DE102022107846A1 - Method, system and computer program product for creating parameterized scenarios - Google Patents

Abstract

The invention relates to a method for creating parameterized scenarios, in particular for testing and training driving assistance systems and automated driving functions, wherein a scenario represents a traffic event in a temporal sequence. The method includes the method steps: - Generating a plurality of parameterized basic building blocks of scenes and events for at least one parameterized scenario; - Passing on the plurality of parameterized basic building blocks to a learning reinforcement module; - Creating at least one parameterized scenario from the learning reinforcement module; - Setting value ranges for the Parameter values of the parameters of the basic building blocks of the parameterized scenario; - Creating a set of concrete scenarios from the at least one parameterized scenario; - Simulating a set of simulated concrete scenarios from the set of concrete scenarios; - Generating a set of real scenarios; - Comparing the set of real scenarios with the set of simulated concrete scenarios; - calculating at least one relevant parameterized scenario from the learning reinforcement module; - outputting the at least one relevant parameterized scenario for testing and training purposes.

Description

The invention relates to a method, a system and a computer program product for creating parameterized scenarios, in particular for testing and training driving assistance systems and automated driving functions.

The trend towards driver assistance systems (Advanced Driver Assistance System, ADAS) and highly automated driving functions (HAF) in motor vehicles, but also in aircraft or watercraft, requires extensive safety strategies, since responsibility for vehicle control no longer lies entirely with the driver, but rather active functions Computer units in the vehicle are taken over. Therefore, it must be ensured that autonomously moving objects have a very low error rate in driving behavior. The detection and classification of objects and the interpretation of traffic scenarios in the area around a vehicle are important prerequisites for the safe functionality of driver assistance systems and automated driving functions. This requires targeted testing and training of driver assistance systems and automated driving functions both in extreme and exceptional situations (corner cases) as well as in everyday situations. Such extreme situations arise from a special combination of different factors. Examples of this are infrastructural features such as the type of road, the development on the edge of a road, the quality of the markings but also environmental conditions such as weather conditions, the time of day and the season. The behavior of other road users, the geographical topography and the weather conditions also play a major role.

However, exceptional traffic situations are difficult to find in the real world because some events and situations occur only rarely. In addition, evaluating available traffic data, for example data from test vehicles used to collect data, is not always easy because the relevant data is often not sufficiently classified and is therefore difficult to find.

Simulated traffic scenarios are therefore increasingly being used to test and train driving functions and automated driving functions. However, these simulated scenarios are currently created manually and this requires a lot of time and effort. Since the manual creation of simulated scenarios is time-consuming, not all possible scenarios that occur in reality are covered, so that the required catalog of available scenarios is incomplete.

The WO 2020/083996 A1 discloses a method for testing a driver assistance system for a vehicle in a test environment. Traffic scenario descriptions are determined, which are stored in a traffic scenario file and can be accessed by a user's input. In addition, a test case is created based on the selected traffic scenario description and a specified value for a parameter.

The US 11137763 B2 discloses an automated driving system that can be simulated and validated using automatically generated simulation scenarios. A large number of simulated scenarios are generated by supplementing a variety of recorded real-world scenarios with one or more extracted data streams.

The EP 3 066 529 B1 discloses a method for validating a driver assistance system of a motor vehicle, with tests defined by test parameters being carried out for a given test scenario. During a first test, at least one test parameter is determined and to create a second test, the first test is modified in such a way that the test parameter is changed within a critical range assigned to it, within which a driver assistance system can actively intervene in the driving situation. For the specified test scenario, the second test is carried out to validate the driving assistance system.

The DE 10 2013 003944 A1 discloses a method for providing motor vehicle driving assistance functionalities based on a catalog of road scenarios.

The DE 10 2007 053500 A1 discloses a method for evaluating a predictive safety system for accident avoidance intended for use in motor vehicles, the safety system being evaluated as a function of at least one reconstruction of an accident that has a reference to this safety system.

The DE 10 2006 044086 B4 discloses a method for simulating traffic situations, wherein virtually generated traffic situations are displayed graphically on a display unit.

The object on which the invention is based is to provide options for the improved creation of parameterized scenarios that cover a wide range of real ones Reproduce traffic scenarios accurately and realistically and can be used in particular for testing and training driving assistance systems and automated driving functions.

This object is achieved according to the invention with regard to a method by the features of patent claim 1, with regard to a system by the features of patent claim 10, and with regard to a computer program product by the features of patent claim 15. The further claims relate to preferred embodiments of the invention.

The present invention enables improved creation of parameterized traffic scenarios by structuring scenarios into basic building blocks representing various scenes and events and using a learning reinforcement algorithm. This creates simulated scenarios that cover a wide range of real traffic scenarios and, in particular, also depict critical traffic situations, so that they can be used in particular for testing and training driver assistance systems and automated driving functions, thereby increasing their safety and possible uses.

• - Generieren einer Mehrzahl von parametrisierten Grundbausteinen von Szenen und Ereignissen für zumindest ein parametrisiertes Szenario durch ein Eingabemodul, wobei das Eingabemodul eine Softwareapplikation verwendet;
• - Weitergeben der Mehrzahl von parametrisierten Grundbausteinen an ein Lernverstärkungsmodul mit einem Lernverstärkungs-Agenten;
• - Erstellen zumindest eines parametrisierten Szenarios durch Auswählen zumindest eines Grundbausteins von dem Lernverstärkungsmodul;
• - Festlegen von Wertebereichen für die Parameterwerte der Parameter der Grundbausteine des parametrisierten Szenarios;
• - Erstellen einer Menge von konkreten Szenarien aus dem zumindest einen parametrisierten Szenario durch Bestimmen von konkreten Parameterwerten der Parameter der Grundbausteine innerhalb der festgelegten Wertebereiche;
• - Simulieren in einer Menge von simulierten konkreten Szenarien aus der Menge von konkreten Szenarien;
• - Generieren einer Menge von realen Szenarien von dem Eingabemodul mittels der Softwareapplikation aus Sensorsignalen von Sensoren, wobei die Sensorsignale der Sensoren in Echtzeit an das Eingabemodul übermittelt werden und/oder aus einer Datenbank abrufbar sind;
• - Vergleichen der Menge von realen Szenarien mit der Menge von simulierten konkreten Szenarien;
• - Berechnen von zumindest einem relevanten parametrisierten Szenario von dem Lernverstärkungsmodul;
• - Ausgeben des zumindest einen relevanten parametrisierten Szenarios für Test- und Trainingszwecke.

According to a first aspect, the invention provides a method for creating parameterized scenarios, in particular for testing and training driving assistance systems and automated driving functions. A parameterized scenario represents a traffic event in a temporal sequence and can be described as a sequence of scenes starting with a starting scene and events, with a scene and an event each being represented by at least one basic module, which is determined by a selection of parameters and associated parameter values is defined. A parameterized scenario is a scenario or a parameterized basic building block is a basic building block in which the parameters and associated parameter values are freely selectable, and a concrete scenario is a scenario or a concrete basic building block is a basic building block in which the parameters and associated Parameter values are set. The procedure includes the following procedural steps:

• - Generating a plurality of parameterized basic building blocks of scenes and events for at least one parameterized scenario through an input module, the input module using a software application;
• - passing on the plurality of parameterized basic building blocks to a learning reinforcement module with a learning reinforcement agent;
• - Creating at least one parameterized scenario by selecting at least one basic building block from the learning reinforcement module;
• - defining value ranges for the parameter values of the parameters of the basic blocks of the parameterized scenario;
• - Creating a set of concrete scenarios from the at least one parameterized scenario by determining concrete parameter values of the parameters of the basic building blocks within the specified value ranges;
• - Simulate in a set of simulated concrete scenarios from the set of concrete scenarios;
• - Generating a set of real scenarios from the input module using the software application from sensor signals from sensors, the sensor signals from the sensors being transmitted to the input module in real time and/or being retrievable from a database;
• - Comparing the set of real scenarios with the set of simulated concrete scenarios;
• - Calculating at least one relevant parameterized scenario from the learning reinforcement module;
• - Outputting the at least one relevant parameterized scenario for testing and training purposes.

• - Zuordnen einer ersten Belohnung für das Vergleichsergebnis zwischen der Menge von realen Szenarien und der Menge von simulierten konkreten Szenarien;
• - Auswählen einer Berechnungsfunktion und/oder einer Aktion von dem Lernverstärkungs-Agenten basierend auf der Belohnung;
• - Berechnen eines modifizierten parametrisierten Szenarios mittels der ausgewählten Berechnungsfunktion und/oder Aktion durch Modifizierung zumindest eines parametrisierten Grundbausteins;
• - Generieren einer Menge von modifizierten simulierten konkreten Szenarien aus dem zumindest einen modifizieren parametrisierten Szenario;
• - Vergleichen der Menge von realen Szenarien mit der Menge von modifizierten simulierten konkreten Szenarien;
• - Zuordnen einer zweiten Belohnung für das Vergleichsergebnis;
• - Anpassen einer Richtlinie des Lernverstärkungs-Agenten basierend auf der zweiten Belohnung, wobei bei einer Konvergenz der Richtlinie keine weitere Berechnungsfunktion und/oder Aktion durchgeführt wird, und bei einer Nicht-Konvergenz der Richtlinie eine weitere Berechnungsfunktion und/oder eine weitere Aktion für ein weiteres modifiziertes parametrisiertes Szenario mit einer Modifizierung zumindest eines weiteren parametrisierten Grundbausteins von dem Lernverstärkungs-Agenten ausgewählt wird, solange bis ein relevantes parametrisierten Szenario berechnet ist.

In a further development it is provided that the calculation of the at least one relevant parameterized scenario includes the following procedural steps:

• - assigning a first reward for the comparison result between the set of real scenarios and the set of simulated concrete scenarios;
• - selecting a calculation function and/or an action from the learning reinforcement agent based on the reward;
• - Calculating a modified parameterized scenario using the selected calculation function and/or action by modifying at least one parameterized basic building block;
• - Generating a set of modified simulated concrete scenarios from the at least one modified parameterized scenario;
• - Comparing the set of real scenarios with the set of modified simulated concrete scenarios;
• - Assign a second reward for the comparison result;
• - Adjusting a policy of the learning reinforcement agent based on the second reward, whereby if the policy converges, no further calculation function and/or action is performed, and if the policy does not converge, another calculation function and/or action is performed for another modified parameterized scenario with a modification of at least one further parameterized basic building block is selected by the learning reinforcement agent until a relevant parameterized scenario is calculated.

In an advantageous embodiment it is provided that a parameter is a moving object such as a motor vehicle, an immovable object such as a building, a road configuration such as a highway, a speed, an acceleration, a street sign, a traffic light, a tunnel, a roundabout, a turning lane , an acceleration, a direction, an angle, a radius, a location, a traffic volume, a topographical structure such as a slope, a time, a temperature, a precipitation value, a weather and/or a season.

In particular, it is intended that the sensors be used as radar systems, LIDAR systems for optical distance and speed measurement, image-recording 2D/3D cameras in the visible, IR and/or UV range, GPS systems, acceleration sensors, speed sensors, capacitive sensors, inductive Sensors, precipitation sensors and / or temperature sensors are formed.

Advantageously, the software application uses artificial intelligence algorithms to generate the parameterized basic modules.

In a further embodiment, it is provided that the artificial intelligence algorithms use at least one neural network to extract and classify parameters and parameter values to create the parameterized basic building blocks from input data and sensor signals from the sensors.

In particular, the neural network is designed as a deep neural network, as a convolutional neural network (CNN) and/or as a feedback network (Recurrent Neural Network (RNN)).

In a further embodiment it is provided that the software application uses at least one encoder and/or at least one decoder.

In a further development, it is provided that at least one algorithm of the learning reinforcement agent (320) is designed as a Markov decision process, temporal difference learning (TD-learning), Q-learning, SARSA, Monte Carlo simulation or actor-critic.

According to a second aspect, the invention provides a system for creating parameterized scenarios, in particular for testing and training driving assistance systems and automated driving functions. A parameterized scenario represents a traffic event in a temporal sequence and can be described as a sequence of scenes starting with a starting scene and events, whereby a scene and an event can each be represented by at least one basic module, which is determined by a selection of parameters and associated parameter values are determinable. A parameterized scenario is a scenario or a parameterized basic building block is a basic building block in which the parameters and associated parameter values are freely selectable, and a concrete scenario is a scenario or a concrete basic building block is a basic building block in which the parameters and associated Parameter values are set. The system includes an input module, a learning reinforcement module with a learning reinforcement agent, a variation module, a simulation module, a comparison module and an output module. The input module is designed to generate a plurality of parameterized basic building blocks of scenes and events for at least one parameterized scenario using a software application and to pass on the plurality of parameterized basic building blocks to the learning reinforcement module. The learning reinforcement module is designed to create at least one parameterized scenario by selecting at least one basic building block and to specify value ranges for the parameter values of the parameters of the basic building blocks of the parameterized scenario. The variation module is designed to create a set of concrete scenarios from the at least one parameterized scenario by determining concrete parameter values of the parameters of the basic building blocks within the specified value ranges. The simulation module is designed to simulate a set of simulated concrete scenarios from the set of concrete scenarios. The input module is designed using a set of real-world scenarios to generate the software application from sensor signals from sensors that are transmitted to the input module in real time and/or can be retrieved from a database. The comparison module is designed to compare the set of real scenarios with the set of simulated concrete scenarios. The learning reinforcement module is designed to calculate at least one relevant parameterized scenario. The output module is designed to output the at least one relevant parameterized scenario for testing and training purposes.

In a further development, a parameter is provided for a moving object such as a motor vehicle, an immovable object such as a building, a road configuration such as a highway, a speed, an acceleration, a street sign, a traffic light, a tunnel, a roundabout, a turning lane, represents an acceleration, a direction, an angle, a radius, a location, a traffic volume, a topographical structure such as a slope, a time, a temperature, a precipitation value, a weather and/or a season.

In a further embodiment it is provided that the sensors are radar systems, LIDAR systems for optical distance and speed measurement, image-recording 2D/3D cameras in the visible range, in the IR and/or UV range, GPS systems, acceleration sensors, speed sensors , capacitive sensors, inductive sensors, precipitation sensors and / or temperature sensors are formed.

In an advantageous embodiment, it is provided that the software application uses artificial intelligence algorithms to extract and classify parameters and parameter values from input data and sensor signals from the sensors for generating the parameterized basic blocks.

Advantageously, at least one algorithm of the learning reinforcement agent is designed as a Markov decision process, temporal difference learning (TD-learning), Q-learning, SARSA, Monte Carlo simulation or actor-critic.

According to a third aspect, the invention provides a computer program product comprising executable program code configured to, when executed, carry out the method according to the first aspect.

The invention is explained in more detail below using exemplary embodiments shown in the drawing.

• 1 ein Blockdiagramm zur Erläuterung eines Ausführungsbeispiels eines erfindungsgemäßen Systems;
• 2 eine schematische Darstellung eines Graphen;
• 3 eine schematische Darstellung einer zeitlichen Sequenz von Grundbausteinen von Szenen und Ereignissen eines Szenarios;
• 4 ein Flussdiagramm zur Erläuterung der einzelnen Verfahrensschritte eines erfindungsgemäßen Verfahrens;
• 5 ein Blockdiagramm eines Computerprogrammprodukt gemäß einer Ausführungsform des dritten Aspekts der Erfindung.

This shows:

• 1 a block diagram to explain an exemplary embodiment of a system according to the invention;
• 2 a schematic representation of a graph;
• 3 a schematic representation of a temporal sequence of basic building blocks of scenes and events of a scenario;
• 4 a flowchart to explain the individual method steps of a method according to the invention;
• 5 a block diagram of a computer program product according to an embodiment of the third aspect of the invention.

Additional features, aspects and advantages of the invention or its embodiments are explained in the following description in conjunction with the claims.

In the context of the invention, a scenario is a traffic event in a temporal sequence. An example of a scenario is driving on a highway bridge, turning in a turning lane, driving through a tunnel, turning into a roundabout or stopping in front of a pedestrian crossing. In addition, specific visibility conditions, for example due to twilight or high levels of sunlight, as well as environmental conditions such as the weather and season, traffic volume and certain geographical topographical conditions can influence a scenario. For example, an overtaking operation can be described as a scenario in which a first vehicle is initially behind another vehicle, then changes lanes to the other lane and increases speed in order to overtake the other vehicle. The season and weather conditions also play a role, as road conditions are different in heavy rain and slippery conditions than on a sunny summer day.

Simulated traffic scenarios that are created through programming are therefore increasingly being used to train driving functions and automated driving functions. In the context of the invention, a parameterized scenario refers to a scenario that is written in particular in a machine-readable code. A parameterized scenario SZp can be defined by various parameters P ₁ , P ₂ , ..., P _n from a set of possible parameters P _i and associated parameter values PV ₁ , PV ₂ , ..., PV _n from a set of possible parameter values PV _i can be defined. The parameter values PV _i determine the value range rich of a parameter P _i . The parameters P _i include, for example, a moving object such as a motor vehicle, an immovable object such as a building, a road configuration such as a highway, a speed, a street sign, a traffic light, a tunnel, a roundabout, a turning lane, an acceleration, a direction, an angle, a radius, a location, a traffic volume, a topographical structure such as a slope, a time, a temperature, a precipitation value, a weather condition, a season. Within the scope of the invention, parameters P _i are therefore referred to as properties and features of a scenario.

In the context of the present invention, a distinction is made between a concrete scenario SZc and a parameterized scenario SZp. A scenario SZ in which the parameters P _i and the associated parameter values PV _i or value ranges of the parameter values PV _i are defined is referred to as a concrete scenario SZc. In the context of this invention, a parameterized scenario SZp is defined as a scenario in which both the parameters P _i and the associated parameter values PV _i are not all fixed. Both scenarios, the concrete scenario SZc and the parameterized scenario SZp, are scenarios that are written in particular in machine-readable code. A simulated concrete scenario SZsc is a collection of simulated measured values that reflect simulated concrete trajectories and signals of a moving object, in particular a motor vehicle. Within the scope of the invention, these simulated concrete trajectories are compared with real trajectories in order to use this comparison to determine whether the simulated concrete trajectories are able to represent real trajectories. According to the invention, a learning reinforcement agent is used for the comparison between the simulated concrete trajectories with the real trajectories.

By combining different parameters P _i and the associated parameter values PV _i , a set of simulated concrete scenarios SZsc _i can be created, which can be used, for example, for testing and training at least one driving assistance system and/or at least one automated driving function. The testing and training of driving assistance systems and automated driving functions can in turn be done using simulation techniques. However, such a catalog is only suitable to a limited extent in practice, since although a large amount of data is created, the relevance of the individual simulated concrete scenarios SZsc, in particular for testing and training driving assistance systems, is not immediately obvious. It is therefore necessary to use a suitable selection to determine those relevant parameterized scenarios SZp _i that cover the highest possible number of scenarios that actually occur in road traffic. This number of traffic situations is also referred to as the “Operational Design Domain”. Driving assistance systems and automated driving functions must be able to handle this specific number of traffic situations.

For this purpose, the invention provides a system 100 for creating parameterized scenarios SZp, which can be used in particular for testing and training driving assistance systems and automated driving functions. The system 100 according to the invention is based on methods of reinforcement learning and is in the 1 shown schematically. The system 100 includes an input module 200, a learning reinforcement module 300 with a learning reinforcement agent (LV agent) 320, a variation module 400, a simulation module 420, a comparison module 440 and an output module 450.

The input module 200, the learning reinforcement module 300, the variation module 400, the simulation module 420, the comparison module 440 and the output module 450 can each be provided with a processor and/or a memory unit.

In connection with the invention, a “processor” can be understood to mean, for example, a machine or an electronic circuit. A processor can in particular be a main processor (Central Processing Unit, CPU), a microprocessor or a microcontroller, for example an application-specific integrated circuit or a digital signal processor, possibly in combination with a memory unit for storing program instructions, etc . A processor can also be understood as a virtualized processor, a virtual machine or a soft CPU. For example, it can also be a programmable processor that is equipped with configuration steps for carrying out the said method according to the invention or is configured with configuration steps in such a way that the programmable processor has the features of the method, the entity, the modules, or other aspects and/or other aspects according to the invention. or partial aspects of the invention are realized. In particular, the processor can contain highly parallel computing units and powerful graphics modules.

In connection with the invention, a “memory unit” or “memory module” and the like can, for example, mean a volatile memory in the form of a random access memory (RAM) or permanent storage such as a hard drive or data carrier or e.g. B. a replaceable memory module can be understood. The storage module can also be a cloud-based storage solution.

In connection with the invention, a “module” can be understood to mean, for example, a processor and/or a memory unit for storing program instructions. For example, the processor is specifically set up to execute the program instructions in such a way that the processor and/or the control unit executes functions in order to implement or realize the method according to the invention or a step of the method according to the invention.

The input module 200 can be integrated into a computer, a tablet, a smartphone, etc. and has a user interface 220 for inputting data such as images, graphics, diagrams, time series, etc.

In addition, the input module 200 is connected to sensors 500, which record data from road users, such as motor vehicles in particular, as well as other objects and a traffic environment and transmit it to the input module 200. The sensors 500 can in particular be designed as radar systems, LIDAR systems for optical distance and speed measurement, image-recording 2D/3D cameras in the visible, IR and/or UV range, and/or GPS systems. Furthermore, acceleration sensors, speed sensors, capacitive sensors, inductive sensors, precipitation sensors and/or temperature sensors, etc. can be provided.

Furthermore, the input module 200 can be connected to at least one database 520. In particular, historical data in the form of images, graphics, time series, parameters, etc. are stored in the database 520. For example, target variables and target values can be stored in the database 520, which define a security standard. “Database” means both a storage algorithm and the hardware in the form of a storage unit. In particular, the database 520 can be designed as a cloud computing infrastructure 700.

In connection with the invention, “data” means both raw data and data that has already been prepared.

Within the scope of the invention, a parameterized scenario SZp comprises a temporal sequence of time intervals Δt ₁ , Δt ₂ .,..., Δt _n , in which different scenes and events take place. A parameterized scenario SZp begins with a starting scene and then develops further through occurring events, from which new subsequent scenes emerge over time. The starting scene is thus changed by one or more events. An event can be a reaction actively triggered by a road user, such as an acceleration, or an event that recurs cyclically, such as the switching of a traffic light. The start scene and the individual subsequent scenes therefore only cover a small time interval Δt or a snapshot, while a parameterized scenario SZp covers a longer period of time. In a graph representation, the events can be represented as edges and the individual scenes as nodes of a graph, as in the 2 is illustrated

Basic building blocks GR are used to describe the elements and properties of the start scene, the events and the subsequent scenes. A basic building block GR ₁ can, for example, contain the description of a street route using online maps. Another basic building block GR ₂ can represent the traffic infrastructure such as traffic lights and traffic signs and a third basic building block GR ₃ can be intended for temporary changes such as construction sites and railway crossings. A fourth basic building block, GR _4, can refer to road users such as motor vehicles, trucks, cyclists and pedestrians. A fifth basic building block GR ₅ can map the environmental conditions such as the weather and the season and a sixth basic building block GR ₆ describes a telecommunications infrastructure. These are examples of possible basic building blocks GR _i . In addition, further basic building blocks GR _k can be provided for scenes that are not mentioned here.

An event can, for example, be an overtaking maneuver by a road user, an acceleration or braking process, a traffic light changing, turning onto another street, etc. However, an event can also include errors in the basic building blocks GR _i of the starting scene or subsequent scenes, such as errors in the online map material. These events are also represented according to the invention by basic building blocks GR _k .

According to the invention it is provided that the various basic blocks GR _i are parameterized by parameters P _i and the associated parameter values PV _i . For example, a lane change maneuver by another road user can be determined by the longitudinal and lateral speed during the lane change as well the direction of the lane change can be parameterized. A basic building block GRp parameterized in this way can be varied by changing the parameters P _i and the associated parameter values PV _i and thus covers a large number of specific basic building blocks GRc. As with the scenarios, a parameterized basic building block GRp is defined as a basic building block in which both the parameters P _i and the associated parameter values PV _i are not all fixed. A concrete basic building block GRc is a basic building block in which the parameters P _i and the associated parameter values PV _i or value ranges of the parameter values PV _i are defined. For example, in a specific basic building block GRc the speed of a vehicle is fixed, while in a parameterized basic building block GRp this can vary within a value range, for example between 50 km/h and 150 km/h.

Since a parameterized scenario SZp changes dynamically and unstructured in a time sequence, it can therefore be described as a sequence of basic building blocks GR _i starting with the start scene and the subsequent events. The composition of the basic building blocks GR _i changes in every time period Δt _i . As exemplified in the 3 shown, a parameterized scenario SZp begins at a time t ₀ with a start scene that can be described with the two basic blocks GR ₁ and GR ₂ . This is followed at a time t ₁ by a first event which consists of the basic building blocks GR ₃ and GR ₄ . At a further time t _2, a second event takes place, which consists of the basic building blocks GR ₄ , GR ₅ and GR ₆ . The basic building block GR ₄ of the second event and the first event are identical, which makes it clear that a basic building block GR _i can extend over more than one time period dt _i , as is shown as an example for the basic building block GR ₄ of the first event. At the next time t _3, a third event occurs, which consists of a basic building block GR ₇ . A parameterized scenario SZp is thus composed in a mosaic manner in a temporal sequence of basic building blocks GR _p , whose parameters P _i and associated parameter values PV _i can be varied within predetermined boundary conditions. The choice of the value of the time period Δt intended for changes in events has an influence on the resolution and accuracy of the presented scenario. The shorter the time period Δt is chosen, the more accurately a parameterized scenario SZp can represent a scenario SZ that actually takes place in reality. Preferably, the length of the time intervals Δt ₁ , Δt ₂ ., ..., Δt _n is constant. However, it can also be provided that the length varies dynamically depending on the respective scenes and events. Typically, the value of a time interval Δt _i is in the range of one hundred milliseconds.

By selecting different parameterized basic building blocks GRp _i of scenes and events and combining them in a temporal sequence, a set of parameterized scenarios SZp _i can be created.

These parameterized basic blocks GRp ₁ , GRp ₂ , ..., GRp _n ∈ GRp _i are generated from input data by the input module 200 using a software application 240. The software application 240 uses in particular artificial intelligence algorithms to generate the parameterized basic building blocks GRp _i , which can be, for example, encoders and decoders with neural networks. By creating a set of parameterized basic building blocks GRp ₁ , GRp ₂ , ..., GRp _n ∈ GRp _i , a construction kit is made available with which parameterized scenarios SZp _i can now be developed. In order to filter out those parameterized scenarios SZp _k from a large number of possible parameterized scenarios that represent real traffic scenarios as precisely and concretely as possible, the construction kit with the set of parameterized basic building blocks GRp ₁ , GRp ₂ , ..., GRp _n ∈ GRp _i is added to the learning reinforcement module 300 handed over.

According to the invention, the learning reinforcement module 300 now calculates from this set of parameterized building blocks GRp _i for the starting scene and the following events of a scenario those parameterized scenarios SZp _k that are relevant for simulation applications, in particular for testing and training driving assistance systems and automated driving functions, and are used sensibly can be. In the following, SZp _i with the index “i” denotes all possible parameterized scenarios that can be assembled from the parameterized building blocks GRp _i , while SZp _k with the index “k” denotes the parameterized scenarios calculated by the learning reinforcement module 300, which are used for testing - and training tasks are relevant.

The parameterized blocks GRp _i generated by the input module 200, in particular in the form of machine-readable data formats, are forwarded to the learning reinforcement module 300 by means of communication connections such as a CAN bus system (Controller Area Network). However, wireless connections can also be provided. A wireless communication connection is designed in particular as a mobile radio connection and/or a near-field communication connection such as ^Bluetooth® , Ethernet, NFC (near field communication) or Wi-Fi®.

The learning reinforcement module 300 calculates relevant parameterized scenarios SZp _k from these parameterized building blocks GRp _i , which are suitable for further applications such as testing and training of driving assistance systems. In addition, other applications such as testing a test site are possible. For this purpose, the learning reinforcement module 300 selects suitable parameterized basic building blocks GRp _i from the modular system and varies the parameters P _i and the associated parameter values PV _i of the respectively selected parameterized basic building blocks GRp _i . The calculated parameterized scenarios SZp _k are output by an output module 400.

The learning reinforcement module 300 creates at least one parameterized scenario SZp _i by selecting at least one basic building block GRp _i and sets value ranges (minimum - maximum) for the parameter values PV ₁ , PV ₂ , ..., PV _n of the parameters P ₁ , P ₂ , .. ., P _n of the basic blocks GRp _i of the parameterized scenario SZp _i . From this parameterized scenario SZp _i, the variation module 400 generates a set of concrete scenarios SZc _i by determining concrete parameter values PV ₁ , PV ₂ , ..., PV _n of the parameters P ₁ , P ₂ , ..., P _n of the basic blocks GRp _i within the value ranges specified by the learning reinforcement module 300. From this set of concrete scenarios SZc _i, the simulation module 420 simulates a set of simulated concrete scenarios SZsc _i .

The set of simulated concrete scenarios SZsc _i is compared with a set of real scenarios SZr _i by the comparison module 440. In particular, the vehicle trajectories of the real scenarios SZr _i are compared with simulated vehicle trajectories of the simulated concrete scenarios SZsc _i in order to determine whether the simulated vehicle trajectories correctly reflect the real driving events. The set of real scenarios SZr _i was generated by the input module 200, in particular by means of the software application 240, from sensor signals from sensors 500, the sensor signals from the sensors 500 being transmitted to the input module 200 in real time and/or being retrievable from a database 520. Based on the comparison, at least one relevant parameterized scenario SZp _k is calculated by the learning reinforcement module 300 using the learning reinforcement agent 320, which is output by the output module 450 for testing and training purposes.

Reinforcement learning relies on the learning reinforcement agent (LV agent) 320 selecting at least one action a _i ∈ A from a set of available actions for a particular state s _i ∈ S. The choice of the selected action a _i is based on a strategy or policy. For the selected action a _i , the LV agent 320 receives a reward r _i ∈ R.

According to the invention, a parameterized scenario SZp _i consisting of a large number of parameterized basic building blocks GRp _i represents a state s _i for the LV agent 320. The strategy is adapted by the LV agent 320 based on the rewards _r i received. The strategy specifies which action a _i ∈ A should be selected from the set of available actions for a certain state s _i ∈ S from the set of available states. This creates a new state s _i+1 , for which the LV agent 320 receives a new reward r _i+1 . A strategy thus specifies the association between a state s _i and an action a _i , so that the strategy specifies the choice of the action a _i to be carried out for a state s _i . The goal of the LV agent 320 is to maximize the achieved rewards r _i , r _i+1 , ..., r _t+n .

An action a _i, for example, results in a selection of a parameterized basic building block GRp _i and/or a modification of a parameter P _i and/or an adjustment of a value range of a parameter value PV _i ∈ PV from the set of parameter values for at least one parameter P _i of a parameterized basic building block GRp _i of the parameterized scenario SZp _i to which the action a _i refers. Preferably, the action a _i is one of the actions A(+), A(0) and A(-). A positive action A(+) can be an action in which a further parameterized basic building block GR _i and/or a parameter P _i is added and/or the value range of a parameter PV _i is increased, in the case of a neutral action A( 0) it can be an action in which the parameterized basic building blocks GR _i and/or the parameters P _i and/or the parameter value ranges PV _i remain unchanged, while in the case of a negative action A(-) a parameterized basic building block GR _i and/or or a parameter P _i can be removed and/or a parameter value range PV _i can be reduced.

In the state module 350, a deviation Δ between a target state s _t and the calculated state s _i is calculated by comparison. A reward r _i is assigned to the degree of deviation Δ between the calculated value for state s _i and the target value for state s _t . The final state is reached when the calculated states s _i are equal to or greater than the target states s _t .

Preferably, a Markov decision process is used as the algorithm for the LV agent 320. However, it can also be planned to use temporal difference learning (TD learning). Algorithm to use. An LV agent 320 with a TD learning algorithm adjusts the actions A(+), A(0), A(-) not only when it receives the reward, but after each action a _i based an estimated expected reward. Furthermore, algorithms such as Q-Learning, SARSA, Actor-Critic or Monte Carlo simulations are also conceivable. With these algorithms, dynamic programming and adaptation of the strategy of the LV agent 320 is possible through iteration processes.

In addition, the LV agent 320 contains calculation methods and algorithms f _i for mathematical regression methods or physical model calculations that establish a correlation between selected parameters P _i ∈ P from a set of parameters and selected parameter value ranges PV _i ∈ PV from a set of parameter values and the target _-Describe states. The mathematical functions f _t can be statistical methods such as mean values, minimum and maximum values, lookup tables, models for expected values, linear regression methods, Gaussian processes, fast Fourier transformations, integral and differential calculations and/or Markov methods. In addition, probability methods such as Monte Carlo methods, temporal difference learning, but also extended Kalman filters, radial basis functions, data fields, and/or one or more convergent neural networks, deep neural networks, feedback/recurrent neural networks or convolutional networks can be used. convolutional neural network) can be used. Based on the actions a _i and the rewards r _{i ,} the LV agent 320 selects one or more of these calculation functions f _i for a state s _i .

A neural network consists of neurons arranged in several layers and connected to each other in different ways. A neuron is able to receive information from outside or from another neuron at its input, evaluate the information in a certain way and forward it in a modified form to another neuron at the neuron output or output it as the end result. Hidden neurons are located between the input neurons and output neurons. Depending on the network type, there may be multiple layers of hidden neurons. They ensure the forwarding and processing of the information. Output neurons ultimately deliver a result and output it to the outside world. The arrangement and connection of the neurons creates different types of neural networks such as feedforward neural networks (RNN) without feedback, recurrent neural networks with feedback or convolutional neural networks networks (CNN)). A convolutional neural network has multiple convolution layers and is very suitable for machine learning and applications in the field of pattern recognition and image recognition. The networks can be trained using unsupervised or supervised learning.

In the context of the present invention, the target states are s _t1 , s _t2 , .... , s _tn around real scenarios SZr _i , which are determined from sensor signals from the sensors 500 in particular by the software application 240 of the input module 200. The sensor signals from the sensors 500 can be measured in real time and transmitted to the input module 200. However, it can also be provided that the data of the sensor signals from the sensors 500 are stored in the database 520 or the cloud computing infrastructure 700, which the input module 200 then accesses. In the input module 200, the software application 240 generates real scenarios SZr _i from the sensor signals. These real scenarios SZr _i represent the target states s _t1 , s _t2 , ...., s _tn of the learning reinforcement module 300.

The goal of the learning reinforcement agent 320 is therefore to select parameterized building blocks GRp _i and to set parameters P _i and parameter value ranges PV _i so that all real scenarios SZr _i can be represented by the calculated relevant parameterized scenarios SZp _k . By setting the parameters P _i and the associated parameter values PV _i, concrete scenarios SZc _k result from the calculated relevant parameterized scenarios SZp _k , which accurately and comprehensively reproduce the real scenarios SZr _i that occur in real road traffic.

For the calculation of the relevant parameterized scenarios SZp _k, the LV agent 320 thus selects a calculation function f _i and/or an action a _i based on a guideline for the parameterized scenario SZp _i , which represents the state s _i for the LV agent 320 . Using the selected calculation function f _i and/or the selected action a _i , at least one modification of a parameterized basic building block GRp _i is carried out, so that a modeled parameterized scenario SZp _i+1 is created, which represents a new state s _i+1 . The new state s _i+1 is now compared with the target state st ₁ and the comparison result produces a deviation Δ, to which a reward r _i+1 is assigned.

If the reward is high, it indicates that the LV agent 320 policy is successful and converges. Therefore, no further action a _i+1 or calculation function f _i+1 needs to be carried out for the calculated state s _i+1 , since the calculated nete state s _i+1 already has sufficient correspondence with a real scenario SZr _i . The modified parameterized scenario SZp _i+1 can therefore be viewed as a relevant parameterized scenario SZp _k .

If the policy does not converge, a further calculation function f _i+1 and/or a further action a _i+1 for a further state s _i+2 with a modification of at least one further parameterized basic building block GRp _i+1 from the LV agent 320 selected until the target state st _i is reached.

In particular, it is provided that the LV agent 320 can access further parameterized basic blocks GRp _i to create a parameterized scenario SZp _k if the deviation between a measured real scenario SZr _i and a created parameterized scenario SZp _i becomes too large. In addition, further links can be provided between different parameterized basic building blocks GRp _i and/or parameterized scenarios SZp _i , each of which forms states s _i for the action of the LV agent 320.

The calculated parameterized scenarios SZp _k and the associated parameterized basic blocks GRp _k are output from the output module 450 as output data 455. In addition, the determined parameters P _k and associated parameter value ranges PV _k can be output. The output module 450 can also be integrated into a hardware device such as a computer, a tablet, a smartphone, etc., in particular the hardware device for the input module 200 and the output module 450 being identical.

Furthermore, it can be provided that the output data 455 is forwarded directly to training software, the training software being designed in particular to test and train driving assistance systems and automated driving functions.

The input module 200 and/or the learning reinforcement module 300 can be designed as independent computing units or as a cloud-based solution. The database 520 can also be integrated into the cloud computing infrastructure 700.

Since a fast calculation result is often sought in order to be able to quickly test and train driving assistance systems and automated driving functions, the calculation speed is crucial for the large number of calculation operations in the learning reinforcement module 300. In order to enable the computing operations to be carried out in real time, the use of a cloud computing infrastructure 700 is advantageous. This ensures fast calculations, as cloud-based solutions offer the advantage of high and therefore fast computing performance. In particular, a 5G mobile connection is used for the communication of the input module 200 with the cloud computing infrastructure 700, since data transmission can take place in real time in this way. This speed of data transfer makes sense when cloud-based solutions are used to process the data.

5G is the fifth generation mobile communications standard and, compared to the 4G mobile communications standard, is characterized by higher data rates of up to 10 Gbit/sec, the use of higher frequency ranges such as 2100, 2600 or 3600 megahertz, increased frequency capacity and thus increased data throughput and real-time data transmission because up to a million devices per square kilometer can be addressed simultaneously. The latency times range from a few milliseconds to less than 1 ms, so that real-time transmission of data and calculation results is possible. The data generated by the input module 200 can be sent in real time to the learning reinforcement module 300 integrated in the cloud computing infrastructure 700.

In order to protect the connection to the cloud computing infrastructure 700 using a mobile radio connection, cryptographic encryption methods are provided in particular.

In particular, the calculation results can be stored in the cloud computing infrastructure 700 in the form of states, actions, rewards and strategies, which are then available via the Internet. The learning reinforcement module 300, the variation module 400, the simulation module 420 and the comparison module 440 have the necessary technical interfaces and protocols for access to the cloud computing infrastructure 700 or are themselves integrated into the cloud computing infrastructure 700. This can increase computing efficiency because the access options and access speeds to already calculated states, actions, rewards and strategies are simplified.

In 4 The procedural steps for creating relevant scenarios for simulation purposes are shown.

In a step S10, a plurality of parameterized basic blocks GRp ₁ , GRp ₂ , ..., GRp _n of scenes and events for at least one parameterized scenario SZp _i is generated by an input module 200, the input module 200 using a software application 240.

In a step S20, the plurality of parameterized basic blocks GRp ₁ , GRp ₂ , ..., GRp _n are passed on to a learning reinforcement module 300.

In a step S30, at least one parameterized scenario SZp _i is created by the learning reinforcement module 300 by selecting at least one basic building block GRp _i .

In a step S40, value ranges for the parameter values PV ₁ , PV ₂ ,..., PV _n of the parameters P ₁ , P ₂ , ..., P _n of the basic blocks GRp _i of the parameterized scenario SZp _i are defined.

In a step S50, a set of concrete scenarios SZc _i is selected from the at least one parameterized scenario SZp _i by determining the concrete parameter values PV ₁ , PV ₂ ,..., PV _n of the parameters P ₁ , P ₂ , ..., P _n of the basic building blocks GRp _i created within the specified value ranges.

In a step S60, the set of concrete scenarios SZc _i is transformed into a set of simulated concrete scenarios SZsc _i .

In a step S70, a set of real scenarios SZr _i is generated by the input module 200 using the software application 240 from sensor signals from sensors 500, which are transmitted in real time to the input module 200 and/or can be retrieved from a database 520.

In a step S80, the set of simulated concrete scenarios SZsc _i is compared with the set of real scenarios SZr _i .

In a step S90, at least one relevant parameterized scenario SZp _k is calculated by the learning reinforcement module 300.

In a step S100, the at least one relevant parameterized scenario SZp _k is output for testing and training purposes.

5 schematically illustrates a computer program product 900 that includes executable program code 950 configured to execute the method according to the first aspect of the present invention.

With the method and the system 100 according to the present invention, scenarios are structured using basic modules for different scenes and events. The scenarios are therefore made up of scenes that are linked to each other by events. The basic blocks are parameterized with the help of parameters and associated parameter values, with the choice of the parameters from a set of available parameters and the parameter values from a set of available parameter values being free within certain boundary conditions. A learning reinforcement agent selects specific parameters and parameter values for the parameterized basic building blocks and scenarios, whereby the learning reinforcement agent is based on real scenarios. Through successive iteration and optimization, those relevant parameterized scenarios SZp _k can be calculated from the pool of possible parameterized scenarios SZp _i , which can be used for test and training tasks, for example for testing and training driving assistance systems. The parameterized basic building blocks and parameterized scenarios differ from real scenarios in that they can be used more universally for test and simulation tasks, as they illustrate the typical features and characteristics of a certain scenario type better than a real scenario, which is always specific includes characteristics that go beyond the respective scenario type.

By using parameterized scenarios, the safety of driving assistance systems and automated driving functions can be increased, since all relevant safety-critical driving situations can be covered by parameterized scenarios. By varying the parameters and parameter values, a large number of concrete scenarios can be developed from the calculated relevant parameterized scenarios SZp _k . In addition, resources can be saved because the driving on test routes can be reduced and the programming of individual concrete scenarios is not necessary since these can be derived from the calculated parameterized scenarios. In addition, structuring the scenarios into basic modules enables a modular approach to calculating relevant parameterized scenarios, which enables efficient use of computing power.

Reference symbols

100

system

200

Input module

220

User interface

240

Software application

300

Learning reinforcement module

320

Learning reinforcement agent

400

Variation module

420

Simulation module

440

Comparison module

450

Output module

455

Output data

500

Sensors

520

Database

700

Cloud computing infrastructure

900

Computer program product

950

Program code

QUOTES INCLUDED IN THE DESCRIPTION

This list of documents listed by the applicant was generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Cited patent literature

• WO 2020/083996 A1 [0005]
• US 11137763 B2 [0006]
• EP 3066529 B1 [0007]
• DE 102013003944 A1 [0008]
• DE 102007053500 A1 [0009]
• DE 102006044086 B4 [0010]

Claims (15)

Method for creating parameterized scenarios (SZp ₁ , SZp ₂ , ..., SZp _n ), in particular for testing and training driving assistance systems and automated driving functions, wherein a parameterized scenario (SZp _i ) represents a traffic event in a temporal sequence and as a sequence of scenes starting with a start scene and events can be described, with a scene and an event each being represented by at least one basic module (GR _i ), which is determined by a selection of parameters (P ₁ , P ₂ , ..., P _n ) and associated parameter values (PV ₁ , PV ₂ , ..., PV _n ), and where a parameterized scenario (SZp _i ) is a scenario or a parameterized basic building block (GRp _i ) is a basic building block in which the Parameters (P ₁ , P ₂ , ..., P _n ) and associated parameter values (PV ₁ , PV ₂ , ..., PV _n ) are freely selectable, and as a concrete scenario (SZc _i ) a scenario or as a concrete one Basic module (GRc _i ) is a basic module in which the parameters (P ₁ , P ₂ , ... , P _n ) and associated parameter values (PV ₁ , PV ₂ , ..., PV _n ) are defined, comprising: - Generating (S10) a plurality of parameterized basic blocks (GRp ₁ , GRp ₂ , ..., GRp _n ) of scenes and events for at least one parameterized scenario (SZp _i ) by an input module (200), the input module (200) uses a software application (240); - Passing on (S 20) the plurality of parameterized basic blocks (GRp ₁ , GRp ₂ , ..., GRp _n ) to a learning reinforcement module (300) with a learning reinforcement agent (320); - Creating (S30) at least one parameterized scenario (SZp _i ) by selecting at least one basic building block (GRp _i ) from the learning reinforcement module (300); - Determination (S40) of value ranges for the parameter values (PV ₁ , PV ₂ , ..., PV _n ) of the parameters (P ₁ , P ₂ , ..., P _n ) of the basic blocks (GRp _i ) of the parameterized scenario ( SZp _i ); - Creating (S50) a set of concrete scenarios (SZc _i ) from the at least one parameterized scenario (SZp _i ) by determining concrete parameter values (PV ₁ , PV ₂ , ..., PV _n ) of the parameters (P ₁ , P ₂ , ...,P _n ) of the basic building blocks (GRp _i ) within the specified value ranges; - Simulating (S60) a set of simulated concrete scenarios (SZsc _i ) from the set of concrete scenarios (SZc _i ); - Generating (S70) a set of real scenarios (SZr _i ) from the input module (200) using the software application (240) from sensor signals from sensors (500), the sensor signals from the sensors (500) being sent to the input module (200) in real time. are transmitted and/or can be accessed from a database (520); - Comparing (S80) the set of real scenarios (SZr _i ) with the set of simulated concrete scenarios (SZsc _i ); - Calculating (S90) at least one relevant parameterized scenario (SZp _k ) from the learning reinforcement module (300); - Outputting (S100) the at least one relevant parameterized scenario (SZp _k ) for testing and training purposes. Procedure according to Claim 1 , wherein calculating the at least one relevant parameterized scenario (SZp _k ) includes the following method steps: - Assigning a first reward (r _i ) for the comparison result between the set of real scenarios (SZr _i ) and the set of simulated concrete scenarios (SZsc _i ); - selecting a calculation function (fi ₎ and/or an action ( _ai ) from the learning reinforcement agent (320) based on the reward (r _i ); - Creating a modified parameterized scenario (SZc _i+1 ) using the selected calculation function (fi ₎ and/or action (a _i ) by modifying at least one parameterized basic building block (GRp _i+1 ); - Generating a set of modified simulated concrete scenarios (SZsc _i+1 ) from the at least one modified parameterized scenario (SZp _i+1 ); - Comparing the set of real scenarios (SZr _i ) with the set of modified simulated concrete scenarios (SZsc _i+1 ); - assigning a second reward (r _i+1 ) for the comparison result; - Adjusting a policy of the learning reinforcement agent (320) based on the second reward (r _i+1 ), whereby no further calculation function (f _i+1 ) and/or action (a _i+1 ) is carried out when the policy converges , and in the event of non-convergence of the guideline, a further calculation function (f _i+1 ) and/or a further action (a _i+1 ) for a further modified parameterized scenario (SZc _i+2 ) with a modification of at least one further parameterized basic building block (GRp _i+2 ) is selected by the learning reinforcement agent (320) until a relevant parameterized scenario (SZp _k ) is calculated. Procedure according to Claim 1 or 2 , where a parameter (P _i ) is a moving object such as a motor vehicle, an immovable object such as a building, a road configuration such as a highway, a speed, an acceleration, a street sign, a traffic light, a tunnel, a roundabout, a turning lane, a Acceleration, a direction, an angle, a radius, a location, a traffic volume, a topographical structure such as a slope, a time, a temperature, a precipitation value, a weather and/or a season. Procedure according to one of the Claims 1 until 3 , whereby the sensors (500) are radar systems, LIDAR systems for optical distance and speed measurement, image-recording 2D/3D cameras in the visible, IR and/or UV range, GPS systems, acceleration sensors, speed sensors, capacitive sensors, inductive Sensors, precipitation sensors and / or temperature sensors are formed. Procedure according to one of the Claims 1 until 4 , whereby the software application (240) uses artificial intelligence algorithms to generate the parameterized basic building blocks (GRp _i ). Procedure according to Claim 5 , wherein the artificial intelligence algorithms for extracting and classifying parameters (P _i ) and parameter values (PV _i ) for creating the parameterized basic building blocks (GRp _i ) from input data and sensor signals from the sensors (500) use at least one neural network. Procedure according to Claim 6 , whereby the neural network is designed as a deep neural network, as a convolutional neural network (CNN) and/or as a feedback network (Recurrent Neural Network (RNN)). Procedure according to one of the Claims 5 until 7 , wherein the software application (240) uses at least one encoder and / or at least one decoder. Procedure according to one of the Claims 1 until 8th , wherein at least one algorithm of the learning reinforcement agent (320) is designed as a Markov decision process, temporal difference learning (TD learning), Q-learning, SARSA, Monte Carlo simulation or actor-critic. System (100) for creating parameterized scenarios (SZp ₁ , SZp ₂ , ..., SZp _n ), in particular for testing and training driving assistance systems and automated driving functions, wherein a parameterized scenario (SZp _i ) represents a traffic event in a temporal sequence and can be described as a sequence of scenes starting with a start scene and events, whereby a scene and an event can each be represented by at least one basic module (GR _i ), which is determined by a selection of parameters (P ₁ , P ₂ , .. ., P _n ) and associated parameter values (PV ₁ , PV ₂ , ..., PV _n ) can be determined, and a scenario is referred to as a parameterized scenario (SZp _i ) or a basic component as a parameterized basic module (GRp _i ), in which the parameters (P ₁ , P ₂ , ..., P _n ) and associated parameter values (PV ₁ , PV ₂ , ..., PV _n ) are freely selectable, and as a concrete scenario (SZc _i ) a scenario or . A concrete basic building block (GRc _i ) is a basic building block in which the parameters (P ₁ , P ₂ , ..., P _n ) and associated parameter values (PV ₁ , PV ₂ , ..., PV _n ) are fixed ; comprising an input module (200), a learning reinforcement module (300) with a learning reinforcement agent (320), a variation module (400), a simulation module (420), a comparison module (440) and an output module (450), wherein the input module (200 ) is designed to generate a plurality of parameterized basic building blocks (GRp ₁ , GRp ₂ , ..., GRp _n ) of scenes and events for at least one parameterized scenario (SZp _i ) by means of a software application (240) and the plurality of parameterized basic building blocks (GRp ₁ , GRp ₂ , ..., GRp _n ) to the learning reinforcement module (300); wherein the learning reinforcement module (300) is designed to create at least one parameterized scenario (SZp _i ) by selecting at least one basic module (GRp _i ) and value ranges for the parameter values (PV ₁ , PV ₂ ,...,PV _n ) of the parameters ( P ₁ , P ₂ , ..., P _n ) of the basic building blocks (GRp _i ) of the parameterized scenario (SZp _i ); wherein the variation module (400) is designed to generate a set of concrete scenarios (SZc _i ) from the at least one parameterized scenario (SZp _i ) by determining concrete parameter values (PV ₁ , PV ₂ , ..., PV _n ) of the parameters ( P ₁ ,P ₂ , ..., P _n ) of the basic building blocks (GRp _i ) within the specified value ranges; wherein the simulation module (420) is designed to simulate a set of simulated concrete scenarios (SZc _i ) from the set of concrete scenarios (SZc _i ); wherein the input module (200) is designed to generate a set of real scenarios (SZr _i ) using the software application (240) from sensor signals from sensors (500), which are transmitted in real time to the input module (200) and / or from a database ( 520) are available to generate; wherein the comparison module (440) is designed to compare the set of real scenarios (SZr _i ) with the set of simulated concrete scenarios (SZsc _i ); wherein the learning reinforcement module (300) is designed to calculate at least one relevant parameterized scenario (SZp _k ); and wherein the output module (540) is designed to output the at least one relevant parameterized scenario (SZp _k ) for testing and training purposes. System (100) after Claim 10 , where a parameter (P _i ) is a moving object such as a motor vehicle, an immovable object such as a building, a road configuration such as a highway, a speed, an acceleration, a street sign, a traffic light, a tunnel, a roundabout, a turning lane, a Acceleration, a direction, an angle, a radius, a location, a traffic volume, a topography phical structure such as a gradient, a time, a temperature, a precipitation value, a weather and/or a season. System (100) after Claim 10 or 11 , whereby the sensors (500) are radar systems, LIDAR systems for optical distance and speed measurement, image-recording 2D/3D cameras in the visible range, in the IR and/or UV range, GPS systems, acceleration sensors, speed sensors, capacitive sensors , inductive sensors, precipitation sensors and / or temperature sensors are formed. System (100) according to one of the Claims 10 until 12 , wherein the software application (240) uses artificial intelligence algorithms to extract and classify parameters (P _i ) and parameter values (PV _i ) from input data and sensor signals from the sensors (500) for generating the parameterized basic blocks (GRp _i ). System (100) according to one of the Claims 10 until 13 , wherein at least one algorithm of the learning reinforcement agent (320) is designed as a Markov decision process, temporal difference learning (TD learning), Q-learning, SARSA, Monte Carlo simulation or actor-critic. Computer program product (900), comprising an executable program code (950) which is configured so that, when executed, it implements the method according to one of Claims 1 until 9 executes.

Images