Modeling of vehicle motion profiles

Abstract

To develop and apply data-driven algorithms in self-driving systems for accurate simulation of self-driving systems; advantageously generate, in simulating the self-driving systems, multiple synthetic data points characterizing an aspect of the performance of a target self-driving system; and, in this way, simulate and statistically evaluate various future scenarios, for example.SOLUTION: The invention relates to a computer-implemented method for training a generative model for machine learning, a computer-implemented method for generating synthetic data series using a generative model for machine learning, and an associated apparatus. The invention also relates to an associated computer program element and computer readable medium.SELECTED DRAWING: Figure 3A

Description

[Technical Field]

[0001] The present invention relates to a computer-implemented method for training a generative model for machine learning, a computer-implemented method for generating artificial data samples using a generative model for machine learning, and related apparatus. The present invention further relates to related computer program elements and computer-readable media. [Background technology]

[0002] Conventional technology The development and application of data-driven algorithms in autonomous driving systems is becoming increasingly important for accurate simulation of these systems. When simulating autonomous driving systems, it is advantageous to generate multiple artificial data points that characterize a particular aspect of the performance of the target vehicle system. In this way, for example, various future scenarios can be simulated and statistically evaluated.

[0003] This thematic area includes, for example, the generation of vehicle trajectory prediction data based on previously acquired vehicle trajectory data. Generative Adversarial Networks (GANs) are typically used to provide generative models. However, this approach can be further improved. [Overview of the project] [Means for solving the problem]

[0004] Summary of the Invention According to a first aspect, a computer-implemented method is provided for training a generative model for machine learning to model the motion profile of a vehicle or robot, which includes adversarial training of a generator model and a discriminator model. · Obtaining a plurality of data sequences from a data space, where each data sequence describes at least one operation profile of a vehicle or a robot, · Generating a plurality of artificial data sequences by sampling from a distribution during at least one generation phase during training of a generative model for machine learning, · During at least one discrimination phase during training of a generative model for machine learning, inputting into a discriminator model (i) a data sequence from the plurality of obtained data sequences, or (ii) an artificial data sequence from the plurality of artificial data sequences, · During at least one discrimination phase during training of a generative model for machine learning, classifying the input to the discriminator model as a data sequence from the data space or, at least, as an artificial data sequence using a first function and a second function of the discriminator model, where a plurality of inputs of the first function include a plurality of consecutive samples of the artificial data sequence, and at least one input of the second function includes at least two data samples of the artificial data sequence obtained from a plurality of different steps of the artificial data sequence, · Repeatedly training a generator model and a discriminator model of a generative model for machine learning to generate a trained model for machine learning including the trained generator model.

[0005] One effect is that the artificial data generated by the model for machine learning trained according to this method matches or approximates the true distribution regarding the specific features of a true vehicle application scenario. As a result, when using a generator trained using a discriminator to generate artificial data samples, the quality of the simulation is improved.

[0006] The model generated according to the above approach can advantageously also be used during the running operation of the vehicle. The engine control unit can, for example, predict fuel consumption or exhaust characteristics based on the recorded speed profile of the vehicle's travel, or based on the vehicle's future route calculated by the navigation system, or based on the most likely route taking into account the vehicle's current position. If the model is trained to predict the future operating mode of the vehicle, it can approximate the exhaust or fuel consumption for a given route more accurately than the engine control unit that uses this model to control the vehicle, whereby, for example, the vehicle can on average reduce its own exhaust emissions.

[0007] Based on the results of this model, the control strategy of the vehicle or robot can be adapted. In the case of an electric vehicle or an electric robot, if a certain route is predicted to consume the battery beyond an acceptable level, another route may be proposed, a driver alarm may be activated, and / or a low power mode (e.g., limiting the speed of the vehicle or robot) may be initiated.

[0008] This model can be configured to evaluate the input to the vehicle route search software with respect to fuel performance or exhaust performance. Routes with more efficient fuel performance or lower exhaust characteristics can be presented to the driver with a higher priority on the output interface.

[0009] This model can similarly be used in the control unit of an electric vehicle or a hybrid vehicle to improve the prediction of battery consumption or charging time based on the predicted route.

[0010] Using the model trained according to the first aspect, the wear of vehicle components that can be caused by driving on one or more predicted routes can be estimated.

[0011] Furthermore, this embodiment of the internal operation of a processor or system used to train a model for machine learning (e.g., cGAN or GAN) is advantageous. In particular, the practical effect of the second module is to provide additional gradient information to the GAN generator or cGAN generator during training. This facilitates the discriminator's task by focusing on domain-specific features such as a sequence of values ​​obtained from distributions resulting from real-world or artificial data sequences (optionally velocity-acceleration histograms), thereby also inducing training and potentially leading to improvements in training stability, speed, and accuracy.

[0012] According to a second aspect, a computer-implemented method is provided for generating artificial data samples that represent an artificial motion profile of a vehicle or robot using a generative model for machine learning. This method is • Constructing a generative model for machine learning according to a plurality of model parameters obtained according to the method of the first embodiment, • Obtaining multiple samples by sampling from a distribution configured to provide a pseudo-random vector, • To obtain multiple artificial data samples within the data space, decode multiple samples of this distribution, • Outputting multiple additional artificial data samples that represent the artificial motion profile of a vehicle or robot, Includes.

[0013] According to a third aspect, a computer-implemented method is provided for training a further model for machine learning to model the operating profile of a vehicle, the method being: • Obtaining a further set of artificial data representing an artificial operating profile of a vehicle generated according to the computer-implemented method of the second aspect, • Inputting multiple additional artificial data sequences representing the vehicle's artificial operating profile into further models for machine learning, • Iteratively training further models for machine learning, • Outputting multiple additional model parameters for further models of machine learning to model the vehicle's operating profile, Includes.

[0014] According to a fourth aspect, an apparatus is provided for training a generative model for machine learning that models the motion profile of a vehicle or robot, including adversarial training of a generator model and a discriminator model stored in memory. The apparatus includes an input interface configured to retrieve multiple data sequences from a data space, where each data sequence describes at least one motion profile of a vehicle. The apparatus further includes a processor configured to generate at least one artificial data sequence by sampling from a distribution during at least one generative phase of training a generative model for machine learning.

[0015] The processor is configured to input the classifier model, during at least one classification phase when training a generative model for machine learning, either (i) a data sequence from multiple acquired data sequences, or (ii) an artificial data sequence from at least one artificial data sequence.

[0016] The processor is configured to classify the input to the classifier model either as a data sequence from the data space or as an artificial data sequence using at least a first and second function of the classifier model, during at least one classification phase when training a generative model for machine learning.

[0017] The multiple inputs of the first function include multiple consecutive samples of an artificial data sequence, and the at least one input of the second function includes at least two data samples of an artificial data sequence obtained from multiple different steps of the artificial data sequence.

[0018] The processor is configured to iteratively train generator and discriminator models for machine learning in order to generate a trained model for machine learning, which includes a trained generator model. The device further includes an output interface configured to output the trained model for machine learning.

[0019] According to the fifth aspect, one or more computer program elements are provided, each computer program element comprising at least: (i) Computer-executable instructions for training a generative model for machine learning according to the method or embodiment of the first aspect, and / or (ii) Computer-executable instructions for generating an artificial data sequence in accordance with the third aspect, and / or (iii) Computer-executable instructions including model parameters for providing a generative model for machine learning to be trained according to the first aspect or an embodiment thereof. Includes.

[0020] According to the sixth aspect, a computer-readable medium is provided which includes one or more computer program elements of the fifth aspect.

[0021] According to a seventh aspect, a vehicle is provided which includes a prime mover configured to provide driving force to operate the vehicle, a plurality of sensors, and a vehicle electronics unit configured to control at least the prime mover based on feedback from the sensors. The vehicle electronics unit includes a processor configured to instantiate and implement a trained model according to a second aspect. The trained model is configured to receive input signals from the sensors and apply these input signals to the trained model. The processor is configured to compute one or more response signals using the trained model. The processor is configured to adapt the configuration of one or more prime movers or other output actuators of the vehicle in response to response signals obtained through the trained model.

[0022] In this patent specification, a vehicle may be any means for transporting or carrying elements (such as passengers or effective load). Examples of vehicles include automobiles, trucks, delivery vehicles, or buses having an internal combustion engine, hybrid engine, gas engine, electric powertrain, or hydrogen hybrid engine. A vehicle may be an autonomous or semi-autonomous passenger car, bus, or truck. A vehicle may be an electric bicycle, electric scooter, factory robot, forklift, or electric wheelchair.

[0023] In this patent specification, the vehicle operation profile is, for example, one or more sampled data sequences describing a series of speeds, engine temperatures, engine speeds, pedal positions, steering angles, or gear shifting functions of a vehicle traveling on a road.

[0024] In one example, the trained model can also be applied to semi-autonomous or autonomous robots, such as warehouse robots, which are considered a type of vehicle in this application.

[0025] In one example, the input to the second function of the classifier model includes a series of velocity-acceleration pairs sampled from a distribution resulting from a real or artificial data sequence. In another example, the operating profile may include an integer number of vehicle stops per unit time, duration of vehicle acceleration to a given speed, duration of vehicle deceleration to a given speed, duration of vehicle deceleration to zero speed, vehicle speed relative to path curvature, vehicle speed relative to path incline, vehicle speed relative to total driving duration, an integer number of gear changes per unit time, gear change function to achieve a predetermined speed, gear change function to achieve zero speed, gear change function relative to path curvature, gear change function relative to path incline, and gear change function relative to total driving duration.

[0026] As a result, in one example, a second function of the classifier can be evaluated to provide additional information about the samples that are influenced by the distribution defined by the expert in the generative model. [Brief explanation of the drawing]

[0027] [Figure 1A] This figure shows an approach for training a conditional adversarial generative network on route / speed data samples. [Figure 1B] This figure shows an approach to generating route / speed data samples using a generative network trained for machine learning. [Figure 2] This figure shows the discriminator of the conditional generative adversarial network in Figure 1B in more detail. [Figure 3A] This figure shows a computer-implemented method for training a generative model for machine learning that models the operating profile of a vehicle according to the first embodiment. [Figure 3B] This figure shows a computer-implemented method for generating artificial data samples representing an artificial operating profile of a vehicle using a generative model for machine learning according to a second aspect. [Figure 4A] This is a diagram illustrating an example of a system structure for machine learning. [Figure 4B] This diagram shows an example structure of a data generator. [Figure 5] This figure shows an example structure of a classifier in a machine learning system. [Figure 6] This figure shows a detailed embodiment of a system classifier for machine learning. [Figure 7] This is a diagram showing an apparatus according to the third embodiment. [Figure 8] This diagram schematically shows a vehicle according to the seventh embodiment, the vehicle being equipped with a control device, which is configured to apply a model trained according to the first embodiment. [Modes for carrying out the invention]

[0028] Detailed explanation Here, an improvement to the generative model is presented. The generative model is a neural network architecture that can generate data samples that represent a set of data samples previously used to train the generative model. In one example, the generative model may be a generator of a generative adversarial network (GAN) or a conditionally generative adversarial network (cGAN).

[0029] Generative models for generating artificial speed curves can (for example) generate samples, which may appear realistic, but the generated trajectories may not always match the statistics of the true data. For instance, when speed trajectories are generated, external factors such as the presence of traffic lights along the route, the behavior of other drivers at specific intersections, the time of day, and the driver's mood can have an effect that conventional speed curve generation does not consider. Such external influences form application-specific distortions that affect the statistics of the generated speed trajectories.

[0030] As an example, sampling from a distribution representing a velocity-acceleration histogram may be used by experts to evaluate the fidelity of a vehicle velocity trajectory generated by a generative network. For vehicle driving representing a real-world scenario, this is expected to follow a given distribution of velocity-acceleration pairs. For real-world driving, this is also expected to represent a specific distribution of several other features, such as the number of segments of driving represented in the velocity trajectory (where segments are identifiable, for example, by the vehicle's stagnant time at zero velocity during the preceding period) or the time spent at zero velocity.

[0031] According to aspects of this patent specification, it is proposed to improve the classifier of a generative adversarial network (GAN) or a conditionally generative adversarial network (cGAN) with additional functionality that improves the quality of statistics resulting from the classification process. In particular, it is proposed to provide at least one additional functionality block in application-specific classifier statistics for the discriminant function. In other words, the generator of a GAN or cGAN database can be enhanced with domain-specific expertise.

[0032] In other words, a modified architecture of GAN or cGAN is proposed that generates trajectories (e.g., velocity trajectories) of data samples that match or are very similar to the distribution of expert features.

[0033] Furthermore, to generate more accurate output data, or as an evaluation metric for assessing the accuracy of data generated by the generative model, the provision of a classifier with a function that applies precise statistics can be applied during the training procedure.

[0034] Figure 1A shows an example of training a conditional generative adversarial network on a sample of path / velocity data. However, the techniques described here are also applicable to generative adversarial networks (GANs) without any conditioning, such as path-based conditioning.

[0035] During the training phase, sample c {1:Ti} and x {1:Ti} This is obtained from the true data distribution. In this case, the environment variable c {1:Ti} This is a matrix that describes the true path, and the time column x {1:Ti} is the vector of velocity observations along this path. Ti is the length of the sequence.

[0036] A pair of true data samples 14, including a path sample 14a and a velocity sample 14b associated with this path sample 14a, is randomly selected from the true data distribution and provided to the classifier 12, thereby forming a sample of true time-sequence data. In the example of training a GAN (unlike cGAN), only a single data sample is needed here instead of a pair of data samples.

[0037] Route sample 14a is a set of environmental points defined in an exemplary route. Speed ​​sample 14b corresponds to each route sample 14a. Speed ​​sample 14b is a set of speed points that define the vehicle speed at points along route sample 14a.

[0038] The generative model 10a from preceding iterations is provided with random samples of a predetermined distribution, such as a Gaussian distribution or a uniform distribution. A vector or a sequence of vectors can be detected when sampling from this distribution. The generative model 10a is further provided with conditional path samples, which in one example are time sequences.

[0039] The generative model 10a is configured to produce one or more outputs, in this case one or more generated velocities 15. The generated velocities 15 and path samples 16b used to condition the generative model 10a form a sample of artificial time-series data 17b.

[0040] The classifier model 12 is trained to distinguish whether the samples in this input represent experimentally observed data (real-world data) 17a or artificial time-sequence data (fake data) 17b from the generative model 10a. Typically, the training process integrates the generative model 10a and the classifier 12 in relation to each other, thereby improving the generative model 10a when the classifier model 12 fails (in other words, the classifier model 12 becomes more likely to distinguish experimentally observed data 17a as artificial data 17b and vice versa).

[0041] In other words, the training process of the cGAN involves optimizing the loss function for the generative model 10a and the loss function for the discriminator model 12. Specifically, this training process is a minimax process that involves minimizing the loss function of the discriminator model 12 and the error loss function of the generative model 10a (which results in maximizing the discriminator's loss on artificial data). The result is a final generative model 10b that can generate an artificial time-series curve.

[0042] Figure 1B illustrates an approach for machine learning to generate route / speed data samples using a trained generative network. A given route may be represented by a set of information, such as gradient, speed limit, traffic density profile, road surface, number of traffic lights, or weather data.

[0043] Route 20 is extracted from the route database 19a and used to condition the trained generation model 10b of the cGAN. Further, samples from the cGAN are extracted from a predefined distribution. Based on the sampling of route 20 and the predefined distribution, the trained generation model 10b provides one or more elements of artificial data, such as an artificial data series of speeds at a specific time 22, for further use. In one example, this further use can be an input of the generated data into system simulation software and analysis software 23.

[0044] In this patent specification, the symbols listed below have the following meanings.

[0045]

Table 1

[0046] The definitions of the names listed above for explaining specific examples in this disclosure are not limitations on the entire technical concept. In particular, a multivariate time series may well be generated with x t accompanied by R D where D is the number of variables to be generated. At the places in the figures where the symbols "x", "c", "z" appear, they indicate the abbreviations of "x 1:T ", "c 1:T ", "z 1:T ".

[0047] The subscript T is used to index the columns listed above. The examples shown here are represented by time series, but this technology is not limited to this. This column is based on, for example, distance steps or gearbox settings.

[0048] Column x 1:T may be univariate or multivariate. In one example, this column is a column based on time series or distance.

[0049] Each element x in multiple columns 1:Ti , c 1:Ti These may have different lengths. In the example in Figure 1A, the actual data is sent to the cGAN target.

number

number

number

number

number

[0050] term

number

[0051] In one example, the Wasserstein distance in a GAN configuration.

number

number

number

number

[0052] Conditional GANs (cGANs) are used for route information and vehicle behavior profiles. Common distribution

number

number

number

number

number

number

[0053] Figure 2 shows the generator and discriminator of the conditional generative adversarial network (cGAN) in Figure 1B in more detail. Identical features are given the same reference numeral. Row 24 is a simplified diagram, and row 25 is a detailed diagram of an exemplary model architecture (network architecture) of the cGAN. In one example, the generative model 10b is performed by a recurrent neural network (LSTM10c). A particular example uses long-short-term memory units to perform the generative model 10b, but other variations can be used. The input to the LSTM is extracted from path 16b.

[0054] In Figure 2, two sets of environmental condition time sequences are extracted as features from path 16b. The first environmental condition time sequence defines the gradient along the path. The second environmental condition time sequence defines the speed limit along the path. Those skilled in the art will understand that an environmental time sequence can be used more or less as input to LSTM10c. The noise sample 16a(z) at each time step is P Z These are extracted and input to LSTM10c. The generative model, including LSTM10c, generates a corresponding artificial velocity time series from these inputs.

[0055] The classifier 12 is subjected to a first environmental condition time sequence that defines the gradient along the path, a second environmental condition time sequence that defines the speed limit along the path, and a generated artificial speed time sequence. In the illustrated example, classifier 12a is also a recurrent neural network such as long-term memory.

[0056] Furthermore, the cGAN shown in Figure 2 can be improved. In particular, the statistical distribution of real-world speed time sequences may be distorted by real-world driving conditions, which can be caused by, for example, traffic lights, weather conditions, driver mood, and the location or time of the speed camera.

[0057] According to a first aspect, a computer-implemented method 30 is provided for training a generative model for machine learning to model the motion profile of a vehicle or robot, which includes adversarial training of a generator model and a discriminator model. This includes retrieving multiple data columns from the data space, where each data column describes at least one motion profile of a vehicle or robot. • During at least one generative phase of training a generative model for machine learning, the process includes generating multiple artificial data sequences by sampling from a distribution. During training a generative model for machine learning, the classifier model is input into (i) a data column from multiple acquired data columns, or (ii) an artificial data column from multiple artificial data columns, 34. During training a generative model for machine learning, the process includes classifying the input to the classifier model as a data sequence from the data space, or as an artificial data sequence using at least a first function and a second function of the classifier model, where the multiple inputs to the first function 10 include multiple consecutive samples of the artificial data sequence, and at least one input to the second function 26 includes at least two data samples of the artificial data sequence obtained from multiple different steps of the artificial data sequence. The method includes iteratively training generator models and discriminator models for machine learning in order to generate trained models for machine learning, including trained generator models.36

[0058] Figure 3A shows a computer-implemented method for training a generative model for machine learning that models the operating profile of a vehicle according to the first embodiment.

[0059] According to one embodiment, the distribution is sampled in at least one generation phase in order to provide multiple artificial data sequences.

[0060] Figure 4A shows an exemplary structure of system 45 for machine learning to train a generative model during the training phase.

[0061] System 45 for machine learning is an example of a Generative Adversarial Network (GAN). Exemplary data, such as routes and associated speed profiles, is obtained from a database containing true data samples 14. A generator 47 is then trained within System 45 for machine learning to generate exemplary output data sequences, such as speed profiles and / or route profiles. The speed profile is determined as a function of a random input variable 16a. Multiple sampled data sequences in data space X, such as speeds extracted from multiple vehicle routes, form multiple sampled data sequences in data space X, but many other variables can be used individually or in combination.

[0062] A distribution, such as a Gaussian distribution, is sampled to obtain a set of multiple artificial data sequences. The distribution parameters are selected, for example, so that the characteristics of the output distribution represent the intended artificial data, such as multiple speed profiles or gearbox settings.

[0063] The noise z at each time step, sampled from this distribution. 1:T The following is input to the generator 47, and the generator 47 can optionally be provided with other variables in the conditional GAN. The generator 47 aims to generate an artificial data sequence that is indistinguishable from the true data sample 14. The iterative optimization of the generator 47 is performed by X and

number

[0064] As a result, the generator 47 is trained to generate artificial data sequences. The generator 47 can be used, for example, as the generative models 10a and 10b shown in Figure 1A.

[0065] As a result, the method of the first embodiment involves step 32 of obtaining multiple data columns X in the data space and at least one artificial data column based on multiple samples in a predetermined distribution.

number

[0066] In one example, step 34, which involves inputting (i) a data sequence from multiple sampled data sequences, or (ii) an artificial data sequence from at least one artificial data sequence, is performed by selecting one or more artificial data sequences, for example, from the output of the generator 47, or by selecting a data sequence from true input data, such as the true input sample 14.

[0067] The GAN classifiers 12 and 12a operate by extracting several latent features (which are not necessarily interpretable), and using these features to determine whether the input sample is true or artificial (generated by generators 10a and 47). Generator 12a, for example, takes a real number f as a score. rnn (x 1:T ) generates.

[0068] A positive score means, for example, that classifiers 12 and 12a have determined this sample to be true. A negative score means, for example, that the classifier has determined this sample to be artificial (fake). Feature extraction can be performed by a neural network.

[0069] Selectively, the score may be a binary value, in which case, for example, "1" indicates a "true" sample and "0" indicates a "false" sample.

[0070] This specification proposes a novel classifier concept aimed at classifying inputs to a classifier model as either true or artificial (false) during the classification phase of a machine learning model.35

[0071] The basic idea is to give the GAN classifier a separate function module f exp The addition is 26. This separate module is configured to extract additional information, such as expert features, and use these as components of the classifier score. This means that classifiers 12, 12a train generator 47 to provide an artificial data sequence with a probability density function that approaches the true data sequence.

[0072] Figure 5 shows an exemplary structure of a classifier 12 for a machine learning system according to the first embodiment. The classifier includes a first function 12a and a second function 26. Multiple inputs to the first function are true data sequences x 1:T or artificial data sequence

number

number

number

number

number

number

[0073] In one example, the input to the second function 26 of the discriminator 12 is obtained from the output of the generator 10, where the first sample from the output of the generator 10 is combined with the adjacent second sample from the output of the generator 10.

[0074] In one example, the input to the second function 26 of the classifier 12 is obtained only from the output of the generator 10. In one example, the input to the second function 26 is the environmental condition c over time. 1:T It does not include examples of (e.g., gradients or speed limits).

[0075] In one example, an artificial data sequence

number

number

number

number

[0076] In one example, the first function 12a of the classifier takes at least one artificial data sequence as its input.

number

[0077] In one example, the first function 12a of the classifier takes an artificial data sequence representing velocity as its input.

number

number

[0078] In one example, the elements of the data sequence and the elements of the artificial data sequence each represent time sequences (arranged in time). In another example, the elements of the data sequence and the elements of the artificial data sequence each represent distance-based sequences (arranged relative to a fixed reference point on a map).

[0079] In one example, the second function does not obtain environmental conditions on its own input side.

[0080] In one example, the second function is an artificial data sequence.

number

[0081] According to one embodiment, the input to the first function is a time-indexed function of the input to the classifier model, and the input to the second function is a differentiable function of the input to the classifier model.

[0082] According to one embodiment, the first aspect further includes combining the output of a first function and the output of a second function using a discriminant neural network.

[0083] The first function is a recurrent neural network, which is optionally a long-term memory network.

[0084] The second function is obtained by projecting at least one artificial data sequence, and further by obtaining the result of the second function of the classifier model based on the combination of the projected at least one artificial data sequence and at least one data sequence from a plurality of data sequences.

[0085] The classifier neural network computes a linear combination of the first function and the second function of the classifier model.

[0086] In one example, a data column included in multiple data columns is either a time column or a distance column. Among multiple artificial data columns, the artificial data column is either an artificial time data column or an artificial distance data column.

[0087] For example, a set of values ​​sampled from a distribution resulting from a true or artificial data sequence can be used as additional information or expert features.

[0088] For example, a series of velocity-acceleration pairs sampled from a distribution resulting from a true data sequence or an artificial data sequence (optionally a velocity-acceleration histogram) can be used as additional information or expert features.

[0089] In cases where expert features are provided by samples extracted from a distribution representing a velocity-acceleration histogram, the second function 26 of the classifier has the following form: that is, the data sequence is x t This includes an element that represents the speed of the target vehicle at time step t.

number

number

number

number

number

[0090] In one example, the input to the second function 26 of the discriminator 12 is obtained from the output of the generator 10, where a sample from the output of the generator 10 is combined with a subsequent second sample from the output of the generator 10.

number

[0091] In one example, each input to the second function 26 of the discriminator 12 is taken from the output of the generator 10, where each sample from the output of the generator 10 is combined with each second sample from the output of the generator 10.

[0092] In one example, each input to the second function 26 of the discriminator 12 is taken from the output of the generator 10, where each sample of the output of the generator 10 is a convolution of the generator output.

[0093] In one example, at least one input to the second function 26 of the classifier 12 is an artificial data sequence

number

number

[0094] In one example, the second function 26 of the classifier applies an expert function (e(.)) to at least one of its inputs. Based on the output of the generator 10, the expert function (e(.)) can evaluate one or more of the following: an integer number of vehicle stops per unit time, the duration of vehicle acceleration to a predetermined speed, the duration of vehicle deceleration to a predetermined speed, the duration of vehicle deceleration to zero speed, vehicle speed relative to path curvature, vehicle speed relative to path incline, vehicle speed relative to total driving duration, an integer number of gear changes per unit time, gear change function to achieve a predetermined speed, gear change function to achieve zero speed, gear change function for path curvature, gear change function for path incline, and gear change function for total driving duration.

[0095] The effect of the example described above is that, as the entire model is trained, the classifier 12 learns to distinguish between false velocity-acceleration pair distributions and true velocity-acceleration pair distributions. Implicitly, the classifier constructs data densities of true and false samples to determine which of the two is more likely for a given input.

[0096] In one embodiment, a second function of the classifier model is configured to sample from a distribution resulting from the true data sequence or the artificial data sequence, based on at least one data sequence representing vehicle speed and at least one artificial data sequence representing artificial vehicle speed.

[0097] In one embodiment, a second function of the classifier model is configured to evaluate the time spent at zero speed based on at least one artificial data sequence representing vehicle speed.

[0098] In one embodiment, the multiple data sequences describe multiple operating profiles of a vehicle, including a series of speeds, engine temperatures, engine speeds, pedal positions, steering angles, or at least one of the gear change functions of a vehicle traveling along a route, and the input to a second function of the classifier model includes multiple speed-acceleration pairs sampled from a series of values ​​sampled from a distribution resulting from a true data sequence or an artificial data sequence.

[0099] In one embodiment, the multiple data sequences describe multiple motion profiles of an autonomous or semi-autonomous robot, each including at least one of a series of displacements, a series of velocities, or a series of positions of the actuators of the autonomous or semi-autonomous robot.

[0100] According to one embodiment, the second function is obtained by projecting at least one artificial data sequence, and further by obtaining the result of the second function of the classifier model based on a combination of the projected at least one artificial data sequence and at least one data sequence from a plurality of data sequences.

[0101] According to one embodiment, a second function of the classifier model is configured to contribute to the classifier score by projecting one or more samples from a plurality of artificial data sequences.

[0102] In one embodiment, D=1. In this case, f proj The output can be considered as the identification output for an expert feature in a single time step. For example, if D=1, the identification output for a velocity-acceleration pair in a single time step is obtained. In one embodiment, D may be greater than 1 for a richer representation (to allow for mixing of expert features across multiple time steps).

[0103] In one example, f proj The projection function can be implemented as a series of dense layers in a multilayer perceptron.

[0104] In one example, f proj The projection function may be manually created to project the input onto the bin of the second function. In one example, f proj The projection function may be manually constructed to project the input onto the expert feature bins of the second function. In one example, f proj The projection function may be manually constructed to project the input onto bins of a distribution representing a set of values ​​sampled from a true or artificial data sequence, such as a velocity-acceleration histogram.

[0105] Many vehicle paths begin with a velocity equal to zero. Therefore, in one example, the velocity in the first time step may be assumed to be zero, and acceleration may be calculated in all subsequent time steps.

[0106] This projection function f proj The output may be optionally averaged in the mean-forming unit 26b, thereby determining the final classification f of the second expert function 26 (expert unit) of the classifier 12. exp (x 1:T ) can be obtained.

[0107] Figure 6 shows an example of a detailed embodiment of a classifier 12 for a machine learning system used in a cGAN. A generator 10c, shown in part 28 of Figure 6, is provided as a reference. The classifier 12 is shown in part 29 of Figure 6. The classifier 12 includes at least a first function 12a and a second function 26.

[0108] In one example, classifier 12 is applied to GAN instead of cGAN. In this case, the environment variables (c1, c) shown in Figure 6 are used. t ,c T ) is unnecessary. This case is not shown in Figure 6.

[0109] Therefore, the generator 10a generates multiple samples from an artificial data sequence, for example, from a series of velocities. The first function 12a of the discriminator 12 is referenced with respect to time. In the specific case where the second function 26 calculates acceleration from a series of velocities, the acceleration is calculated from pairs of adjacent time steps, as shown in Figure 6, and the input to the second function 26 is extracted in a predetermined order. Acceleration

number

number

[0110] In other cases where the second function is not based on acceleration, the second function 26 of the classifier may be independent of the time structure. In the illustrated example, the second function 26 calculates acceleration from velocity and f exp (x 1:T ) is calculated. Those skilled in the art will understand that, by the second function 26, many other functions can be calculated with or without relation to the time structure of a data sequence or an artificial data sequence. The purpose of the second function 26 is to make multiple artificial data sequences more realistic by ensuring that the identification device is given expert information regarding specific vehicle or robot characteristics.

[0111] According to one embodiment, a trained generative model for machine learning is configured to generate a plurality of artificial data sequences having a statistical distribution more similar to the statistical distribution of a plurality of artificial data sequences compared to the statistical distribution of a plurality of artificial data sequences generated without using a second function in the classifier model.

[0112] The outputs of the first function 12a and the second function 26 are combined in the joining block 27. This is optionally performed after the application of projections Φ(.) and Ψ(.). Optionally, projections Φ(.) and Ψ(.) can be performed by a multilayer perceptron MLP. Optionally, the joining block 27 may be a sigmoid function, a linear combination, or another form of neural network.

[0113] According to one embodiment, the output of the classifier model is:

number

[0114] The above example considers the use of a second function 26 in a classifier 12 configured to sample from a distribution representing a velocity-acceleration histogram.

[0115] A second example is discussed in relation to vehicle speed generation for exhaust simulation or load collection analysis.

[0116] In the second example, vehicle speed is generated for the purpose of emission or changes in load on various vehicle components. The data thus generated can be used to train a model that can be used while the passenger car is in motion. In practice, such a model would be trained to characterize fuel consumption, exhaust characteristics, energy demand, or wear on vehicle components such as pumps or brakes in real time.

[0117] In this example, the target variable is a time series of velocity. In this example, the second function 26 of the classifier 12 is a series of values ​​sampled from a distribution, which results from a true or artificial data series of time spent at zero velocity relative to the total travel time. This is advantageous because a vehicle run that includes multiple episodes spent at zero velocity will show higher emissions or cause more parts wear compared to a vehicle run that remains at a substantially constant speed from start to finish.

[0118] An exemplary execution of the second function 26 of the classifier 12 is as follows in this case:

number

[0119] The terms in this sum map zero velocity to 1 and each non-zero velocity to approximately zero (assuming only non-negative velocities occur). Temperature takes a baseline value of 1, but can be increased to obtain a steeper approximation of the step function. In such cases, temperature is a term specific to the sigmoid function and is not related to the vehicle's engine temperature. The function ρ(.) maps the input (time spent at zero velocity and total travel time) to the feature space for discrimination. In one example, ρ(.) may be a multilayer perceptron.

[0120] Accordingly, two detailed examples of constructing the second function 26 of the classifier are presented to accommodate different applications. However, those skilled in the art will understand that the second function 26 of the classifier can be constructed to address many cases of statistical skew relevant to the field of application.

[0121] According to one embodiment, a second function of the classifier model is configured to sample a set of values ​​sampled from a distribution. This distribution results from a true data sequence or an artificial data sequence, for example, at least one artificial data sequence representing vehicle speeds to provide a speed-acceleration histogram based thereon.

[0122] According to one embodiment, a second function of the classifier model is configured to evaluate the time spent at zero speed based on at least one artificial data sequence representing vehicle speed.

[0123] According to one embodiment, the multiple data sequences describe multiple operating profiles of a vehicle, including a series of speeds, engine temperatures, engine speeds, pedal positions, steering angles, or at least one of the gear change functions of a vehicle traveling along a route, and the input to a second function of the classifier model is sampled from a series of values ​​resulting from a true or artificial data sequence having a distribution representing a velocity-acceleration function.

[0124] The example in Figure 6 considers the application to be a vehicle speed trajectory, and the second function 26 (expert feature) should be sampled from a set of values ​​resulting from true or artificial data that have a distribution representing a speed-acceleration histogram, but this setting can be extended to many other features such as other environment variables c.

[0125] In one example, the second function 26 of the classifier 12 (if present) is f exp (x 1:T ) should be coded in a differentiable manner.

[0126] Further functions of the classifier, which add more expert features to the classifier, may similarly be added to the second function 26. For example, classifier 12 may include a third function, a fourth function, a fifth function, and a sixth function, in which case the combination block 27 that executes the function r(.) must be modified to combine any third, fourth, fifth, and sixth functions. The second function 26 of classifier 12 (if present) is f exp (x 1:T ) should be coded in a differentiable manner.

[0127] In one example, the input to the second function 26 of the classifier 12 is differentiable.

[0128] According to one embodiment, the input to the second function of the classifier model defines one or more of the following: (i) an integer number of vehicle stops per unit time, (ii) duration of vehicle acceleration to a predetermined speed, (iii) duration of vehicle deceleration to a predetermined speed, (iv) duration of vehicle deceleration to zero speed, (v) vehicle speed relative to path curvature, (vi) vehicle speed relative to path incline, (vii) vehicle speed relative to total driving duration, (viii) an integer number of gear changes per unit time, (ix) gear change function to achieve a predetermined speed, (x) gear change function to achieve zero speed, (xi) gear change function for path curvature, (xii) gear change function for path incline, and (xiii) gear change function for total driving duration.

[0129] A variation of the second function 26 can be used as an evaluation metric for optimization iterations. In addition to using expert features to train a GAN (cGAN), the second function 26 can be used to evaluate the quality of the output generated by the model for machine learning. Non-differentiable expert features or non-differentiable functions, such as a random forest classifier, can also be used in such evaluation steps.

[0130] In this way, these features also function as stopping criteria for the trained model. While the stopping criteria can be used to assess the quality of the model, they can also guide the training process, allowing for early termination of the training procedure if the stopping criteria fall below a desired threshold. Without expert feature-based stopping criteria, such a procedure would be impossible, as the GAN's loss function would often fluctuate.

[0131] According to one embodiment, the step of stopping the iterative training of a generator model is provided using a stopping criterion, where the stopping criterion is obtained by evaluating at least one artificial data sequence generated by the iterations of the generator model using a second function of a discriminator model, where the second function outputs the stopping criterion.

[0132] As a first example of using the second function 26 of the classifier 12 to stop the training process, the case of generating vehicle speed for exhaust simulation or load collection analysis (parts wear) is considered.

[0133] Vehicle speed is generated to model exhaust or loads on various vehicle components. The target variable is a time series of speed. The second function 26 of classifier 12 is configured to sample from a distribution representing a speed-acceleration histogram for the vehicle or scenario under consideration. Evaluation metrics are obtained by obtaining sets of true time series samples and artificial time series samples. A classification model (such as a random forest) is trained to predict the class of speed trajectories (true vs. artificial).

[0134] In a well-trained model, the accuracy of classification between two options should be nearly random, i.e., 50%.

[0135] This tracking of accuracy during the training of the discriminator and generator also provides a good indicator of the performance of the generator 10.

[0136] Other examples of using the second function 26 of the classifier 12 as an evaluation index include similar cases of generating vehicle speeds for exhaust simulation or load collection analysis (parts wear).

[0137] As a first example of using the second function 26 of the classifier 12 to stop the training process, consider the case of generating vehicle speed for exhaust simulation or load collection analysis (parts wear).

[0138] Vehicle speed is generated to model exhaust or loads on various vehicle components. The target variable is a time sequence of speed. The second function 26 of the classifier 12 is configured to obtain the time spent at zero speed for the vehicle or scenario in question.

[0139] The evaluation metrics are the mean and standard deviations of the time spent at zero velocity for the true and artificial velocity data sets. In a well-trained model, these metrics should be equivalent for both the true and artificial velocity data sets. In addition to the mean and standard deviations, other metrics (or measurements) can be used for comparing probability distributions (e.g., KL divergence).

[0140] It is not essential that the model for machine learning, trained according to the method of the first embodiment, be a GAN or a cGAN. Any model for adversarial machine learning having a generator can be applied. An adversarial autoencoder including a discriminator including a second function 26 may be provided. Thus, in one embodiment, the generative model for machine learning according to the first embodiment is one of a GAN, a conditional GAN, or an adversarial autoencoder.

[0141] According to a second aspect, a computer-implemented method is provided for generating artificial data samples that represent an artificial motion profile of a vehicle or robot using a generative model for machine learning. This method is • Constructing a generative model for machine learning according to a plurality of model parameters obtained according to the method of the first embodiment, • Obtaining multiple samples by sampling from a distribution configured to provide a pseudo-random vector, • To obtain multiple artificial data samples within the data space, decode multiple samples of this distribution, • Outputting multiple additional artificial data samples that represent the artificial motion profile of a vehicle or robot, Includes.

[0142] Figure 3B shows a computer-implemented method for generating artificial data samples representing an artificial behavioral profile of a vehicle using a generative model for machine learning, according to a second embodiment.

[0143] Figure 4B shows an exemplary structure of a data generator 49 that can be used while generating an artificial data sequence.

[0144] The sample generator 48 generates noise (z) from a predetermined probability distribution. 1:T ) is sampled. The probability density can be determined, for example, using a density estimator. This is obtained by the noise (z) obtained in one or more time steps, for example, as the density estimator was trained in Figure 4A. 1:T This is done by obtaining the noise (z) as a function of the obtained noise (z) in one or more time steps. The trained generator 47 (trained adversarially against a discriminator having a second function 26 that performs expert inputs, as described below) randomly takes samples and obtains the noise (z) in one or more time steps. 1:T The generator 47 extracts the noise (z) obtained from the parameters of the generator 47 acquired during training and the noise obtained at one or more time steps. 1:T Based on the sample of ), an artificial data column

number

[0145] The generator 47 is pre-trained, for example, according to the first embodiment. This involves applying a second function 26 to adversarial training in the classifier. The second function 26 influences the distribution that the classifier tests, based on specific technical requirements, in one example, based on a velocity-acceleration histogram. Thus, the artificial data sequence generated according to the second embodiment more faithfully reproduces real-world data.

[0146] For example, the profile generated according to the second embodiment is used to simulate exhaust emissions and for probabilistic evaluation for compliance with exhaust standards.

[0147] For example, the generated speed profile is used to optimize the software application.

[0148] For example, the generated speed profile is used to simulate the vehicle's exhaust. In this example, the generated speed profile and / or simulated vehicle exhaust may be applied to a software representation (also called a digital twin) of a vehicle component (such as an internal combustion engine) or the entire vehicle. The software representation may be configured to optimize various parts of the vehicle or vehicle component according to the input speed profile. The size, layout, and component configuration of the internal combustion engine or gearbox parts can be modified using the software representation of the vehicle or vehicle component based on the generated speed profile, which has improved accuracy for the real use case. Thus, the application of a machine learning model trained according to the first aspect results in a concrete conceptual change of the technical components included in the vehicle or robot.

[0149] For example, a classifier 12 including a second function 26 can be applied to evaluate a potential speed profile in order to determine whether a given route is related to actual driving behavior.

[0150] For example, a data generation method according to the second embodiment is used to determine the load (wear) on individual components of a vehicle, such as a pump, during the vehicle's service life.

[0151] For example, the data generation method according to the second embodiment is applicable when the vehicle is in motion. The engine control unit can generate a speed profile according to the second embodiment. Selectively, the engine control unit can retrieve a pre-generated speed profile calculated according to the second embodiment from memory. The engine control unit can control engine commands in relation to the generated speed profile.

[0152] For example, the data generation method of the second embodiment can acquire a speed profile or an acceleration profile. The speed profile or acceleration profile may be used to generate a trajectory that defines the available battery capacity of an electric vehicle, the driving management of a hybrid vehicle, the regenerative management of a regenerative braking system, or to improve the battery management of an electric bicycle.

[0153] According to a third aspect, a computer-implemented method is provided for training a further model for machine learning to model the operating profile of a vehicle, the method being: • Obtaining a further set of artificial data representing an artificial operating profile of a vehicle generated according to the computer-implemented method of the second aspect, • Inputting multiple additional artificial data sequences representing the vehicle's artificial operating profile into further models for machine learning, • Iteratively training further models for machine learning, • Outputting multiple additional model parameters for further models of machine learning to model the vehicle's operating profile, Includes.

[0154] The data generated by the computer-implemented method for generating artificial samples according to the second embodiment is more realistic when training a GAN or cGAN according to the method of the first embodiment, due to the effect of at least the second function 26 of the discriminator 12. Accordingly, the artificial data generated according to the method of the second embodiment is used in the third embodiment to train further models for machine learning.

[0155] It is not necessary that the further models for machine learning, trained according to the third aspect, be GANs or cGANs. Each model for machine learning that requires data for training can use artificial data generated according to the method of the second aspect. One effect is that the further models for machine learning are trained more accurately because the artificial data used for their own training is generated according to the adversarial model. This adversarial model has a classifier 12 which has a second function 26 that applies expert data to the classifier.

[0156] According to a fourth aspect, an apparatus is provided for training a generative model for machine learning that models the motion profile of a vehicle or robot, including adversarial training of a generator model and a discriminator model stored in memory. The apparatus includes an input interface configured to retrieve a plurality of data sequences from a data space, where each data sequence describes at least one motion profile of a vehicle. The apparatus further includes a processor configured to generate at least one artificial data sequence by sampling from a distribution during at least one generative phase of training a generative model for machine learning.

[0157] The processor is configured to input the classifier model, during at least one classification phase when training a generative model for machine learning, either (i) a data sequence from multiple acquired data sequences, or (ii) an artificial data sequence from at least one artificial data sequence.

[0158] During at least one discrimination phase when training a generative model for machine learning, the processor is configured to classify the input to the classifier model either as a data sequence from the data space or as an artificial data sequence using at least a first function and a second function of the classifier model.

[0159] The multiple inputs of the first function include multiple consecutive samples of an artificial data sequence, and the at least one input of the second function includes at least two data samples of an artificial data sequence obtained from multiple different steps of the artificial data sequence.

[0160] The processor is configured to iteratively train generator and discriminator models for machine learning in order to produce a trained model for machine learning, which includes a trained generator model. The device further includes an output interface configured to output the trained model for machine learning.

[0161] In one example, device 50 is a personal computer, a server, a cloud-based server, or an embedded computer. It is not essential that processing takes place on a physical processor. For example, processing tasks may be divided across multiple processor cores on the same processor, or across multiple different processors. The processor may be a Hadoop® cluster, or it may be provided on a commercial cloud processing service. Some processing may be performed on non-traditional processing hardware, such as a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), one or more graphical processors, or an application-specific processor for machine learning.

[0162] The memory 58 of the device 50 stores a computer program that, when executed by the processor 54, causes the processor 54 to perform a function described by a computer-implemented method according to the first, second, or third aspect.

[0163] For example, the input interface 52 and / or output interface is one of the following: a USB interface, an Ethernet interface, a WLAN interface, or other suitable hardware that can enable the input and output of data samples from the device 50.

[0164] In one example, the device 50 further includes a volatile memory system 58 and / or a non-volatile memory system 58 configured to receive input comments as input data from an input interface 52.

[0165] In one example, the device 50 is an automotive embedded computer included in a vehicle, in which case the automotive embedded computer may be connected to sensors and actuators provided in the vehicle. The input interface 52 of the device 50 may be connected to one or more engine control unit ECUs that provide, for example, speed data, fuel consumption data, battery data, position data, etc. The output interface 56 of the device 50 may be connected to one or more of, for example, multiple brake actuators, throttle actuators, fuel mixture actuators or fuel-air mixture actuators, turbocharger control devices, battery management systems, vehicle lighting or entertainment systems, etc.

[0166] According to the fifth aspect, one or more computer program elements are provided, each computer program element comprising at least: (i) Computer-executable instructions for training a generative model for machine learning according to the method or embodiment of the first aspect, and / or (ii) Computer-executable instructions for generating an artificial data sequence in accordance with the third aspect, and / or (iii) Computer-executable instructions including model parameters for providing a generative model for machine learning to be trained according to the first aspect or an embodiment thereof. Includes.

[0167] The computer program elements of the fifth embodiment may include, for example, machine-readable instructions that can be stored in computer memory.

[0168] According to the sixth aspect, a computer-readable medium is provided which includes one or more computer program elements of the fifth aspect.

[0169] Figure 8 schematically shows a vehicle according to the seventh embodiment, the vehicle being equipped with a control device, which is configured to apply a model trained according to the first embodiment.

[0170] According to a seventh aspect, a vehicle 60 is provided, which includes a prime mover 61 configured to provide driving force to operate the vehicle 60, a plurality of sensors 62a, 62b, and a vehicle electronics unit 63 configured to control at least the prime mover 61 based on feedback from the sensors 62a, 62b. The vehicle electronics unit 63 includes a processor configured to instantiate and implement a trained model that responds to at least one operating profile, according to a second aspect.

[0171] The trained model is configured to receive input signals from sensors 62a and 62b and apply the input signals to the trained model. The processor is configured to compute one or more response signals using the trained model. The processor is configured to adapt the configuration of one or more prime movers 61 or other output actuators of the vehicle 60 in response to the response signals obtained through the trained model.

[0172] According to one embodiment, the prime mover 61 may be an internal combustion engine coupled with a fuel system and an exhaust system, or an electric motor or hybrid powertrain coupled with a battery.

[0173] The vehicle's operating profile may, advantageously, include one, some, or all of the following variables that can be detected by or derived from multiple sensors 62a, 62b of the vehicle: the position of the vehicle's accelerator pedal, the position of the vehicle's brake pedal, the position of the clutch of the vehicle's transmission, the gears of the transmission, the vehicle's speed, the vehicle's rolling resistance, the traction force of the internal combustion engine, the traction force of the vehicle's electric drive, the rotational speed of the components of the internal combustion engine, the intake air volume per unit time of the internal combustion engine, the pressure in the intake manifold of the internal combustion engine, the amount of high-pressure EGR (exhaust gas recirculation), the amount of low-pressure EGR, the timing of the intake valve closing, the timing of the exhaust valve opening, the position of the compression modification system of the internal combustion engine, the injection timing of the injection unit, the pressure in the high-pressure fuel accumulator (common rail), the coolant temperature of the internal combustion engine, or the temperature in the intake system of the internal combustion engine.

[0174] Further parameters applicable to the model, or parameters that may be considered during training, include vehicle mass, vehicle gear ratio, maximum driving force of the vehicle's drive system, maximum torque of the drive system, transmission type, fuel type, hybridization specifications, engine type specification, vehicle type specification, and characteristics of the driving route, such as GPS recording, ambient temperature, gradient parameters, and ambient pressure.

[0175] The trained model can, for example, be used during operation, by using the detected speed profile to simulate potential future exhaust, thereby allowing the vehicle electronics unit 63 to set the actuators of the prime mover 61 to ensure that control commands are provided to the prime mover 61 to ensure that the vehicle's exhaust never exceeds a predetermined limit. In another example, the trained model can be used during operation to predict wear load scenarios in components such as pumps or brake discs during operation.

[0176] The examples shown in the drawings and described above are intended to aid in understanding the concepts of this disclosure and are not intended to limit the scope of protection of this disclosure. This disclosure describes changes and modifications to the illustrated examples. Only favorable examples are presented, and all changes, modifications, and further applications within the scope of protection of this disclosure should be protected.

Claims

1. A computer-driven, computer-implemented method (30) for training a generative model for machine learning to model the motion profile of a vehicle or robot, including adversarial training of a generator model and a discriminator model, wherein the method is - Includes retrieving multiple data columns from the data space (32), where each data column describes at least one motion profile of a vehicle or robot. - During at least one generative phase of training the generative model for machine learning, the method includes generating multiple artificial data sequences by sampling from a distribution (33), - During at least one classification phase when training the generative model for machine learning, the classifier model is input (i) a data sequence from the acquired plurality of data sequences, or (ii) an artificial data sequence from the plurality of artificial data sequences (34), - During at least one classification phase of training the generative model for machine learning, classifying the input to the classifier model as a data sequence from the data space, or as an artificial data sequence using at least a first function and a second function of the classifier model (35), wherein the plurality of inputs to the first function (10) comprise a plurality of consecutive samples of the artificial data sequence, and at least one input to the second function (26) comprises at least two data samples of the artificial data sequence obtained from a plurality of different steps of the artificial data sequence. - Including the iterative training of the generator model and the discriminator model of the generative model for machine learning in order to generate a trained model for machine learning, which includes a trained generator model, The input to the first function is the time-integrated function of the input to the classifier model, and the input to the second function is a differentiable function of the input to the classifier model. Computerized method (30).

2. - Further includes combining the output of the first function and the output of the second function using a neural classifier network, The first function is a recurrent neural network, optionally a long-term memory network, and / or The second function is obtained by projecting at least one of the artificial data sequences, and further by obtaining the result of the second function of the classifier model based on the combination of the projected at least one artificial data sequence and at least one of the plurality of data sequences. The neural classifier network optionally computes a linear combination of the first function and the second function of the classifier model. The computer-implemented method (30) according to claim 1.

3. The data column included in the plurality of data columns is either a time column or a distance column, and the artificial data column among the plurality of artificial data columns is either an artificial time data column or an artificial distance data column. The computer-implemented method (30) according to claim 1 or 2.

4. In order to provide the aforementioned multiple artificial data sequences, the latent space is sampled in at least one generation phase. The computer-implemented method (30) according to any one of claims 1 to 3.

5. The method further includes stopping the iterative training of the generator model using at least one stopping criterion, the at least one stopping criterion being obtained by evaluating at least one artificial data sequence generated by the iteration of the generator model using the second function of the discriminator model, the second function outputting the stopping criterion. The computer-implemented method (30) according to any one of claims 1 to 4.

6. The second function of the classifier model is evaluated on a true data sequence or a sequence of values ​​sampled from a distribution resulting from an artificial data sequence, based on at least one data sequence representing vehicle speed and at least one artificial data sequence representing artificial vehicle speed, or The second function of the classifier model is configured to evaluate the time spent at zero speed based on at least one artificial data sequence representing vehicle speed, or The plurality of data sequences describe a plurality of vehicle operation profiles, each including at least one of a series of speeds, engine temperature, engine speed, pedal position, steering angle, or gear change function of the vehicle traveling along the route, and the input to the second function of the classifier model includes a plurality of speed-acceleration pairs sampled from a speed-acceleration histogram, or The aforementioned plurality of data sequences describe a plurality of motion profiles of the autonomous or semi-autonomous robot, each including at least one of a series of displacements, a series of velocities, or a series of positions of the actuators of the autonomous or semi-autonomous robot. The computer-implemented method (30) according to any one of claims 1 to 5.

7. The output of the aforementioned discriminator model is [Math 1] It is defined by, where f ｒｎｎ (x １：Ｔ ) is the first function, and f ｅｘｐ (x １：Ｔ ) is the second function, and Φ(.) and Ψ(.) are projections that are arbitrarily executed as multilayer perceptrons. The computer-implemented method (30) according to any one of claims 1 to 6.

8. The data sequence input to the second function of the classifier model defines one or more of the following: an integer number of vehicle stops per unit time, duration of vehicle acceleration to a predetermined speed, duration of vehicle deceleration to a predetermined speed, duration of vehicle deceleration to zero speed, vehicle speed relative to path curvature, vehicle speed relative to path incline, vehicle speed relative to total travel duration, an integer number of gear changes per unit time, gear change function to achieve a predetermined speed, gear change function to achieve zero speed, gear change function for path curvature, gear change function for path incline, gear change function relative to total travel duration, and the position of the robot's actuator. The computer-implemented method (30) according to any one of claims 1 to 7.

9. The generative model trained for machine learning is configured to generate a plurality of artificial data sequences having a statistical distribution more similar to the statistical distribution of the plurality of artificial data sequences compared to the statistical distribution of the plurality of artificial data sequences generated in the classifier model without using the second function. The computer-implemented method (30) according to any one of claims 1 to 8.

10. A computer-driven, computer-implemented method (30) for generating artificial data samples representing an artificial motion profile of a vehicle or robot using a generative model for machine learning, - Constructing a generative model for machine learning according to a plurality of model parameters obtained according to the method described in any one of claims 1 to 9, - Obtaining multiple samples by sampling from a distribution configured to provide a pseudo-random vector, - Decoding the multiple samples of the distribution in order to obtain further multiple artificial data samples within the data space, - Outputting the further multiple artificial data samples that represent the artificial motion profile of the vehicle or robot, Computerized methods (30), including the above.

11. An apparatus (50) for training a generative model for machine learning to model the motion profile of a vehicle or robot, including adversarial training of a generator model and a discriminator model stored in memory, wherein the apparatus - Includes an input interface (52) configured to retrieve multiple data columns from a data space, where each data column describes at least one motion profile of a vehicle or robot. The system includes a processor (54) configured to generate at least one artificial data sequence by sampling from a distribution during at least one generative phase of training the generative model for machine learning, wherein, The processor is configured to input to the classifier model (i) a data sequence from the acquired plurality of data sequences, or (ii) an artificial data sequence from the at least one artificial data sequence, during at least one classification phase of training the generative model for machine learning. The processor is configured to classify the input to the classifier model as a data sequence from the data space or as an artificial data sequence using at least a first function and a second function of the classifier model during the training of the generative model for machine learning, wherein the plurality of inputs to the first function include a plurality of consecutive samples of the artificial data sequence, and at least one input to the second function includes at least two data samples of the artificial data sequence obtained from a plurality of different steps of the artificial data sequence. The processor is configured to iteratively train the generator model and the discriminator model of the machine learning model in order to generate a trained model for machine learning, which includes a trained generator model. The aforementioned device is The system further includes an output interface (56) configured to output the trained model for machine learning, The input to the first function is the time-integrated function of the input to the classifier model, and the input to the second function is a differentiable function of the input to the classifier model. Apparatus (50).

12. One or more computer program elements, wherein each computer program element comprises at least: (i) Computer-executable instructions for training a generative model for machine learning according to any one of claims 1 to 9, in the use of machine learning, and / or (ii) A computer-executable instruction for generating an artificial data sequence according to claim 10, and / or (iii) A computer-executable instruction including model parameters for providing a generative model for machine learning to be trained according to any one of claims 1 to 9. One or more computer program elements, including [the specified element].

13. A computer-readable medium comprising one or more of the computer program elements described in claim 12.

14. A computer-operated, computer-implemented method (40) for training further models for machine learning to model the operating profile of a vehicle, - Obtaining a further number of artificial data sequences representing an artificial operating profile of a vehicle generated according to the computer-implemented method described in claim 10 (41), - Inputting the further multiple artificial data sequences representing the artificial operating profile of the vehicle into a further model for machine learning (42), - Iteratively training the aforementioned further models for machine learning (43), - Outputting a further number of model parameters for the further model for machine learning to model the vehicle's operating profile (44), Computerized methods including (40).

Images