DE102020216202A1 - Method for training a learning algorithm for converting sensor data from a sensor system into output data for controlling an actuator and method for controlling an actuator of a vehicle - Google Patents

Abstract

The present invention relates to a method for training a learning algorithm (102) for converting sensor data (108, x) of a sensor (106) into output data (108, y) for controlling an actuator (110). The method comprises the following steps: inputting the sensor data (104, x) into the learning algorithm (102), the learning algorithm (102) having a first learning module (112) parameterized by first parameters (114) and a first learning module (112) a second learning module (116) connected in series and parameterized by second parameters (118) for converting the sensor data (104, x) into the output data (108, y); outputting the output data (108,y) by the learning algorithm (102); and optimizing the learning algorithm (102) by an optimization module (120) based on the output data (108, y), wherein a first objective function (122) is optimized by changing the first parameters (114) and the second parameters (118) during the optimization of the first objective function (122) are kept constant.

Description

field of invention

The invention relates to a method for training a learning algorithm for converting sensor data from a sensor system into output data for controlling an actuator and a method for controlling an actuator of a vehicle. Furthermore, the invention relates to a device, a computer program and a computer-readable medium for executing at least one of the methods mentioned.

State of the art

When training an artificial neural network, it is possible, by optimizing a cost function, to optimize one or more target variables or also a target variable taking into account a marginal variable. For example, a trade-off parameter of the cost function can be defined to compensate for a conflict of objectives between the variables to be optimized. However, determining a suitable trade-off parameter can involve considerable effort. Furthermore, such a cost function involving multiple quantities to be optimized may require a relatively complex network architecture.

Such an artificial neural network can be used, for example, in the context of autonomous driving for processing sensor data from a sensor system and for providing corresponding inputs for a control unit for controlling an actuator system, for example for a speed controller or the like.

Disclosure of Invention

Against this background, with the approach presented here, a method for training a learning algorithm, a method for controlling an actuator system, a device, a computer program and a computer-readable medium are presented according to the independent claims. Advantageous developments and improvements of the approach presented here result from the description and are described in the dependent claims.

Advantages of the Invention

Advantageously, embodiments of the present invention make it possible to optimize multiple competing (target) variables when training a learning algorithm such as an artificial neural network without the need for the often difficult and a priori unclear manual determination of a trade-off between these variables is. On the one hand, this enables a targeted optimization of the learning algorithm. It can thus be achieved, for example, that the learning algorithm fulfills several goals at once without being driven in different directions that may conflict with one another. On the other hand, the learning process can be significantly simplified. Furthermore, the use of learning algorithms with relatively simple architectures is made possible, which saves time and money.

A first aspect of the invention relates to a computer-implemented method for training a learning algorithm for converting sensor data from a sensor system into output data for controlling an actuator system. The method comprises at least the following steps: inputting the sensor data into the learning algorithm, the learning algorithm comprising a first learning module parameterized by first parameters and a second learning module connected in series with the first learning module and parameterized by second parameters for converting the sensor data into the output data; outputting the output data by the learning algorithm; and optimizing the learning algorithm by an optimization module based on the output data, wherein a first objective function is optimized by changing the first parameters and the second parameters are kept constant while optimizing the first objective function.

The method can be executed automatically by a processor, for example.

A learning algorithm can generally be understood as an algorithm for machine learning of a model that converts inputs into specific outputs, for example by classification and/or regression. Examples of possible learning algorithms are artificial neural networks, genetic algorithms, support vector machines, k-means, kernel regression or discriminant analysis. The learning algorithm can also include a combination of several of the examples mentioned.

It is possible, for example, for the two learning modules to be alternately optimized for different target values, with the target value of the learning module that is not currently being optimized being kept constant (see below).

The sensor system can be configured, for example, to detect an environment and/or a driving state of a vehicle. The vehicle may be an automobile, such as a car, truck, bus, or motorcycle. In a broader sense, a vehicle can also be understood as an autonomous, mobile robot. For example, the sensor system can have at least one environment sensor such as an ultrasonic, radar, lidar sensor or a camera and/or at least one vehicle dynamics sensor such as a yaw rate, acceleration, wheel speed or steering wheel angle sensor. In addition, the sensors can use a location sensor to determine an absolute position of the vehicle using a global navigation satellite system such as GPS, GLONASS or similar. include.

The sensor data used to train the learning algorithm can be real, i. H. act through the sensors generated by measurement sensor data. However, it is also possible for sensor data calculated by a computer, for example as part of a simulation of the vehicle and its surroundings, to be used to train the learning algorithm.

The output data can be an acceleration signal or a steering angle signal, for example. However, the output data can also include any other information suitable for controlling the actuators, such as recognized objects in the vicinity of the vehicle, positions and/or locations of the recognized objects or a trajectory of the vehicle.

The actuator system can be configured, for example, to steer, accelerate, brake and/or navigate the vehicle. For this purpose, the actuator can include a steering actuator, a brake actuator, an engine control unit, an electric motor or a combination of at least two of the examples mentioned. It is possible for the vehicle to be equipped with a driver assistance system or robot control system for partially or fully automated control of the actuators based on the sensor data. The driver assistance system or robot control system can be implemented as hardware and/or software in a control unit of the vehicle.

For example, the learning algorithm for classifying accelerations can be configured for a vehicle controller. In this case, for example, image sizes of one or more target objects can be entered as the sensor data in the learning algorithm. Such an image size can be, for example, a position in the image of the object, for example in the form of a u or v pixel value, a time-to-contact value, or TTC value for short, information about an optical flow or around Trade turn signal or brake light information from objects. Accordingly, the learning algorithm output data may include an acceleration requested by the vehicle controller.

As an alternative or in addition to the aforementioned example, the learning algorithm can also be configured for vehicle lateral guidance. For this purpose, information about detected lane boundaries of an ego vehicle can be entered as the sensor data in the learning algorithm, for example in the form of pixel coordinates of a right and/or left lane boundary. The learning algorithm can be configured to predict a steering angle based on the sensor data. Accordingly, the output data of the learning algorithm can include a predicted steering angle.

However, other implementations are also conceivable. In principle, the approach presented here is suitable for predicting any type of signal from which a smooth course is expected.

A target function can be understood, for example, as a cost or loss function to be minimized or a reward function to be maximized.

The first parameters can be variables, i. H. changeable weights during a runtime of the learning algorithm in an optimization process. On the other hand, the second parameters can be either variables or constants. For example, the second parameters can have been predefined as constants prior to the training of the learning algorithm. However, it is also possible for the second parameters to be changed in the same way as the first parameters when training the learning algorithm, for example as part of the optimization of a second target function that differs from the first target function (see below). For example, the second parameters can be filter coefficients of a filter with a finite impulse response, also known as an FIR filter (FIR=finite impulse response). The FIR filter can be configured, for example, as a low-pass filter, high-pass filter or band-pass filter (see also below).

The first target function can be optimized, for example, in a gradient method by backpropagation. However, other optimization methods are also conceivable.

For example, the steps of inputting, outputting and optimizing can be repeated cyclically, in particular until a local or global optimum is reached when optimizing the learning algorithm.

The approach presented here enables, for example, in the case of a cruise controller of an autonomous vehicle, a behavioral approximation of the cruise controller to a predetermined target curve, which can be generated, for example, by a pro professional driver, and also ensures that the learned, well-approximated journey is not characterized by jerky changes in vehicle acceleration, which the vehicle occupants could find uncomfortable.

For example, the learning algorithm can include a combination of an artificial neural network as the first learning module and a frequency filter as the second learning module, as used in classical signal processing. The frequency filter can be connected to the output of the artificial neural network and, for example, be designed in advance of the learning procedure in such a way that a resulting overall signal from the learning algorithm does not contain undesirable properties of a marginal variable, such as high frequency components of an acceleration signal, or at best contains them in a greatly weakened form. In the simplest case, the parameterization of the frequency filter can remain constant during the learning procedure, i. That is to say, the second parameters are not optimized using the first objective function, in contrast to the first parameters, here the parameters of the artificial neural network. In this way, compliance with conditions that the marginal size is supposed to fulfill can be automatically guaranteed without having to introduce a trade-off parameter into the cost function.

Depending on the boundary size, this can also simplify the learning procedure, since all those signals at the output of the artificial neural network that differ only in the frequency components filtered out by the frequency filter from signals at the output of the frequency filter lead to the same solution. This results in an increase in the number of (local) optima of the first objective function, which allows the learning procedure to converge faster and more reliably.

Ultimately, the reliability and recognition accuracy of the learning algorithm can thus be increased. In particular when used for safety-related functions such as driving functions, this can contribute to increasing traffic safety.

A second aspect of the invention relates to a computer-implemented method for controlling an actuator of a vehicle. In addition to the actuator system, the vehicle also includes a sensor system for detecting a driving state and/or an environment of the vehicle. The method comprises at least the following steps: receiving sensor data generated by the sensor system; inputting the sensor data into a learning algorithm trained in a method according to an embodiment of the first aspect of the invention to convert the sensor data into output data; and generating a control signal for controlling the actuator based on the output data.

In addition to the sensors and actuators, the vehicle can include a control unit for controlling the actuators based on the sensor data. The learning algorithm can be implemented as hardware and/or software in the control unit. The method can be executed automatically by a processor of the control device, for example.

Features of the method according to an embodiment of the first aspect of the invention can also be features of the method according to an embodiment of the second aspect of the invention and vice versa.

A third aspect of the invention relates to a device for data processing. The device comprises a processor configured to carry out the method according to an embodiment of the first aspect of the invention and/or the method according to an embodiment of the second aspect of the invention. A device for data processing can be understood to mean, for example, a computer, a computer system or a control device. The device may include hardware and/or software modules. In addition to the processor, the device may include memory and data communication interfaces for data communication with peripheral devices. Features of the method according to an embodiment of the first or second aspect of the invention can also be features of the device and vice versa.

A fourth aspect of the invention relates to a computer program. The computer program comprises instructions which, when the computer program is executed by the processor, cause a processor to carry out the method according to an embodiment of the first aspect of the invention and/or the method according to an embodiment of the second aspect of the invention.

A fifth aspect of the invention relates to a computer-readable medium on which the computer program according to an embodiment of the fourth aspect of the invention is stored. The computer-readable medium can be volatile or non-volatile data storage. For example, the computer-readable medium can be a hard drive, USB storage device, RAM, ROM, EPROM, or flash memory. The computer-readable medium can also be a data communication network such as the Internet or a data cloud (cloud) enabling a download of a program code.

Features of the method according to an embodiment of the first or second aspect of the invention can also be features of the computer program and/or the computer-readable medium and vice versa.

Ideas for embodiments of the present invention can be regarded as being based, among other things, on the ideas and findings described below.

According to one embodiment, to optimize the learning algorithm, a second objective function that differs from the first objective function is further optimized by changing the second parameters based on the output data. The first objective function and the second objective function can be different cost or reward functions, for example. The second target function can be optimized, for example, in a gradient method by backpropagation. However, other optimization methods are also conceivable. In this way, the learning algorithm can be specifically adapted to competing target values.

According to one embodiment, the first parameters are kept constant during the optimization of the second objective function. Additionally or alternatively, a function value of the first target function can be kept constant during the optimization of the second target function. A function value of the first target function can be understood as a dependent variable of the first target function. Undesirable interactions between the first target function and the second target function can thus be avoided when optimizing the learning algorithm.

It is also possible that a function value of the second target function is kept constant in addition to the second parameters during the optimization of the first target function.

According to one embodiment, the first objective function and the second objective function are alternately optimized in a plurality of loops, for example until a local or global optimum is reached. Thus, the learning algorithm can be optimized very efficiently. In addition, undesired interactions between the first target function and the second target function can thus be avoided.

According to one embodiment, the sensor data is input into the first learning module and the output data is output by the second learning module. In this case, outputs from the first learning module can be entered as inputs into the second learning module. This allows output data from the first learning module (configurable by optimizing the first objective function) to be filtered by the second learning module (preconfigured, for example).

Alternatively, the sensor data can be entered into the second learning module and the output data can be output by the first learning module. In this case, outputs from the second learning module can be entered as inputs into the first learning module. This allows output data from the second (preconfigured, for example) learning module to be filtered by the first learning module (configurable by optimizing the first objective function).

According to one embodiment, the first learning module is configured as a first filter for filtering outputs of the second learning module. The first parameters are filter coefficients of the first filter. For example, the sensor data can first be entered into the second learning module and the outputs of the second learning module can then be entered into the first learning module for further filtering. Accordingly, the output data from the learning algorithm can be outputs from the first learning module. This has the advantage that the filter coefficients can be selected independently of the parameters of the second learning module. In this way, target variable conflicts can be avoided when optimizing the learning algorithm.

According to an alternative embodiment, the second learning module is configured as a second filter for filtering outputs of the first learning module. The second parameters are filter coefficients of the second filter. For example, the sensor data can first be entered into the first learning module and the outputs of the first learning module can then be entered into the second learning module for further filtering. Accordingly, the output data from the learning algorithm can be outputs from the second learning module. This has the advantage that the filter coefficients can be selected independently of the parameters of the first learning module. In this way, target variable conflicts can be avoided when optimizing the learning algorithm.

The first or second filter can be, for example, a filter for digital signal processing, such as a filter with a finite impulse response, also known as an FIR filter (see above). Depending on the filter coefficients, the first or second filter can be configured as a high-pass, low-pass or band-pass filter, for example.

According to one embodiment, the first filter is a low-pass filter. This has the effect that the Outputs of the second learning module, such as an acceleration or steering angle signal, can be smoothed using the first learning module.

According to an alternative embodiment, the second filter is a low-pass filter. This has the effect that the outputs of the first learning module, such as an acceleration or steering angle signal, can be smoothed using the second learning module.

According to one embodiment, target values associated with the sensor data are also input into the learning algorithm. The first target function quantifies a deviation between the output data and the target values. The first objective function is then minimized by changing the first parameters based on the output data and the target values. The target values can be understood as labels with which the sensor data are annotated. In other words, the learning algorithm can be optimized in such a way that the output data matches the target values as closely as possible.

It is possible that the learning algorithm includes at least one third learning module parameterized by third parameters, in addition to the first and second learning module, which connects the first learning module to the second learning module. The first and/or second parameters can be changeable independently of the third parameters and vice versa. Analogously to the first or second parameters, a third target function that deviates from the first and/or second target function can also be optimized by changing the third parameters based on the output data, for example to optimize the learning algorithm. For example, the second and/or third parameters can be kept constant during the optimization of the first objective function, the first and/or third parameters during the optimization of the second objective function, and/or the first and/or second parameters during the optimization of the third objective function. Thus, the learning algorithm can be specifically adapted to at least three competing target variables.

According to one embodiment, the learning algorithm includes at least one artificial neural network. The first parameters are trainable weights of at least one first layer of the artificial neural network. Additionally or alternatively, the second parameters can be trainable weights of at least a second layer of the artificial neural network. The artificial neural network can be, for example, a multi-layer perceptron with at least one intermediate layer that can be trained, a convolutional neural network, also known as a convolutional neural network (CNN), a recurrent neural network, also known as a recurrent neural network (RNN), an LSTM network (LSTM = long short-term memory) or a combination of at least two of the above examples. An artificial neural network with a large number of trainable intermediate layers, also known as a deep neural network, is particularly advantageous. It is also possible, for example, for the first layer and the second layer to belong to different artificial neural networks that are linked to one another.

character list

• 1 zeigt schematisch eine Vorrichtung zur Datenverarbeitung gemäß einem Ausführungsbeispiel der Erfindung.
• 2 zeigt schematisch ein Fahrzeug mit einem Steuergerät gemäß einem Ausführungsbeispiel der Erfindung.
• 3 zeigt schematisch ein Flussdiagramm zur Veranschaulichung eines Verfahrens zum Trainieren eines Lernalgorithmus gemäß einem Ausführungsbeispiel der Erfindung.
• 4 zeigt ein Diagramm zur Veranschaulichung eines Beschleunigungssignals am Ausgang eines einfachen künstlichen neuronalen Netzes.
• 5 zeigt ein Diagramm zur Veranschaulichung eines Beschleunigungssignals am Ausgang eines Lernalgorithmus, der in einem Verfahren gemäß einem Ausführungsbeispiel der Erfindung trainiert wurde.

Embodiments of the invention are described below with reference to the accompanying drawings, neither the drawings nor the description being to be construed as limiting the invention.

• 1 shows schematically a device for data processing according to an embodiment of the invention.
• 2 shows schematically a vehicle with a control device according to an embodiment of the invention.
• 3 FIG. 12 schematically shows a flowchart to illustrate a method for training a learning algorithm according to an embodiment of the invention.
• 4 shows a diagram illustrating an acceleration signal at the output of a simple artificial neural network.
• 5 shows a diagram to illustrate an acceleration signal at the output of a learning algorithm that was trained in a method according to an embodiment of the invention.

The figures are merely schematic and not true to scale. In the figures, the same reference symbols denote the same features or features that have the same effect.

Embodiments of the invention

1 shows a device 100 for training a learning algorithm 102 for converting sensor data 104 of a sensor system 106 into output data 108 for controlling an actuator 110. The learning algorithm 102 comprises a first learning module 112, which is parameterized by first parameters 114, and a second learning module 116, which is parameterized by second parameter 118. In this example, the output of the first learning module 112 is connected to an input of the second learning module 116, with an output of the second learning module 116 representing an output of the learning algorithm 102, at which the output data 108 is provided will. In this case, the sensor data 104 are input into the first learning module 112 .

The two learning modules 112, 116 are configured in such a way that the first parameters 114 can be optimized independently of the second parameters 118, i. H. are changeable. Conversely, for example, the second parameters 118 can also be changed independently of the first parameters 114 . Alternatively, unlike the first parameters 114, the second parameters 118 may be fixed.

The first parameters 114 and the second parameters 118 can be, for example, trainable weights of different layers of one and the same artificial neural network 119 or trainable weights of different artificial neural networks 119a, 119b. More specifically, the first parameters 114 may be trainable weights of one or more consecutive first layers and the second parameters 118 may be trainable weights of one or more second layers subsequent to the first layers.

The learning algorithm 102 can, in addition to the first learning module 112 and the second learning module 116, an optimization module 120 for optimizing a first objective function 122 by changing the first parameters 114 and, optionally, for optimizing a second objective function 124 by changing the second parameters 118 based on the output data 108 have.

For this purpose, for example, corresponding target values y can be entered into the optimization module 120 . For example, the sensor data 104 can include a plurality of real or simulated measured values x, wherein each of the measured values x can be assigned a specific target value y and the learning algorithm 102 can output a prediction value y for each of the measured values x. The output data 108 can thus include a plurality of prediction values y.

In order to optimize the first target function 122, the first parameters 114 can be changed, for example, in a gradient method by backpropagation based on the prediction values y and the corresponding target values y. For example, the first objective function 122 can be a cost or loss function that quantifies a deviation between the prediction values y and the target values y, for example a quadratic loss function or a 0-1 loss function. In this case, the prediction values y, the target values y and the first parameters 114 can be independent variables of the first target function 122, while a function value, i. H. a dependent variable of the first objective function 122 to be minimized by changing the first parameters 114 accordingly.

The optimization of the second target function 124 by changing the second parameters 118 can take place analogously to the optimization of the first target function 122 described above. However, the second objective function 124 can also be optimized in a different optimization method. It is important that optimization takes place under boundary conditions, i.e. the other target function remains constant or better.

It is pointed out that it is not absolutely necessary to change the second parameters 118 in addition to the first parameters 114 in order to train or optimize the learning algorithm 102 . As already mentioned, it is also possible, for example, for the second parameters 118 to be specified as fixed values, i. H. remain the same during the runtime of the learning algorithm 102 independently of the optimization of the first objective function 122.

The device 100 also includes a processor 126 for executing a computer program and a memory 128 on which the computer program is stored. The execution of the computer program causes the learning algorithm 102 in a method as exemplified with reference to 1 until 5 is described, is trained. The modules of the device 100 mentioned above and below can be software and/or hardware modules.

The completely trained learning algorithm 102 can be used, for example, in a control unit 200 of a vehicle 202 (see 2 ) can be implemented as a software and/or hardware module and used to automatically control actuators 110 of vehicle 202, such as a steering or brake actuator or a drive motor of vehicle 202. For example, vehicle 202 can be equipped with a suitable driver assistance function for this purpose. However, the vehicle 202 can also be an autonomous robot with a suitable robot control program. At its input, control unit 200 is connected to sensor system 106, which can include one or more environment sensors for detecting an environment of vehicle 202 and/or one or more driving dynamics sensors for detecting a driving state of vehicle 202.

For the automated control of the actuator 110 , the sensor data 104 , x provided by the sensor 106 are first received in the control unit 200 and then entered into the learning algorithm 102 . The learning algorithm 102 can be executed by a processor of the control device 200 (not shown) are executed in the form of a corresponding computer program. Based on the output data 108, y output by the learning algorithm 102, the control unit 200 generates a corresponding control signal 204 and outputs it to the actuator system 110. For example, output data 108, y can be input into a speed controller of control unit 200, control signal 204 being generated by the speed controller based on output data 108, y, and outputting control signal 204 to actuator system 110 causes vehicle 202 to move its speed changes.

As already mentioned, it is possible that when training the learning algorithm 102, in addition to the parameterization of the first learning module 112, the parameterization of the second learning module 116 is adapted, for example in order to obtain output data 108, y, which in addition to (main) conditions of a target variable ( Fulfill secondary conditions of a boundary size. In this case, the second learning module 116 can only be responsible for the marginal variable, but not for the target variable. The adjustment can, for example, be carried out in a two-stage process with different cost functions, as described in 3 is shown schematically and is described below.

For this purpose, in a first optimization step, the first parameters 114 can be optimized for the target variable with constant parameterization of the second learning module 116, with an approximation error, which indicates the deviation between the prediction values y and the target values y, being determined. Then, in a second optimization step, the second parameters 118 can be optimized for the edge size with constant parameterization of the first learning module 112 and/or constant approximation error, as was determined in the first optimization step. A preferred frequency range of the output data 108, y, for example, can be defined by the edge size.

The first optimization step and the second optimization step can, for example, be carried out alternately until they converge.

3 shows an example of a coupled pair of the artificial neural network 119a as the first learning module 112 with feature inputs 300 for inputting the sensor data 104, x and a frequency filter as the second learning module 116, whose output is the approximation y of a predetermined target variable 302, y, for example an acceleration signal. An approximation error 304 results from the target variable 302, y and the approximation y. The frequency filter is parameterized in such a way that compliance with a specific value range of a further target variable or the marginal variable is automatically ensured.

For example, the frequency filter can be parameterized once before the training of the artificial neural network 119a and then remain fixed in these values, also for the further course of the training of the learning algorithm 102.

It is also conceivable that the first learning module 112 and the second learning module 116 are interchanged, so that the sensor data 104, x are input into the second learning module 116 and the output data 108, y are output by the first learning module 112. Thus, input data filtered by the second learning module 118 can be provided for the first learning module 112 .

5 FIG. 12 shows, by way of example, a time profile of the target values y in the form of a recorded acceleration signal together with a time profile of the prediction values y in the form of an approximated acceleration signal learned by the learning algorithm 102.

4 shows the corresponding time curves of the target values y and the prediction values y when using a simple artificial neural network instead of the learning algorithm 102.

It can be seen that the acceleration signal approximated by the learning algorithm 102 in 5 with sufficient approximation much smoother than in 4 runs.

$$\text{NMARuck:}=\frac{{ƒ}_R\left(y\right)-{ƒ}_R\left(\widehat{y}\right)}{{ƒ}_R\left(\widehat{y}\right)},$$

$${ƒ}_R\left(\widehat{y}\right):={\displaystyle \sum {}_i\left|{\widehat{y}}^i-{\widehat{y}}^{i-1}\right|},$$

$${ƒ}_R\left(y\right):={\displaystyle \sum {}_i\left|y^i-y^{i-1}\right|},$$

wobei yⁱ die Zielwerte und ŷⁱ die Prädiktionswerte zum Zeitpunkt i sind. Der visuell glattere Eindruck des approximierten Beschleunigungssignals in 5 im Vergleich zu 4 wird durch den verringerten Wert von NMARuck bestätigt.For a better comparison of the approximation results, 4 and 5 the quality criteria NMAE (a measure of the approximation quality) and NMARuck (a measure of the jerky change in the acceleration signal), which are defined as follows: $$\text{NMAE:=}\frac{{\displaystyle \sum {}_i\left|y^i-{\widehat{y}}^i\right|}}{{\displaystyle \sum {}_i\left|{\widehat{y}}^i\right|}},$$

$$\text{NMA jerk:}=\frac{{ƒ}_R\left(y\right)-{ƒ}_R\left(\widehat{y}\right)}{{ƒ}_R\left(\widehat{y}\right)},$$

$${ƒ}_R\left(\widehat{y}\right):={\displaystyle \sum {}_i\left|{\widehat{y}}^i-{\widehat{y}}^{i-1}\right|},$$

$${ƒ}_R\left(y\right):={\displaystyle \sum {}_i\left|y^i-y^{i-1}\right|},$$

where y ⁱ are the target values and ŷ ⁱ are the prediction values at time i. The visually smoother impression of the approximated acceleration signal in 5 compared to 4 is confirmed by the decreased value of NMARuck.

Finally, it is noted that terms such as "comprising,""comprising," etc. do not exclude other elements or steps, and terms such as "a" or "an" do not exclude a plurality. Any reference signs in the claims should not be construed as limiting.

Claims (13)

Method for training a learning algorithm (102) for converting sensor data (104, x) of a sensor (106) into output data (108, y) for controlling an actuator (110), the method comprising: Entering the sensor data (104, x) into the learning algorithm (102), the learning algorithm (102) having a first learning module (112) parameterized by first parameters (114) and a first learning module (112) connected in series and by second parameters (118) comprises a parameterized second learning module (116) for converting the sensor data (104, x) into the output data (108, y); outputting the output data (108,y) by the learning algorithm (102); and Optimizing the learning algorithm (102) by an optimization module (120) based on the output data (108, y), wherein a first objective function (122) is optimized by changing the first parameters (114) and the second parameters (118) while optimizing the first objective function (122) are kept constant. procedure after claim 1 , In order to optimize the learning algorithm (102), a second objective function (124) deviating from the first objective function (122) is also optimized by changing the second parameters (118). procedure after claim 2 , wherein the first parameters (114) are kept constant during the optimization of the second objective function (124); and/or wherein a function value of the first objective function (122) is kept constant during the optimization of the second objective function (124). procedure after claim 2 or 3 , wherein the first objective function (122) and the second objective function (124) are alternately optimized in a plurality of loops. Method according to one of the preceding claims, wherein the sensor data (104, x) is input to the first learning module (112) and the output data (108, y) is output by the second learning module (116); or wherein the sensor data (104, x) is input to the second learning module (116) and the output data (108, y) is output by the first learning module (112). Method according to one of the preceding claims, wherein the first learning module (112) is configured as a first filter for filtering outputs of the second learning module (116) and the first parameters (114) are filter coefficients of the first filter; or wherein the second learning module (116) is configured as a second filter for filtering outputs of the first learning module (112), and the second parameters (118) are filter coefficients of the second filter. procedure after claim 6 , wherein the first filter is a low-pass filter; or wherein the second filter is a low-pass filter. Method according to one of the preceding claims, wherein target values (y) associated with the sensor data (104, x) are also input into the learning algorithm (102); wherein the first objective function (122) quantifies a deviation between the output data (108, y) and the target values (y); wherein the first objective function (122) is minimized by changing the first parameters (114) based on the output data (108, y) and the objective values (y). Method according to one of the preceding claims, wherein the learning algorithm (102) comprises at least one artificial neural network (119, 119a, 119b); wherein the first parameters (114) are trainable weights of at least a first layer of the artificial neural network (119, 119a); and or wherein the second parameters (118) are trainable weights of at least a second layer of the artificial neural network (119, 119b). Method for controlling an actuator (110) of a vehicle (202), the vehicle (202) in addition to the actuator (110) including a sensor (106) for detecting a driving condition and / or an environment of the vehicle (202), the method includes: receiving sensor data (104, x) generated by the sensor system (106); inputting the sensor data (104,x) into a learning algorithm (102) trained in a method according to any one of the preceding claims to convert the sensor data (104,x) into output data (108,y); and Generating a control signal (204) for driving the actuator (110) based on the output data (108, y). Apparatus (100; 200) for data processing, comprising a processor (126) configured to carry out at least one of the methods according to any one of the preceding claims. A computer program comprising instructions which cause a processor (126) to execute the computer program by the processor (126), at least one of the methods according to any one of Claims 1 until 10 to execute. Computer-readable medium on which the computer program claim 12 is saved.

Images