DE102019125960B3 - System and method for calibrating the cylinder air charge in at least one cylinder in an internal combustion engine - Google Patents

Abstract

A method for calibrating a cylinder air charge in at least one cylinder of an internal combustion engine, in particular a reciprocating piston engine, comprising: - determining (S10) a state (si) of the internal combustion engine by a state module (300), a state (si) being determined by data and measured values of at least a parameter (pi) of the internal combustion engine is defined and the value of the cylinder air charge depends at least partially on this parameter (pj); - selecting (S20) a calculation function (fi) and / or an action (ai) for the parameter (pi) from a Learning reinforcement agents (200); calculating (S30) a modeled value (ZM) for the cylinder air charge from the learning reinforcement agent (200) using the selected calculation function (fi) and / or the action (aj); providing (S40) a real value (ZR) for the cylinder air charge from an environmental module (400) based on measurement results; - comparison (S50) of the modeled value (ZM) for the cylinder air charge treatment with the real value (ZR); forwarding (S60) the comparison result to a reward module (500) and determining a reward for the comparison result; forwarding (S70) the reward for the comparison result to the learning reinforcement agent (200) and based on this reward, again selecting another action (aj) from the learning reinforcement agent (200).

Description

The invention relates to a system and method for calibrating the cylinder air charge in at least one cylinder in an internal combustion engine which is designed in particular as a reciprocating piston engine.

It is known that internal combustion engines burn an air-fuel mixture within cylinders of the engine to drive pistons, thereby generating drive torque. On some types of engines, the air flow into the engine can be regulated by a throttle valve. The throttle valve can adjust the throttle area, thereby increasing or decreasing the airflow in the engine, as the throttle area increases, the airflow in the engine increases. In the case of fuel injection, the desired fuel-air mixture is controlled by a fuel control system. Generally, increasing the amount of air and fuel provided to the cylinders increases engine torque

In exhaust technology, the lambda (λ) symbol stands for the ratio of air to fuel compared to a combustion stoichiometric mixture. With the stoichiometric fuel ratio, exactly the amount of air or the oxygen concentration is available that is theoretically required to completely burn the fuel. This is referred to as A = 1. In the case of gasoline, this mass ratio is 14.5: 1, which means that 14.5 kg of air are needed to completely burn 1 kg of fuel. The ratio for ethanol is 9: 1 and for diesel fuel and heating oil 14.7: 1. If there is more fuel, it is called a rich mixture (λ <1), and if there is excess air, it is called a lean mixture (λ> 1).

The oxygen concentration in a gas mixture is determined by means of an oxygen-sensitive gas probe, a lambda probe. The lambda probe provides an actual probe signal which is dependent on the oxygen content of the gas mixture and which, for example, can be a probe voltage in the case of step lambda probes or a current intensity in the case of linear lambda probes. This probe signal is converted into the lambda value using a stored characteristic curve or a corresponding calculation rule.

Maintaining a certain lambda value has a major influence on the quality of the combustion and the possibility of complete catalytic exhaust gas cleaning. The fuel consumption usually reaches its minimum at a value of λ = 1.1. For a maximum engine torque, albeit with increased fuel consumption (incomplete combustion due to lack of air), a value of approx. Λ = 0.85 is optimal. With high engine performance, rich engine operation and thus colder exhaust gas prevent overheating and destruction of exhaust gas components such as manifolds, turbochargers and catalytic converters. Modern engines achieve a lower cylinder air charge through design measures (such as water-cooled manifolds, direct injection or cooled exhaust gas recirculation) so that enrichment of the mixture is not necessary or only necessary in a small area for reasons of component protection.

In a motor vehicle, a large number of parameters, such as the torque generated, are therefore dependent on the cylinder air charge. A control unit therefore calculates the amount of fuel that is injected based on the cylinder air charge. Modeling the cylinder air charge is therefore of considerable importance for engine control.

There are mainly two approaches to determining and predicting the cylinder air charge in an internal combustion engine. The first approach is based on a physical model in which the pressure and volume of the air are calculated depending on the temperature, the load and the number of revolutions of the engine. The corresponding values depending on the state of the engine are stored in the form of tables, characteristic curves and diagrams, preferably in an electronic control unit of the vehicle.

Another approach uses regression analysis to model and predict cylinder air charge. Regression methods are statistical analysis methods that aim to model relationships between a dependent and one or more independent variables. The regression models receive, for example, the gas charge, the number of revolutions or the intake temperature as input and calculate an estimated value for the cylinder air charge using mathematical methods such as the method of least squares or neural networks.

However, these tables and diagrams or the mathematical procedures used are based on expert knowledge and are no longer changed after they have been implemented. Furthermore, they are not adapted for an individual engine, but are defined for a specific model series.

The DE 10 2011 085 898 A1 and the US 8 762 307 B2 describe a control / regulating system for cylinder filling in reciprocating piston engines, in which a neural network trained by means of reinforcement learning controls the cylinder filling.

The US 8 676 476 B2 relates to a system and method for the self-learning real-time characterization of fuel injector performance during engine operation. The system includes an algorithm for an engine controller that enables the controller to learn the correlation between fuel mass and pulse width for each injector in the engine in real time while the engine is running. The controller progressively detects those pulse widths that achieve the desired mass of fuel, while continuously applying what it has learned based on various input changes such as B. temperature and fuel rail pressure can adjust. The controller then uses the learned actual power of each injector to command the pulse width required to achieve the desired amount of fuel for each cylinder in each cycle.

The US 2016363083 A1 describes an engine control system of a vehicle having a cylinder control module, an air-per-cylinder (APC) module, an adjustment module and a fuel control module. The air-per-cylinder (APC) module determines an APC value based on a mass air flow rate, an intake manifold pressure, an air temperature, a volumetric efficiency of the engine, and a model calibrated based on the ideal gas equation. The engine control system can use this as a self-learning basis to determine the fuel supply and other engine operating parameters.

The US 8 880 321 B2 describes a model for estimating the air charge in an internal combustion engine that uses a linear regression method.

The US 2020/0063681 A1 describes an engine system with an air supply and fuel system, the states of which are managed by a management unit. The engine system has several sensors, the sensor signals of which at least partially define a current state of the engine system. The management unit includes a controller that controls the air supply and fuel system of the engine system, and a processing unit that is coupled to the sensors and the controller. The processing unit includes an agent that learns a policy function and is trained to process the current state, determine air supply references and fuel system references using the policy function after receiving the current state as input, and output the air supply references and fuel. Then the agent receives a next status and a reward value from the processing unit and updates the policy function using a policy evaluation and a policy improvement algorithm based on the received reward value. The controller then controls the engine's air supply and fuel systems in response to receiving the air supply references and the fuel system references.

The object on which the invention is based is now to create a method and a system for automatically calibrating the cylinder air charge of at least one cylinder in an internal combustion engine, which is designed in particular as a reciprocating piston engine, which is characterized by high reliability and accuracy and is simple can be implemented.

According to the present invention, a method is proposed by means of which an automatic calibration of the cylinder air charge in at least one cylinder of an internal combustion engine is made possible in order to create the basis for reliable and precise control by a control and regulating device.

This object is achieved according to the invention with regard to a method by the features of claim 1, with regard to a system by the features of claim 8, and with regard to a computer program product by the features of claim 15. The further claims relate to preferred embodiments of the invention.

According to a first aspect, the invention relates to a method for calibrating a cylinder air charge in at least one cylinder in an internal combustion engine, in particular a reciprocating piston engine. The method includes the determination of a state s _{i of} the internal combustion engine by a state module, a state s _{i being} defined by data and measured values of at least one parameter p _{i of} the internal combustion engine and the value of the cylinder air charge at least partially dependent on this parameter p _i . A calculation function f _i and / or an action a _i for the parameter p _i is selected by a learning reinforcement agent and a modeled value Z _M for the cylinder _air charge is calculated using the selected calculation function f _i and / or the action a _i . A real value Z _R for the cylinder _air charge is provided by an environmental module based on measurement results. The modeled value Z _M for the cylinder _air charge is compared with the real value Z _R. The comparison result is passed on to a reward module and a reward is determined for the comparison result. The reward for the comparison result is passed on to the learning reinforcement agent and, based on this reward, another action a _j is selected again by the learning reinforcement agent.

Advantageously, sensors and / or measuring devices are provided for acquiring measured values of the parameters p _i .

In an advantageous further development, there is a positive action A + which increases the value for a parameter p _i , a neutral action A0 in which the value of the parameter p _i remains the same, and a negative action A- in which the value of the parameter p _{i changes} reduced, provided.

In a further embodiment, the reward module comprises a database or matrix for evaluating the actions a _i .

The learning reinforcement agent is advantageously designed to use an algorithm from reinforcement learning.

In a further embodiment, the algorithm is designed as a Markov decision process or as temporal difference learning (TD learning) or as Q-learning or as SARSA or as Monte Carlo simulation.

At least one parameter p _i advantageously represents an engine speed, the composition of a gas mixture, the temperature of the cylinder, and / or the lambda value of a lambda probe.

According to a second aspect, the invention provides a system for calibrating a cylinder air charge in at least one cylinder in an internal combustion engine, in particular a reciprocating piston engine. The system comprises a state module which is designed to determine a state s _{i of} the internal combustion engine, a state s _{i being} defined by data and measured values of at least one parameter p _{i of} the internal combustion engine and the value of the cylinder air charge at least partially from this parameter p _i depends. In addition, the system has a learning reinforcement agent which is designed to select a calculation function f _i and / or an action a _i for the parameter p _i and a modeled value Z _M for the cylinder _air charge using the selected calculation function f _i and / or the To calculate action a _i . Furthermore, an environment module is provided which is designed to provide a real value Z _R for the cylinder _air charge based on measurement results. The status module is designed to compare the modeled value Z _M for the cylinder _air charge with the real value Z _R. Furthermore, the system comprises a reward module which is designed to determine a reward for the comparison result and to pass on the reward for the comparison result to the learning reinforcement agent, the learning reinforcement agent being designed to carry out another action a _j based on this reward to select.

In an advantageous development, sensors and / or measuring devices are provided for recording measured values of the parameters p _i .

Advantageously, there are a positive action A +, which increases the value for a parameter p _i , a neutral action A0, in which the value of the parameter p _i remains the same, and a negative action A-, in which the value of the parameter p _{i is} reduced, intended.

In a further embodiment, the reward module comprises a database or matrix for evaluating the actions a _i .

In a further embodiment, the learning reinforcement agent is designed to use an algorithm from reinforcement learning.

The algorithm is advantageously designed as a Markov decision process or as temporal difference learning (TD learning) or as Q-learning or as SARSA or as DYNAQ or as Monte Carlo simulation.

In one development, at least one parameter p _{i represents} an engine speed, the composition of a gas mixture, the temperature of the cylinder and / or the lambda value of a lambda probe.

According to a third aspect, the invention provides a computer program product comprising executable program code which is configured such that it executes the method according to the first aspect when it is executed by a processor.

The invention is explained in more detail below with reference to the exemplary embodiments shown in the drawing.

• 1 ein Blockdiagram zur Erläuterung eines Ausführungsbeispiels eines erfindungsgemäßen Systems;
• 2 ein Flussdiagramm zur Erläuterung der einzelnen Verfahrensschritte eines erfindungsgemäßen Verfahrens;
• 3 zeigt schematisch ein Computerprogrammprodukt gemäß einer Ausführungsform des dritten Aspekts der Erfindung.

It shows:

• 1 a block diagram to explain an embodiment of a system according to the invention;
• 2 a flow chart to explain the individual method steps of a method according to the invention;
• 3 shows schematically a computer program product according to an embodiment of the third aspect of the invention.

Additional features, aspects and advantages of the invention or its exemplary embodiments will be apparent from the detailed description in conjunction with the claims.

1 shows a system according to the invention 100 for calibrating the cylinder air charge in an internal combustion engine, in particular a reciprocating piston engine. The system according to the invention is based on a reinforcement learning agent (LV) 200, which selects at least one action a _i ∈ A from a set of available actions for a specific state s _i ∈ S from a set of available states. For the selected action a _i , the agent receives 200 a reward r _i ∈ ℝ. The agent receives the states s _i ∈ S 200 from a state module 300 on which the LV agent 200 can access and preferably a data memory 320 with data from various sensors and data sources as well as calculation results. The state module 300 processes the data of various parameters p _i from a data memory 320 or another data source and assigns states s _i ∈ S to these processed data.

A state s _i ∈ S is thus defined by the selection of certain parameter values of parameters such as the number of revolutions of the engine, the inlet temperature of the gas mixture, the composition of the gas mixture, the temperature of the cylinder, etc. A state s _i is thus characterized by measured and / or calculated values of selected parameters p _i and thus represents a specific operating situation of the engine. The data or measured values are from an environmental module 400 have been determined and were preferably in the data memory 320 saved. The environmental module 400 contains at least one sensor 420 , but can have additional sensors 440 included or associated with them. In a reward module 500 a reward r _i eine ℝ is assigned to the selected action a _i , which is given to the LV agent 200 is transmitted.

With the sensors 440 it can be, for example, pressure sensors, piezo sensors, speed sensors, electrical sensors, temperature sensors, Hall sensors and / or image-recording sensors such as cameras, and a lambda probe. The sensors 420 , 440 determine measured values such as the cylinder air charge or a lambda value that are sent to the status module 300 or another processor. The state module 300 advantageously also has a processor that processes the data using a software application with an algorithm. However, it is also conceivable that the data are initially stored in a further memory unit or a software module and only at a later point in time in the data memory 320 transmitted and / or from the status module 300 and / or a processor.

In connection with the invention, a “processor” can be understood to mean, for example, a machine or an electronic circuit. A processor can in particular be a main processor (Central Processing Unit, CPU), a microprocessor or a microcontroller, for example an application-specific integrated circuit or a digital signal processor, possibly in combination with a memory unit for storing program commands, etc. . A processor can also be understood to be a virtualized processor, a virtual machine or a soft CPU. For example, it can also be a programmable processor that is equipped with configuration steps for executing the above-mentioned method according to the invention or is configured with configuration steps in such a way that the programmable processor uses the inventive features of the method, the component, the modules, or other aspects and / or implemented partial aspects of the invention.

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

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

In connection with the invention, “measured values” are to be understood as meaning both the raw data and already processed data from the measurement results of the sensors.

The learning reinforcement agent 200 contains calculation methods and algorithms f _i for mathematical regression methods or physical model calculations, which establish a correlation between selected parameters p _i ∈ P from a set of parameters and a cylinder air charge of one or more cylinders of an internal combustion engine, in particular a reciprocating engine, describe. The mathematical functions f _t can be mean values, minimum and maximum values, Fast Fourier Transformations, integral and differential calculations, linear regression methods, Markov methods, probability methods such as Monte Carlo methods, Tempora Difference Learning, but also extended Kalman -Filter, radial basic functions, data fields, convergent neural networks, artificial neural networks and / or feedback neural networks. The LV agent 200 selects one or more of these calculation functions f _i for a state s _i . It calculates a modeled value of the cylinder air charge Z _M for this specific state s _i and using the selected calculation functions f _i , the measured or calculated value for at least one parameter p _{i representing} an input value for the calculation functions f _i .

In the reinforcement learning agent 200, actions A +, A0 and A- are also defined as actions a _i , which are selected on the basis of the calculation method f _i used in order to adapt a measured or calculated parameter value of a parameter p _i ∈ P from the set of parameters in a table of values and the like. The parameters p _i ∈ P are, for example, the lambda value, the number of revolutions of the engine, the inlet temperature of the gas mixture, the gas density and / or the composition of the gas mixture. A positive action A + is an action that increases the value for a parameter p _t , and a neutral action A0 it is an action in which the value of the parameter p _t remains the same, and in the case of a negative action A- the value of the parameter p _{t is} reduced. The LV agent 200 selects for the state s _i defined by these measured and calculated data _in particular one of the actions A +, A0 or A- for adapting the parameter value of a parameter p _i in a table of values or the like. The LV agent now calculates for the selected action a _i 200 now the modeled value of the cylinder air charge Z _M , which depends on the selected action a _i for the parameter p _i .

This calculated value Z _{M of} the cylinder air charge for a specific action a _i is now sent to the status module 300 passed on. The environmental module 400 determines a real value Z _{R of} the cylinder air charge based on a direct measurement or a calculation. A neural network can advantageously be used to calculate the real value Z _{R of} the cylinder air charge. However, other calculation methods such as regression methods can also be used. The real cylinder air charge Z _R is preferably provided by a sensor 420 measured.

As already stated, in the status module 300 a current or future state s _{i is} defined by various data values of parameters p _i , for example a certain lambda value, the current number of revolutions of the engine, the current inlet temperature of the gas mixture, the current gas density and / or the composition of the gas mixture. In particular, the status module calculates 300 the current deviation Δ between the measured (real) value of the cylinder air charge Z _R and the calculated model value Z _M for a selected action a _i by the LV agent 200 .

In the state module 300 the calculated model value Z _{M of} the cylinder air charge is now compared with the actually measured value Z _{R of} the cylinder air charge. The calculated value Z _{M of} the cylinder air charge is based on the changed values for the selected parameters p _i by the LV agent 200 . In the reward module 500 the degree of the deviation Δ between the calculated model value Z _M and the measured value Z _{R of} the cylinder air charge is compared and a reward B _{i is} assigned to the degree of the deviation Δ. Since the degree of the deviation Δ is dependent on the selection of the respective action (A +), (A0), (A-), it is preferably in a matrix or some other database 520 the respective selected action (A +), (A0), (A-) is assigned the reward B _i . The reward B _i preferably has the values +1 and -1, with a slight deviation Δ between the calculated value Z _M and the measured value Z _{R being} rewarded with +1 and thus reinforced, while a considerable deviation Δ with -1 is rewarded and therefore rated negatively. However, it is also conceivable that values> 1 and values <1 are used.

A second cycle for calibrating the cylinder air charge now begins. Here, another action a _i and / or another calculation function f _i and / or another parameter p _i can be selected. The result is in turn sent to the status module 300 supplied and the result of the comparison in the reward module 500 rated. The LV agent 200 repeats the calibration of the cylinder air charge for all intended actions a _i and parameters p _i until the greatest possible correspondence between the calibrated model value Z _{M of} the cylinder air charge and the measured value W _{R is} achieved. The final state of the calibration of the cylinder air charge is preferably reached when the deviation is in the range of +/- 5%. A very large deviation in the cylinder charge, which could lead to the destruction of the internal combustion engine, is rewarded very negatively.

The preferred algorithm for the LV agent is 200 uses a Markov decision-making process. However, provision can also be made to use a temporal difference learning (TD learning) algorithm. An LV agent 200 with a TD learning algorithm, the adaptation of the actions A +, A0, A- does not only take place when he receives the reward, but after each action a _i on the basis of an estimated expected reward. In addition, algorithms such as Q-Learning and SARSA or Monte Carlo simulations are also conceivable.

According to the method and the system of the present invention, reinforcement learning is used to automatically calibrate the cylinder air charge of an internal combustion engine, in particular of a reciprocating piston engine. Since various parameters p _i such as the number of revolutions or the oxygen content of the gas mixture are included in the calculation process and actions a _{i are} selected for this, in particular non-linear relationships between these parameters p _{i can be} recorded and taken into account, which is hardly possible in conventional control methods. It is therefore an optimized calibration method, since a variation of various parameters p _i , which have an influence on the cylinder air charge and interact with one another, is taken into account. Because the LV agent 200 selects the actions a _i himself and receives a reward for each of them, he can provide an optimal calibration strategy for the cylinder air charge.

Since the cylinder air charge is calibrated automatically and at the same time during operation of the internal combustion engine, the performance of the internal combustion engine can be increased, since the cylinder air charge can be individually adapted to the respective engine. This can lead to a reduction in pollutant emissions and thus to an improved environmental balance. In addition, the fuel consumption of the internal combustion engine can be reduced. In addition, costs for a previous manual and therefore time-consuming calibration can be reduced.

In 2 the method steps for calibrating the cylinder air charge in at least one cylinder of an internal combustion engine are shown.

In one step S10 receives a learning reinforcement agent 200 a state s _{i of} an internal combustion engine from a state module 300 , wherein a state s _{i is} defined by data and measured values of at least one parameter p _{i of} the internal combustion engine and the value of the cylinder air _charge depends at least partially on this parameter p _i .

In one step S20 the course agent chooses 200 for the state s _i at least one calculation function f _i and / or an action a _i for the parameter p _i .

In one step S30 the LV agent calculates a modeled value Z _M for the cylinder _air charge using the selected calculation function f _i and / or the action a _i .

In one step S40 becomes a real value Z _R for the cylinder _air charge from an environmental module 400 provided based on measurement results.

In one step S50 the modeled value Z _M for the cylinder _air charge and the real value Z _R for the cylinder air charge are compared with one another.

In one step S60 the comparison result is sent to the reward module 500 passed on and a reward for the comparison result is in the reward module 500 determined.

In one step S70 becomes the reward for the comparison result to the learning reinforcement agent 200 passed and another action a _j is taken by the learning reinforcement agent 200 is selected based on the reward for the comparison result for the state s _i .

3 schematically illustrates a computer program product 700 represents an executable program code 750 that is configured to carry out the method according to the first aspect of the present invention when carried out.

With the method according to the present invention, reliably optimized states s _{i can be found} by selecting suitable actions a _i , for example to reliably and automatically calibrate a control and regulating device for setting the cylinder air charge of an internal combustion engine, in particular a reciprocating piston engine. Through the use of a learning reinforcement agent 200 With an algorithm of reinforcing learning, an autonomous adjustment and calibration of the cylinder air charge for an individual motor vehicle is made possible autonomously and automatically.

Claims (15)

A method for calibrating a cylinder air charge in at least one cylinder of an internal combustion engine, in particular a reciprocating piston engine, comprising: determining (S10) a state (s _i ) of the internal combustion engine by a state module (300), a state (s _i ) being determined by data and measured values is defined by at least one parameter (p _i ) of the internal combustion engine and the value of the cylinder air charge depends at least in part on this parameter (p _j ); - Selecting (S20) a calculation function (f _i ) and / or an action (a _i ) for the parameter (p _i ) from a learning reinforcement agent (200); - Calculating (S30) a modeled value (Z _M ) for the cylinder air charge from the learning reinforcement agent (200) by means of the selected calculation function (f _i ) and / or the action (a _j ); - Providing (S40) a real value (Z _R ) for the cylinder air charge from an environmental module (400) based on measurement results; - comparing (S50) the modeled value (Z _M ) for the cylinder air charge with the real value (Z _R ); - Forwarding (S60) of the comparison result to a reward module (500) and determining a reward for the comparison result; - Forwarding (S70) of the reward for the comparison result to the learning reinforcement agent (200) and, based on this reward, again selecting a further action (a _j ) by the learning reinforcement agent (200). Procedure according to Claim 1 , sensors (420, 440) and / or measuring devices for detecting measured values of the parameters (p _i ) being provided. Procedure according to Claim 1 or 2 , wherein a positive action (A +), which increases the value for a parameter (p _i ), is provided, wherein a neutral action (A0), in which the value of the parameter (p _i ) remains the same, is provided, and where a negative action (A-), in which the value of the parameter (p _i decreases, is provided. Method according to one of the preceding Claims 1 to 3 wherein the reward module (500) comprises a database (520) or matrix for the evaluation of the actions (a _i ). Method according to one or more of the preceding Claims 1 to 4th wherein the learning reinforcement agent (200) is configured to use an algorithm from the reinforcement learning. Procedure according to Claim 5 , whereby the algorithm is designed as a Markov decision process or as Temporal Difference Learning (TD-Learning) or as Q-Learning or as DYNAQ or as SARSA or as Monte Carlo simulation. Method according to one or more of the Claims 1 to 6th , wherein at least one parameter (p _i ) represents an engine speed, the composition of a gas mixture, a temperature of the cylinder, and / or the lambda value of a lambda probe. A system (100) for calibrating a cylinder air charge in at least one cylinder of an internal combustion engine, in particular a reciprocating piston engine, with a status module (300) which is designed to determine a state (s _i ) of the internal combustion engine, a state (s _i ) through Data and measured values of at least one parameter (p _i ) of the internal combustion engine is defined and the value of the cylinder air charge depends at least partially on this parameter (p _i ), and with a learning reinforcement agent (200) which is designed, a calculation function (f _i ) and / or to select an action (a _i ) for the parameter (p _i ) and to calculate a modeled value (Z _M ) for the cylinder air charge using the selected calculation function (f _i ) and / or the action (a _i ); and with an environment module (400) which is designed to provide a real value (Z _R ) for the cylinder air charge based on measurement results, the state module (300) being designed to provide the modeled value (Z _M ) for the cylinder air charge with the real value (Z _R ), and with a reward module (500) which is designed to determine a reward for the comparison result and to pass on the reward for the comparison result to the learning reinforcement agent (200), the learning reinforcement agent (200) is designed to select another action (a _j ) based on this reward. System (100) according to Claim 8 , sensors (420, 440) and / or measuring devices for detecting measured values of the parameters (p _i ) being provided. System (100) according to Claim 8 or 9 , wherein a positive action (A +), which increases the value for a parameter (p _i ), is provided, wherein a neutral action (A0), in which the value of the parameter (p _i ) remains the same, is provided, and where a negative action (A-), in which the value of the parameter (p _i ) is reduced, is provided. System (100) according to one of the preceding Claims 8 to 10 wherein the reward module (500) comprises a database (520) or matrix for the evaluation of the actions (a _i ). System (100) according to one or more of the preceding Claims 8 to 11 wherein the learning reinforcement agent (200) is configured to use an algorithm from the reinforcement learning. System (100) according to Claim 12 , whereby the algorithm is designed as a Markov decision process or as Temporal Difference Learning (TD-Learning) or as Q-Learning or as DYNAQ or as SARSA or as Monte Carlo simulation. System (100) according to one or more of the Claims 8 to 13 , wherein at least one parameter (p _i ) represents an engine speed, the composition of a gas mixture, a temperature of the cylinder, and / or the lambda value of a lambda probe. A computer program product (700) comprising an executable program code (750) which is configured such that, when it is executed by a processor, it executes the method according to one of the Claims 1 to 7th executes.

Images