Method and apparatus for training a machine learning system
The novel loss function addresses inconsistent performance in machine learning systems by minimizing maximum errors in sub-ranges, enhancing robustness and accuracy across varying data distributions.
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- ROBERT BOSCH GMBH
- Filing Date
- 2025-11-11
- Publication Date
- 2026-05-21
AI Technical Summary
Conventional machine learning system training methods, such as empirical risk minimization and distributionally robust optimization, fail to ensure consistent model performance across varying data distributions and are susceptible to bias from outliers or uneven data.
A novel loss function that minimizes the maximum error in sub-ranges of the input space by dividing the input space into discrete areas and optimizing the machine learning system's parameters based on the maximum loss value, using techniques like gridding and binning.
This approach enhances the robustness and accuracy of machine learning systems by ensuring consistent performance across the entire input range, even with noisy or unevenly distributed data, leading to improved model reliability and adaptability.
Smart Images
Figure EP2025082657_21052026_PF_FP_ABST
Abstract
Description
[0001] R.416415
[0002] - 1 -
[0003] Description
[0004] title
[0005] Method and apparatus for training a machine learning system
[0006] State of the art
[0007] Zhao et al. “Virtual Sensor for Real-Time Bearing Load Prediction Using Heterogeneous Temporal Graph Neural Networks”, 02.04.2024, https: / / ar-xiv.org / pdf / 2404.02304 reveals a method for training a machine learning system.
[0008] Advantages of the invention
[0009] Machine learning systems can be used in a wide variety of engineering fields, particularly in the development of virtual sensors. Current approaches to training these systems are mostly based on empirical risk minimization. This approach implicitly assumes that training and test datasets are uniformly distributed and that the mean error is representative of the risk to be minimized. To improve the robustness of the models, distributionally robust optimization can be employed. This involves optimizing across a family of distributions to improve the model's performance across different data distributions. Another important aspect is ensuring sufficient model quality across all relevant data slices.
[0010] Conventional loss functions used in such training focus on either mean evaluation or maximum evaluation. Mean evaluation considers only the average and is therefore susceptible to bias from outliers (R.416415).
[0011] - 2 -
[0012] or uneven data distributions. Max Evaluation, on the other hand, focuses on the largest error, but is difficult to optimize.
[0013] The present invention proposes a novel loss function that addresses the shortcomings of previous approaches. The aim is to achieve consistent (i.e., similarly good) model performance across the entire input space. In this context, "input space" or "input range" refers to the range of all possible input values that the machine learning system can process. Each point in this space represents a specific combination of input variables. In the example of the virtual sensor, the input values are the measurement data from one or more real sensors. The input space thus encompasses all possible combinations of measured values that the sensor(s) can provide. The dimension of the input space corresponds to the number of input variables. For example, if three sensors measure three variables (e.g., temperature, pressure, humidity), then the input space is three-dimensional.Each point in three-dimensional space represents a specific combination of temperature, pressure, and humidity.
[0014] The aim of the invention is to train a model that exhibits consistently good behavior across the entire input range. This means that the predictions of the virtual sensor should be as accurate and reliable as possible for all possible combinations of input values. This is achieved by minimizing the maximum error in sub-ranges of the input range. Advantageously, the inventors were able to find that this increases the robustness of the machine learning system to changes in the input distribution, and thus a corresponding virtual sensor exhibits better performance.
[0015] Unlike mean evaluation, which only considers the average, and max evaluation, which is difficult to optimize, the new loss function offers a middle ground. It makes it possible to improve average performance across the entire input range while also limiting the maximum error (in subranges). R.416415
[0016] - 3 -
[0017] The invention is based on a "gridding" approach, in which the input space is divided into discrete areas. For each area, an average error is calculated, and then the maximum of these errors is minimized across all areas. This approach enables efficient implementation and optimization during the training process.
[0018] In addition to the described gridding, extensions are also possible, such as the use of non-equidistant grids or the application of gridding to transformed data. The grid resolution can be optimized as a hyperparameter.
[0019] Disclosure of the invention
[0020] In a first aspect, the invention relates to a computer-implemented method for training a machine learning system, wherein the machine learning system is configured to determine an output based on an input, wherein the input characterizes at least one value of a sensor signal, and the output comprises a classification and / or a regression result with respect to the input, wherein the method comprises the following steps:
[0021] • Providing multiple inputs;
[0022] • Determining a binning of the sensor signal values;
[0023] • Determining a plurality of outputs, whereby for each input of the plurality of inputs an output is determined and each output is assigned to the bin of the binning to which the input was assigned;
[0024] • Determining a first loss value for a plurality of bins of the binning, preferably all bins of the binning, wherein the first loss value characterizes an average or median of a second loss value, wherein a second loss value characterizes a deviation with respect to an output assigned to the bin from correspondingly desired outputs;
[0025] • Determining a maximum loss value, where the maximum loss value corresponds to the numerically largest of the first loss values; R.416415
[0026] - 4 -
[0027] • Adjusting parameters (Φ) of the machine learning system based on the determined maximum loss value.
[0028] In the context of this invention, the term "machine learning system" can be understood as a computer system capable of learning from data. It is based on algorithms that recognize patterns and relationships in data and use them to make predictions or decisions. The machine learning system is optimized by training with sample data, whereby the system's parameters are adjusted to maximize prediction accuracy. Examples of machine learning systems include neural networks, support vector machines, decision trees, Bayesian networks, and k-nearest neighbor classifiers / regressors. The choice of the specific machine learning system depends on the application and the characteristics of the data.In the context of this invention, the machine learning system is preferably a neural network, in particular a deep neural network with multiple layers, since these are capable of learning complex relationships in the sensor data and achieving high predictive accuracy. However, it is not limited to this and can comprise any other suitable machine learning system that fulfills the desired functionality.
[0029] In the context of this invention, the term "sensor signal" can be understood as the digitized representation of the information generated by a sensor, which reflects the physical measurement quantity of the sensor. Since the method is executed by a computer, the sensor signal must necessarily be in digital form, typically as an ordered sequence of binary values. The digitization of the analog sensor signal can be performed, for example, by an analog-to-digital converter (ADC).
[0030] The digitized sensor signal can take on various formats, including:
[0031] Time series: These are sequences of discrete measurements recorded over time. Each measurement is represented as a digital value that reflects the measured quantity at a specific point in time. R.416415
[0032] - 5 -
[0033] Digitized images: These can be viewed as two-dimensional arrays of digital values that spatially represent the measured quantity. An example of this is pixel values in a grayscale or color image.
[0034] Digital individual values: These represent discrete digital measurements that represent the measured quantity at a specific point in time.
[0035] Feature vectors: These can be understood as vectors of digital values representing various features of the measured signal, such as mean, standard deviation, or frequency components. The invention is not limited to specific sensor types and can be used with a variety of sensors whose output signals can be digitized. Examples of sensors that can be used within the scope of this invention are (taking into account the digitization of the signal):
[0036] Temperature sensors: Measure the temperature and provide a digitized temperature value.
[0037] Pressure sensors: Measure the pressure and deliver a digitized value.
[0038] Light sensors: Measure light intensity and provide a digital value. Accelerometers: Measure acceleration and provide digital values.
[0039] Gyroscopes: Measure rotational speed and provide digitized values. Magnetometers: Measure magnetic fields and provide digitized values.
[0040] Proximity sensors: Detect the presence of objects and provide digitized status values (e.g., 0 or 1).
[0041] Image sensors: Capture images, which are represented as digital pixel values. Acoustic sensors (microphones): Capture sound waves, which are represented as digital audio data.
[0042] Position sensors: Measure the position and provide digitized coordinates. R.416415
[0043] - 6 -
[0044] Chemical sensors: Measure the concentration of chemical substances and deliver digitized values.
[0045] Biological sensors: Measure biological parameters and deliver digitized values.
[0046] This list is not exhaustive, and the invention can be used with any sensor that provides a digitizable signal representing the physical quantity to be measured. The invention is particularly advantageous in applications where the digitized sensor data is noisy, incomplete, or unevenly distributed. The method according to the invention makes it possible to train a robust and precise machine learning system even under these challenging conditions. In particular, the input can also characterize multiple sensor signals, either from sensors of the same type and / or from sensors of different types.
[0047] The term "the input characterizes the sensor signal" can be understood to mean that the input includes or consists of the sensor signal. Alternatively, it is also possible that the input characterizes a result of preprocessing the sensor signal by a corresponding preprocessing method.
[0048] The term "binning" can be understood as a method for discretizing the values of the sensor signal or the derived input (x). This can be understood as dividing the range of values of the input variables into a finite number of intervals, so-called "bins." Each bin represents a specific range of values.
[0049] Binning can be performed in various ways, including:
[0050] Equidistant binning: Here, the bins are chosen so that they all have the same width. This is the simplest form of binning and is particularly suitable for evenly distributed data. R.416415
[0051] - 7 -
[0052] Quantile-based binning: Here, the boundaries of the bins are chosen so that each bin contains the same number of data points. This method is robust against outliers and is well-suited for unevenly distributed data.
[0053] K-means-based binning: This uses a clustering algorithm such as K-means to group the data points into K clusters. Each cluster then corresponds to a bin. This method is well-suited for identifying non-linear relationships in the data.
[0054] The number of bins can be understood as a hyperparameter of binning. The optimal number of bins depends on the application and the properties of the data and can be determined through experimentation or cross-validation.
[0055] In the context of the invention, binning can be applied to the raw values of the sensor signal or to features derived therefrom. It can also be multidimensional, meaning that several input variables can be binned simultaneously. Advantageously, this allows complex relationships between the input variables to be captured. For example, in the case of two-dimensional binning, temperature and pressure could each be divided into their own bins, resulting in a matrix of bins.
[0056] In advantageous embodiments of the method, a second loss value can be determined using a loss function, wherein the loss function is an average squared error loss function, a cross-entropy loss function, a Kullback-Leibler divergence, a hinge loss function, a Huber loss function, an L1 loss function, or an L2 loss function.
[0057] The advantages of each loss function are as follows:
[0058] • Mean Squared Error (MSE): Particularly suitable for regression problems with continuous target variables. The quadratic term penalizes large errors more severely, leading to a focus on minimizing significant deviations. R.416415
[0059] - 8 -
[0060] It is also continuously differentiable, which facilitates the application of gradient-based optimization methods.
[0061] • Cross-entropy loss function: Ideal for classification problems, especially with multiple classes. It measures the discrepancy between the predicted probability distribution and the true distribution, thus promoting convergence to a precise classification probability.
[0062] • Kullback-Leibler divergence (KL divergence): Offers advantages in modeling probability distributions. It quantifies the loss of information when one probability distribution is approximated by another, thus enabling the model to be fitted to the underlying data distribution.
[0063] • Hinge Loss Function: Frequently used in Support Vector Machines. It focuses on the data points that are closest to the decision boundary, thus maximizing the separation margin between the classes.
[0064] • Huber loss function: Combines the advantages of the MSE and L1 loss functions. It is more robust against outliers than the MSE because it punishes large errors linearly rather than quadratically. At the same time, it retains the advantages of the MSE's continuous differentiability for small errors.
[0065] • L1 loss function (Mean Absolute Error - MAE): Exhibits high robustness against outliers, as it penalizes large errors less severely than the MSE. This is particularly advantageous for noisy data or data containing outliers.
[0066] In summary, each loss function offers specific advantages for different problems and data properties.
[0067] In various preferred embodiments, it is possible that for binning, values of the sensor signal are normalized to a predefinable numerical range, preferably -1 to 1, or that the value of the sensor signal is normalized to the predefinable numerical range when the input is provided.
[0068] This is advantageous because sensor data can have different units and measurement ranges. Normalization to a uniform numerical range R.416415
[0069] - 9 -
[0070] This ensures that all input data is considered on the same scale. This prevents features with larger numerical values from dominating features with smaller numerical values and improves data comparability. Normalization allows model parameters to be learned faster and more effectively, leading to better performance and generalization capabilities of the machine learning system.
[0071] In various preferred embodiments, it is possible to divide the number range into equidistant bins.
[0072] Equidistant binning is computationally efficient, both in creating the bins and in assigning the data points to the bins. This accelerates the training process of the machine learning system and allows the method to be used even in repeated iterations of iterative optimization methods, such as stochastic gradient descent.
[0073] In various embodiments, it is possible for the input to characterize a plurality of sensor signals.
[0074] Advantageously, the value range of each input variable, derived from a separate sensor signal, can be divided into bins. The combination of bins for all input variables then spans a multidimensional space in which each range is represented by a unique combination of bins. The number of bins per input dimension can also be chosen differently.
[0075] In various embodiments, it is possible that the machine learning system is a virtual sensor.
[0076] The term "virtual sensor" in the context of this invention can be understood as a software-based model that mimics the function of a physical sensor. Instead of directly measuring a physical quantity, a virtual sensor estimates this quantity based on the readings of other, actually existing sensors and / or other available data. It uses R.416415 for this purpose.
[0077] - 10 -
[0078] A trained machine learning system that has learned the relationship between the available data and the quantity to be estimated. A concrete example of a virtual sensor in the context of this invention is the estimation of the engine temperature in a vehicle. Instead of installing an expensive and potentially difficult-to-access temperature sensor directly on the engine, a virtual sensor could estimate the engine temperature based on more readily available measurements. Exemplary inputs for the virtual sensor include the engine speed, the coolant temperature, the ambient temperature, the engine load state (e.g., torque), and / or the time since the engine was last started.
[0079] In various embodiments of the invention, it is possible that the input characterizes a position of a robot arm and the output characterizes a combination of positions of actuators of the robot arm, wherein the positions are assumed by the robot arm when the positions of the actuators are set.
[0080] Instead of calculating the robot's joint angles using complex inverse kinematics, the machine learning system can advantageously learn the relationship between the desired position of the robot arm and the necessary actuator settings. The desired position, defined, for example, by Cartesian coordinates and the orientation of an endpoint of the robot arm, such as a gripper, serves as input for the virtual sensor. The output of the virtual sensor can consist of, or include, the combination of the actuator positions or settings, such as the angles of the individual joints that move the robot arm into the desired position. The virtual sensor thus acts as an inverse kinematic model trained through machine learning.To train this virtual sensor, training data can first be collected by having the robot arm assume various positions and recording the corresponding actuator settings. The input (desired position) and output data (actuator settings) can then be normalized and binned as described above. If the robot arm's position is multidimensional, multidimensional binning is used. This prepared data is then used to train the machine learning system, for example, R.416415.
[0081] - 11 -
[0082] A neural network is trained. By using the maximum loss value in the training process, it is ensured that the virtual sensor reliably estimates the actuator settings even with unfavorable input data, such as at the edge of the robot's workspace. In operation, the virtual sensor receives the desired position of the robot arm as input and provides the corresponding actuator settings as output, which are then used to control the robot arm. The advantages of this approach lie in the simplified control of the robot arm, as no complex inverse kinematics need to be calculated. Additionally, training with the maximum loss value increases the robustness of the virtual sensor against inaccuracies and variations in the input data. The virtual sensor can also adapt to changes in the robot system by being retrained with new data.
[0083] Advantageously, estimating the actuator settings using the virtual sensor is faster than analytically calculating the inverse kinematics.
[0084] A robot arm can also be understood to be, in particular, the arm of an excavator, to which the above example is equivalently applicable.
[0085] In another aspect, the invention relates to a computer-implemented method for determining a control signal of an actuator of a technical system, comprising the following steps:
[0086] • Receiving a sensor signal from a sensor;
[0087] • Providing an input characterizing the sensor signal to a machine learning system, wherein the machine learning system has been trained according to one of the embodiments of the method described above;
[0088] • Determining an output in relation to the input using the machine learning system;
[0089] • Determining the control signal based on the output.
[0090] By using a machine learning system trained with an embodiment of the method described above, it is ensured that R.416415
[0091] - 12 -
[0092] The determined control signal is precise and reliable even with noisy, incomplete or unevenly distributed sensor data. Advantageously, this enables robust and efficient control of technical systems in a variety of applications.
[0093] In another aspect, the invention relates to a training device which is set up to carry out an embodiment of the training method described above.
[0094] In another aspect, the invention relates to a control device which is configured to execute an embodiment of the method for determining the control signal.
[0095] In another aspect, the invention relates to a computer program which is configured to execute an embodiment of one of the above-mentioned methods.
[0096] In another aspect, the invention relates to a machine-readable storage medium on which the computer program is stored.
[0097] Embodiments of the invention are explained in more detail below with reference to the accompanying drawings. The drawings show:
[0098] Figure 1 schematically depicts a training system;
[0099] Figure 2 schematically shows the structure of a control system for controlling an actuator;
[0100] Figure 3 schematically shows an embodiment of the control system for controlling a robot arm;
[0101] Figure 4 schematically shows an embodiment of the control system for controlling a vehicle. R.416415
[0102] - 13 -
[0103] Description of the exemplary implementations
[0104] Figure 1 shows an embodiment of a training system (140) for training a machine learning system (60), which can in particular represent a virtual sensor for estimating a physical quantity based on measured values from at least one sensor signal, using a training data set (T). The training data set (T) comprises a plurality of inputs (x), each of which characterizes measured values from different sensors. The training data set (T) further comprises, for each input (x), a desired output (t), which corresponds to the input (x) and can characterize the physical quantity.
[0105] For training, a training data unit (150) accesses a computer-implemented database (St2), which provides the training data set (T). The training data unit (150) preferably randomly determines a plurality of inputs (x) and the desired outputs (t) corresponding to the inputs (x) from the training data set (T). The inputs (x) and the desired outputs (t) are transmitted to a modification unit (180). The modification unit (180) performs binning of the values of the inputs (x) and determines the bins to which the inputs (x) are assigned. The inputs (x) are also transmitted to the machine learning system (60). The machine learning system (60) determines outputs (y) based on the inputs (x), each of which represents a prediction regarding the desired outputs (t).
[0106] The desired outputs (t) and the determined outputs (y) are received by the change unit (180).
[0107] Based on the desired outputs (t) and the determined outputs (y), the change unit (180) then determines new parameters (Φ') for the machine learning system (60). For this purpose, the change unit (180) determines a first loss value for each bin of the binning, which characterizes an average or median of second loss values. Each second loss value characterizes a deviation from the respective R.416415
[0108] - 14 -
[0109] The outputs determined by the bin (y) are compared to the corresponding desired outputs (t). In the exemplary embodiment, an average squared error loss function is chosen as the loss function for determining the second loss value. In alternative embodiments, a cross-entropy loss function, a Kullback-Leibler divergence, a hinge loss function, a Huber loss function, or an L1 loss function are also conceivable.
[0110] The change unit (180) determines the maximum loss value of the initial loss values and, based on this maximum loss value, determines the new parameters (Φ'). In the exemplary embodiment, this is done using a gradient descent method, preferably Stochastic Gradient Descent, Adam, or AdamW. In other exemplary embodiments, the training can also be based on an evolutionary algorithm or second-degree optimization.
[0111] The determined new parameters (Φ') are stored in a model parameter memory (St1). Preferably, the determined new parameters (Φ') are provided as parameters (Φ) to the machine learning system (60).
[0112] In further preferred embodiments, the described training is repeated iteratively for a predefined number of iteration steps or iteratively until the maximum loss value falls below a predefined threshold. Alternatively or additionally, it is also conceivable that the training is terminated when an average maximum loss value with respect to a test or validation dataset falls below a predefined threshold. In at least one of the iterations, the new parameters (Φ') determined in a previous iteration are used as parameters (Φ') of the machine learning system (60).
[0113] Furthermore, the training system (140) can comprise at least one processor (145) and at least one machine-readable storage medium (146) containing instructions which, when executed by the processor (145), cause the training system (140) to perform a training procedure according to one of the aspects of the invention. The instructions can therefore be understood as a computer program. R.416415
[0114] - 15 -
[0115] Figure 2 shows a control system (40) which uses a machine learning system (60) to determine control signals (A) of an actuator. Preferably at regular intervals, the environment (20) of the actuator (10) is captured by a sensor (30) or a plurality of sensors (30). The sensor signal (S) – or, in the case of multiple sensors, a single sensor signal (S) – from the sensor (30) is transmitted to the control system (40). The control system (40) thus receives a sequence of sensor signals (S). From this, the control system (40) determines control signals (A), which are then transmitted to the actuator (10).
[0116] The control system (40) receives the sequence of sensor signals (S) from the sensor (30) in an optional receiver unit (50), which converts the sequence of sensor signals (S) into a sequence of input signals (x) (alternatively, the sensor signal (S) can be directly received as the input signal (x)). The input signal (x) can, for example, be a segment or a further processing of the sensor signal (S). In other words, the input signal (x) is determined depending on the sensor signal (S). The sequence of input signals (x) is fed to a machine learning system (60).
[0117] The machine learning system (60) is preferably parameterized by parameters (O) which are stored in and provided by a parameter memory (P).
[0118] The machine learning system (60) derives output signals (y) from the input signals (x). The output signals (y) are fed to an optional conversion unit (80), which derives control signals (A) from them. These control signals are then fed to the actuator (10) to control the actuator (10) accordingly.
[0119] The actuator (10) receives the control signals (A), is controlled accordingly, and performs a corresponding action. The actuator (10) may include (not necessarily integrated) control logic that derives a second control signal from the control signal (A), which is then used to control the actuator (10). R.416415
[0120] - 16 -
[0121] In further embodiments, the control system (40) includes the sensor (30). In still further embodiments, the control system (40) alternatively or additionally also includes the actuator (10).
[0122] In further preferred embodiments, the control system (40) comprises at least one processor (45) and at least one machine-readable storage medium (46) on which instructions are stored which, when executed on the at least one processor (45), cause the control system (40) to execute the method according to the invention.
[0123] In alternative embodiments, a display unit (10a) is provided as an alternative or in addition to the actuator (10).
[0124] Figure 3 shows how the control system (40) can be used to control a robot arm (100). A desired position of the robot arm (100), which can in particular characterize the positioning and rotation of a tool (102) at the end of the robot arm, is provided as input (x) and transmitted to the control system (40). The machine learning system (60) is configured to output positions of the actuator (10) of the robot arm that must be assumed to achieve the desired position of the robot arm. The positions of the actuator (10) can in particular be joint angles, linear deflections, or other position-related values of the actuators of the robot arm.
[0125] The machine learning system (60) of the control system (40) was previously trained using the training system (140) described in Figure 1. For training, the positions of the robot arm and the corresponding positions of the actuators (10) were measured to determine the inputs (x) and the desired outputs (t). The machine learning system (60) determines outputs (y) from the input (x), which represent the corresponding positions of the actuators (10) in order to move the robot arm (100) to the desired position.
[0126] The outputs (y) are fed to the forming unit (80). The forming unit (80) determines the final control signals (A) for the actuators (10) from the outputs (y). R.416415
[0127] - 17 -
[0128] The actuators (10) of the robot arm receive the control signals (A) and execute the movement.
[0129] In alternative embodiments, a display unit (10a) is provided as an alternative or additional measure to the actuators (10), which displays the determined positions of the actuators.
[0130] Figure 4 shows an embodiment in which the machine learning system (60) acts as a virtual sensor for estimating the engine temperature in a vehicle (200). The inputs (x) for the virtual sensor (60) characterize the measured values of various sensors (30) in the vehicle, such as engine speed, coolant temperature, ambient temperature, and engine load.
[0131] The virtual sensor (60) receives the inputs (x) and determines an output (y). The output (y) represents the estimated engine temperature.
[0132] The output (y) can be displayed by a display unit (10a). Alternatively, the conversion unit (80) can control a component of the vehicle (200) or engine based on the output (y), for example, the engine cooling. For this purpose, the control signal (A) can be transmitted to an actuator (10), which then controls the engine cooling.
[0133] In further embodiments (not shown), the virtual sensor (60) can be used for various other applications. For example, the control system (40) can be used as an electric tool for detecting objects in walls, e.g., electrical wires, water pipes, wood, or the like. The tool can, in particular, comprise one or more radar sensors, one or more Hall effect sensors, and / or one or more magnetoresistive sensors, the sensor signals of which are provided as input (x) to the virtual sensor (60), the virtual sensor (60) outputting as output (y) a classification indicating whether an object is located in the wall near the tool. R.416415
[0134] - 18 -
[0135] Alternatively, the control system (40) can be used to determine the current fuel consumption of an internal combustion engine. For this purpose, appropriate sensors can measure the opening of the engine's injectors, the pressure within a fuel line of the engine, the degree of injector contamination based, for example, on exhaust gas recirculation, the total running time of the engine, and / or the number of engine rotations, with these measurements being provided as input (x). The output (y) can then characterize the current fuel consumption of the engine. The inventors found that this form of fuel consumption estimation is significantly more accurate, robust, and reliable.
[0136] The term "computer" encompasses any device for executing predefined computational instructions. These instructions can be in the form of software, hardware, or a hybrid of both.
[0137] In general, a plural can be understood as indexed, i.e., each element of the plural is assigned a unique index, preferably by assigning consecutive integers to the elements contained in the plural. Preferably, if a plural comprises N elements, where N is the number of elements in the plural, the elements are assigned the integers from 1 to N.
Claims
R.416415 - 19 - Claims 1. Computer-implemented method for training a machine learning system (60), wherein the machine learning system (60) is configured to determine an output (y) based on an input (x), wherein the input (x) characterizes at least one value of a sensor signal (S), and the output (y) comprises a classification and / or a regression result with respect to the input, wherein the method comprises the following steps: • Providing a plurality of inputs (x; • Determining a binning of the values of the sensor signals (S); • Determining a plurality of outputs (y), where for each input (x) of the plurality of inputs an output (y) is determined and each output is assigned to the bin of the binning to which the input was assigned; • Determining a first loss value for a plurality of bins of the binning, preferably all bins of the binning, wherein the first loss value characterizes an average or median of a second loss value, wherein a second loss value characterizes a deviation with respect to an output assigned to the bin from correspondingly desired outputs; • Determining a maximum loss value, where the maximum loss value corresponds to the numerically largest of the first loss values; • Adjusting parameters (Φ) of the machine learning system (60) based on the determined maximum loss value.
2. The method of claim 1, wherein a second loss value is determined using a loss function, wherein the loss function is an average squared loss function (MSE loss), a cross entropy loss function, a Kullback-Leibler divergence, a hinge loss function, a Huber loss function, a R.416415 - 20 - L1 loss function or an L2 loss function is 3. Method according to claim 1 or 2, wherein, for binning, the values of the sensor signal are normalized to a predefinable numerical range, preferably -1 to 1, in the binning step, or the value of the sensor signal is already normalized to the predefinable numerical range when the input is provided.
4. Method according to claim 3, wherein the number range is divided into equidistant bins.
5. Method according to any of the preceding claims, wherein the input is characterized by a plurality of sensor signals (S).
6. Method according to any of the preceding claims, wherein the machine learning system (60) is a virtual sensor.
7. Method according to one of the preceding claims, wherein the input characterizes a position of a robot arm and the output characterizes a combination of positions of actuators of the robot arm, wherein the positions are assumed by the robot arm when the positions of the actuators are set.
8. Computer-implemented method for determining a control signal (A) of an actuator (10) of a technical system (100, 200) comprising the steps: • Receiving a sensor signal (S) from a sensor (30); • Providing an input (x) characterizing the sensor signal (S) to a machine learning system (60), wherein the machine learning system (60) was trained according to any one of claims 1 to 7; • Determining an output (y) with respect to the input (x) using the machine learning system (60); • Determining the control signal (A) based on the output (y). R.416415 - 21 - 9. Training device (140) which is configured to perform the method according to any one of claims 1 to 7.
10. Control device (40) which is configured to carry out the procedure according to claim 8.
11. Computer program configured to execute the method according to any one of claims 1 to 8 when executed by a processor (45, 145).
12. Machine-readable storage medium (46, 146) on which the computer program according to claim 11 is stored.