefore based on the problem of providing an optimized method for training machine learning algorithms, in particular for retraining machine learning algorithms with uncertainty. [Means for solving the problem]

[0009] This problem is solved by a method for training a machine learning algorithm according to the features of claim 1.

[0010] This problem is further solved by a control device having the features of claim 7.

[0011] Furthermore, the problem is also solved by a system for training a machine learning algorithm according to claim 12.

[0012] Preferred embodiments and developments result from the dependent claims and the description with reference to the drawings.

[0013] Disclosure of the Invention According to one embodiment of the present invention, this problem is solved by a method for training a machine learning algorithm, where the machine learning algorithm is an algorithm for controlling at least one controllable system, the machine learning algorithm assigns possible output values ​​to input values, and the machine learning algorithm has an uncertainty for each assignment of input values ​​and output values, each uncertainty indicating how well the assignment of input values ​​and possible output values ​​has been trained so far, wherein the method comprises a step of pre-training the machine learning algorithm based on data initially collected by a control device that trains the machine learning algorithm to obtain an initial model. Further, during operation of the at least one controllable system, data characterizing a current state of the at least one controllable system is detected at a plurality of time points, and for each of the plurality of time points, the influence that the data detected at the corresponding time point has on the uncertainty currently included in the initial model is respectively determined, and for each of the plurality of time points, a resolution of the data detected at the corresponding time point is determined based on the determined data level and the respective influence that the data detected at the corresponding time point has on the uncertainty currently included in the initial model, in which case, for each of the plurality of time points, the data detected at the corresponding time point is transmitted to a control device based on or together with the corresponding determined resolution, and the initial model is re-trained by the control device based on the data transmitted to the control device.

[0014] Uncertainty is understood here to mean a value or variable that indicates the quality of the assignment between an input value and a corresponding output value, i.e., how much information this assignment is based on or how much training data has been recorded so far in the vicinity of this assignment.

[0015] Furthermore, initially collected data is understood to mean data or training data that can be initially used to train a machine learning algorithm, particularly in this case data that describes situations that occur frequently or regularly. The machine learning algorithm trained with this initially collected data is referred to herein as an initial model.

[0016] Furthermore, the influence that the detected data at a particular point in time has on the uncertainty currently contained in the initial model is understood to mean the information content of the corresponding data, which indicates to what extent these data are relevant for retraining the initial model, i.e., whether or to what extent the uncertainty in the initial model can be reduced by retraining the initial model on this data. In this case, the uncertainty contained in the initial model, which may have already been retrained at a particular point in time, is referred to as the currently contained uncertainty.

[0017] Furthermore, resolution is understood to mean the accuracy with which the detected data is transmitted to the control device and displayed or mapped. For example, resolution can indicate with what quality the detected image data is transmitted or how many pixels of the detected image data are transmitted. In the case of dynamic measurements, resolution can also indicate, for example, at what point in time the detected data is transmitted to the control device. In this case, resolution also affects the memory size of the corresponding data.

[0018] Furthermore, the established data level represents a defined level of information in the data transmitted to the control device, such as a maximum amount of data to be transmitted or at least how much the total amount of information in the transmitted data should be, etc. In this case, the established data level can be set, for example, by the model manufacturer or the manufacturer of the at least one controllable system, or by the operator of the control device or the corresponding data center.

[0019] Thus, overall, the resolution of the detected data, and thus the memory size, can be adapted so that the detected data is not transmitted at its full resolution or memory size. This prevents problems during data transmission, such as latency and interference effects. Furthermore, the data transmission can be configured to take into account the given conditions of the corresponding data transmission system, such as the capacity of a CAN bus or the available bandwidth of wireless data transmission. Furthermore, the method ensures dynamic refinement or active learning of the algorithm, whereby the transmitted data can be trained even in situations that are extremely rare in practice, based on the fact that the transmitted data simultaneously contains the desired amount of information. This further allows the machine learning algorithm to increase the accuracy of assigning input values ​​to output values. Overall, therefore, an optimized method for training a machine learning algorithm, and in particular, an optimized method for retraining a machine learning algorithm with uncertainty, is presented.

[0020] The at least controllable system here may for example be a driver assistance function of the autonomous vehicle, with functions corresponding to driving the autonomous car, such as gear selection, speed selection or temperature setting. Furthermore, the at least one controllable system may for example also be a further system that is controllable based on a machine learning algorithm, such as a kitchen appliance or a washing machine.

[0021] The individual time points may furthermore be, for example, equally distributed across the time axis, although other distributions are also possible.

[0022] The input value may further be, for example, a sensor signal. The input value may further be set, for example, by a user. The output value characterizes a corresponding drive control signal for controlling the at least one controllable system or a value to which the at least one controllable system is set.

[0023] In one embodiment, the initial model is then a Gaussian process.

[0024] A Gaussian process is understood in this case to mean a stochastic process in which each finite subset of random variables has a multidimensional normal distribution (Gaussian distribution). In general, Gaussian processes represent time, space or any other function whose values ​​can only be modeled with certain uncertainties and probabilities based on incomplete information.

[0025] Therefore, this kind of Gaussian process is suitable for obtaining or training an initial model in a fast and simple manner.

[0026] Furthermore, the initial model may be any other machine learning algorithm that involves uncertainty, such as a Bayesian neural network.

[0027] The determined data level can then indicate at least how much influence the data transmitted to the control device should have on the uncertainty currently contained in the initial model, i.e., at least how much information is desired in the transmitted data for retraining the initial model. In particular, the determined data level can be determined so that the uncertainty in the machine learning algorithm or initial model is reduced as quickly as possible, thereby reducing the computation time and corresponding required resources when improving or retraining the initial model.

[0028] In this case, for each of the plurality of time points, determining a resolution of the detected data at the corresponding time point based on the determined data level and the respective influence that the detected data at the corresponding time point has on the uncertainty currently included in the initial model may include determining a resolution such that the memory size of the data to be transmitted to the control device is minimized while the amount of information in the data transmitted to the control device is at least as large as the influence of the data transmitted to the control device on the uncertainty currently included in the initial model. Therefore, a constrained optimization technique can be used to determine the corresponding resolution.

[0029] The resolution of the data may further be the spatial resolution, areal resolution or temporal resolution of the data.

[0030] Spatial resolution is understood to mean the precision of the representation or transmission of a three-dimensional plane.

[0031] Surface resolution is further understood to mean the precision of the display or transmission of a two-dimensional plane, where, for example, the resolution at the bottom left edge of the two-dimensional plane is higher than at the top right edge.

[0032] Time resolution is further understood to mean the point in time at which the data detected during the dynamic measurement are transmitted to the control device.

[0033] According to a further embodiment of the present invention, there is also provided a method for controlling at least one controllable system, wherein the method comprises the steps of training a machine learning algorithm for controlling the at least one controllable system according to the above-mentioned method, and controlling the at least one controllable system based on the trained machine learning algorithm.

[0034] Thus, an optimized method for controlling at least one controllable system is presented, in which the at least one controllable system is controlled based on a machine learning algorithm trained based on the optimized method. Therefore, by first training an initial model, the machine learning algorithm can be used to control the at least one controllable system relatively quickly, without the need to first painstakingly accurately train all possible assignments, especially those that occur rarely. This further saves computation time and corresponding computational capacity when training the machine learning algorithm, so that the method can also be performed by a controller configured in the at least one controllable system itself, which has less capacity than a comparable controller typically configured in a backend. Furthermore, to reduce uncertainty in the initial model, the machine learning algorithm or the initial model can be retrained during operation of the at least one controllable system, in which case the resolution of the data to be transmitted, and thus the memory size, are adapted so that the data to be transmitted is not transmitted at its full resolution or memory size. This can prevent problems during data transmission, such as latency and interaction effects, such as interference. Furthermore, the data transmission can be configured to take into account the given conditions of the corresponding data transmission system, for example the capacity of the CAN bus or the available bandwidth of wireless data transmission. Furthermore, the method ensures dynamic improvement or active learning of the algorithm, which can be trained even in situations that are very rare in practice based on the selected and transmitted data. This further allows the machine learning algorithm to increase the accuracy of assigning input values ​​to output values.

[0035] According to a further embodiment of the present invention, there is also provided a control device for selecting training data for training a machine learning algorithm, wherein the machine learning algorithm is an algorithm for controlling at least one controllable system, the machine learning algorithm assigns possible output values ​​to input values, the machine learning algorithm has uncertainties for each assignment of input values ​​and output values, the uncertainties respectively indicating how well the assignment of input values ​​and possible output values ​​has been trained so far, the machine learning algorithm is pre-trained based on initially collected data to obtain an initial model, and the control device selects data characterizing a current state of the at least one controllable system at multiple time points during operation of the at least one controllable system. a detection unit configured to detect, for each of the plurality of time points, an influence that the data detected at the corresponding time point has on the uncertainty currently included in the initial model; a first determination unit configured to determine, for each of the plurality of time points, a resolution of the data detected at the corresponding time point based on the determined data level and the respective influence that the data detected at the corresponding time point has on the uncertainty currently included in the initial model; and a transmission unit configured to transmit, for each of the plurality of time points, the data detected at the corresponding time point based on or together with the corresponding determined resolution to a control device for training a machine learning algorithm.

[0036] Overall, therefore, a control device for transmitting or selecting training data for training a machine learning algorithm is presented, which is configured to adapt the resolution of detected data and thus the memory size so that the detected data is not transmitted at its full resolution or memory size. This prevents problems during data transmission, such as latency and interaction effects, such as interference. Furthermore, the data transmission can be configured to take into account the given conditions of the corresponding data transmission system, such as the capacity of a CAN bus or the available bandwidth of wireless data transmission.

[0037] In one embodiment, the initial model is also a Gaussian process in this case. This type of Gaussian process is suitable for obtaining or training the initial model in a fast and easy way. Furthermore, the initial model may be any other machine learning algorithm with uncertainty, such as a Bayesian neural network.

[0038] The determined data level can then also indicate at least how much influence the transmitted data should have on the uncertainty currently contained in the initial model, i.e., at least how much information is desired in the transmitted data for retraining the initial model. In particular, the determined data level can be determined so that the uncertainty in the machine learning algorithm or initial model is reduced as quickly as possible, thereby reducing the computation time and corresponding required resources when improving or retraining the initial model.

[0039] In this case, the second determination unit may be configured to determine, for each of the plurality of time points, a resolution of the data detected at the corresponding time point such that the memory size of the data transmitted to the control device is minimized while the information content of the data transmitted to the control device is at least as large as the influence of the data transmitted to the control device on the uncertainties currently included in the initial model. Accordingly, the second determination unit may be configured to use a constrained optimization method to determine the corresponding resolution.

[0040] The resolution of the data may further be the spatial, areal or temporal resolution of the data, however, other types of resolution may also be used, such as for example radiometric resolution.

[0041] According to a further embodiment of the present invention, there is also provided a system for training a machine learning algorithm, wherein the machine learning algorithm is an algorithm for controlling at least one controllable system, the machine learning algorithm assigning possible output values ​​to input values, the machine learning algorithm having uncertainties for each assignment of input values ​​and output values, the uncertainties each indicating how well the assignment of input values ​​and possible output values ​​has been trained so far, the system comprising: the above-mentioned control device for selecting training data for training the machine learning algorithm; and a control device for training the machine learning algorithm, wherein the control device for training the machine learning algorithm is configured to pre-train the machine learning algorithm based on initial collected data to obtain an initial model, and to re-train the initial model based on data obtained from the control device for selecting training data for training the machine learning algorithm for controlling the controllable system.

[0042] Thus, an optimized system for training machine learning algorithms, particularly for retraining machine learning algorithms with uncertainty, is presented. The system can be configured to adapt the resolution of detected data and thus the memory size so that the detected data is not transmitted at its full resolution or memory size. This can prevent problems during data transmission, such as latency and interference. Furthermore, the data transmission can be configured to take into account the given conditions of the corresponding data transmission system, such as the capacity of a CAN bus or the available bandwidth of wireless data transmission. Furthermore, the system ensures dynamic refinement or active learning of the algorithm, which can be trained even in rare situations based on selected and transmitted data. This further allows the machine learning algorithm to increase the accuracy of assigning input values ​​to output values.

[0043] According to a further embodiment of the present invention, there is also provided a system for controlling a controllable system, wherein the system includes at least one controllable system and a control device for controlling the at least one controllable system based on a machine learning algorithm trained by the above-mentioned system for training a machine learning algorithm.

[0044] Thus, an optimized system for controlling at least one controllable system is presented, in which the at least one controllable system is controlled based on a machine learning algorithm trained according to the optimized method. Therefore, by first training an initial model, the machine learning algorithm can be used to control the at least one controllable system relatively quickly, without the need to first painstakingly accurately train all possible assignments, especially those that occur rarely. This further saves computation time and corresponding computational capacity when training the machine learning algorithm. Furthermore, the machine learning algorithm or the initial model can be retrained during operation of the at least one controllable system to reduce uncertainty in the initial model. In this case, the resolution of the detected data, and thus the memory size, is adapted so that the detected data is not transmitted at its full resolution or memory size. This prevents problems during data transmission, such as latency and interference. Furthermore, the data transmission can be configured to take into account the given conditions of the corresponding data transmission system, such as the capacity of the CAN bus or the available bandwidth of wireless data transmission. Furthermore, the system ensures dynamic improvement or active learning of the algorithm, which can be trained even in situations that are very rare in practice based on selected and transmitted data, thereby further enabling the machine learning algorithm to increase the accuracy in assigning input values ​​to output values ​​when controlling at least one controllable system.

[0045] The at least one controllable system here may be an autonomous vehicle, i.e. a system in which a number of controllable functions or controllable actuators are controllable simultaneously and independently of one another during operation, for example for speed setting, gear selection or temperature control. Furthermore, the at least one controllable system may also be further systems, for example kitchen appliances or washing machines, that are controllable based on machine learning algorithms.

[0046] Overall, therefore, it is determined that the present invention presents an optimized method for training machine learning algorithms, and in particular, an optimized method for retraining machine learning algorithms with uncertainty.

[0047] The described embodiments and developments can be combined with one another in any combination.

[0048] Further possible embodiments, developments and implementations of the invention are intended to include not explicitly mentioned combinations of the features of the invention that are explained with respect to the preceding and following examples.

[0049] The accompanying drawings are intended to provide a further understanding of embodiments of the present invention, they illustrate embodiments and, together with the description, serve to explain the principles and concepts of the invention.

[0050] Other embodiments and many of the advantages discussed above will become apparent in conjunction with the drawings, in which elements shown are not necessarily drawn to scale relative to each other. [Brief explanation of the drawings]

[0051] [Figure 1] 1 is a flowchart of a method for training a machine learning algorithm according to an embodiment of the present invention. [Figure 2] FIG. 1 is a schematic block diagram of a system for training a machine learning algorithm according to an embodiment of the present invention.

[0052] In the figures of the drawings, the same reference numbers, unless otherwise stated, indicate identical or functionally equivalent elements, parts or components. DETAILED DESCRIPTION OF THE INVENTION

[0053] FIG. 1 shows a flowchart of a method 1 for training a machine learning algorithm according to an embodiment of the present invention.

[0054] Digital controllers are used in many applications for open-loop and closed-loop control of technical systems, also generally referred to as controllable systems below, by processing sensor signals or other input values ​​according to predetermined control algorithms to determine one or more output values ​​for functions, e.g., manipulated variables for controlling actuators.

[0055] The control algorithm here can be, for example, a machine learning algorithm. In this case, each individual controllable system collects data about its use, and these data are then used to train the machine learning algorithm accordingly. Typically, this type of machine learning algorithm is based on using statistical methods to train a data processing system to perform a specific task without the data processing system being explicitly programmed for this purpose from the beginning. In this case, the goal of machine learning is to build an algorithm that can learn from data and make predictions. These algorithms create mathematical models that can be used, for example, to classify data.

[0056] In this case, a method is known in which an algorithm is first pre-trained on initially collected data to obtain an initial model with uncertainties. This has the advantage that the machine learning algorithm can be used relatively quickly to control at least one controllable system, without the need to first train all possible assignments, especially those that occur rarely, with great effort. This further saves computation time and corresponding computational capacity when training the machine learning algorithm. Then, during operation of the at least one controllable system, the machine learning algorithm or the initial model can be re-trained based on corresponding data detected during operation of the at least one controllable system to eliminate uncertainties.

[0057] However, it is known that the problem here is that the controllable system has an increasing number of functions to be controlled or corresponding actuators, but if during the operation of the controllable system data relating to all these functions to be controlled are detected here and thus a large number of data or a large amount of data is transmitted to the corresponding control device for retraining the machine learning algorithm, this may lead to problems in the data transmission, such as interaction effects, for example latency or interference.

[0058] 1 , a method 1 for training a machine learning algorithm is shown, where the machine learning algorithm is an algorithm for controlling at least one controllable system, the machine learning algorithm assigns possible output values ​​to input values, and the machine learning algorithm has uncertainties for each assignment of input values ​​and output values, each uncertainty indicating how well the assignment of input values ​​and possible output values ​​has been trained so far. In this case, the method 1 includes a step 2 of pre-training the machine learning algorithm based on data initially collected by a control device that trains the machine learning algorithm to obtain an initial model. Further, in a subsequent step 3, during operation of the at least one controllable system, data characterizing a current state of the at least one controllable system is detected at a plurality of time points, respectively; in step 4, for each of the plurality of time points, the influence that the data detected at the corresponding time point has on the uncertainty currently included in the initial model is determined, respectively; in step 5, for each of the plurality of time points, the resolution of the data detected at the corresponding time point is determined based on the determined data level and the respective influence that the data detected at the corresponding time point has on the uncertainty currently included in the initial model, respectively; wherein, in a subsequent step 6, for each of the plurality of time points, the data detected at the corresponding time point is transmitted to a control device based on the corresponding determined resolution, and subsequently, in step 7, the initial model is re-trained by the control device based on the data transmitted to the control device.

[0059] Uncertainty is understood here to mean a value or variable that indicates the quality of the assignment between an input value and a corresponding output value, i.e., how much information this assignment is based on or how much training data has been recorded so far in the vicinity of this assignment.

[0060] Initially collected data is further understood to mean data or training data that can be initially used to train a machine learning algorithm, particularly in this case data that describes situations that occur frequently or regularly. The machine learning algorithm trained with this initially collected data is referred to herein as an initial model.

[0061] Furthermore, the influence that the detected data at a particular point in time has on the uncertainty currently contained in the initial model is understood to mean the information content of the corresponding data, which indicates to what extent these data are relevant for retraining the initial model, i.e., whether or to what extent the uncertainty in the initial model can be reduced by retraining the initial model on this data. In this case, the uncertainty contained in the initial model, which may have already been retrained at a particular point in time, is referred to as the currently contained uncertainty.

[0062] Furthermore, resolution is understood to mean the accuracy with which the detected data is transmitted to the control device and displayed or mapped. For example, resolution can indicate with what quality the detected image data is transmitted or how many pixels of the detected image data are transmitted. In the case of dynamic measurements, resolution can also indicate, for example, at what point in time the detected data is transmitted to the control device. In this case, resolution also affects the memory size of the corresponding data.

[0063] Furthermore, the established data level represents a defined level of information in the data transmitted to the control device, such as a maximum amount of data to be transmitted, or at least what the overall amount of information in the transmitted data should be, etc. In this case, the established data level can be set, for example, by the model manufacturer or the manufacturer of the at least one controllable system, or by the operator of the control device or the corresponding data center.

[0064] Thus, overall, the resolution of the detected data, and thus the memory size, is adapted so that the detected data is not transmitted at its full resolution or memory size. This prevents problems during data transmission, such as latency and interference. Furthermore, data transmission can be configured to take into account the given conditions of the corresponding data transmission system, such as the capacity of a CAN bus or the available bandwidth of wireless data transmission. Furthermore, method 1 ensures dynamic refinement or active learning of the algorithm, whereby the transmitted data can be trained even in situations that are extremely rare in practice, based on the desired amount of information being simultaneously transmitted. This further allows the machine learning algorithm to increase the accuracy of assigning input values ​​to output values. Thus, overall, an optimized method 1 for training a machine learning algorithm, particularly an optimized method 1 for retraining a machine learning algorithm with uncertainty, is presented.

[0065] Method 1 can here, for example, be repeatedly executed at specific time intervals, for example every 10 minutes. Furthermore, it is also possible to adaptively adapt the time intervals between the individual repetitions of method 1. In this case, a predictive model can be used that indicates which data is expected at a certain point in time, and based on this predictive model and possibly its uncertainty, it can be determined at what point in time or after what period of time method 1 is executed again.

[0066] Furthermore, the points in the plurality of points in time may be equidistantly distributed across the time axis, for example, corresponding points in time may occur every 1 second, every 10 seconds, every 30 seconds, although other distributions are also possible.

[0067] If the at least one controllable system has multiple channels of data, e.g., multiple channels of data from different controllable functions of the at least one controllable system, then a similar resolution can be determined for each channel, or alternatively, different or unique resolutions can be determined for at least some of the channels.

[0068] Therefore, overall, based on the method 1, it is possible to realize an optimized selection of information on which resolution of data to transmit.

[0069] According to the embodiment of FIG. 1 , the initial model is a Gaussian process, which here refers to a multivariate normal distribution with functional correlation. Gaussian processes are useful as methods in machine learning because a Gaussian process, given a covariance function that controls its properties, can also be understood as a prior prediction for the properties of an unknown function. This prior prediction can be efficiently conditioned on the data, thereby generating a posterior distribution that can be used to predict unknown data points. Gaussian processes therefore provide a complete Bayesian framework for reasoning on functions.

[0070] Therefore, in this case, an indication of the uncertainty or quality, which is also frequently called the tolerance, can be determined as the variance, in particular the prediction variance of the Gaussian process.

[0071] Furthermore, the determined data level indicates at least how much the transmitted data should contribute to the uncertainty currently contained in the initial model, i.e., at least how much information is desired in the transmitted data to retrain the initial model.

[0072] In this case, the resolution can be determined based on a constrained optimization problem, respectively. According to the embodiment of Fig. 1, for each of the plurality of time points, step 5 of determining the resolution of the data detected at the corresponding time point based on the determined data level and the respective influence that the data detected at the corresponding time point has on the uncertainty currently contained in the initial model includes a step of determining the resolution so that the memory size of the data to be transmitted to the control device is minimized while the amount of information in the data transmitted to the control device is at least as large as the influence of the data transmitted to the control device on the uncertainty currently contained in the initial model.

[0073] In this case, furthermore, in order to reduce the calculation time, the determination of the resolution can also be interrupted as soon as the memory size falls below the threshold value for the memory size in this case.

[0074] Resolution may also be the spatial, areal, and / or temporal resolution of the data.

[0075] FIG. 2 shows a schematic block diagram of a system 10 for training machine learning algorithms according to an embodiment of the present invention.

[0076] In this case, the machine learning algorithm is also an algorithm for controlling at least one controllable system, where the machine learning algorithm assigns possible output values ​​to input values, and where the machine learning algorithm has an uncertainty for each assignment of input values ​​and output values, each uncertainty indicating how well the assignment of input values ​​and possible output values ​​has been trained so far.

[0077] As shown in FIG. 2 , the system 10 includes a control device 11 for selecting training data for training the machine learning algorithm and a control device 12 for training the machine learning algorithm, wherein the control device 12 for training the machine learning algorithm is configured to pre-train the machine learning algorithm based on initial collected data to obtain an initial model, and to re-train the initial model based on data obtained from the control device 11 for selecting training data for training the machine learning algorithm to control a system controllable by the control device 11.

[0078] The control device 11 for selecting training data for training the machine learning algorithm may in this case in particular be configured in or integrated into the at least one controllable system itself.

[0079] The control device 12 for training the machine learning algorithm may further be configured in the at least one controllable system itself, but may also be configured in a backend, where a predictive model is used for what data is expected in the near future and / or at what resolution it will be transmitted in the future, and the predictive model may be a machine learning algorithm that is trained during the operation of the at least one controllable system, and the predictive model may be fixed or adaptive. If the predictive model is adaptive, a probabilistic predictive model is also required, which can be used to evaluate the corresponding uncertainty, for example, to evaluate at what point in time a corresponding amount of information in the data can be expected. In this case, the control device 12 can, for example, predict which data will be available in the near future based on the predictive model and accordingly inform the control device 11, thereby reducing the uncertainty currently included in the initial model based on these data expected in the near future.

[0080] In this case, according to the embodiment of FIG. 2, the at least one controllable system is furthermore an autonomous vehicle or a function related to the operation of the autonomous vehicle, in particular a speed setting unit, a gear selection unit or a temperature control unit.

[0081] As Figure 2 further shows, the control device 11 for selecting training data for training the machine learning algorithm in this case comprises a detection unit 13 configured to detect data characterizing a current state of at least one controllable system at each of a plurality of time points during operation of the at least one controllable system; a first determination unit 14 configured to determine, for each of the plurality of time points, the influence that the data detected at the corresponding time point has on the uncertainty currently included in the initial model; a second determination unit 15 configured to determine, for each of the plurality of time points, the resolution of the data detected at the corresponding time point based on the determined data level and the respective influence that the data detected at the corresponding time point has on the uncertainty currently included in the initial model; and a transmission unit 16 configured to transmit, for each of the plurality of time points, the data detected at the corresponding time point to the control device 12 for training the machine learning algorithm based on the corresponding determined resolution.

[0082] The detection unit and the transmission unit here may for example both be integrated in a transceiver. Furthermore, both the first and second determination unit may each be realized for example on the basis of code stored in a memory and executable by a processor.

[0083] Furthermore, the control device 12 for training the machine learning algorithm also comprises a detection unit 17 configured to detect or receive the data transmitted by the transmission unit 16 .

[0084] The respective data can in this case be detected, for example, by corresponding sensors or corresponding control devices arranged in the at least one controllable system or in the autonomous vehicle.

[0085] According to the embodiment of FIG. 2, the initial model is a Gaussian process.

[0086] Additionally, the determined data level also indicates how much the data transmitted to the controller should at least influence the uncertainty currently contained in the initial model.

[0087] According to the embodiment of Fig. 2, the second determining unit 15 is further configured to determine, for each of the plurality of time points, a resolution of the data detected at the corresponding time point such that the memory size of the data to be transmitted to the control device is minimized while at the same time the information content of the data transmitted to the control device 12 for training the machine learning algorithm is at least as large as the influence of the data transmitted to the control device 12 on the uncertainties currently contained in the initial model. Accordingly, the second determining unit 15 is also configured to determine the resolution based on a respective constrained optimization problem.

Claims

1. A method (1) for training a machine learning algorithm, wherein the machine learning algorithm is an algorithm for controlling at least one controllable system, the machine learning algorithm assigns possible output values to input values, the machine learning algorithm has an uncertainty for each assignment of the input value and the output value, and the uncertainty indicates how well the assignment of the input value and the possible output value has been trained so far, in method (1). - A step (2) of pre-training a machine learning algorithm based on data initially collected by a control device that trains the machine learning algorithm to obtain an initial model. - A step (3) of respectively detecting data characterizing the current state of the at least one controllable system at a plurality of time points during the operation of the at least one controllable system. - A step (4) of respectively determining, for each of the plurality of time points, the influence that the data detected at the corresponding time point has on the uncertainty currently included in the initial model. - A step (5) of determining the resolution of the data detected at the corresponding time point based on the determined data level and the respective influence that the data detected at the corresponding time point has on the uncertainty currently included in the initial model for each of the plurality of time points. - A step (6) of transmitting, for each of the plurality of time points, the data detected at the corresponding time point to the control device based on the corresponding determined resolution. - A step (7) of re-training the initial model by the control device based on the data transmitted to the control device. The method (1) comprising.

2. The method (1) according to claim 1, wherein the initial model is a Gaussian process.

3. The method (1) according to claim 1, wherein the determined data level indicates at least how much the data transmitted to the control device should affect the uncertainty currently included in the initial model.

4. For each of the plurality of time points, based on the determined data level and the respective impact on the uncertainty currently included in the initial model of the data detected at the corresponding time point, the step (5) of determining the resolution of the data detected at the corresponding time point further includes the following steps, namely, - The method (1) according to claim 3, including the step of determining the corresponding resolution such that, simultaneously with the memory size of the data to be transmitted to the control device being minimized, the amount of information of the data transmitted to the control device is at least as much as the degree to which the data transmitted to the control device should affect the uncertainty currently included in the initial model.

5. The method (1) according to claim 1, wherein the resolution of the data is the spatial resolution, surface resolution, or temporal resolution of the data.

6. A method for controlling at least one controllable system, comprising: - Training a machine learning algorithm for controlling the at least one controllable system by the method (1) according to any one of claims 1 to 5; and - Controlling the at least one controllable system based on the trained machine learning algorithm. A method including the above steps.

7. A control device for selecting training data for training a machine learning algorithm, wherein the machine learning algorithm is an algorithm for controlling at least one controllable system; the machine learning algorithm assigns possible output values to input values; the machine learning algorithm has uncertainty for each assignment of input values and output values; the uncertainty indicates, respectively, the degree to which the assignment of the input values and the possible output values has been trained well so far; the machine learning algorithm is pre-trained based on initially collected data to obtain an initial model; the control device (11) is configured to respectively detect data characterizing the current state of the at least one controllable system at a plurality of time points during the operation of the at least one controllable system. a first determination unit (14) configured to determine, for each of the plurality of time points, the influence on the uncertainty currently included in the initial model of the data detected at the corresponding time point; a second determination unit (15) configured to determine, for each of the plurality of time points, the resolution of the data detected at the corresponding time point based on the determined data level and the respective influence on the uncertainty currently included in the initial model of the data detected at the corresponding time point; a transmission unit (16) configured to transmit, for each of the plurality of time points, the data detected at the corresponding time point to a control device for training the machine learning algorithm based on the corresponding determined resolution; A control device comprising the above.

8. The control device according to claim 7, wherein the initial model is a Gaussian process.

9. The control device according to claim 7, wherein the determined data level indicates at least to what extent the data transmitted to the control device should affect the uncertainty currently included in the initial model.

10. For each of the plurality of time points, the second determination unit (15) is configured to determine the corresponding resolution such that, at the same time as the memory size of the data transmitted to the control device for training the machine learning algorithm is minimized, the amount of information of the data transmitted to the control device for training the machine learning algorithm is at least as great as at least to what extent the data transmitted to the control device for training the machine learning algorithm should affect the uncertainty currently included in the initial model. The control device according to claim 9.

11. The control device according to claim 7, wherein the resolution of the data is the spatial resolution, surface resolution or temporal resolution of the data.

12. A system for training a machine learning algorithm, comprising the machine learning algorithm is an algorithm for controlling at least one controllable system; the machine learning algorithm assigns possible output values to input values; the machine learning algorithm has uncertainty for each assignment of the input values and the output values; The uncertainty indicates, respectively, how well the assignment between the input value and the possible output value has been trained so far. The system (10) includes a control device (11) for selecting training data for training the machine learning algorithm according to any one of claims 7 to 11, and a control device (12) for training the machine learning algorithm. The control device (12) for training the machine learning algorithm is configured to pre-train the machine learning algorithm based on the initially collected data to obtain an initial model, and re-train the initial model based on the data obtained from the control device (11) for selecting training data for training the machine learning algorithm.

13. A system for controlling a controllable system, at least one controllable system, and a control device for controlling the at least one controllable system based on a machine learning algorithm trained by a system for training the machine learning algorithm according to claim 12. A system including the above.

14. The system according to claim 13, wherein the at least one controllable system is an autonomous vehicle.

Images