Method and device for providing an explainable boosting machine (EBM) model for a control device

Abstract

The invention relates to a method (100) for providing, in particular in a memory-space-optimised manner, an explainable boosting machine (EBM) model for a control device, in particular for an embedded system, the method (100) having the steps of: - providing (S1) a trained EBM model; - extracting (S2) one-dimensional and / or two-dimensional lookup table raw data for the trained EBM model; - approximating (S3) the one-dimensional and / or two-dimensional lookup table raw data by means of at least one approximation algorithm for generating interpolated lookup table data by optimising a cost function, which is dependent in particular on the model performance and memory requirements of the EBM model; and - providing (S4) an approximated expandable boosting machine (EBM) model for a control device on the basis of the interpolated lookup table data.

Description

[0001] Description

[0002] title

[0003] Method and apparatus for providing an Explainable Boosting Machine (EBM) model for a control unit

[0004] The invention relates to a method and a device for providing, in particular in a memory-space-optimized, an Explainable Boosting Machine (EBM) model for a control unit, in particular for an embedded system.

[0005] State of the art

[0006] Computation using physical models represents a central approach in the prediction and analysis of physical systems. These models are based on equations of state that calculate the output taking into account specific assumptions and simplifications. Such simplifications are necessary to reduce the complexity of physical reality, but this often leads to a reduction in accuracy. This reduction arises from the generalizing approach, which cannot capture all the intricacies of the real system.

[0007] To improve the accuracy of predictions, data-driven approaches such as machine learning (ML) and artificial intelligence (AI) are becoming increasingly important. The goal is to optimize the predictive accuracy of physical models by applying these technologies. This is achieved by improving model performance and robustness.

[0008] A promising approach in this area is the Explainable Boosting Machine (EBM), which belongs to the class of supervised learning and can be used for both regression and classification. EBM is one of the interpretable machine learning methods and makes it possible to explain and visualize the relationships between the input and output variables of the model. It represents a further development of Generalized Additive Models (GAM) and offers performance comparable to conventional state-of-the-art machine learning methods.

[0009] EBM's functionality is based on an additive model, where each feature contributes to the predictions in a modular fashion. Training is performed using a tree-based, cyclic gradient boosting approach with automatic interaction detection. The results are presented in tabular form, with 2D graphics for each feature and 3D heatmaps for feature interactions. The final output of EBM is the sum of its tabular components.

[0010] However, implementing EBM presents challenges, particularly regarding storage requirements. The generated lookup tables consume a significant amount of memory, which can lead to extremely high storage demands when there are a large number of interactions. The goal is therefore to reduce storage requirements without significantly compromising the model's accuracy and interpretability, and to create the simplest and most flexible pipeline possible for different datasets.

[0011] It is an object of the invention to provide a method and / or a device improved in this respect.

[0012] The problem is solved by a method according to the features of claim 1. The problem is solved by a device according to the features of claim 10.

[0013] Disclosure of the invention

[0014] According to a first aspect, a method for providing an Explainable Boosting Machine (EBM) model for an electronic control unit (ECU), particularly for an embedded system, is proposed, with a focus on optimized memory space. The method comprises the following steps:

[0015] - Providing a trained EBM model; - Extracting one-dimensional and / or two-dimensional lookup table raw data for the trained EBM model;

[0016] - Approximating the one-dimensional and / or two-dimensional lookup table raw data by at least one approximation algorithm to generate interpolated lookup table data by optimizing a cost function that depends in particular on the model performance and memory requirements of the EBM model; and

[0017] - Providing an approximate Explainable Boosting Machine (EBM) model for an ECU based on interpolated lookup table data.

[0018] It is understood that the steps according to the invention, as well as further optional steps, do not necessarily have to be carried out in the sequence shown, but can also be carried out in a different sequence. Furthermore, additional intermediate steps may be provided. The individual steps may also comprise one or more sub-steps without thereby departing from the scope of the method according to the invention.

[0019] According to a second aspect, a device is proposed for providing an Explainable Boosting Machine (EBM) model for a control unit, particularly for an embedded system, in a particularly memory-space-optimized manner. The device includes an evaluation and computing unit configured to perform the following steps:

[0020] - Providing a trained EBM model;

[0021] - Extracting one-dimensional and / or two-dimensional lookup table raw data to the trained EBM model;

[0022] - Approximating the one-dimensional and / or two-dimensional lookup table raw data by at least one approximation algorithm to generate interpolated lookup table data by optimizing a cost function that depends in particular on the model performance and memory requirements of the EBM model; and

[0023] - Providing an approximate Explainable Boosting Machine (EBM) model for an ECU based on interpolated lookup table data.

[0024] The statements made regarding the procedure apply accordingly to the device. It is understood that linguistic modifications of procedurally formulated features can be reformulated for the device according to common linguistic practice, without such formulations needing to be explicitly listed here.

[0025] The present method is extended compared to a conventional EBM model with regard to its process chain, making it suitable for use on a memory-limited control unit, such as an embedded system. In a conventional EBM model, model training is performed based on input data, followed by model performance optimization. The performance-optimized EBM model is then ready for use / inference.

[0026] This conventional process chain cannot be used for embedded systems because they do not have sufficient storage space. Therefore, the conventional process chain is extended here by adding an optimization step for EMB model optimization specifically for embedded systems, as well as an implementation step for implementing the optimized model on an embedded system for later use.

[0027] The cost function or loss function is preferably optimized in such a way as to achieve a trade-off between maintaining sufficiently good model performance and the memory requirements or overhead of the approximated EBM model. This optimization is particularly relevant for embedded systems where only limited memory is available. The method can preferably be implemented using Python code, which can also be compiled to C code.

[0028] The loss function preferably comprises the following components: accuracy achieved by minimizing the root mean squared error (RMSE), and regularization to prevent overfitting and keep the EBM model as compact as possible. Regularization preferably incorporates both the number of models and the degree of the polynomial. The weighting of these components is preferably adjustable by the user via weights.

[0029] Extracting one-dimensional and / or two-dimensional lookup table raw data for the trained EBM model preferably involves extracting all lookup tables of the Explainable Boosting Machine (EBM) model, preferably performing a case distinction:

[0030] If no interactions exist, a table or lookup table is created for each input feature. This means that a separate lookup table exists for each individual feature, representing its contributions to the predictions. If interactions exist, in addition to the tables for the individual input features, a table for each interaction is also created. These additional tables capture the combined effects of the feature interactions and thus contribute to the overall predictions.

[0031] This ensures that both the individual and interactive influences of the features on the predictions are considered and transparently presented. Interactions in the EBM model refer to the interactions between different input variables (features). These interactions occur when the combined influence of two or more variables on the target variable (output) cannot be explained simply by the sum of their individual influences.

[0032] The Explainable Boosting Machine (EBM) is a machine learning model designed to be both powerful and explainable. EBM combines the advantages of boosting algorithms with high interpretability, allowing users to better understand the model's decision-making processes. EBM is based on an ensemble of weak learners, typically decision trees, which are trained iteratively. Each new tree attempts to correct the errors of the previous trees, resulting in a strong and accurate overall model.

[0033] In another aspect, it is proposed that the one-dimensional and / or two-dimensional lookup table raw data contain lookup tables for input features and / or interactions of the trained EBM model.

[0034] These lookup tables preferably contain the raw data used to describe the contributions of individual features or feature interactions to the model's predictions. One-dimensional lookup tables preferably represent the effects of individual input features, while two-dimensional lookup tables preferably represent the effects of interactions between different input features.

[0035] In a further aspect, it is proposed that the at least one approximation algorithm includes a, in particular, domain-wise regression and / or at least a piecewise linear and / or polynomial interpolation and / or an equation learner and / or local models for decomposing an interaction space represented by the two-dimensional lookup table raw data and / or a pointwise interpolation and / or a clusterwise interpolation and / or an axis-skewed division (HILOMOT) and / or an axis-orthogonal subdivision (LOLIMOT) and / or a Delaunay triangulation and / or a grid-based approach.

[0036] To approximate the EBM model, several regression models are preferably applied to different subsets of the lookup table raw data. The models are preferably adapted to the respective data types of the raw data, each of which can represent a specific segment or interval.

[0037] Individual sections of the lookup table raw data can be interpolated using linear or polynomial regression to reduce the number of data points or support locations in the relevant section of the lookup table. By dividing the lookup table raw data into regions that can each be interpolated by an approximation algorithm, the model's performance can be improved.

[0038] Furthermore, each data segment can preferably be optimally approximated.

[0039] The selection of the approximation algorithm is preferably carried out by variation and / or trial and / or random selection and variation, taking into account the optimization of the cost function. For each input variable and learned interaction of the EBM model, a number of local models and a degree of an interpolation polynomial can thus be determined.

[0040] In the section-wise linear interpolation of individual features, each individual input feature is preferably interpolated. The inputs for this are the EBM lookup table of the input feature and preferably a number of local models by which the input data can be subdivided. The output consists of support points that define the beginning and end of each data section.

[0041] In piecewise polynomial interpolation of individual features, each input feature of the EBM model is interpolated. The inputs for this are the EBM lookup table of the input feature, the number of local models, and the polynomial degree. The outputs generated are breakpoints or support points and the coefficients of the polynomials.

[0042] An equation learner is an algorithm that learns mathematical equations to approximate the underlying relationships in the data. The goal of the equation learner is to find an equation or formula that represents the relationship between the input data and the target variables in the most efficient way possible.

[0043] For input features, all lookup tables can be combined into a single equation, or a separate equation can be created for each individual input feature. This depends on the complexity of the input data. During interaction, the lookup tables can be interpreted either using the Equation Learner (EQL) or by linear interpolation.

[0044] Local models are models based on specific subsets of the data to represent particular local relationships. Pointwise interpolation is a method in which intermediate values ​​are determined by interpolating between individual data points. Clusterwise interpolation is a method in which the data is divided into clusters and interpolated within these clusters. Skewed division (HILOMOT) is an algorithm that divides space using planes oblique to the axis.

[0045] Axis-orthogonal subdivision (LOLI MOT) is an algorithm that subdivides space by planes orthogonal to the axes. Delaunay triangulation is an algorithm that subdivides space by triangles based on specific geometric criteria.

[0046] A grid-based approach is an algorithm that divides the space into a grid of points. All these algorithms can preferably be used individually or in combination, at least for data segments, to interpolate lookup tables. These features describe various methods and algorithms that can be used for approximating and decomposing data spaces.

[0047] In another aspect, it is proposed that the two-dimensional lookup table raw data be approximated, in particular by layer-wise approximation of several one-dimensional lookup table raw data layers extracted from the two-dimensional lookup table raw data.

[0048] If a two-dimensional relationship exists in the lookup tables, each EBM interaction can be interpolated individually. The inputs for this are the two lookup tables of the input features and the number of ranges for an axis. The method involves keeping one axis as it appears in the EBM lookup table, while the other axis is interpolated layer by layer. The same applies to polynomial interpolation.

[0049] In another aspect, it is proposed that approximation involves reducing the number of function support points represented by the lookup tables and / or at least piecewise replacing functions represented by the lookup tables with an approximation function. Support points are selected points in a domain where certain values ​​of a function are known or calculated. These values ​​serve as the basis for interpolating or approximating the function in other regions of the domain.

[0050] Interpolation uses data points to define a function that passes precisely through them. A curve is then constructed between these points that accurately reflects the function values ​​at those locations. Approximation uses data points to approximate a function. The approximated function does not necessarily have to pass through all data points, but should be close to them to accurately represent the overall properties of the function.

[0051] In a further aspect, it is proposed that the trained EBM model and the approximated EBM model be used to estimate the rack force for a vehicle's steering system. Preferably, the method includes a step of estimating a rack force of a vehicle's steering system using the trained EBM model and the approximated EBM model.

[0052] The trained EBM model is a machine learning model developed and refined using training data to make accurate predictions about rack force. The approximate EBM model is a simplified version of the trained model, used for faster computations and real-time applications.

[0053] Rack force estimation is the estimation of the force acting on the rack of the steering system, based on input data and relationships learned in the model. A vehicle's steering system is the system within the vehicle responsible for controlling steering movements.

[0054] In another aspect, it is proposed that the one-dimensional and / or two-dimensional lookup table raw data represent functional relationships of input data of the EBM model, where the input data includes data about a steering angle and / or steering angle velocity and / or longitudinal acceleration of a vehicle and / or vehicle speed and / or motor torque of a steering system servo motor and / or rack position and / or rack velocity and / or rack acceleration of a rack and / or steering wheel torque.

[0055] The steering angle is the angle by which the steering wheel is turned. The steering angle rate is the rate at which the steering angle is changed. The longitudinal acceleration of a vehicle is the acceleration of the vehicle in the longitudinal direction. The vehicle speed is the speed at which the vehicle is moving.

[0056] The motor torque of a steering system servo motor is the torque generated by the servo motor of the steering system. The rack position is the position of the rack in the steering system. The rack velocity is the speed at which the rack moves. The rack acceleration is the acceleration of the rack. The steering wheel torque is the torque exerted on the steering wheel.

[0057] In another aspect, a control unit, in particular an embedded system, is also claimed, which is included in a vehicle with an autonomous driving function and / or a robotics system and / or an industrial machine, and on which the approximately generated Explainable Boosting Machine (EBM) model is executable. The control unit can be an engine control unit, a steering angle control unit, or another type of control unit.

[0058] According to a further aspect, the present invention relates to a vehicle with a control unit described above, which is in particular configured to carry out the steps of the method described above and below.

[0059] In another aspect, a computer program with program code is claimed to execute at least parts of the present method in one of its aspects when the computer program is executed on a computer. In other words, a computer program (product) comprising instructions that, when executed by a computer, cause it to execute the method(s) in one of its aspects.

[0060] In a further aspect, a computer-readable data carrier containing the program code of a computer program is proposed to execute at least parts of the present method in one of its aspects when the computer program is executed on a computer. In other words, the invention relates to a computer-readable (storage) medium comprising instructions which, when executed by a computer, cause it to execute the method / steps of the method in one of its aspects.

[0061] The described configurations and training programs can be combined in any way desired.

[0062] Further possible embodiments, developments and implementations of the invention also include combinations of features of the invention described previously or subsequently with regard to the exemplary embodiments that are not explicitly mentioned.

[0063] Brief description of the drawings

[0064] The accompanying drawings are intended to provide a further understanding of the embodiments of the invention. They illustrate embodiments and, in conjunction with the description, serve to explain the principles and concepts of the invention.

[0065] Other embodiments and many of the aforementioned advantages become apparent with reference to the drawings. The elements depicted in the drawings are not necessarily shown to scale.

[0066] Figure 1 shows a schematic flowchart of one aspect of the present procedure;

[0067] Fig. 2 shows a schematic block diagram for a training session.

[0068] EBM model;

[0069] Fig. 3 shows a schematic block diagram of one aspect of the present procedure;

[0070] Fig. 4 a schematic block diagram of a piecewise linear

[0071] Interpolation of two-dimensional lookup table raw data;

[0072] Fig. 5 shows a schematic block diagram of a piecewise polynomial interpolation of two-dimensional lookup table raw data;

[0073] Fig. 6 shows a schematic block diagram of one aspect of the present procedure;

[0074] Fig. 7 shows a schematic representation of an application of a

[0075] Aspect of the present procedure; and

[0076] Fig. 8 shows a diagram of a functional relationship that has been interpolated using linear interpolation.

[0077] In the figures of the drawings, identical reference symbols denote identical or functionally equivalent elements, parts or components, unless otherwise stated.

[0078] Fig. 1 shows a schematic flowchart of a method for providing an Explainable Boosting Machine (EBM) model for a control unit, particularly on an embedded system, in a manner optimized for memory space. The method can be carried out, in any embodiment, at least partially by a device 100, which may comprise several components not shown in detail, for example, one or more provisioning devices and / or at least one evaluation and computing device.

[0079] It is understood that the provisioning device may be designed together with the evaluation and computing device, or it may be different from it. Furthermore, the device 100, which may be part of a system, may comprise a storage device and / or an output device and / or a display device and / or an input device.

[0080] The computer-implemented procedure includes at least the following steps:

[0081] In step S1, a trained EBM model is provided.

[0082] In step S2, one-dimensional and / or two-dimensional lookup table raw data is extracted for the trained EBM model.

[0083] In step S3, the one-dimensional and / or two-dimensional lookup table raw data is approximated by at least one approximation algorithm to generate interpolated lookup table data by optimizing a cost function that depends in particular on the model performance and storage requirements of the EBM model.

[0084] In step S4, an approximate Explainable Boosting Machine (EBM) model for a control unit is provided based on the interpolated lookup table data.

[0085] Figure 2 shows a schematic block diagram for training an EBM model. The training can be simplified into an input section 200, a process section 202, and an output section 204. In the input section 200, training data 206 (for example, in .csv format) and further hyperparameters 208 are provided for training 205 of the EBM model. Based on this input data, the amount of memory required later by the (unapproximated) model can be calculated. This memory requirement depends, among other things, on the number of "bins" or the resolution of the input data for the individual features, the number of interactions, and the number of "bins" (resolution) of the interactions. The output consists of the trained EBM model 210 and training information 212. The hyperparameters 208 can be optimized, if necessary, by empirical testing and / or automated hyperparameter optimization.

[0086] Fig. 3 shows a schematic block diagram of an extended process chain for generating an approximate, memory-optimized EBM model. Training data 300 is input into an EBM model 302 to be trained, followed by performance optimization 304. This results in a trained EBM model 306.

[0087] The process chain is now supplemented by an optimization step 308, which serves to optimize the trained EBM model 306 for an embedded system. Following the optimization 308, an implementation 310 of such an optimized, approximated model 324 is carried out on an embedded system for the use 312 of the approximated model 3024. The optimization step 308 comprises, in this case, providing the trained EBM model 306, extracting one-dimensional and / or two-dimensional lookup table raw data 314, and approximating the one-dimensional and / or two-dimensional lookup table raw data 314 by at least one approximation algorithm 316.

[0088] The approximation algorithm 316 can be implemented by manually interpolating 318 individual lookup tables, automatically interpolating 320 lookup tables and / or applying an equation learner 322 to obtain an optimized, approximate model 324.

[0089] Fig. 4 shows a schematic block diagram of a piecewise polynomial

[0090] Interpolation of two-dimensional lookup table raw data. This interpolation can be roughly divided into a data input (400), a function (402), and an interpolation output (404).

[0091] EBM-extracted two-dimensional lookup table raw data 406 is provided, from which an interaction between two features 408, 410 can be extracted. Furthermore, a number of areas 412 into which the lookup table raw data 406 is divided can be specified. Based on the features 408, 410 and the number 412, a linear interaction interpolation 414 can then be performed, whereby individual layers of the 2D lookup table are retained while another layer is interpolated.

[0092] The output 404 then provides, for each area, interpolated support points 416 and coefficients 418 of the interpolation 414. If the interpolation is a polynomial interpolation 500, the polynomial degree 502 is also specified as input data (see Fig. 5).

[0093] Fig. 6 shows another exemplary block diagram. One-dimensional lookup tables 600 of input features and two-dimensional lookup tables 602 of interactions between input features are provided. The lookup tables are interpolated either by linear or polynomial interpolation 604, 606, whereby in the case of interpolation based on a loss function 608, a loss 610 within an optimization loop 812 is to be optimized or minimized.

[0094] The optimization algorithm seeks the optimal composition 614 from the number of breakpoints / support points (linear) or the optimal composition 616 from the number of breakpoints and the polynomial degrees (polynomial). Based on the optimization, the approximate EBM model 618 can then be obtained.

[0095] Fig. 7 shows a schematic view of an application where an approximate EBM model is used for rack force estimation for a steering system of a vehicle 700. The steering system is controlled by an embedded control unit 702, on which the approximate EBM model for rack force estimation is executed. For rack force estimation, the approximate EBM model receives input data 704.

[0096] The input data includes data on a steering angle 706, a steering angle velocity 708, a longitudinal acceleration 710 of the vehicle 700, a vehicle speed 712, a motor torque 714 of a steering system servo motor, a rack position 715, a rack speed 716, a rack acceleration 718 of a rack and / or a steering wheel torque 720.

[0097] Fig. 8 shows a diagram of a function 800, whose data points can be represented as a lookup table. The function shows a dependence of a steering angle 802 (on the ordinate) on a rack force contribution 804 (on the abscissa).

[0098] In this case, the function 800 can be approximated by linear interpolation 805 between individual support points 806, so that the number of data points in an approximated lookup table can be significantly reduced compared to the original function, which was represented as a lookup table. This reduces the storage requirement for the EBM model approximated on the basis of this linear interpolation.

Claims

Claims 1. Method (100) for providing, in particular in a memory-space-optimized, an Explainable Boosting Machine (EBM) model for an electronic control unit, in particular for an embedded system, the method (100) comprising the steps: Provision (S1) of a trained EBM model (210, 306); Extracting (S2) one-dimensional and / or two-dimensional lookup table raw data (314, 406) to the trained EBM model (210, 306); Approximating (S3) the one-dimensional and / or two-dimensional lookup table raw data (314, 406) by at least one approximation algorithm (316) to generate interpolated lookup table data by optimizing a cost function that depends in particular on a model performance and a memory cost of the EBM model (210, 306); and Providing (S4) an approximate Explainable Boosting Machine (EBM) model (324) for an electronic control unit based on the interpolated lookup table data.

2. Method (100) according to claim 1, wherein the one-dimensional and / or two-dimensional lookup table raw data (314, 406) comprise lookup tables for input features and / or interactions of the trained EBM model (210, 306).

3. Method (100) according to claim 1 or 2, wherein the at least one approximation algorithm (316) comprises, in particular, a region-wise regression and / or at least a piecewise linear and / or polynomial interpolation and / or an equation learner and / or local models for decomposing a two-dimensional Lookup table raw data (314, 406) represent interaction space and / or pointwise interpolation and / or clusterwise interpolation and / or axis oblique division (HILOMOT) and / or axis orthogonal subdivision (LOLI MOT) and / or Delaunay triangulation and / or a grid-based approach.

4. Method (100) according to one of the preceding claims, wherein the two-dimensional lookup table raw data (314, 406) are approximated by, in particular, layer-wise approximation of several one-dimensional lookup table raw data layers extracted from the two-dimensional lookup table raw data (314, 406).

5. Method (100) according to one of the preceding claims, wherein the approximation (S3) comprises a reduction of a number of function support points represented by the lookup tables and / or at least a section-wise replacement of functions represented by the lookup tables by means of an approximation function.

6. Method (100) according to any of the preceding claims, further comprising a step of estimating a rack force of a steering system of a vehicle using the trained EBM model (210, 306) and the approximated EBM model (324).

7. Method (100) according to one of the preceding claims, wherein the one-dimensional and / or two-dimensional lookup table raw data (314, 406) represent functional relationships of input data of the EBM model (210, 306), wherein the input data comprise data about a steering angle and / or a steering angle velocity and / or a longitudinal acceleration of a vehicle and / or a vehicle velocity and / or a motor torque of a steering system servo motor and / or a rack position and / or a rack velocity and / or a rack acceleration of a rack and / or a steering wheel torque.

8. Computer program with program code to execute at least parts of a method (100) according to any one of claims 1 to 7 when the computer program is executed on a computer.

9. Computer-readable data carrier containing program code of a computer program for executing at least parts of a method (100) according to any one of claims 1 to 7 when the computer program is executed on a computer.

10. Device (100) for providing, in particular in a memory-space-optimized, an Explainable Boosting Machine (EBM) model for a control unit, in particular for an embedded system, wherein the device (100) has an evaluation and computing unit configured to perform the following steps: Providing a trained EBM model (210, 306); Extracting one-dimensional and / or two-dimensional lookup table raw data (314, 406) to the trained EBM model (210, 306); Approximating the one-dimensional and / or two-dimensional lookup table raw data (314, 406) by at least one approximation algorithm (316) to generate interpolated lookup table data by optimizing a cost function that depends in particular on a model performance and a memory cost of the EBM model (210, 306); and Providing an approximate Explainable Boosting Machine (EBM) model (324) for an electronic control unit based on the interpolated lookup table data.

Images