Computer-implemented method and test unit for approximating a subset of test results

Abstract

The invention relates to a computer-implemented method for approximately calculating a subset of test results for a virtual test of a device for at least partially autonomously driving a motor vehicle. The invention also relates to a test unit (1) for approximately calculating a subset of test results for a virtual test of a device for at least partially autonomously driving a motor vehicle. The invention also relates to a computer program and a computer-readable data carrier.

Description

Technical Field

[0001] The invention relates to a computer-implemented method for approximating a subset of test results for a virtual test of an apparatus for at least partially autonomous driving of a motor vehicle.

[0002] The invention further relates to a computer-implemented test unit for identifying a subset of test results for a virtual test of a device for at least partially autonomously driving a motor vehicle. The invention further relates to a computer program and a computer-readable data medium. Background Art

[0003] Driver assistance systems, such as adaptive cruise control and / or functions for highly automated driving, can be verified or validated by means of various testing methods, in particular hardware-in-the-loop methods, software-in-the-loop methods, simulations and / or test drives.

[0004] The outlay, in particular the time and / or cost outlay, for testing such vehicle functions when using the above-described testing methods is typically very high, since a large number of potentially possible driving situations must be tested.

[0005] This can lead to high costs, in particular for test drives and simulations. DE 10 2017 200 180 A1 proposes a method for verifying and / or validating vehicle functions, which is provided for autonomously driving a vehicle in a longitudinal and / or transverse direction.

[0006] The method includes ascertaining a test control command for a vehicle function of a vehicle actuator based on environmental data relating to an environment of the vehicle, wherein the test control command is not executed by the actuator.

[0007] The method further comprises simulating a fictitious traffic situation that would exist if the test control command was executed, based on the surroundings data and using a traffic participant model for at least one traffic participant in the vehicle surroundings.

[0008] The method further comprises providing test data about a fictitious traffic situation. In this case, a vehicle function is passively operated in the vehicle in order to determine the test control commands.

[0009] A disadvantage of this method is that, in order to verify and / or validate the vehicle function, the vehicle must be actually operated to determine the necessary data.

[0010] There is therefore a need to improve existing methods and testing devices in such a way that so-called critical test situations can be efficiently ascertained within the scope of scenario-based testing for systems and system components during highly automated driving. Summary of the Invention

[0011] The object of the present invention is therefore to provide a method, a test unit, a computer program, and a computer-readable data medium which enable efficient identification of critical test situations within the scope of scenario-based testing for systems and system components during highly automated driving.

[0012] This object is achieved according to the invention by a computer-implemented method for approximating a subset of test results of a virtual test of a device for at least partially autonomously driving a motor vehicle, a test unit for identifying a subset of test results of a virtual test of a device for at least partially autonomously driving a motor vehicle, a computer program, and a computer-readable data carrier.

[0013] The invention relates to a computer-implemented method for approximating a subset of test results for a virtual test of an apparatus for at least partially autonomous driving of a motor vehicle.

[0014] The method includes: providing a data set that defines a state space, wherein each state is formed by a parameter group of driving situation parameters, for which one or more actions can be performed in order to realize another parameter group from this parameter group, wherein each parameter group has at least one environmental parameter that describes the vehicle environment and at least one intrinsic parameter that describes the vehicle state.

[0015] In this case, the realization of a further parameter set in the state space may be understood to mean the ascertainment of the further parameter set.

[0016] The method further comprises carrying out an approximate calculation step, in which a function value of at least another parameter group is approximately calculated using an artificial neural network, wherein, if the approximate calculated function value of the at least another parameter group is greater than or equal to a predetermined threshold value, the at least another parameter group is identified as belonging to a subset of the test results.

[0017] If the function value of the at least one further parameter set is less than a predetermined threshold value, the artificial neural network performs at least one further approximation step starting from the respective last approximated further parameter set until the function value of the further parameter set is greater than or equal to the predetermined threshold value.

[0018] The method advantageously uses an artificial neural network whose task is to approximate a subset of test results. The subset of test results is of interest and critical test results that are to be the subject of a virtual test of a device, such as a control unit, for autonomously driving a motor vehicle.

[0019] In scenario-based testing of systems and system components for autonomous vehicles, scenarios are defined, which can be referred to as abstractions of traffic situations. A logical scenario is an abstraction of a traffic situation with no specific parameter values ​​for roads, driving behavior, and adjacent traffic.

[0020] By selecting specific parameter values, a specific scenario is obtained from the logical scenario. Such a specific scenario corresponds to the respective individual traffic situation.

[0021] The autonomous driving function is implemented by a system, for example a control unit, which is conventionally tested in a real vehicle under real traffic conditions or alternatively verified by virtual testing.

[0022] In this document, the method approximates critical test results or traffic situations, that is, a subset of all test results that is considered critical. Critical test situations are, for example, all parameter combinations of specific driving situation parameters that lead to a critical driving situation, such as a vehicle collision or a near-vehicle collision.

[0023] In order to avoid having to test an unnecessarily large number of parameter combinations and traffic situations using conventional simulation methods, the method approximates the aforementioned subset of the test results corresponding to the critical test situations using an artificial neural network.

[0024] The test results thus approximately calculated can then advantageously be verified within the scope of a virtual test of the control unit, so that the method according to the present invention enables a more efficient virtual verification of a control unit for autonomously driving a motor vehicle.

[0025] The definition of the predetermined threshold value therefore defines, according to the invention, when an approximation-calculated test result can be identified as belonging to the desired subset.

[0026] The method is designed in this case such that it carries out any number of approximation calculation steps until a relevant parameter group of driving situation parameters is identified, which belongs to the subset of test results of interest.

[0027] Other embodiments of the present invention are the technical solutions described below with reference to the accompanying drawings.

[0028] According to one aspect of the present invention, the method according to the present invention also includes selecting an initial parameter group from a plurality of parameter groups of driving situation parameters, wherein, in an approximation calculation step, the function value of each adjacent parameter group that can be achieved by action by the initial parameter group is approximated using an artificial neural network, wherein a selection step is performed in which the parameter group having the smallest or highest function value approximated in the approximation calculation step is selected.

[0029] If the function value of the selected parameter set is less than a predetermined threshold value, at least one further selection step is carried out starting from the respective last selected parameter set until the function value of the selected parameter set is greater than or equal to the predetermined threshold value.

[0030] This method uses reinforcement learning to identify critical test cases. Unlike supervised learning with artificial neural networks, reinforcement learning does not utilize provided training data. Instead, there are two parties involved in reinforcement learning: the network, often referred to as the agent, and the environment. The environment can also be thought of as a playing field, on which the agent's current state or position is read.

[0031] The network performs an action based on its current state. This action changes the state of the environment. The network in turn receives a new state and an evaluation of the action it performed from the environment.

[0032] The goal is to achieve the best possible evaluation, i.e., to maximize it. Therefore, during the learning process, the weights of the neural network are adapted based on the evaluation of each executed step, and new actions are implemented. By gradually adapting the weights, the network learns a strategy for implementing the best possible actions, i.e., achieving the best possible evaluation. Similarly, if a minimum is desired, the evaluation can also be minimized through simple adaptation.

[0033] One possible approach for identifying critical test cases is the Q-learning principle, in which all possible actions and their evaluations are considered starting from a certain state. The action with the greatest benefit is selected and executed. For large state and action spaces, as in the present case, the Q function is implemented as a neural network. Such networks are known as DQNs (Deep Q Networks). Neural networks approximate the Q function.

[0034] According to another aspect of the present invention, the method according to the present invention further comprises generating a plurality of parameter sets of driving situation parameters using an artificial neural network or a simulation. The parameter sets of driving situation parameters can thus be generated in a simple manner within a predetermined definition domain, for example by applying a random function when using an artificial neural network.

[0035] According to another aspect of the present invention, the method according to the present invention further includes: the artificial neural network has four hidden layers each including 128 neurons and an ELU activation function; and a coefficient γ of 0.8 is used to attenuate the function value of the another approximate calculation step.

[0036] The approximate calculation of the critical test results is thus carried out in such a way that the artificial neural network is fed with the current position as input in each training step and the function values ​​or Q values ​​for the adjacent positions are approximately calculated.

[0037] The best neighboring position is identified based on the highest gain. The current position is then swapped to this best neighboring position. Network training therefore involves predicting the Q values ​​of neighboring positions based on the given position. The neighboring position with the highest Q value is selected for swapping. The Q value or function value of this position is then decayed by a decay factor γ.

[0038] To determine the payoff, the direct payoff of the position is added to the decayed Q value. For example, a theoretical value of 0 is assigned to all positions except the selected neighboring positions as a theoretical value for the neural network to determine the error and update the weights. The determined payoff is assigned to the selected neighboring positions.

[0039] According to another aspect of the present invention, the method according to the present invention further comprises the step of selecting an initial parameter set from a plurality of parameter sets of driving situation parameters and evaluating the further parameter set approximately calculated by the artificial neural network by another artificial neural network if the function value of the further parameter set is less than a predetermined threshold value.

[0040] The artificial neural network is then adapted based on this evaluation. The artificial neural network adapted in this way performs at least one further approximation step starting from the respective last approximated further parameter set until the function value of the further parameter set is greater than or equal to a predetermined threshold value.

[0041] The above method is an actor-critic approach. In the actor-critic model, there are two parties: the actor and the judge. The actor receives a state and performs actions based on that state, as in Q-learning. Unlike Q-learning, this method does not require discretization, allowing actions to be selected from a continuous set of actions. Furthermore, there is no need to discretize each state.

[0042] The judges evaluate the actor's actions. To do this, they need an evaluation of the environment and the new state. By approximating these evaluations, the judges learn to predict the evaluation of the action. The actor is adapted through the updates of the judges. The actor and judges are trained, for example, simultaneously.

[0043] The judges are given an evaluation of the state and environment, which serves as theoretical values. Errors can be calculated based on the theoretical values ​​and the actual values ​​determined by the judges. The judges are updated using backpropagation. A special feature of actor training is that the errors determined by the judges are used to update the actor using backpropagation.

[0044] According to another aspect of the present invention, the method also includes: the artificial neural network has four hidden layers, each including 256 neurons, and a PReLU activation function; the other artificial neural network has four hidden layers, each including 256 neurons, and an ELU activation function; and the artificial neural network and the other artificial neural network apply the Adam optimization method.

[0045] According to another aspect of the present invention, the method further comprises: the own parameter including the speed of the motor vehicle, and the environmental parameter including the speed of the other motor vehicle and the distance between the motor vehicle and the other motor vehicle.

[0046] Using these parameters, for example, a so-called cut-in scenario can be approximately calculated. A cut-in scenario can be described as a traffic situation in which a highly automated or autonomous vehicle is traveling in a predetermined lane and another vehicle moves from another lane into the lane of the vehicle at a predetermined distance at a reduced speed compared to the vehicle itself.

[0047] The speed of the host vehicle and the other vehicle, also referred to as the following vehicle, is constant here. Because the speed of the host vehicle is higher than the speed of the following vehicle, the host vehicle must be braked in order to avoid a collision of the two vehicles.

[0048] Based on the aforementioned intrinsic and environmental parameters, the method according to the invention can therefore approximately calculate critical traffic situations within a predetermined definition range of the aforementioned parameters.

[0049] According to another aspect of the present invention, the method further comprises: the function on which the function value is based is a safety target function, the safety target function having the following value, the safety distance between the motor vehicle and the other motor vehicle ≥ V FELLOW × 0.55, has a minimum value, and has a maximum value in the case of a collision between the motor vehicle and the other motor vehicle, and the safety distance between the motor vehicle and the other motor vehicle is ≤ V FELLOW In the case of ×0.55, the value is greater than the minimum value.

[0050] The safety objective function specifies how safe the traffic situation is for the vehicle. The safety objective function is described in detail as follows: if the distance between the vehicle and the following vehicle is greater than or equal to the safety distance, the function value of the safety objective function is 0.

[0051] The safety distance may be defined as a distance at which the vehicle can always be safely braked without colliding with the following vehicle, based on the speed difference between the vehicle and the following vehicle and the distance between the vehicle and the following vehicle.

[0052] Such a distance is defined in this example by a value in meters, which corresponds to the speed V FELLOW ×0.55.

[0053] As the distance between the host vehicle and the following vehicle decreases or falls below the safety distance, the objective function value gradually approaches the value 1. If a collision between the host vehicle and the following vehicle occurs, the distance between the host vehicle and the following vehicle is less than or equal to zero and the objective function value is 1.

[0054] According to another aspect of the present invention, the method according to the present invention further comprises: the function on which the function value is based is a comfort target function or an energy consumption target function, the comfort target function or the energy consumption target function having the following values, which have a minimum value when the acceleration of the motor vehicle does not change and a maximum value in the event of a collision between the motor vehicle and the other motor vehicle, and have a value between the minimum and maximum value depending on the amount of the acceleration change when the acceleration of the motor vehicle changes.

[0055] The comfort objective function can be used to draw conclusions about how comfortable the driving maneuver is for the driver of the vehicle. Strong acceleration or braking and the frequent repetition of these processes are considered uncomfortable.

[0056] The change in acceleration is called ruck. The smaller the calculated value of the comfort objective function, the more comfortable the driving situation. In the event of a collision between the vehicle and the following vehicle, fuel consumption is set to 1, meaning that fuel consumption is set to a specific maximum value. This is because the vehicle's fuel tank is no longer available in the event of an accident.

[0057] With regard to the cut-in scenario, the potentially critical test situations therefore lie at the boundary between collision and non-collision situations, which can be defined according to the corresponding objective functions, ie, a safety objective function, a comfort objective function, and an energy consumption objective function.

[0058] According to another aspect of the present invention, the method includes generating a plurality of driving situation parameters, in particular the speed of the motor vehicle and the speed of the other motor vehicle, within a predetermined domain using a stochastic algorithm. This allows for a simple and time-saving generation of a plurality of driving situation parameters, which form a data set for approximating the calculation of critical test results.

[0059] According to another aspect of the invention, the method further comprises using a separate artificial neural network for approximating the value range of each function on which the function value is based, wherein the individual hyperparameters of each artificial neural network are stored in a database.

[0060] The use of a separate artificial neural network for approximating the value range of each individual objective function, ie, the safety objective function, the comfort objective function, and / or the energy consumption objective function, advantageously enables more accurate approximation results.

[0061] According to a further aspect of the present invention, a testing unit is provided for identifying a subset of test results of a virtual test of a device for at least partially autonomously driving a motor vehicle.

[0062] The test unit comprises means for providing a data set defining a state space, wherein each state is formed by a parameter group of driving situation parameters, for which one or more actions can be performed in order to implement another parameter group from this parameter group, wherein each parameter group has at least one environmental parameter describing the vehicle environment and at least one intrinsic parameter describing the vehicle state.

[0063] The test unit further comprises an artificial neural network which implements an approximate calculation step in which a function value of at least another parameter group can be approximately calculated, wherein if the approximate calculated function value of the at least another parameter group is greater than or equal to a predetermined threshold value, then the at least another parameter group is identified as belonging to the subset of the test results.

[0064] If the function value of the at least one other parameter set is less than a predetermined threshold, the artificial neural network is configured to perform at least one other approximation step starting from the corresponding last approximated other parameter set until the function value of the other parameter set is greater than or equal to the predetermined threshold.

[0065] Within the scope of the present test unit, an artificial neural network is therefore advantageously used, which has the task of approximating a subset of the test results, ie, the critical test results of interest.

[0066] The test results calculated approximately in this way can then advantageously be verified within the scope of a virtual test of the control unit, so that a more efficient virtual verification of a control unit for autonomously driven motor vehicles can be achieved by the test unit according to the present invention.

[0067] According to another aspect of the invention, the device is formed by a control unit, and the approximate calculation of the test results of the virtual test of the control unit is based on a driving situation in which another motor vehicle changes lanes into the lane of the motor vehicle using a plurality of driving situation parameters.

[0068] The test unit is therefore advantageously able to approximately calculate corresponding test results of a virtual test with respect to, for example, a cut-in scenario.

[0069] According to another aspect of the present invention, a computer program is provided, which includes a program code for carrying out the method according to the present invention when the computer program is executed on a computer. According to another aspect of the present invention, a data carrier is provided, which includes the program code of the computer program for carrying out the method according to the present invention when the computer program is executed on a computer.

[0070] The features of the method described herein can be used to approximate critical test results for a plurality of different scenarios or driving situations. Similarly, the test unit according to the invention is suitable for testing a plurality of different devices or control units, for example, of automobiles, trucks and / or commercial vehicles, ships or aircraft, with respect to critical test results. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] For a better understanding of the present invention and its advantages, reference is now made to the following description in conjunction with the accompanying drawings. The present invention is further explained below based on exemplary embodiments, each of which is shown in the schematic drawings of the accompanying drawings. The accompanying drawings show:

[0072] Figure 1 A flow chart showing a method for approximating a subset of test results of a virtual test of a device for at least partially autonomously driving a motor vehicle according to a preferred embodiment of the present invention;

[0073] Figure 2 A schematic diagram showing an approximate calculation method according to the present invention according to a preferred embodiment of the present invention;

[0074] Figure 3 A schematic diagram showing an approximate calculation method according to the present invention according to a preferred embodiment of the present invention;

[0075] Figure 4 A flow chart showing a DQN network according to the present invention according to a preferred embodiment of the present invention;

[0076] Figure 5 A flow chart showing a method for approximating a subset of test results of a virtual test of a device for at least partially autonomously driving a motor vehicle according to another preferred embodiment of the present invention;

[0077] Figure 6 Another preferred embodiment of the present invention is shown in Figure 5 Another flow chart of the method shown in;

[0078] Figure 7 A 3-dimensional diagram showing the target function according to the present invention according to another preferred embodiment of the present invention;

[0079] Figure 8 Another preferred embodiment of the present invention is shown in Figure 7 2-dimensional diagram of a cross section of the objective function according to the present invention shown in FIG; and

[0080] Figure 9 Another preferred embodiment of the present invention is shown in Figure 7 2D diagram of a cross section of the target function according to the invention is shown in FIG.

[0081] Unless otherwise specified, the same reference numerals denote the same elements of the drawings. DETAILED DESCRIPTION

[0082] Figure 1 A flow chart of a method for approximating a subset of test results of a virtual test of a device for at least partially autonomously driving a motor vehicle is shown according to a preferred embodiment of the present invention.

[0083] The method includes providing S1 a data set D defining a state space Z. Each state Z1, Z2, ... Zn is formed by a parameter set P1, P2, ... Pn of driving situation parameters. For each state Z1, Z2, ... Zn, one or more actions can be performed in order to implement another parameter set P1, P2, ... Pn from the parameter set P1, P2, ... Pn, wherein each parameter set P1, P2, ... Pn comprises at least one environmental parameter describing the vehicle environment and at least one intrinsic parameter describing the vehicle state.

[0084] The method further comprises carrying out an approximation calculation step S2 in which function values ​​F1 , F2 . . . Fn of at least one further parameter set P1 , P2 . . . Pn are approximately calculated using the artificial neural network K1 .

[0085] If the approximated function values ​​F1, F2 ... Fn of the at least one further parameter group P1, P2 ... Pn are greater than or equal to a predetermined threshold value W, the at least one further parameter group P1, P2 ... Pn is identified (S3) as belonging to the subset of the test results. If the function values ​​F1, F2 ... Fn of the at least one further parameter group P1, P2 ... Pn are less than the predetermined threshold value W, the artificial neural network K1 performs at least one further approximation step S4 starting from the corresponding last approximated further parameter group P1, P2 ... Pn until the function values ​​F1, F2 ... Fn of the at least one further parameter group P1, P2 ... Pn are greater than or equal to the predetermined threshold value W.

[0086] A plurality of parameter sets (P1, P2 . . . Pn) of driving situation parameters are generated by the artificial neural network K1. Alternatively, the plurality of parameter sets (P1, P2 . . . Pn) can be generated, for example, by simulation.

[0087] In this embodiment, the artificial neural network K1 has four hidden layers, each comprising 128 neurons, and an ELU activation function. Furthermore, a coefficient γ of 0.8 is used to attenuate the function values ​​F1, F2 . . . Fn of the further approximation step.

[0088] Figure 2 A schematic diagram of the approximate calculation method according to the invention is shown according to a preferred embodiment of the invention.

[0089] Figure 2 The diagram shows a Q-learning game field consisting of 10×10 grids. The target point is located in the center of the game field and is marked in black. Within the scope of this Q-learning method, a neural network, in particular a DQN (Deep Q-Network), is used to approximate the corresponding target function for a given scenario or traffic situation to be tested.

[0090] The initial position is randomly assigned. An action moves the current position to an adjacent position. Accordingly, a position can be moved to the adjacent upper, right, lower, or left space. As a direct benefit, a predetermined value, such as 100, is assigned to the target space, while another value, such as 0, is assigned to each other space on the field.

[0091] During a game, the transition from the initial position to the new position is repeated until the target position is reached. A new game is then started from a new random initial position. The training of the neural network for approximating the Q function is completed after a predetermined number of game sessions, for example, 1,000 game sessions. The value of the coefficient γ for attenuating the gains is set to 0.8, for example.

[0092] Figure 3 A schematic diagram of the approximate calculation method according to the invention is shown according to a preferred embodiment of the invention.

[0093] exist Figure 3 The example in which the driving situation parameter V is used is shown in FIG. EGO , that is, the speed of the vehicle and V on the vertical axis FELLOW , that is, the speed of the following vehicle in front.

[0094] exist Figure 3 The function shown in forms the boundary between critical and non-critical test results and essentially corresponds to Figure 2 The objective function shown in . The points shown are test results of approximate calculations. Alternatively, the points shown can be, for example, test results of simulations.

[0095] The function shown is a safety target function having the value of a safety distance between the motor vehicle and the other motor vehicle ≥ V FELLOW × 0.55, has a minimum value, and has a maximum value in the case of a collision between the motor vehicle and the other motor vehicle, and the safety distance between the motor vehicle and the other motor vehicle is ≤ V FELLOW The case of ×0.55 has a value larger than the minimum value.

[0096] Alternatively to the safety objective function, for example, a comfort objective function or an energy consumption objective function can be approximately calculated, wherein the comfort objective function or the energy consumption objective function has a value that has a minimum value when the acceleration of the motor vehicle does not change, has a maximum value in the event of a collision between the motor vehicle and the other motor vehicle, and has a value between the minimum and maximum value depending on the amount of the acceleration change when the acceleration of the motor vehicle changes.

[0097] A plurality of driving situation parameters, in particular the speed V of the motor vehicle EGO and the speed V of the other motor vehicle FELLOW The plurality of driving situation parameters may be generated by a random algorithm within a predetermined definition domain. Alternatively, the plurality of driving situation parameters may be generated, for example, by simulation.

[0098] A separate artificial neural network is used to approximate the value range of each function on which the function value is based. The individual hyperparameters of each artificial neural network are stored in a database.

[0099] Figure 4 A flow chart of a DQN network according to the present invention is shown according to a preferred embodiment of the present invention.

[0100] An initial parameter set SP is selected from the plurality of parameter sets P1, P2 ... Pn of driving situation parameters. In an approximation step S2, function values ​​F1, F2 ... Fn of each adjacent parameter set P1, P2 ... Pn that can be realized from the initial parameter set SP via actions A1, A2 ... An are approximated using an artificial neural network K1.

[0101] Subsequently, a selection step S2A is carried out, in which the parameter group P1 , P2 . . . Pn approximated in the approximation step to have the smallest or highest function value F1 , F2 . . . Fn is selected.

[0102] If the function value F1, F2...Fn of the selected parameter group is less than the predetermined threshold W, then at least another selection step S2B is performed starting from the corresponding last selected parameter group P1, P2...Pn until the function value F1, F2...Fn of the selected parameter group is greater than or equal to the predetermined threshold W.

[0103] Figure 5 A flow chart of a method for approximating a subset of test results of a virtual test of a device for at least partially autonomously driving a motor vehicle is shown according to another preferred embodiment of the present invention.

[0104] Another alternative implementation currently involves an actor-judge method or model compared to the Q-learning method. In the actor-judge method, the discretization of states and actions is not necessary. For applications such as cut-in scenarios, the state is V EGO and V FELLOW Value pairs. These are Figure 5 For example, parameter groups P1 and P2 are shown in FIG.

[0105] From a specific parameter set, it is possible to transform to any other parameter set. Therefore, it is not necessary to transform to specific adjacent value pairs or parameter sets, as is the case with Q-learning. The step size is arbitrary and can achieve value pairs that would not be possible in Q-learning due to discretization.

[0106] As in Q-learning, two application cases are considered. Critical test cases are to be identified in which a collision occurs or is on the borderline between a collision and a non-collision situation. For example, a safety objective function is used as an evaluation of the environment.

[0107] An initial parameter set SP is selected from the plurality of parameter sets P1 , P2 . . . Pn of driving situation parameters.

[0108] If the function value F1, F2 . . . Fn of the further parameter set P1, P2 . . . Pn is less than a predetermined threshold value W, the further artificial neural network K2 evaluates the further parameter set P1, P2 . . . Pn approximated by the artificial neural network K1 and adapts the artificial neural network K1 based on the evaluation BW in step S5.

[0109] The artificial neural network K1 adapted in this way carries out at least one further approximation step S4 starting from the respective last approximated further parameter set P1, P2 . . . Pn until the function value F1, F2 . . . Fn of the further parameter set P1, P2 . . . Pn is greater than or equal to a predetermined threshold value W.

[0110] In step S6, the second artificial neural network K2 or the jury network is trained or learned. The jury network is trained using backpropagation. The jury network or the other neural network K2 is fed with states and evaluations of the environment, which serve as theoretical values. Errors can be calculated based on the theoretical values ​​and the actual values ​​ascertained by the jury network. Subsequently, the jury network is updated using backpropagation.

[0111] Figure 6Another preferred embodiment of the present invention is shown in Figure 5 Another flow chart of the method shown in .

[0112] The artificial neural network K1 has four hidden layers, each containing 256 neurons, and a PReLU activation function. The other artificial neural network K2 has four hidden layers, each containing 256 neurons, and an ELU activation function. The artificial neural network K1 and the other artificial neural network K2 use the Adam optimization method.

[0113] The own parameters FP3 include the speed V of the motor vehicle EGO The environmental parameters FP1, FP2 include the speed V of the other motor vehicle. FELLOW and the distance d between the motor vehicle and the other motor vehicle SPUP .

[0114] The artificial neural network K1 receives the value pair V EGO and V FELLOW and spacing d SPUP As input parameter and the value of V EGO and V FELLOW Convert to a new value pair V' EGO and V' FELLOW .

[0115] The further artificial neural network K2 evaluates the new value pair V' EGO and V' FELLOW The artificial neural network K1 is adapted by means of the evaluation of the further artificial neural network K2 .

[0116] Figure 7 A three-dimensional diagram of the target function according to the present invention is shown according to another preferred embodiment of the present invention.

[0117] The function shown is a truncated cone with a constant peak. The purpose of the approximate calculation method is to realize a point that lies on the plane of the cone. The given parameter pairs P1, P2 belong to the corresponding function to be determined. These parameter pairs can be, for example, the speed V of the vehicle. EGO and the speed V of the following vehicle FELLOW .

[0118] The other artificial neural network K2 has been pre-trained or previously trained. The neural network or actor network is updated based on the evaluation of the other neural network or the judges network to achieve the best possible result. The evaluation of the judges network is the actual value, and the theoretical value is the maximum value of the function.

[0119] For training, a random initial position is generated, similar to Q-learning. The goal is to transform this initial position into the target region, i.e., the plane of the cone. This initial position is submitted to the actor network, which transforms the initial position into the new position.

[0120] The actor network is updated based on the new position using the evaluations from the judges network. The current position is then transferred back to the new position by the actor network. This process is repeated multiple times. The actor network is thus updated at each step based on the evaluations from the judges network.

[0121] Figure 8 Another preferred embodiment of the present invention is shown in Figure 7 2D diagram of a cross section of the target function according to the invention is shown in FIG.

[0122] exist Figure 8 The circular surface shown in FIG includes or corresponds to the target area of ​​the function. The points shown are test results approximately calculated by the method according to the present invention.

[0123] Figure 9 Another preferred embodiment of the present invention is shown in Figure 7 2D diagram of a cross section of the target function according to the invention is shown in FIG.

[0124] In this function, the target region is relatively narrowly defined and corresponds to the Figure 8 The edge region of the target region is shown in FIG. The points arranged along the linear edge region correspond to the test results approximately calculated by the method.

[0125] As Figure 9 It can be seen that the approximately calculated test results lie within the specified target area and therefore correspond to a subset of the test results, ie the critical test results of interest.

[0126] exist Figure 1 and Figure 5 1 also shows a test unit 1 according to the invention for identifying a subset of test results from a virtual test of a device for at least partially autonomously driving a motor vehicle. The test unit 1 comprises corresponding means 2 for providing a data set D defining a state space Z, as well as an artificial neural network K1 and / or a further artificial neural network K2.

[0127] Although specific implementations are shown and described herein, it is clear to those skilled in the art that there are multiple alternative and / or equivalent implementations. It should be noted that one or more exemplary implementations are merely examples and are not intended to limit the scope, applicability, or configuration in any way.

[0128] Moreover, the above summary and detailed description provide a person skilled in the art with comfortable guidance for implementing at least one exemplary embodiment, wherein it is clear that different changes can be made in the functional scope and arrangement of the elements without departing from the scope of protection of the technical solution and its legal equivalents.

[0129] Generally, this application is intended to cover adaptations or variations of the embodiments disclosed herein.

Claims

1. A computer-implemented method for approximating a subset of test results of a virtual test of an apparatus for at least partially autonomous driving of a motor vehicle, the method comprising the steps of: Provide (S1) a data set (D) defining a state space (Z) where, Each state (Z1, Z2 . . . Zn) is formed by a parameter set (P1, P2 . . . Pn) of driving situation parameters, for which one or more actions (A1, A2 . . . An) can be performed in order to implement a further parameter set (P1, P2 . . . Pn) from the parameter set (P1, P2 . . . Pn), wherein each parameter set (P1, P2 . . . Pn) comprises at least one environmental parameter (FP1, FP2) describing the vehicle environment and at least one intrinsic parameter (FP3) describing the vehicle state; An approximation calculation step (S2) is performed, in which a function value (F1, F2 ... Fn) of at least one other parameter group (P1, P2 ... Pn) is approximately calculated using an artificial neural network (K1); if the function value (F1, F2 ... Fn) of the at least one other parameter group (P1, P2 ... Pn) after the approximation calculation is greater than or equal to a predetermined threshold value (W), then the at least one other parameter group (P1, P2 ... Pn) is identified (S3) as belonging to a subset of the test results; if the function value (F1, F2 ... Fn) of the at least one other parameter group (P1, P2 ... Pn) is less than the predetermined threshold value (W), then the artificial neural network (K1) performs at least one other approximation calculation step (S4) starting from the corresponding last approximated other parameter group (P1, P2 ... Pn) until the function value (F1, F2 ... Fn) of the other parameter group (P1, P2 ... Pn) is greater than or equal to the predetermined threshold value (W).

2. The computer-implemented method of claim 1 , wherein: An initial parameter set (SP) is selected (S1) from a plurality of parameter sets (P1, P2 . . . Pn) of driving situation parameters, wherein in an approximation calculation step (S2) function values ​​(F1, F2 . . . Fn) of each adjacent parameter set (P1, P2 . . . Pn) that can be realized by the initial parameter set (SP) via actions (A1, A2 . . . An) are approximately calculated using an artificial neural network (K1), wherein a selection step (S2A) is performed in which the function values ​​(F1, F2 . . . Fn) of the adjacent parameter sets (P1, P2 . . . Pn) are selected. ) and, if the function value (F1, F2 . . . Fn) of the selected parameter group (P1, P2 . . . Pn) is less than a predetermined threshold value (W), then at least one further selection step (S2B) is performed starting from the corresponding last selected parameter group (P1, P2 . . . Pn) until the function value (F1, F2 . . . Fn) of the selected parameter group (P1, P2 . . . Pn) is greater than or equal to the predetermined threshold value (W).

3. The computer-implemented method according to claim 1 or 2, wherein: A plurality of parameter sets (P1, P2 . . . Pn) of driving situation parameters are generated by means of the artificial neural network (K1) or by simulation.

4. The computer-implemented method according to claim 1 or 2, wherein: The artificial neural network (K1) has four hidden layers each including 128 neurons and an ELU activation function; and a coefficient γ of 0.8 is used to attenuate the function values ​​(F1, F2 . . . Fn) of the other approximate calculation step.

5. The computer-implemented method of claim 1 , wherein: An initial parameter set (SP) is selected from a plurality of parameter sets (P1, P2 . . . Pn) of driving situation parameters, and if a function value (F1, F2 . . . Fn) of the other parameter set (P1, P2 . . . Pn) is less than a predetermined threshold value (W), then a further artificial neural network (K2) evaluates the other parameter set (P1, P2 . . . Pn) approximately calculated by the artificial neural network (K1) and adapts the artificial neural network (K1) based on the evaluation (BW), and the artificial neural network (K1) adapted in this way performs at least one further approximation step (S3) starting from the respective last approximately calculated other parameter set (P1, P2 . . . Pn) until the function value (F1, F2 . . . Fn) of the other parameter set (P1, P2 . . . Pn) is greater than or equal to the predetermined threshold value (W).

6. The computer-implemented method of claim 5, wherein: The artificial neural network (K1) has four hidden layers each including 256 neurons and a PReLU activation function; the other artificial neural network (K2) has four hidden layers each including 256 neurons and an ELU activation function; and the artificial neural network (K1) and the other artificial neural network (K2) apply the Adam optimization method.

7. The computer-implemented method of claim 1 or 2, wherein: The own parameters (FP3) include the speed V of the motor vehicle EGO , and the environmental parameters (FP1, FP2) include the speed V of the other motor vehicle FELLOW and the distance (d SPUR ).

8. The computer-implemented method of claim 7, wherein: The function on which the function values ​​(F1, F2 . . . Fn) are based is a safety target function having a value such that the safety distance between the motor vehicle and the other motor vehicle is ≥ V FELLOW × 0.55, has a minimum value, and has a maximum value in the case of a collision between the motor vehicle and the other motor vehicle, and the safety distance between the motor vehicle and the other motor vehicle is ≤ V FELLOW The case of ×0.55 has a value larger than the minimum value.

9. The computer-implemented method of claim 1 or 2, wherein: The function on which the function values ​​(F1, F2 . . . Fn) are based is a comfort target function or an energy consumption target function, which has the following values: when the acceleration of the motor vehicle does not change, it has a minimum value; when a collision occurs between the motor vehicle and the other motor vehicle, it has a maximum value; and when the acceleration of the motor vehicle changes, it has a value between the minimum and maximum value depending on the amount of the acceleration change.

10. The computer-implemented method of claim 7, wherein: A plurality of driving situation parameters are generated within a predetermined definition domain by a random algorithm.

11. The computer-implemented method of claim 7, wherein: The speed V of the motor vehicle EGO and the speed V of the other motor vehicle FELLOW Generated by a random algorithm within a predetermined domain.

12. The computer-implemented method of claim 8, wherein: For the approximate calculation of the value range of each function on which the function values ​​(F1, F2 . . . Fn) are based, a separate artificial neural network is used, wherein the individual hyperparameters of each artificial neural network are stored in a database.

13. A test unit (1) for approximating a subset of test results of a virtual test of an apparatus for at least partially autonomous driving of a motor vehicle, the test unit comprising: Means (2) for providing a data set (D) defining a state space (Z), wherein each state (Z1, Z2, ... Zn) is formed by a parameter set (P1, P2, ... Pn) of driving situation parameters, for which one or more actions (A1, A2, ... An) can be performed in order to implement another parameter set (P1, P2, ... Pn) from the parameter set (P1, P2, ... Pn), wherein each parameter set (P1, P2, ... Pn) has at least one environmental parameter (FP1, FP2) describing the vehicle environment and at least one intrinsic parameter (FP3) describing the vehicle state; An artificial neural network (K1) is provided, wherein the artificial neural network performs an approximate calculation step, in which a function value (F1, F2 ... Fn) of at least another parameter group (P1, P2 ... Pn) can be approximately calculated, and if the function value (F1, F2 ... Fn) of the at least another parameter group (P1, P2 ... Pn) after the approximate calculation is greater than or equal to a predetermined threshold value (W), the at least another parameter group (P1, P2 ... Pn) can be identified as belonging to a subset of test results; if the function value (F1, F2 ... Fn) of the at least another parameter group (P1, P2 ... Pn) is less than the predetermined threshold value (W), the artificial neural network (K1) is configured to perform at least another approximate calculation step starting from the corresponding last approximately calculated another parameter group (P1, P2 ... Pn) until the function value (F1, F2 ... Fn) of the another parameter group (P1, P2 ... Pn) is greater than or equal to the predetermined threshold value (W).

14. The test unit according to claim 13, characterized in that The device is formed by a control unit, and the approximate calculation of the test results of the virtual test of the control unit is based on a driving situation in which another motor vehicle changes lanes into the lane of the motor vehicle using a plurality of driving situation parameters. 15 . A computer-readable data carrier comprising a program code of a computer program for carrying out the method according to claim 1 , when the computer program is executed on a computer.

Images