Method and system for optimally training machine learning algorithms

Abstract

The invention relates to a method for optimally training a machine learning algorithm, comprising the following steps: providing a domain model having domain parameters and / or domain values for at least one domain; providing a data model having a training data set having training data for at least one domain; removing and / or hiding and / or changing the at least one training data from the training dataset according to the at least one domain parameter and / or domain value to provide a reduced training dataset; training a neural network based on the reduced training data set to determine model performance according to the reduced data set; comparing the determined model performance with the model performance assigned to the training data set; selecting training data from the training data set according to comparison of model performance; training a machine learning algorithm based on the selected training data; and providing a trained machine learning algorithm.

Description

Technical Field

[0001] The invention relates to a method and a system for optimally training a machine learning algorithm. The invention also relates to a computer program with a program code and a computer-readable data carrier with a program code of the computer program. Background Art

[0002] A domain model in the form of an ontology is a structured semantic mapping of a specific knowledge domain. In the context of machine learning models, the development of such a domain model plays a vital role in improving the understanding and interpretation of the model. Domain models also enable effective communication between experts, developers, and application personnel.

[0003] Building a domain model in the form of an ontology of machine learning models is usually an iterative process that requires a systematic collection and classification of relevant knowledge. First, different concepts and entities related in the domain are identified, such as data sources, models, algorithms, metrics, and evaluation methods. On this basis, relationships between concepts are defined to capture dependencies and associations of concepts. In this case, classification, aggregation, association, and hierarchy can be used to design the structure of the ontology. The definition of attributes and properties of concepts enables a detailed description and characterization of individual elements. For example, these attributes can contain information about the parameters, characteristics, or scope of application of a machine learning model.

[0004] In order to validate and improve the ontology, it is preferable to involve experts and practitioners. Their expertise and experience can help ensure the completeness and correctness of the model. Using existing ontologies and standards can also help ensure consistency and interoperability.

[0005] A mature domain model in the form of an ontology enables a structured and unified representation of knowledge within the scope of machine learning models. This mature domain model supports the understanding, interpretation, and reusability of the model and promotes efficient collaboration between experts, developers, and application personnel.

[0006] In the publication "Using ontologies for dataset engineering in automotive AI applications" (http: / / dx.doi.org / 10.23919 / DATE54114.2022.9774675) a method for constructing a domain model in the form of an ontology is described. The domain model describes the effects and / or properties of the scope or domain of use of the machine learning model (the so-called "dimensions" or domain parameters) and the possible expressions in which these effects and / or properties may occur (the so-called "options" or domain values).

[0007] The publication “Datamodels: Predicting Predictions from Training Data” describes the use of data models, i.e. using machine learning models to observe the impact of data points (and in the case of category extensions) on the performance of a trained machine learning model.

[0008] Furthermore, a method for using combinatorial tests developed for testing software is known. In this case, a concept for incrementally building a data set with the aid of a domain model is described. In this case, the dimensions and options of the domain model are covered, but the performance of the trained machine learning model is not observed.

[0009] To date, the performance of trained machine learning models has not been fully observed. The importance of the dimensions and options observed has not been discussed. Summary of the invention

[0010] The object of the present invention is therefore to specify a method for training a machine learning algorithm by which the performance of a machine learning algorithm or a machine learning model is improved in particular with respect to the dimensions and options of a domain model. In addition, in the present case there is also the object of taking into account the importance of the dimensions and options of the observed domain model.

[0011] This object is achieved by a method for optimally training a machine learning algorithm according to the features of claim 1. Furthermore, this object is achieved by a system for optimally training a machine learning algorithm according to the features of claim 11.

[0012] In the present case, a method for optimally training a machine learning algorithm is described. In the present case, the method comprises at least the following steps:

[0013] - providing a domain model having domain parameters and / or domain values ​​for at least one domain;

[0014] - providing a data model having a training data set, wherein the training data set has training data for at least one domain;

[0015] - removing and / or hiding and / or changing at least one training data from the training data set according to at least one domain parameter and / or domain value to provide a reduced training data set, in particular for evaluating the sensitivity of at least one removed and / or hidden domain parameter and / or domain value;

[0016] - training a neural network based on the reduced training data set to determine model performance based on the reduced data set;

[0017] - comparing the determined model performance with the model performance assigned to the training dataset;

[0018] - selecting training data from the training dataset, at least temporarily, based on a comparison of model performance;

[0019] - training a machine learning algorithm based on the selected training data; and

[0020] -Provide trained machine learning algorithms.

[0021] The data model preferably has image data and / or video data, by means of which different combinations of domain parameters and / or domain values ​​for at least one domain are preferably mapped. In principle, not all combinations of domain parameters and / or domain values ​​need to be mapped by the data model. For example, if at least one domain parameter and / or domain value is at least temporarily hidden and / or removed and there is no image data and / or video data for the at least one domain parameter and / or domain value, the absence can be detected by the method according to the invention.

[0022] The model performance assigned to the training data set preferably describes the model performance determined based on the entire training data set. The domain model does not necessarily have to be introduced in this determination.

[0023] In the present case, the data model is linked to the domain model so that it is possible to derive in this way which domain parameters and domain values ​​are important and therefore necessary in the training phase of the machine learning algorithm. In addition, it is thereby possible to derive, in particular also to derive, based on the domain parameters and / or domain values, what degree of interaction between the domain parameters and the domain values ​​must be used. Using the data model obtained in this way, a (training) data set is derived to train the machine learning algorithm or the machine learning model. Thus, the model trained in this way is preferably robust to the diversity of the data model. Here, in particular based on computing power, it is always preferred to keep the training data set as small as possible. Here, for example, there should be little redundancy. In addition, there are also economic reasons for keeping the training data set small, especially in the case of (partial) supervised training methods, because the size of the training data set will also increase the amount of labeling work that is usually performed manually. However, the training data set should preferably not be selected too small in order to contain sufficient variability with respect to domain characteristics.

[0024] In the scope of computer vision, "domain model" preferably refers to the structured and / or semantic mapping of the knowledge scope related to the processing and / or analysis of visual information. The domain model is used to improve the understanding of images, videos or other visual data by using computer vision techniques. The domain model can include various aspects, such as object recognition, face recognition, image segmentation, motion analysis, image classification, etc. It preferably has a collection of concepts, entities and / or relationships representing relevant visual characteristics, visual features and / or visual structures in the corresponding field. Such a model can, for example, define the category of objects, such as cars, people or animals, and / or the characteristics of objects and / or the relationship between each other. The model can also include algorithms and techniques for processing and analyzing visual data, such as convolutional neural networks (CNNs), feature extraction or clustering methods. The development of the domain model is preferably carried out by introducing expert knowledge in the fields of image processing, pattern recognition and / or artificial intelligence. Here, experts preferably identify relevant concepts and define the relationships between these concepts to achieve a comprehensive understanding of the field. The domain model preferably supports the development of powerful computer vision applications, such as object recognition systems, autonomous vehicles, medical imaging and / or video surveillance systems. Domain models can also help integrate computer vision techniques into other application areas such as robotics, security, and / or augmented reality.

[0025] The domain model preferably describes effects and / or properties from the domain or scope of use of the machine learning model. These effects and / or properties are also referred to as "dimensions" and are described, for example, as domain parameters. The possible forms in which the effects and / or properties may occur are preferably referred to as "options" and are described, for example, as domain values.

[0026] The method can ensure during the training phase and / or when determining whether there is enough data in the training data set for each dimension and option of the domain model and for their interaction. The invention no longer requires dense sampling of the domain, which is still required for sensitivity analysis for some known machine learning algorithms, and no longer requires the generation of simulated or synthetic (training) data to iteratively improve the (training) data set.

[0027] This method allows "discrete" conclusions to be drawn about whether additional data are likely to lead to an increase in model performance with respect to a specific effect. In addition, the effects of interactions between different effects can be captured and further evaluated how these interactions affect model performance. This method increases the efficiency or quality of training. Overall, this makes the training phase of the machine learning algorithm cheaper. Because in the current case, dedicated data models are used in a targeted manner, so that a complete combination and / or a complete sampling of dimensions is not required in the training data. Instead, well-founded estimates can already be obtained using a subset of the data. Overall, the training workload is significantly reduced in the current case.

[0028] The method can be used in principle in all applications based on image processing, in particular in applications where there is no or only insufficient labeled image data for the target area. The method can be applied in principle to analyze and / or process sensor data. This is especially true in driver assistance systems and / or fully automated driving and / or surveillance cameras and / or automated facilities and / or other image processing areas, especially in the case where a large amount of data is preferably used for training and / or application assistance functions. In addition, the present invention also extends to the application of multimodal systems, which are based on image data generated by cameras, lidar sensors and / or radar sensors or any combination thereof, for example. Here, the data used by the machine learning algorithm can come from at least one sensor. The sensor can determine a measurement value of the environment in the form of a sensor signal, which can come from, for example, the following sources: digital images (e.g., video, radar, lidar, ultrasound, motion), thermal images and / or audio signals. Based on the sensor signal, information about the elements encoded by the sensor signal can be obtained (i.e., indirect measurements can be performed based on the sensor signal used as a direct measurement).

[0029] The method for training a machine learning algorithm is used in particular within the scope of active learning and / or testing and / or evaluation or data management. In this context, the method can be used in particular for the active selection of (training) data, which are transmitted by preferably any technical system, in particular an autonomous vehicle and / or a robotic system and / or an industrial machine to a backend computer. This can reduce the data traffic. The freed-up bandwidth can be used in particular for the efficient and effective use of information for training a machine learning system.

[0030] Based on the improved domain model, in the current situation, the training data set and / or test data set of the machine learning algorithm can be better and / or more effectively managed. In addition, a split can be set. Through this method, unlabeled or unmarked data can also be selected for labeling or marking, so that these data can be used for supervised learning or training of machine learning algorithms, for example. In addition, the field model can be implemented and / or executed in, in particular, autonomous vehicles and / or robotic systems and / or industrial machines to effectively record (training) data based on at least one trigger (Trigger) provided by the field model. From a technical point of view, the present method for training a machine learning algorithm is particularly aimed at the technical implementation of a mathematical method, so that the mathematical method is executed as efficiently and effectively as possible on a computer and / or control device. Here, in the current situation, the internal technical working mode of a computer and / or a processor plays a role in the implementation design of the method to optimize the internal working mode of a computer and / or a control device. This optimization is achieved in particular by using the present data model, so as to avoid the combination "explosion" associated with the combination test in terms of calculation.

[0031] At least the following technical effects are achieved by this training model, which are particularly important when deriving (training) data sets from domain models. On the one hand, sensitivity analysis is made possible, through which it can be determined which "dimensions" and / or which "options" in the dimensions are essential for a good or effectively usable training data set. In addition, it can be determined how much interaction is required between different dimensions, thereby explaining the necessary combined interactions. In addition, the results of this method can support the selection of hitherto unlabeled or unidentified (training) data, for which labeling is "worthwhile" especially based on sensitivity analysis and / or performance analysis. It may be advantageous that, for example, the performance of a trained machine learning algorithm or machine learning model in terms of the effects included in the (training) data has not yet been saturated. Therefore, in the current situation, a training method is generally described for selecting the most important effects or dimensions and / or interactions or options of dimensions from the description of the domain, so that the performance of the machine learning model can be optimized, especially in terms of selectable metrics. By using a data model, it can be determined which effects are important in the (training) data and / or which of these effects should be considered in the (training) data set, and in particular, to what extent.

[0032] It should be understood that the steps according to the present invention and further optional steps do not necessarily have to be performed in the order shown, but may also be performed in other orders. In addition, further intermediate steps may be specified. Each step may also include one or more sub-steps without departing from the scope of the method according to the present invention.

[0033] In a preferred embodiment, the training data for at least one domain are each identified according to at least one combination of domain parameters and / or domain values. Thus, each training data preferably has a label corresponding to at least one specific combination of a domain value and an associated domain parameter.

[0034] In a preferred embodiment, the training data set of the data model has at least one prediction domain identifier. The prediction domain identifier preferably corresponds to a "standard marker" in the prior art, by means of which the machine learning algorithm to be trained draws conclusions about in which domains the training data are included. After the algorithm has evaluated and / or classified and / or segmented the image data and / or video data, the result is compared with the prediction domain identifier to check whether the algorithm has correctly captured the meaning included in the corresponding image data and / or video data.

[0035] In a preferred embodiment, the removal and / or hiding and / or changing of at least one training data from the training data set according to at least one domain parameter and / or domain value is performed continuously and / or iteratively by varying at least one domain parameter and / or domain value. The method comprises varying at least one domain parameter and / or domain value to perform the removal, hiding or changing of the training data. The exact manner in which the variation and / or stepwise occurs is not specified and can be implemented in various forms.

[0036] In a preferred embodiment, the training of the respective neural network, the determination of the respective model performance and the comparison of the model performance are performed for each training data set that is continuously and / or iteratively reduced in this manner, so that a subset is selected from the training data set in this manner, based on which the machine learning algorithm is trained. The neural network is trained for each domain parameter and / or domain value that is removed and / or hidden and / or changed, so that the model performance can be determined in this manner and compared with the comparative model performance. The respective neural network does not yet correspond to the machine learning algorithm to be trained, which is only trained when the training data set is selected in the present manner and / or the training data set is reduced based on the original larger training data set.

[0037] In a preferred embodiment, the subset from the training data set has training data that have a predetermined influence on the model performance, the predetermined influence being in particular dependent on at least one limit value. In contrast, training data that do not show an influence on the model performance in the current model performance comparison are preferably hidden from the training data set, since these training data only require computing power during training, but have no positive effect on the quality and / or performance of the training.

[0038] In a preferred embodiment, the interaction between at least two domain parameters and / or domain values ​​can be determined based on a comparison of the determined model performance with the model performance assigned to the training data set. Based on the data model, an embedding in the data model can preferably be calculated for a specific instance. In the case of a linear data model, this is an embedding in a transformed space described by the coefficients of the data model. The embedding preferably makes it possible to identify commonalities and / or interactions between the training data. Here, it can be determined which combinations of domain parameters and / or domain values ​​should preferably be observed together to achieve a predetermined model performance.

[0039] In a preferred embodiment, the machine learning algorithm comprises a neural network, in particular a deep neural network. A deep neural network is preferably an artificial neural network architecture consisting of multiple layers of neurons. Each layer processes input data and passes the input data to the next layer, thereby learning complex patterns and relationships. Deep neural networks are preferably used for tasks such as image and speech recognition, machine translation, and other complex data processing tasks. The machine learning algorithm in the preferred embodiment preferably uses such a deep neural network.

[0040] In a preferred embodiment, a production line comprising an equipment assembly for producing a predetermined product is also provided. The production line is preferably a series of production stations and / or work areas arranged to work together to manufacture at least one product. The production line may include various equipment, machines and / or facilities configured to produce the predetermined product. It is emphasized in this embodiment that there is a dedicated equipment assembly on the production line in this preferred embodiment. The equipment assembly may include, for example, machines, robots, automated assembly lines, tools and / or other equipment required to produce the predetermined product.

[0041] In a preferred embodiment, after the production line is provided, the method further comprises the step of producing at least one pre-determinable product using the equipment components. After the production line is provided, a method is carried out according to this embodiment to manufacture at least one pre-determinable product. This step is carried out using the equipment components available on the production line. In the present case, it is not specified in detail what type of product is or what the exact manufacturing process flow is. The point is that the method in the preferred embodiment is to produce at least one pre-determinable product with the aid of the equipment components provided on the production line.

[0042] According to the invention, protection is also claimed for a control device which is included in an autonomous vehicle and / or a robotic system and / or an industrial machine and on which a machine learning algorithm trained according to the method in one of its embodiments can be executed.

[0043] According to the present invention, a system for optimally training a machine learning algorithm is also claimed. The system has a providing device, which is configured to provide a domain model having domain parameters and / or domain values ​​for at least one domain, and a data model having a training data set, wherein the training data set has training data for at least one domain. In addition, the system has an evaluation and calculation device, which is configured to remove and / or hide and / or change at least one training data from the training data set according to at least one domain parameter and / or domain value to provide a reduced training data set; train a neural network based on the reduced training data set to determine model performance based on the reduced data set; compare the determined model performance with the model performance assigned to the training data set; select training data from the training data set based on the comparison of the model performance; and train the machine learning algorithm based on the selected training data; wherein the providing device is also configured to provide a trained machine learning algorithm.

[0044] According to the invention, a computer program with a program code is also claimed for carrying out at least some parts of the method according to the invention in one of the embodiments of the method according to the invention when the computer program is executed on a computer. In other words, according to the invention, a computer program (computer program product) is provided comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method in one of the embodiments of the method according to the invention / the steps of the method according to the invention.

[0045] According to the invention, a computer-readable data carrier with a program code of a computer program is also proposed for carrying out at least some parts of the method according to the invention in one of its embodiments when the computer program is executed on a computer. In other words, the invention relates to a computer-readable (storage) medium comprising instructions which, when executed by a computer, cause the computer to carry out the method in one of its embodiments / the steps of the method according to the invention.

[0046] The described embodiments and developments can be combined with one another as desired.

[0047] Further possible embodiments, developments and implementations of the invention also include combinations of features of the invention which have been described above or below with reference to the exemplary embodiments, which have not been explicitly mentioned. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The accompanying drawings should provide a further understanding of the embodiments of the present invention. The accompanying drawings illustrate embodiments and are used to describe and explain the principles and concepts of the present invention in context.

[0049] Further embodiments and many of the mentioned advantages are derived with reference to the drawings. The elements shown in the drawings are not necessarily shown to scale with respect to one another.

[0050] Figure 1 A schematic flow chart showing an embodiment of the present method for optimally training a machine learning algorithm; and

[0051] Figure 2 A schematic block diagram showing a comparison between a conventional training method (a) and the present method for optimally training a machine learning algorithm (b).

[0052] In the various figures of the drawings, unless otherwise indicated, the same reference numerals indicate the same or functionally identical elements, components or assemblies. DETAILED DESCRIPTION

[0053] Figure 1 A schematic flow chart of a method for optimally training a machine learning algorithm is shown.

[0054] In any embodiment, the method can be at least partially performed by a system 1, which can include a plurality of components not shown in detail, such as one or more providing devices and / or at least one evaluating and calculating device. It is understood that the providing device can be constructed together with the evaluating and calculating device or can be different from the evaluating and calculating device. In addition, the system can include a storage device and / or an output device and / or a display device and / or an input device.

[0055] In the present case, a computer-implemented method for optimally training a machine learning algorithm comprises at least the following steps:

[0056] In step S1 , a domain model having domain parameters and / or domain values ​​for at least one domain is provided.

[0057] In step S2 , a data model having a training data set is provided, wherein the training data set has training data for at least one domain.

[0058] In step S3, at least one training data is removed and / or hidden and / or changed from the training data set according to at least one domain parameter and / or domain value to provide a reduced training data set.

[0059] In step S4, a neural network is trained based on the reduced training data set to determine model performance based on the reduced data set.

[0060] In step S5, the determined model performance is compared with the model performance assigned to the training data set.

[0061] In step S6, training data is selected from the training data set based on the comparison of model performance.

[0062] In step S7, a machine learning algorithm is trained based on the selected training data.

[0063] In step S8, a trained machine learning algorithm is provided.

[0064] Removing and / or hiding and / or changing S3 at least one training data from the training data set according to at least one domain parameter and / or domain value is particularly preferably performed continuously and / or iteratively by changing at least one domain parameter and / or domain value. Particularly preferably, for each training data set continuously and / or iteratively reduced in this way, a corresponding neural network is trained, a corresponding model performance is determined, and a comparison of the model performance is performed, so that a subset is selected from the training data set in this way, on the basis of which a machine learning algorithm is trained.

[0065] Figure 2 The traditional training method ( Figure 2 (a)) and the present method for optimizing training ( Figure 2 (b)) comparison.

[0066] refer to Figure 2 (a), a neural network 200 is trained with the aid of a training data set 202 of training data. The model trained in this way has a certain model performance 205, which is schematically shown in a graph 204. Optionally, the training data of the training data set 202 can be identified or labeled according to a predetermined category 206 (e.g., a predicted class). This labeling is indicated by reference numeral 208.

[0067] Figure 2(b) shows an embodiment of the method for optimally training a machine learning algorithm as claimed in the present invention. A domain model 300 is provided, which can be defined, for example, by a large number of domain parameters P1, P2, P3 and domain values ​​v11, v12, v13; v21, v22; v31, v32, v33 associated with each domain parameter. In addition, a data model 302 with a training data set 304 is provided, which has training data for at least one domain. Optionally, the training data of the training data set 304 can be identified or marked according to a predetermined category or domain 306 (e.g., a prediction class). The marking is indicated by reference numeral 308. In the present case, at least one training data 310 is removed and / or hidden from the training data set 304 at least temporarily, exemplarily according to at least one domain parameter P3 and / or domain value v33, to provide a reduced training data set 312. Removal and / or hiding is shown by reference numeral 314. The influence on the model performance can thus be determined based on the isolated domain parameters and / or domain values, in particular. In the present case, the individual training data of the training data set 304 are labeled according to the domain parameters P1, P2, P3 and the domain values ​​v11, v12, v13; v21, v22; v31, v32, v33 associated with each domain parameter. Thus, for example, at least one identifier for a specific configuration of domain parameters and domain values ​​is assigned to each training data of the training data set 304. This identifier is shown by reference numeral 316. In the present case, for example, at least one training data having a domain value v33 for a domain parameter or labeled as such is removed. Thus, a neural network 318 is trained based on the reduced training data set 312 in each case to determine the model performance based on the reduced data set 312. The model performance 319 determined for the neural network 318 trained with the reduced data set 312 is schematically shown in a diagram 320. The model performance 319 determined in this way or illustrated by the neural network 318 trained based on the reduced data set 312 is compared with the model performance 205 assigned to the (entire) training data set. This comparison is indicated by reference numeral 322 and corresponds to step S5 .

[0068] Based on this comparison, in the present case, training data can be selected from the training dataset, which Figure 2 The data model 302 enriched with the preferred semantic domain model makes it possible to determine subsets of the training data and / or combinations of domain parameters and domain values ​​and thereby preferably determine the global performance behavior, in particular by continuously and / or partially removing and / or adding data points X'\in X with specific domain values ​​from the domain model.

[0069] The machine learning algorithm can then be trained based on the selected training data.

[0070] Therefore, in the present case, the learned data model is particularly used to enrich the preferred semantic domain model. Here, the domain model S preferably has a plurality of input variables D N And it can be expressed as follows.

[0071] S={D1,...,D N}, where each D i ={D i1 ,...,D ik}

[0072] X,Y,M can be defined as (training) data, where x\inX is the input data, y\inY is the optional label for prediction, and m\in M is the description of x in the domain model S.

[0073] The data model can be defined as follows: f(x)→y'

[0074] The cost function is: L(y'), L(y,y') if necessary

[0075] In the present case, it is particularly preferred to calculate and / or specify the data model D M =g(s)=g(D 11 ,...,D Nk )→R.

[0076] For the comprehensive combination of domain models, a prediction is preferably formed for the model performance of the resulting neural network. The calculation of the data model is preferably performed according to methods known in principle, but in contrast to the prior art, the data model is calculated based on a specific instance s\in S. In other words, a mapping to the domain model is performed for calculating the data model.

[0077] As a model output S', where |S'|<|S|, it is particularly preferred that the following subset of training data can be determined and / or output, which can be effectively used to train a machine learning algorithm. This subset of training data preferably maps those scenarios that have a significant impact on model performance to domain parameters and / or domain values ​​that have a significant impact on model performance. In addition, in the current case, an estimate of the degree of interaction k can be output, especially for combinatorial testing. Here, for example, (unimportant) combinations of domain parameters and / or domain values ​​can be checked. By means of the data model reduced in this way and / or the determined degree of interaction, (training) data X' that meets the predetermined quality of the domain model is preferably selected from X. Optionally, if not present, a label for supervised learning can be created for the reduced (training) data set. The resulting (training) data set is then used to train a machine learning algorithm that is robust to the data model.

Claims

1. A method for optimally training a machine learning algorithm, the method comprising the following steps: - providing (S1) a domain model having domain parameters and / or domain values ​​for at least one domain; - providing (S2) a data model having a training data set, wherein the training data set has training data for the at least one domain; - removing and / or hiding and / or changing (S3) at least one training data from said training data set according to at least one domain parameter and / or domain value to provide a reduced training data set; In particular for evaluating the sensitivity of at least one removed and / or hidden domain parameter and / or domain value; - training (S4) a neural network based on the reduced training data set to determine model performance based on the reduced data set; - comparing the determined model performance with the model performance assigned to the training data set (S5); - selecting (S6) training data from the training data set based on the comparison of model performance; - training (S7) the machine learning algorithm based on the selected training data; as well as -Providing (S8) a trained machine learning algorithm. 2 . The method according to claim 1 , wherein the training data for the at least one domain are each identified according to at least one combination of domain parameters and / or domain values.

3. The method according to claim 1 or 2, wherein the training data set of the data model has at least one prediction domain identifier.

4. A method according to any of the preceding claims, wherein removing and / or hiding and / or changing (S3) at least one training data from the training data set according to at least one domain parameter and / or domain value is performed continuously and / or iteratively by changing the at least one domain parameter and / or domain value.

5. A method according to claim 4, wherein training the corresponding neural network, determining the corresponding model performance and comparing the model performance are performed for each training data set that is continuously and / or iteratively reduced in this way, so as to select a subset from the training data set in this way, based on which to train the machine learning algorithm.

6. The method according to claim 5, wherein a subset from the training data set has training data that has no predetermined influence on the performance of the model, the predetermined influence being in particular dependent on at least one limit value.

7. The method according to any of the preceding claims, wherein an interaction between at least two domain parameters and / or domain values ​​can be determined based on a comparison of the determined model performance with a model performance assigned to the training data set.

8. The method according to any of the preceding claims, wherein the machine learning algorithm comprises a neural network, in particular a deep neural network.

9. The method according to claim 1, wherein a production line comprising a plant assembly for producing a predeterminable product is also provided.

10. The method according to claim 9, wherein after providing the production line, the method further comprises the following steps: At least one predefinable product is produced using the plant components.

11. A control device, which is included in an autonomous vehicle and / or a robotic system and / or an industrial machine, and on which a machine learning algorithm trained by the method according to any one of claims 1 to 10 can be executed.

12. A system (1) for optimally training a machine learning algorithm, the system (1) comprising: - a providing device, the providing device being configured to provide a domain model having domain parameters and / or domain values ​​for at least one domain, and to provide a data model having a training data set, the training data set having training data for the at least one domain; as well as - an evaluation and calculation device, which is configured to remove and / or hide and / or change at least one training data from the training data set according to at least one domain parameter and / or domain value, in order to provide a reduced training data set; Training a neural network based on the reduced training data set to determine a model performance based on the reduced data set; comparing the determined model performance with a model performance assigned to the training data set; selecting training data from the training data set based on the comparison of the model performance; and training the machine learning algorithm based on the selected training data; wherein the providing device is also constructed to provide the trained machine learning algorithm.

13. A computer program having a program code for executing at least some parts of the method according to any one of claims 1 to 10 when the computer program is executed on a computer.

14. A computer-readable data carrier having a program code of a computer program for executing at least parts of the method according to any one of claims 1 to 10 when the computer program is executed on a computer.

Images