Apparatus and method for generating first agent and apparatus and method for training at least one model to generate first agent
Through multi-layer neural network model and reinforcement learning, the problem of inefficient generation and training of first agent interaction behavior in the prior art is solved, and efficient behavior generation and interpretability under rare or extreme conditions is achieved.
Patent Information
- Application Number
- CN202510145007.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-02-09
- Filing Date
- 2025-02-10
- Publication Date
- 2025-08-12
AI Technical Summary
The prior art is difficult to efficiently generate and train the behavior of the first agent in interaction with the second agent, especially in rare or extreme conditions, and the existing methods are inefficient.
A multi-layer neural network model is adopted, including a pre-trained first model, a second model and a third model, and the behavior of the second agent is described by mapping the behavior of the representation and ultimately affecting the first agent, using natural language or formal language to describe the interactive behavior, and training the behavior of the second agent through reinforcement learning.
The behavior of efficiently generating the first agent under rare or extreme conditions is achieved, which improves the interpretability and data efficiency of interactive behavior and reduces the training data requirements.
Smart Images
Figure CN120470148A_ABST
Abstract
Description
Technical Field
[0001] The present invention is based on a device and a method for generating a first agent, in particular for an interaction between the first agent and a second agent, and a device and a method for training at least one model to generate the first agent. Summary of the Invention
[0002] The method for generating a first agent, in particular for an interaction between the first agent and the second agent, provides for mapping a description of the behavior of the first agent, in particular in the interaction between the first agent and the second agent, into a first representation using a first model. wherein the first model is designed to map the description to the first representation, wherein the first representation is mapped to the second representation using a second model, wherein the second model is designed to map the first representation to the second representation, wherein the second representation is mapped to an output variable using a third model, wherein the third model is designed to map the second representation to the output variable in order to influence the behavior of the first agent, wherein the description is predetermined in natural language, in particular in text form or audio form, or in a formal language (in formeller Sprache) or in the form of digital graphics, wherein the behavior of the first agent, in particular in the interaction between the first agent and the second agent, is predetermined in a manner that depends on the output variable. The corresponding model can be a stochastic model, a probabilistic model or a deterministic model.
[0003] For example, anomalies in the behavior of the second agent in an interaction between a first agent and a second agent are identified in a manner that depends on the interaction.
[0004] It may be provided that the output variable comprises a trajectory of the first agent; and / or that the output variable comprises a controller parameter of a controller for the first agent, wherein the behavior of the first agent is determined in a manner dependent on the behavior of the controller in the first agent.
[0005] It may be provided that the first model comprises a pretrained artificial neural network and / or the second model comprises a pretrained artificial neural network and / or the third model comprises a pretrained artificial neural network.
[0006] A method for training at least one model to generate a first agent, in particular for an interaction between the first agent and the second agent, provides for mapping a description of the behavior of the first agent, in particular in an interaction between the first agent and the second agent, to a first representation using a first model, wherein the first model is designed to map the description to the first representation, wherein the first representation is mapped to a second representation using a second model, wherein the second model is designed to map the first representation to the second representation, wherein the second representation is mapped to an output variable using a third model, wherein the third model is designed to map the second representation to the output variable, wherein the description is predetermined in a natural language or in a formal language or in the form of a digital graphic, wherein a reference for the output variable and the description are predetermined, wherein the reference characterizes a behavior of the first agent that is realistic in the real world and conforms to the description, and wherein the second model is trained based on the difference between the output variable and the reference. Thus, the second model is trained for the output of the output variable, wherein the output variable defines a behavior of the first agent that is as physically realistic as possible in the real world and conforms to the description.
[0007] It can be provided that the first model comprises a pretrained artificial neural network and / or the third model comprises a pretrained artificial neural network. This means that the training is based on a model that is already available.
[0008] It can be provided that the first model and / or the third model remain unchanged during training. Thus, the second model is trained exclusively. This requires less training data than if the second model were trained together with the first model and / or the second model.
[0009] It may be provided that the reference includes the trajectory of the first agent and / or controller parameters for a controller of the first agent.
[0010] The apparatus for generating an interaction or for training at least one model or for training a first agent to perform an interaction between a first agent and a second agent comprises at least one processor and at least one memory, wherein the at least one processor is designed to execute instructions, and when the at least one processor executes these instructions, the apparatus performs the method, wherein the at least one memory stores these instructions.
[0011] A data structure comprises at least one data field for describing in a natural language or a formal language the behavior of a first agent, in particular in an interaction between the first agent and a second agent, wherein the data structure comprises at least one data field for a first representation of the description, wherein the data structure comprises at least one data field for a second representation of the description, wherein the data structure comprises at least one data field for an output variable for influencing the behavior of the first agent.
[0012] It can be provided that the data structure comprises at least one data field for a first model, which is designed to map the description to a first representation, wherein the data structure comprises at least one data field for a second model, which is designed to map the first representation to a second representation, and / or wherein the data structure comprises at least one data field for a third model, which is designed to map the second representation to an output variable.
[0013] A computer program may be provided, comprising computer-executable instructions which, when executed by a computer, perform the method on the computer. BRIEF DESCRIPTION OF THE DRAWINGS
[0014] Further advantageous embodiments can be found in the following description and the accompanying drawings. In the drawings:
[0015] Figure 1 A schematic diagram showing an apparatus for machine learning or generating behavior of a first agent;
[0016] Figure 2 A schematic diagram showing a model for machine learning or generating behavior of a first agent,
[0017] Figure 3 A schematic diagram showing an exemplary interaction,
[0018] Figure 4 showing a flowchart having steps for a method for generating behavior of a first agent,
[0019] Figure 5 A flow chart with steps of a method for machine learning is shown. DETAILED DESCRIPTION
[0020] exist Figure 1 Schematically shown in FIG. 1 is an apparatus 100 for machine learning or generating a behavior of a first agent, in particular in an interaction between the first agent and a second agent.
[0021] The apparatus 100 includes at least one processor 102 and at least one memory 104 .
[0022] The at least one processor 102 is designed to execute the following instructions. When the instructions are executed by the at least one processor 102, the apparatus 100 performs the method described below for machine learning or for generating the behavior.
[0023] The at least one memory 104 stores the instructions. The at least one memory 104 includes, for example, a non-volatile memory. The at least one memory 104 includes, for example, a volatile memory.
[0024] In this example, the apparatus 100 comprises an interface 106. The interface 106 is designed, for example, to receive a description of the behavior.
[0025] The interface 106 can be designed to record the description in text form, audio form or digital graphic form.
[0026] The interface 106 may be designed to request the description to be output in text form or audio form. For example, the device 100 may be designed to request and record the description in a conversation with the user.
[0027] For example, the apparatus 100 is designed to automatically query the information required for the description and add it to the description. For example, the apparatus 100 is designed to generate the description when the apparatus 100 recognizes that the information required for the description has been recorded.
[0028] Figure 2 A model for machine learning or generative behavior is schematically shown in FIG.
[0029] In this example, the first model 202 is designed to map input variables 204 of the first model 202 to output variables 206 of the first model 202. The first model 202 is, for example, an encoder, or a sequential and / or autoregressive encoder or transformer architecture.
[0030] In this example, the second model 208 is designed to map the output variable 206 of the first model 202 to the output variable 210 of the second model 208. The second model 208 is, for example, a converter It converts the output variables 206 of the first model 202 into output variables 210 of the second model 208 , ie, input variables for the third model 212 .
[0031] In this example, the third model 212 is designed to map the output variables 210 of the second model 208 to the output variables 214 of the third model 212. The third model 212 is, for example, a decoder.
[0032] In this example, input variables 204 of first model 202 include descriptions in natural language, formal language, or digital graph form. In this example, output variables 206 of first model 202 include a first representation. For example, the first representation is a first sequence of tokens or a first embedding. In this example, output variables 210 of second model 208 include a second representation. For example, the second representation is a second sequence of tokens or a second embedding. Output variables 214 of third model 212 define behavior.
[0033] The first model 202 includes, for example, an artificial neural network. The second model 208 includes, for example, an artificial neural network. The third model 212 includes, for example, an artificial neural network.
[0034] The apparatus 100 comprises, for example, a model. For example, the apparatus 100 is designed to receive a description via the interface 106 and to determine output variables for the description using the model.
[0035] It can be provided that apparatus 100 is designed to perform a test. For example, apparatus 100 is designed to check, during a test, the behavior of the second agent in an interaction with the first agent that is dependent on the interaction. Apparatus 106 is designed, for example, to identify an anomaly in the interaction during the test and, as a result of the test, output, via interface 106, the presence of an anomaly or output, via interface 106, that no anomaly was identified during the test.
[0036] The interaction is not limited to agents to be tested in the test or agents that can be moved synthetically in the interaction. A plurality of agents can be provided that are tested by the interaction in the test. A plurality of agents can be provided that can be moved synthetically in the interaction during the test. The description, for example, describes the behavior of the respective agent to be moved synthetically. The behavior of the respective agent that can be moved synthetically is defined, for example, by output variables 214.
[0037] exist Figure 3 , an exemplary scene 300 is illustrated as an example of the interaction, including a first vehicle 302 as an example of the first agent and a second vehicle 304 as an example of the second agent. Scene 300 includes a trajectory 306 of first vehicle 302 and a trajectory 308 of second vehicle 304.
[0038] The scene 300 is illustrated graphically for the following exemplary description in natural language:
[0039] A highway entrance scenario where a first vehicle moving at the highway entrance must merge behind a second vehicle traveling on the highway
[0040] In this example, first vehicle 302 includes at least one sensor for recording sensor data characterizing an environment of vehicle 302 .
[0041] In this example, the first vehicle 302 includes a regulator for regulating the behavior of the first vehicle 302 based on sensor data. The regulator is parameterized, for example by regulator parameters.
[0042] In the test, for example, second vehicle 304 is examined in interaction with first vehicle 302 .
[0043] The graphical representation of scenario 300 shows a trajectory 306 in digital graphical form as an exemplary description of the behavior of first vehicle 302. This graphical representation is exemplary. Apparatus 100 can be designed to use scenario 300 in a formal language that can be automatically processed in a test environment, such as a test bench, or in a simulation environment where the test is performed.
[0044] It can be provided that the description provided as input variable 204 of first model 202 is supplemented by information about framework conditions in natural language. These frameworks define, for example, the geometry of the highway, weather conditions such as dry, windy, rainy, snowy, icy, visibility conditions such as foggy, dark, bright, and traffic regulations such as speed limits and prohibitions on overtaking in natural language.
[0045] For example, scenario 300 includes a description of a map, including boundary conditions. For example, the description of the map can be performed in a formal language that can be automatically processed by a test bench or simulation environment.
[0046] Figure 4 A flow chart with the steps of the method is shown in FIG.
[0047] In this example, the agents are traffic participants. A first agent is, for example, a first vehicle 302 . A second agent is, for example, a second vehicle 304 .
[0048] In an example in robotics, the first agent is a robot and the second agent is a human model.
[0049] The method is based on a predefined description.
[0050] The description is predefined, for example, in natural language, particularly text or audio form, or in a formal language or digital graphic form. For example, the user enters the description in text form, or speaks the description in audio form. The user draws the description in digital graphic form (e.g., in the form of a sketch).
[0051] For example, the description is requested via output in text or audio form, for example, the description is requested and recorded in a conversation with the user.
[0052] For example, the information required for the description is automatically retrieved and added to the description.
[0053] For example, the description is generated only when it is recognized that information required for the description has been recorded.
[0054] For example, descriptions of interactions that are rare in the real world are recorded, in particular through language. An example of a rare interaction is when the second agent is an autonomous vehicle, where the first agent is another vehicle, and the other vehicle unexpectedly drives in front of the autonomous vehicle from a parking space that the autonomous vehicle cannot see. An example of a rare interaction is an interaction with an aggressive driving style of an agent, such as when the agent merges into traffic of other agents on a highway. An example of a rare interaction is an interaction under extreme weather conditions in which the first agent and / or the second agent is located. An example of a rare interaction is an interaction under very rare environmental conditions, such as on a day when the first agent represents a human in costume, such as on Halloween or a carnival or a fun day. In this context, rare means that the interaction occurs very rarely in the training data recorded in the real world.
[0055] The method is based on a first model 202, a second model 208, and a third model 212. In this example of the method, the first model 202, the second model 208, and the third model 212 are predetermined. For example, the corresponding artificial neural network is pretrained.
[0056] The first model 202 is, for example, a model pre-trained for textual input and / or audio input, and is designed to map a description in textual or audio form to a first representation. Compared to a single model trained to directly map the description to the output variable 214, using the pre-trained first model 202 can achieve higher data efficiency.
[0057] The behavior of the first agent is generated, for example, in a specific domain, such as transportation or robotics. For example, the first model 202 and / or the third model 212 are pre-trained in other domains or without content specialization (spezialisierung) in the domain in which the behavior of the agent is generated, and the domain in which the first model 202 and / or the third model 212 are pre-trained is a domain different from the domain in which the behavior of the first agent is generated, for example a more general domain.
[0058] The pre-trained first model 202 and / or the pre-trained third model 212 enable knowledge transfer, in particular from the domain in which the first model 202 or the third model 212 was pre-trained to the domain in which the first agent is generated. For example, the respectively pre-trained model may include general knowledge about the behavior of the first agent. For example, if the respectively pre-trained model was pre-trained with more data containing human behavior, the behavior of the pedestrian represented by the first agent can be better generated based on general knowledge. In the subsequent training of the respectively pre-trained model for generating the first agent, less data containing pedestrian behavior is then needed to generate the behavior of the first agent in a relatively good manner, which would otherwise be possible only with more data containing pedestrian behavior.
[0059] The method includes step 402 .
[0060] In step 402 , a predetermined description is mapped to a first representation using the first model 202 .
[0061] The method includes step 404 .
[0062] In step 404 , the first representation is mapped to a second representation using the second model 208 .
[0063] The method includes step 406 .
[0064] In step 406 , the second representation is mapped to the output variable 214 using the third model 212 .
[0065] The method may comprise further steps for training and / or testing the behavior of the second agent.
[0066] Steps 402 to 406 are repeated, for example, to generate training data including output variables 214 for training and / or testing the behavior of the second agent. For example, a large number of output variables 214 are determined from a description pair catalog (Katalog) and added to the training data.
[0067] For example, the catalog includes descriptions of interactions in the domain. An example of a description of a domain is the operational design domain described in "Koopman, P., Osyk, B., Weast, J. (2019) Autonomous Vehicles Meet the Physical World: RSS, Variability, Uncertainty, and Proving Safety, In: Romanovsky, A., Troubitsyna, E., Bitsch, F. (eds) Computer Safety, Reliability, and Security. SAFECOMP 2019. Lecture Notes in Computer Science (), Volume 11698, Springer, Cham. https: / / doi.org / 10.1007 / 978-3-030-26601-1_17".
[0068] For example, the method for training includes step 408 .
[0069] In step 408 , the behavior of the second agent is trained in the interaction between the first agent and the second agent.
[0070] The behavior of the first agent is determined, for example, by a trajectory predetermined for the first agent in the scene.
[0071] The first agent may include a regulator, and the regulator is designed to determine the behavior of the first agent according to the scenario. The behavior of the first regulator is determined by the regulator in the first agent, for example.
[0072] For example, the second agent is trained by reinforcement learning, i.e., re-reinforcement learning. For example, if the second agent does not collide with the first agent moving on a predetermined trajectory, the second agent will receive a reward, otherwise it will not receive a reward.
[0073] The second agent may include a sensor and a regulator, wherein the regulator is configured to determine the behavior of the second agent based on information about the first agent measured by the sensor. The behavior of the second regulator is determined, for example, by the regulator in the second agent. For example, the regulator of the second agent is trained using reinforcement learning.
[0074] For example, the method for testing includes step 410 .
[0075] In step 410, the behavior of the second agent is monitored. Provision may be made for monitoring the behavior of the first agent, or for the behavior of the first agent generated using the model to serve as the basis for testing.
[0076] The behavior of the first agent is determined, for example, by a regulator or trajectory in the first agent.
[0077] For example, the method for testing includes step 412 .
[0078] In step 412, an anomaly in the behavior of the second agent is identified based on the interaction between the second agent and the first agent. For example, an anomaly in the behavior of the second agent is identified based on the interaction.
[0079] For example, when a first agent collides with a second agent, an anomaly in the behavior of the second agent is identified.
[0080] It may be provided that a plurality of agents are generated using the method, including their interaction behavior with each other and with respect to potential third parties. It may be provided that the generated agents interact with one or more external agents in a real test environment or in a simulation in order to test how the latter interact with the generated agents.
[0081] The test itself is performed, for example, in a simulation in which the agents interact or in a real test environment. The simulation is particularly preferred for tests that could lead to collisions, so as not to endanger human life or prevent damage to real agents.
[0082] For example, by describing a merging scenario in which two agents are driving with a small space between each other in a highway entrance area, behaviors of the two agents are generated, which drive on the highway according to the described merging scenario.
[0083] For example, external agents whose behavior is not generated by a description but is controlled by externally predefined controllers are tested to see whether they can still be imported without problems.
[0084] Additionally, it can also be provided that a test is carried out for the domain in which the first agent was generated and for a further description specified by the user.
[0085] It can be provided that, for example, further tests are carried out using interactions recorded in the real world.It can be provided that these interactions recorded in the real world are randomly selected from a set of predetermined interactions recorded in the real world.
[0086] Figure 5 A flowchart with steps of a method for training at least one of these models is shown in .
[0087] The training method is based on first model 202, second model 208, and third model 212. In this example, first model 202 and third model 212 are predefined in the method for training at least one of these models. For example, the corresponding neural networks of first model 202 and third model 212 are pretrained.
[0088] Due to the information stored in the pre-training and due to the density of information contained in language, the pre-trained model results in: knowledge transfer and / or higher data efficiency compared to a model learned only based on the domain from which the first agent was generated.
[0089] The method for training at least one of the models is based on training data. In an example, the training data comprises training data points, each of which comprises a predetermined description and a reference associated with the predetermined description. Furthermore, the method can be based on a pre-trained model.
[0090] The method for training at least one of the models includes step 502 .
[0091] In step 502 , descriptions and references from training data points are specified.
[0092] In step 504 , for this training data point, a predetermined description is mapped to a first representation using first model 202 .
[0093] The method for training at least one of the models includes step 506 .
[0094] In step 506 , for the training data point, the first representation is mapped to a second representation using the second model 208 .
[0095] The method for training at least one of the models includes step 508 .
[0096] In step 508 , for the training data point, the second representation is mapped to an output variable 214 using the third model 212 .
[0097] The reference includes, for example, output variables for a behavior that is realistic in the real world and corresponds to the description. The output variables include, for example, a trajectory defining the behavior of the first agent, controller parameters for a controller of the first agent, and / or a map defining boundary conditions for the behavior of the first agent.
[0098] In this example, steps 502 to 508 are performed for training data points from the training data.
[0099] The method for training at least one of the models includes step 510 .
[0100] In step 510 , the second model 208 is trained based on the difference between the output variable 214 and the reference.
[0101] In this example, the second model 208 is trained according to an objective function including the respective differences determined for the training data points. For example, the second model 208 is determined in which the objective function depending on the sum of the differences is as small as possible, in particular minimum.
[0102] For example, the parameters of the neural network comprising the second model 208 are determined using a gradient descent method based on the differences.
[0103] In this example, first model 202 and third model 212 remain unchanged during the training of second model 202. Provision may be made to also train first model 202 and / or third model 212 during the training.
[0104] The output variables of these models are summarized, for example, in the form of vectors. For example, the controller parameters or trajectories in the scenario are described by the vector output by the third model 212.
[0105] For example, the map is described by a vector output by the third model 212 , wherein the vector includes parameters of a formal language describing the map.
[0106] The first model 202 , the second model 208 , and / or the third model 212 may be stochastic models, probabilistic models, or deterministic models.
[0107] In this example, the first model 202, the second model 208, and the third model 212 are designed to perform the mapping in reverse order. Because models are language-dependent, performing the mapping in reverse order enables interpretability of behavior. This means that the third model 212 is designed to map the output variables 214 of the third model 212 to the output variables 210 of the second model 108. The second model 208 is designed to map the output variables 210 of the second model 208 to the output variables 206 of the first model 202. The first model 202 is designed to map the output variables 206 of the first model 202 to the input variables 204 of the first model 202.
[0108] For the reverse mapping, it can be provided that a plurality of images of a video are used successively for the output variable 214 of the third model 212, wherein the video is mapped to a textual description of the content of the video. This means that a suitable textual description is generated from the action video.
[0109] The method for explaining the behavior of a second agent, in particular in interaction with a first agent, provides that output variables 214 of third model 212 comprise the behavior to be explained.
[0110] The method for interpretation provides that output variables 214 of third model 212 are mapped to output variables 210 of second model 108 .
[0111] The method for interpretation provides that output variables 210 of second model 208 are mapped to output variables 206 of first model 202 .
[0112] The method for explanation provides that output variables 206 of the first model 202 are mapped to input variables 204 of the first model 202. The input variables 204 of the first model 202 include a description of the behavior to be explained, in particular of the interaction between a first agent and a second agent to be explained.
Claims
1. A method for generating a first agent (302), in particular for an interaction (300) between the first agent (302) and a second agent (304), characterized in that A description of a behavior of a first agent (302), in particular in an interaction (300) between the first agent (302) and a second agent (304), is mapped (402) to a first representation using a first model (202), wherein the first model is designed to map the description to the first representation, wherein the first representation is mapped (404) to the second representation using a second model (208), wherein the second model is designed to map the first representation to a second representation, wherein the second representation is mapped (406) to an output variable using a third model (212). quantity (214), wherein the third model (212) is designed to map the second representation to an output variable (214) in order to influence the behavior of the first agent (302), wherein the description is predetermined in natural language, in particular in text form or audio form, or in formal language or in digital graphic form, wherein the behavior of the first agent (302), in particular in the interaction (300) between the first agent (302) and the second agent (304), is predetermined (408) in a manner dependent on the output variable (214).
2. The method according to claim 1, characterized in that Anomalies in the behavior of the second agent (304) in an interaction between the first agent (302) and the second agent (304) are identified (412) in a manner that is dependent on the interaction (300).
3. The method according to claim 1 or 2, characterized in that The output variable (214) includes a trajectory of the first agent (302); and / or the output variable (214) includes a regulator parameter for a regulator of the first agent (302), wherein the behavior of the first agent (302) is determined (410) in a manner that depends on the behavior of the regulator in the first agent (302).
4. The method according to any one of the preceding claims, characterized in that The first model (202) includes a pre-trained artificial neural network and / or the second model (208) includes a pre-trained artificial neural network and / or the third model (212) includes a pre-trained artificial neural network.
5. A method for training at least one model to generate a first agent (302), in particular for an interaction (300) between the first agent (302) and a second agent (304), characterized in that A description of the behavior of the first agent (302), in particular in an interaction (300) between the first agent (302) and the second agent (304), is mapped (504) to a first representation (304) using a first model (202), wherein the first model is designed to map the description to the first representation, wherein the first representation is mapped (506) to a second representation using a second model (208), wherein the second model is designed to map the first representation to a second representation, wherein the second representation is mapped (508) to the output variable (212) using a third model (212). 14), wherein the third model is designed to map the second representation to the output variable (214) to influence the behavior of the first agent (302), wherein the description is predetermined in a natural language or in a formal language or in the form of a digital graphic, wherein a reference for the output variable (214) and the description are predetermined (502), wherein the reference characterizes the following behavior of the first agent (302), wherein the behavior is realistic in the real world and conforms to the description, and wherein the second model is trained (510) based on the difference between the output variable (214) and the reference.
6. The method according to claim 5, characterized in that The first model (202) includes a pre-trained artificial neural network and / or the third model (212) includes a pre-trained artificial neural network.
7. The method according to claim 6, characterized in that The first model (202) and / or the third model (212) remain unchanged during training (510).
8. The method according to any one of claims 5 to 7, characterized in that The reference includes a trajectory of the first agent (302) and / or regulator parameters for a regulator of the first agent (302).
9. A device (100) for generating an interaction or for training at least one model or for training a first agent for an interaction (300) between said first agent (302) and a second agent (304), characterized in that The device (100) comprises at least one processor (102) and at least one memory (104), wherein the at least one processor (102) is designed to execute instructions, and when the at least one processor (102) executes the instructions, the device (100) performs the method according to any one of claims 1 to 8, wherein the at least one memory (104) stores the instructions.
10. A data structure characterized in that The data structure comprises at least one data field for describing in a natural language or a formal language the behavior of the first agent (302), in particular in an interaction between the first agent (302) and the second agent (304), wherein the data structure comprises at least one data field for a first representation of the description, wherein the data structure comprises at least one data field for a second representation of the description, wherein the data structure comprises at least one data field for an output variable (214) for influencing the behavior of the first agent (302).
11. The data structure according to claim 10, characterized in that The data structure comprises at least one data field for a first model (202), the first model being designed to map the description to the first representation, wherein the data structure comprises at least one data field for a second model (208), the second model being designed to map the first representation to the second representation, and / or wherein the data structure comprises at least one data field for a third model (212), the third model being designed to map the second representation to the output variable (214).
12. A computer program, characterized in that The computer program comprises computer-executable instructions, which, when executed by a computer, perform the method according to any one of claims 1 to 8 on the computer.