Method for generating a model ensemble for calibrating a control device

Abstract

A method for generating a model ensemble that estimates at least one output variable of a physical process as a function of at least one input variable, the model ensemble being formed from a sum of model outputs from a plurality of models that have been weighted with a weighting factor.

Description

TECHNICAL FIELD

The present invention relates to a method for generating a model ensemble that estimates at least one output variable of a physical process as a function of at least one input variable, the model ensemble being formed from a sum of model outputs from a plurality of models that have been weighted with a weighting factor.

BACKGROUND

In the development of internal combustion engines, there are legal requirements that must be taken into account with respect to emissions, in particular of NOx, soot, CO, CO₂, etc., and with respect to consumption. A combustion engine control device is calibrated for this purpose during development in such a manner that these requirements are met during operation of the combustion engine. Calibration here means that certain parameters of the combustion engine, such as the air/fuel ratio, the recirculation of waste gas in the cylinder, firing intervals, etc., are specified as a function of a particular state of the combustion engine, such as torque, speed, engine coolant temperature, etc. For example, corresponding engine maps are stored in the control device that are readout during operation of the combustion engine in order to identify the control parameters for a particular state. Because of the many influencing variables, calibration is a very time-consuming and expensive process that is primarily carried out on specialized test benches. For this, a combustion engine is mounted on a test bench and connected to a loading machine that simulates specific, predetermined load conditions. In the course of this, predetermined load cycles (driving cycles) are usually run on the test bench. During the load cycles, the emissions values and/or the consumption of the combustion engine are measured as a function of the prevailing state. After evaluation of the recorded measurement values, the control parameters are changed in the control device and the process is repeated until a satisfactory calibration was achieved. The time on the test bench is very expensive, however, and should be reduced as much as possible.

Therefore, methods have been developed to simplify the calibration, in particular to reduce test bench times. These methods are often based on models of the emissions or consumption behaviors of the combustion engine, or on models of a physical process in general. What is required here, therefore, is to determine a sufficiently precise model of the physical process that can then be used for calibration of the control device. Methods for automated model identification of non-linear processes (for example, of NOx emission or of consumption) have already become known, as described in WO 2013/131836 A2, for example. These methods are each based on a predetermined model structure, such as a neural network, a Kriging model or a linear model network. The model structure chosen thus defines model parameters that are determined by the method for automated model identification. Data in the form of measurement values on a test bench are collected for this, and the model is parameterized or trained using these data. Consequently, only a small number of test runs using an actual combustion engine on an actual test bench is necessary. Using the trained model, the influence of particular control parameters on emissions or on consumption can be examined without the need for further test bench tests.

The dissertation by Hartmann B. “Local model networks for identification and test design of non-linear systems,” Siegen University January 2014, goes into more detail on linear model networks as model structure. In the case of linear model networks, valid local models are defined in a known way over partial ranges of the input variable range. The output of a linear model network over the whole input range then results from the sum of the outputs of the local models weighted using validity functions. A local model thus estimates only a locally valid output value, and hence only a portion of the output value of the model network.

However, the selection of the best model structure for modeling a specific behavior of a combustion engine (NOx emission, for example) is already ambitious and not immediately obvious.

For this reason, so-called model ensembles have already been employed. Here, different models are trained that are then weighted in order to achieve the best possible estimation for a specific behavior of a combustion engine (emission or consumption, for example). The output of the complete model (of the model ensemble) results from a weighted sum of the outputs of the individual models. The weighting factors must therefore be determined for a model ensemble. A method often used for determining the weighting factors is based on the Akaike Information Criterion, as described in Akaike, H. “Information theory and an extension of the maximum likelihood principle,” Proceedings 2nd International Symposium on Information Theory, Budapest 1973, pp. 267-281. Hartmann's dissertation also describes model ensembles using weighting factors according to an Akaike Information Criterion.

The plausibility of a model M_j is evaluated using the Akaike Information Criterion. Thus, for each model M_j, the model error E_j and the complexity of model M_j are evaluated in the form

${\mathrm{AIC}}_j=N·\log \underset{\underset{E_j}{︸}}{{\mathrm{MSE}}_j}+2·\left[\alpha \right]·p_j.$ Using model error E_j, the deviation of the model output variable ŷ_j, of model M_j from the actual measured output y of the process is evaluated. In the Akaike Information Criterion AIC, the mean square error MSE_j of model M_j at N different known input variables u of model M_j is used as model error E_j. The mean square error MSE_j of the jth model M_j is calculated in well-known manner from

${\mathrm{MSE}}_j=\frac{1}{N}\sum _{i=1}^N{\left(y\left(u_i\right)-{\hat{y}}_j\left(u_i\right)\right)}^2.$

Complexity is evaluated in the Akaike Information Criterion simply in the form of the number p_j of the model parameters of the jth model M_j. There are also known modifications, for example, in the case of a neural network as a model structure for the jth model M_j, the number of effective parameters, that can be calculated according to known methods not more closely detailed here, is often employed for evaluating the complexity. It is also known to weight the complexity using a factor α (indicated in the above equation for AIC_j), the so-called risk aversion parameter.

The weighting factors w_j for the individual models M_j are then calculated according to the associated determined plausibility according to the Akaike Information Criterion AIC_j using

$w_j=e^{-\frac{1}{2}{\mathrm{AIC}}_j}$ and normalized to 1.

The problem with this Akaike Information Criterion AIC is that, although it can be calculated quickly, it is designed for a large number N of known data points (measured output variables y at a specified input variable u). The known data points are thus often also those data points by which the models M_j were trained. Moreover, the model structures (for the number of parameters p_j) of models M_j must be known so that the plausibility and thus the weighting factors w_j can be calculated.

For this application, however, the lowest number possible of known data points is desired because the expense for test bench tests and measurements on the test bench should be reduced as much as possible.

An adapted Akaike Information Criterion in the form

${\mathrm{AIC}}_{j,C}={\mathrm{AIC}}_j+2\frac{\left(p_j+1\right)·\left(p_j+2\right)}{N-p_j-2}$ has already been proposed for a smaller number N of data points. This adapted Akaike Information Criterion, however, delivers also unsatisfactory results for the small number of data points available. Within the meaning of the present invention, a small number of data points is understood as a number N that is comparable to the number p_j of the model parameters, thus N≈p_j, whereas N and p_j preferably having the same order of magnitude. In particular, it is a goal of the invention to keep the number N of measured data points as low as possible in order to keep the necessary number of test bench tests or measurements on the test bench as low as possible.

Further, it is often also the case in the present application that the models M_j are available as already trained models, whose model structure is completely unknown. The models M_j of the model ensemble can thus partly also be present as unknown black box. The known Akaike Information Criterion AIC cannot be used on such unknown models M_j for generating a model ensemble, however, because the number p_i of the model parameter is not known. Because of these disadvantages of the Akaike Information Criterion AIC under the present terms and conditions (small number of available data points, possibly no knowledge about the model structure), this cannot be used, or at least not satisfactorily used, for the generation of a model ensemble.

SUMMARY

It is therefore an object of the present invention to specify a method for the generation of a model ensemble that can manage using a low number of available data points and thus a low number of test bench tests and that requires no knowledge of the model structures of the models in the model ensemble.

In order to be able to determine for such a small number of actually measured, available data points and for partially unknown models the weighting factors for a good model ensemble, according to the invention for each model an empirical complexity measurement, that evaluates over a predetermined range of input variables evaluates the deviation of the model output variables from the output variables of the actual physical process, and a model error are determined and a surface information criterion is formed from the empirical complexity measurement and the model error, from which the weighting factor for the model ensemble can be derived. The model structure is thus not evaluated as in the Akaike Information Criterion, but an empirical complexity measurement is instead used that evaluates the complexity of a model based on the deviation of the model from the underlying physical process. Not only the deviation of the measured data points (model error) is evaluated here, but also a deviation between these data points, thus over a complete input variable range, that is reflected in the empirical complexity measurement. Knowledge of the model structures of the models in the model ensemble is thus no longer necessary. Through the use of the empirical complexity measurement, the number of necessary data points can also be greatly reduced so that the time on the test bench for measuring the necessary data points can also be greatly shortened.

The model error of a jth model can be calculated simply and quickly as a mean square error between the output variables measured at an input variable (data points) of the physical process and the model output variables calculated at these input variables according to the equation

${\mathrm{MSE}}_j=\frac{1}{N}\sum _{i=1}^N{\left(y\left(u_i\right)-{\hat{y}}_j\left(u_i\right)\right)}^2.$

Especially advantageously, the empirical complexity measurement of a jth model is calculated using the formula

$c_j={\int }_U▽{{\hat{y}}_j\left(u\right)}^T▽{\hat{y}}_j\left(u\right)\mathrm{du}$ or the formula

$c_j={\int }_U{{\hat{y}}_j\left(u\right)}^T{\hat{y}}_j\left(u\right)\mathrm{du}-\frac{1}{N}\sum _{i=1}^N{{\hat{y}}_j\left(u_i\right)}^T{\hat{y}}_j\left(u_i\right).$ Using these empirical measures of complexity allows an especially good model ensemble to be determined that is in particular better than any individual model of the model ensemble.

In order to have a degree of freedom for the determination of the weighting factors, the empirical complexity measurement is preferably weighted with a complexity aversion parameter.

In a simple embodiment of the invention, the weighting factors for each model of the model ensemble can be calculated from the formula

$w_j=e^{-\frac{1}{2}{\mathrm{SIC}}_j}.$

In a particularly advantageous embodiment of the invention, it is provided that the surface information criterion is formed for the model ensemble from an error matrix that includes the model error of the models and from a complexity measurement matrix that includes the empirical complexity measurement of the models M_j whereas the error matrix and the complexity measurement matrix according to the formula SIC={w^TFw+w^TCw} each being weighted twice with a weighting vector that includes the weighting factors of the model, and the surface information criterion of the model ensemble being minimized with respect to the weighting factors. Weighting factors of the model ensemble can be determined by optimization, which results in an especially small error between the model output variables of the model ensemble and the actual physical process.

It is advantageous in this context if the error matrix is calculated as a product of a matrix E, the matrix being calculated using the formula E=(y(u_i)−ŷ_j (u)).

It is especially advantageous if the complexity measurement matrix is weighted by a complexity aversion parameter because one can thereby obtain a degree of freedom by which it is possible to reduce the error between model output variables of the model ensemble and the actual process even further.

For this purpose, it can be provided in an especially advantageous embodiment of the invention that the weighting vectors for different complexity aversion parameters can be calculated and the weighting vector associated with a selected complexity aversion parameter is chosen as the optimum weighting vector for the model ensemble, or that the weighting vector for different complexity aversion parameters is calculated and the relationship

$\left\{w_{{\alpha }_K}^T{\mathrm{Fw}}_{{\alpha }_K}+\frac{2}{N}{\sigma }^2{w}_{{\alpha }_K}^{T}p\right\}$ is minimized with respect to the weighting vectors calculated for the different complexity aversion parameters in order to determine the optimum weighting vector.

BRIEF DESCRIPTION OF THE DRAWINGS

The present invention is explained below with reference to FIGS. 1 to 5, showing exemplary, schematic and non-restrictive advantageous embodiments of the invention. Shown are

FIG. 1, a model ensemble having a plurality of models weighted with weighting factors,

FIG. 2, the convergence of data points via different complex models,

FIG. 3, the training and validation errors as a function of the number of model parameters and

FIGS. 4 and 5, the effect of the calculation of weighting factors according to the invention.

DETAILED DESCRIPTION

A model ensemble 1, as illustrated in FIG. 1, is made of a number j of models M_j. Each model M_j is defined by a model structure, for example a neural network, a Kriging model, a linear model network, a polynomial, etc. and by a fixed number p_j of model parameters P_j={p_1,j, . . . , p_pj,j}. Model parameters P_j were or are trained or determined using an appropriate method and are generally known. Each model M_j maps an input variable vector u={u₁, . . . , u_k} onto an estimated value of the output variable ŷ_j of the modeled physical process. The actual output variable y of the physical process, which can be measured, for example, is approximated by model M_j. For example, input variable u includes the input variables torque and speed of a combustion engine, as well as the engine coolant temperature of the combustion engine, and an emission or consumption variable is estimated. Input variables u_i in input variable vector u can vary within a predetermined or specified input variable range U, with uεU.

Using model ensemble 1, or models M_j included therein, as a physical process e.g. an emission or consumption variable of a combustion engine, such as the NOx emission, the CO or CO₂ emission or the fuel consumption, is estimated as an output variable ŷ of model ensemble 1, or as a model output variable ŷ_j of model M_j. In the following description, for the sake of simplicity, a single output variable y will be assumed without limitation of general applicability, whereas an output variable vector y made up of a plurality of output variables y also being possible of course.

In model ensemble 1, each model output variable ŷ_j is weighted with a weighting factor w_j and output variable ŷ of model ensemble 1 is the weighted sum of model output variables ŷ_j of individual models M_j in the form

$\hat{y}\left(u\right)=\sum _j{w}_j{\hat{y}}_j\left(u\right).$ In the description, for simplicity's sake, ŷ and ŷ_j are also used, respectively, instead of the correct notation ŷ(u) and ŷ_j(u). With respect to weighting factors w_j, boundary conditions w_j∈[0,1]and

$\sum _j{w}_j=1$ are preferably to be taken into consideration. The problem is thus presented of how to best determine weighting factors w_j so that output variables y of the physical process are approximated by model ensemble 1 or by its output variable ŷ, as best as possible. The goal here, of course, is for model ensemble 1 to estimate output variable y of the physical process over the complete input variable range U, or the range of interest, better than the best model M_j of model ensemble 1.

FIG. 2 shows model output variable ŷ_j of a jth model M_j as a function of a single input variable u (that is without limitation of generality the simplest case). In it the points are measured data points, thus in each case a measured output variable y(u) at an input variable u. A first model M₁ having the associated model output variable ŷ_j=M1 approximates the measured output variable y using a simple model. Model M₂ with the associated model output variable ŷ_j=M2, represents a somewhat more complex model that better approximates the measured output variable y (in the sense of a smaller deviation of model M_j from the underlaying physical processes). The better approximation via model M₂, however, comes at the cost of greater model complexity due to a greater degree of freedom in modeling in the form of a larger number p_j of model parameters. In general, more complex models better approximate the actual physical processes, but require more model parameters and also more data to train the model and are also more sensitive to changes in the model parameters.

This basic relationship is illustrated in FIG. 3. In it, for example, model error E (for example MSE as above) of model M_j is applied across the number of model parameters p_j. In one case as model error E_T during the training of model M_j using the available training data (all or part of the measured data points available). And the other time as model error E_V, which was determined using predefined validation data (preferably available measurement values of the process output y(u) at specified input variables u). For the present application, however, there is also the problem that no or very few validation data are available.

In order to evaluate the complexity of the jth model M_j, an empirical complexity measurement c_j is used according to the invention that does not evaluate the model structure as in the prior art, but instead evaluates the deviation of model output variable ŷ_j from the output variable y of the physical process over a specified input variable range U. In contrast to a model error E, which relates to the deviation between model M_j and the physical process at specific measured data points, empirical complexity measurement c_j evaluates the deviation over a complete input variable range U, thus specifically also between the measured data points. Different approaches are available for an evaluation of this sort.

In a first approach, the surface of the model output variable ŷ_j over the input variable range U is used for evaluation. The inventive idea behind this can also be explained in reference to FIG. 2. As shown, (in a one-dimensional case illustrated here) the greater the length of the model output variable ŷ_j (which corresponds in the generalized case to the surface) over input variable range uεU, the more complex model M_j becomes, that is, the better output variable y is thus approximated by model M_j. Naturally, this can be generalized to any desired dimension (number of input variables u; in input vector u). The empirical complexity measurement c_j for evaluating the deviation of model M_j from the physical process over input variable range U on the basis of the surface is determined according to the following relationship:

$c_j={\int }_U▽{{\hat{y}}_j\left(u\right)}^T▽{\hat{y}}_j\left(u\right)\mathrm{du}.$

In this, ∇ is the known Nabla operator with respect to the input variables in the input variable vector u, therefore

$▽=\left(\frac{\partial }{\partial u_1},\dots ,\frac{\partial }{\partial u_i}\right).$ The integral is determined over a specified input variable range uεU, preferably over the whole range. This integral increases monotonically with the surface of model output variable ŷ_j. The surface of model output variable ŷ_j over input variable range U is thus evaluated here as empirical complexity measurement c_j.

As an alternative empirical complexity measurement c, which evaluates the deviation of model M_j or of model output variable ŷ_j from output variable y of the physical process, the variance of the model output variables ŷ_j can be employed. The variance (also designated as the second moment of a random variable) is, as is well known, the expected square deviation of a random variable from its expected value. Applied to the present invention, the model output variable ŷ_j at the available N data points is compared, using the variance, to the model output variable ŷ_j between these data points, which is designated here as variability. The idea behind this is that a model M_j having an increased variability generally predicts the basic physical process over input variable range U worse than a model M_j having a lower variability. This lies in the fact that the better model M_j approximates the measured data points, i.e. the more complex the model M_j becomes, the greater the probability of an increased variability becomes. However, if the variability becomes too large, the risk of overfit for model M_j therefore also increases. The typical behavior of such an overfilled or too-complex model M_j is a greatly varying model output variable ŷ_j over input variable range U, which in turn can lead to a larger deviation between actual output variable y and model output variable ŷ_j. This variability based on the variance can be mapped onto empirical complexity measurement c_j if empirical complexity measurement c_i is calculated according to the following formula.

$c_j={\int }_U{{\hat{y}}_j\left(u\right)}^T{\hat{y}}_j\left(u\right)\mathrm{du}-\frac{1}{N}\sum _{i=1}^N{{\hat{y}}_j\left(u_i\right)}^T{\hat{y}}_j\left(u_i\right)$

It is clear that there are additional possibilities for evaluating the deviation between model M_j and the physical process, or output variable y of the process and the model output variable ŷ_j. The basic idea remains unaltered, namely, the idea that the larger the empirical complexity measurement c_j, the more complex basic model M_j is. Empirical complexity measurement c_j therefore also evaluates the complexity of model M_j.

According to the invention, a surface information criterion SIC_j of jth model M_j is derived from empirical complexity measurement cj, which, analogous to the above Akaike Information Criterion AIC in the prior art, is again formed from model error E_j of model M_j and empirical complexity measurement c_j, therefore SIC_j=(E_j+α_K·c_j^s). Mean square error

${\mathrm{MSE}}_j=\frac{1}{N}\sum _{i=1}^N{\left(y\left(u_i\right)-{\hat{y}}_j\left(u_i\right)\right)}^2,$ for example, can again be used as model error E_j, wherein also any other model error E_j, such as in the form of the mean absolute deviation, could obviously also be used.

The preferably used parameter α_Kε[0, ∞[ in surface information criterion SIC_j is used as a complexity aversion parameter. This represents the only degree of freedom with which the complexity of model M_j of model ensemble 1 can be further penalized. The larger the complexity aversion parameter α_K becomes, the more complexity enters into the surface information criterion SIC_j. Small complexity aversion parameters α_K therefore favor more complex models M_j, meaning models M_j having more degrees of freedom (number of model parameters p_j).

Analogous to the known Akaike Information Criterion, weighting factors w_j can again be determined from

$w_j=e^{-\frac{1}{2}{\mathrm{SIC}}_j},$ wherein w_j∈[0,1] and

$\sum _j{w}_j=1$ can be preferably be considered as boundary conditions. Although a model ensemble 1 can already be formed by using this, which, under the given conditions, better approximates the actual process, meaning with fewer errors than a model formed using the Akaike Information Criterion AIC, the quality of model ensemble 1 can be further improved according to the invention. The method involves the approach as explained below.

It can be shown that the mean square model error MSE and the empirical complexity measurement c of model ensemble 1 with respect to a weighting vector w, which includes weighting factors w_j of j models M_j can each be represented as a quadratic function of model error E_j and empirical complexity measurement c_j of models M_j in the form SIC={w^TFw+α_Kw^TCw}. Within this, optional complexity aversion parameter α_K represents a degree of freedom in the determination of weighting factors w_j of j models M_j.

In this context, F designates an error matrix that includes model error E_j of models M_j and C a complexity measurement matrix that includes empirical complexity measurement c_j of models M_j. In the case of mean square error MSE; as model error E_j and with a matrix E=(y(u_i)−ŷ_j(u_i)), for all i ∈ N data points and j, error matrix F results as the product of matrix E with itself, according to F=E^TE. Depending upon the empirical complexity measurement c_j chosen, complexity measurement matrix C results in, for example,

$C={\int }_U\nabla {{\hat{y}}_a\left(u\right)}^T\nabla {\hat{y}}_a\left(u\right)\mathrm{du}\mathrm{or}$ $C={\int }_U{{\hat{y}}_a\left(u\right)}^T{\hat{y}}_a\left(u\right)\mathrm{du}-\frac{1}{N}\sum _{i=1}^N{{\hat{y}}_a\left(u_i\right)}^T{\hat{y}}_a\left(u_i\right),$ each having model output variable vector ŷ_a, which contains model output variables ŷ_j of j models, thus ŷ_a={ŷ₁ . . . ŷ_j}. Matrices F and C can thus be calculated in advance and, above all, without knowledge of models M_j or their model structure or the number of model parameters p_j.

For determining weighting factors w_j (or, analogously, weighting vector w), surface information criterion SIC of model ensemble 1 for a specified complexity aversion parameter α_K can be optimized with regard to weighting factors w_j, in particular minimized. An optimization problem in the form

$w_{\alpha }=\arg \underset{w}{\min }\left\{w^T\mathrm{Fw}+{\alpha }_K{w}^T\mathrm{Cw}\right\}$ can be derived from this.

As can be easily recognized, this is a quadratic optimization problem that can be solved quickly and efficiently using available standard solution algorithms for a predetermined complexity aversion parameter α_K, w_j∈[0,1] and

$\sum _j{w}_j=1$ preferably apply as boundary conditions for optimization. Any initial weighting vector w can be specified.

The result of the optimization of Surface Information Criterion SIC of model ensemble 1 for determination of weighting factors w_j is described in reference to FIG. 4. The exemplary embodiment is based on j models M_j trained using a small number of data points. Within this, the two diagrams to the left show training error E_TE and validation error E_VE of model ensemble 1, which is determined using the optimization of surface information criterion SIC described above. Plotted on the abcissa is an empirical model ensemble complexity p_eff^w, which is derived from the complexities of each model M_j in the form p_eff^w=w^Tp. w is the weighting vector determined for a specified complexity aversion parameter α_K, and vector p includes the number of model parameters p_j for every model M_j. If complexity aversion parameter α_K is varied and optimization is resolved for each complexity aversion parameter α_K, whereby one obtains an associated weighting vector w_αK for each, one obtains the curves in the two diagrams to the left. The points in the diagram respectively represent training error E_Tj and validation error E_vj of a jth model M_j of model ensemble 1 (the number p_j of the model parameter of the jth model M_j is applied for this in each case). As can easily be recognized, model ensemble 1 determined according to the invention is always better in each case, meaning having smaller errors, than the best single model M_j.

In FIG. 4, validation error E_V is represented as a function of training error E_T by the diagram on the right. The points again represent the individual models M_j. In this, model ensemble 1 that is determined using the standard Akaike Information Criterion AIC is also compared to model ensemble 1 that is determined according to the invention using the surface information criterion SIC. As can be clearly recognized, surface information criterion SIC not only performs far better than Akaike Information Criterion AIC, but also better than each individual model M_j.

It can also be deduced from the diagram on the right in FIG. 4 that there is a complexity aversion parameter α_K,opt that minimizes the model error of model ensemble 1. It can now be attempted to manually find this optimum complexity aversion parameter α_K,opt or at least to manually approximate it. In a second step following the optimization of surface information criterion SIC, an attempt can also be made, however, to determine optimum complexity aversion parameter α_K,opt, and thus also the associated optimum weighting vector w_opt, with the procedure being described below.

To accomplish this, the associated weighting vectors w_αK are first determined for a plurality of complexity aversion parameters α_K. A set of weighting vectors is thus obtained

${\left\{w_{{\alpha }_K}\right\}}_{{\alpha }_K\ge 0}.$

Using the known Mallow equation, complexity aversion parameter α_K is chosen as optimum complexity aversion parameter α_K,opt, which solves the following optimization problem

$w_{\mathrm{opt}}=\arg \underset{w_{{\alpha }_K}}{\min }\left\{w_{{\alpha }_K}^T{\mathrm{Fw}}_{{\alpha }_K}+\frac{2}{N}{\sigma }^2{w}_{{\alpha }_K}^{T}p\right\}.$

Within this, F is again the error matrix (F=E^TE) and a is the standard deviation of the available data points, but which is generally not known. There are, however, known methods (as described in Hansen, B. E. “Least squares model averaging,” Econometrica, 75(4), 2007, pp. 1175-1189, for example) to estimate the standard deviation σ from the available data points. Vector p again includes for all j models M_j the number of model parameters p_j. The knowledge of models M_j or their model structures is, therefore, required for this step.

This optimization is not, however, solved directly, but with respect to the initially determined set of weighting vectors

${\left\{w_{{\alpha }_K}\right\}}_{{\alpha }_K\ge 0}.$ This means that there is selected the weighting vector w associated to a specific complexity aversion parameter α_K as optimum weighting vector w_opt, which yields the minimal expression

$\left\{w_{{\alpha }_K}^T{\mathrm{Fw}}_{{\alpha }_K}+\frac{2}{N}{\sigma }^2{w}_{{\alpha }_K}^{T}p\right\}.$

In FIG. 5, the effect of the method according to the invention for determining weighting factors w_j for a small number N of available data points is demonstrated using an example. To do this, 5,000 data points for NOx emission as well as for soot emission are measured on a specific combustion engine, thus in each case a measurement value for NOx and soot at 5,000 input vectors u, in order to have sufficient data for the demonstration example. Input vector u incorporated here, for example, five input variables u_i, namely torque, speed, engine coolant temperature, position of a turbocharger having variable turbine geometry and position of an exhaust gas recirculation. Fifteen different models M_j (different model structures and/or different number of model parameters and/or different model parameters) were trained using a random selection of data points from the 5,000 available data points. The random selection was used as a small number N of available data points. Number N of available data points was thereby increased from 10 to 150, thus 10≤N≤150. The data points remaining in each case (5,000-N) were used as validation data for validation of the example in order to verify the effectiveness of the invention. In FIG. 5, the validation error for NOx emission E_V,NOx and soot emission E_V,Soot is illustrated for each case. Validation error E_V is the mean square error between model output variable ŷ_j or output variable ŷ of model ensemble 1 and the validation data. For each number N of data points, a best and a worst model M_k results (dashed lines in FIG. 5). This bandwidth of model M_j is represented in FIG. 5. In addition, for each number N of data points, a model ensemble according to the known Akaike Information Criterion AIC and a model ensemble 1 using the surface information criterion SIC according to the invention were determined. The validation errors for this are shown in the diagrams in FIG. 5, as well. It is immediately obvious from this that model ensemble 1 determined according to the invention is not only usually better than the best model M_j in each case, but also better than the model ensemble that was determined using Akaike Information Criterion AIC.

A model ensemble determined according to the invention is used, for example, for calibrating a technical system, such as a combustion engine. In the calibration—in order to optimize at least one output variable of the technical system—control variables of the technical system, by which the technical system is controlled, are varied in a specified operational state of the technical system that is defined by state variables or a state variable vector. The optimization of output variables by variation of the control variables is generally formulated and solved as an optimization problem. There are sufficient known methods for accomplishing this. The control variables determined in this manner are stored as a function of the respective operational conditions, for example in the form of characteristic maps or tables. This relationship can then be used to control the technical system as a function of the actual operational state (which is measured or otherwise determined (for example, estimated)). This means that the stored control variables for the relevant operational state are readout from the stored relationship and used to control the technical process. In the case of a combustion engine as technical system, the operational condition is often described using measurable variables such as speed and torque, wherein other variables such as engine coolant temperature, ambient temperature, etc., can also be used. In a combustion engine, the position of a variable-turbine-geometry turbocharger, the position of an exhaust-gas recirculation system or the injection timing are often used as control variables. The output variable to be optimized in a combustion engine is typically the consumption and/or emission variable (for example, NOx, CO, CO₂, etc.). Calibration of a combustion engine thus ensures by setting correct control variables that consumption and/or emission during operation are minimal.

Claims

1. A method for calibrating a technical system controllable by control variables with a model ensemble, that estimates at least one output variable (y) of the technical system as a function of at least one input variable (u), comprising: forming the model ensemble from a sum of model outputs (ŷj) from a plurality (j) of models (Mj) that have been weighted with a weighting factor (wj),determining for each model (Mj) an empirical complexity measurement (cj), that evaluates the deviation of the model output variable (ŷj) from the output variable (y) of the actual physical process over a specified input variable range (U), and a model error (Ej), wherein the empirical complexity measurement (cj) is weighted with a complexity aversion parameter (aK),forming a surface information criterion (SICj, SIC) from the empirical complexity measurement (cj) and the model error (Ej) from which the weighting factor (wj) for the model ensemble is determined,calibrating the technical system using the model ensemble by setting the control variables of the technical system to ensure an optimized at least one output variable (y) of the technical system during operation.

2. The method according to claim 1, wherein the mean square error (MSEj) between the output variables (y) of the physical process measured at N input variables (u) and the model output variables calculated at these N input variables is used as model error (Ej) of a model (Mj) according to the relationship

3. The method according to claim 1, wherein the empirical complexity measurement (cj) of a model (Mj) is calculated using the formula

4. The method according to claim 1, wherein the weighting factors (wj) of each model (Mj) of the model ensemble are calculated using the surface

5. The method according to claim 1, wherein for the model ensemble the surface information criterion (SIC) is formed from an error matrix (F) that includes the model error (Ej) of the models (Mj) and a complexity measurement matrix (C) that includes the empirical complexity measurement (cj) of the models (Mj), whereas the error matrix (F) and the complexity measurement matrix (C) according to the formula SIC={wTFw+wTCw} each being weighted twice using a weighting vector (w) that includes the weighting factors (wj) of the models (Mj), and the surface information criterion (SIC) of the model ensemble being minimized with respect to the weighting factors (wj).

6. The method according to claim 5, wherein the error matrix (F) is calculated as a matrix product of a matrix (E), whereas the matrix (E) being calculated using the formula E=(y(ui)−ŷj(uj)).

7. The method according to claim 5, wherein the complexity measurement matrix (C) is weighted using a complexity aversion parameter (aK).

8. The method according to claim 7, wherein the weighting factors (wj) are calculated for different complexity aversion parameters (aK) and the weighting vector (wak) belonging to a selected complexity aversion parameter (aK) is chosen as the optimum weighting vector (wopt) for the model ensemble.

9. The method according to claim 7, wherein weighting vectors (wak) are calculated for different complexity aversion parameters (aK) and, in order to determine the optimum weighting vector (wopt), the relationship

10. The method according to claim 6, wherein the complexity measurement matrix (C) is weighted using a complexity aversion parameter (aK).

11. A combustion engine comprising a control device calibrated by a model ensemble generated by the method of claim 1.

12. The method according to claim 1, wherein the output variable comprises an emission variable.

13. A method of calibrating a technical system controllable by control variables using a model ensemble that estimates an output variable of the technical system as a function of an input variable, comprising: forming the model ensemble from a sum of model outputs of a plurality of models that have been weighted with a weighting factor;determining an empirical complexity measurement for each of the plurality of models that evaluates a deviation of the model output from the output variable over a specified input variable range, wherein the empirical complexity measurement is weighted with a complexity aversion parameter;determining a model error for each of the plurality of models;forming a surface information criterion from the empirical complexity measurement and the model error from which the weighting factor is determined;calibrating the technical system using the model ensemble by setting the control variables of the technical system to ensure an optimized output variable of the technical system during operation.

14. A combustion engine comprising a control device calibrated by the method of claim 13.

15. The method according to claim 13, wherein the output variable comprises an emission variable.