Device and method for controlling a physical system
A method for controlling a physical system. The method includes training a neural network to output, for a plurality of tasks, a result of the task carried out, in each case in response to the input of a control configuration of the physical system and the input of a value of a task input parameter; ascertaining a control configuration for a further task with the aid of Bayesian optimization, the neural network, parameterized by the task input parameter, being used as a model for the relationship between control configuration and result; and controlling the physical system according to the control configuration to carry out the further task.
1. A method for controlling a physical system, comprising the following steps:
training a neural network to output, for a plurality of tasks carried out by the physical system, a result of the task carried out, in each case in response to input of a control configuration of the physical system and input of a value of a task input parameter, the training including the ascertaining of weights of the neural network and, for each of the tasks carried out, the value of the task input parameter;
ascertaining a control configuration for a further task using Bayesian optimization, successive evaluations of control configurations being carried out, during each of the evaluations, a result of an execution of the further task being ascertained for a respective control configuration, the neural network, parameterized by the task input parameter distributed according to a probability distribution, being used as a model for a relationship between control configuration and result and, using the evaluations, being successively updated in that the probability distribution of the task input parameter is conditioned on the evaluations; and
controlling the physical system according to the control configuration to carry out the further task, wherein the probability distribution of the task input parameter for the further task is conditioned on the evaluations in that probabilities for task input parameter values for which an output of the neural network is closer to results supplied by the evaluations is increased compared to probabilities for task input parameter values for which the output of the neural network is less close to the results supplied by the evaluations.
2. The method as recited in claim 1 , wherein the training of the neural network is carried out by supervised learning using a loss function which depends on the weights of the neural network and the values of the task input parameter for the tasks carried out.
3. The method as recited in claim 2 , wherein the loss function includes a regularization term which causes an empirical distribution of the values of the task input parameter which are trained for the tasks carried out to approximate a predefined probability distribution.
4. The method as recited in claim 3 , wherein the predefined probability distribution is a Gaussian distribution.
5. The method as recited in claim 1 , wherein the physical system includes one or multiple actuators, and the control of the physical system according to the control configuration for carrying out the further task includes control of the one or multiple actuators according to control parameter values given by the control configuration.
6. The method as recited in claim 1 , wherein the control configuration includes hyperparameters of a machine learning model which is implemented by the physical system.
7. The method as recited in claim 1 , wherein the further task is an image classification of digital images or a manufacture of a product.
8. A control unit configured to control a physical system, the control unit configured to:
train a neural network to output, for a plurality of tasks carried out by the physical system, a result of the task carried out, in each case in response to input of a control configuration of the physical system and input of a value of a task input parameter, the training including the ascertaining of weights of the neural network and, for each of the tasks carried out, the value of the task input parameter;
ascertain a control configuration for a further task using Bayesian optimization, successive evaluations of control configurations being carried out, during each of the evaluations, a result of an execution of the further task being ascertained for a respective control configuration, the neural network, parameterized by the task input parameter distributed according to a probability distribution, being used as a model for a relationship between control configuration and result and, using the evaluations, being successively updated in that the probability distribution of the task input parameter is conditioned on the evaluations; and
control the physical system according to the control configuration to carry out the further task, wherein the probability distribution of the task input parameter for the further task is conditioned on the evaluations in that probabilities for task input parameter values for which an output of the neural network is closer to results supplied by the evaluations is increased compared to probabilities for task input parameter values for which the output of the neural network is less close to the results supplied by the evaluations.
9. A non-transitory computer-readable memory medium on which are stored program instructions for controlling a physical system, the program instructions, when executed by one or more processors, causing the one or more processors to perform the following steps:
training a neural network to output, for a plurality of tasks carried out by the physical system, a result of the task carried out, in each case in response to input of a control configuration of the physical system and input of a value of a task input parameter, the training including the ascertaining of weights of the neural network and, for each of the tasks carried out, the value of the task input parameter;
ascertaining a control configuration for a further task using Bayesian optimization, successive evaluations of control configurations being carried out, during each of the evaluations, a result of an execution of the further task being ascertained for a respective control configuration, the neural network, parameterized by the task input parameter distributed according to a probability distribution, being used as a model for a relationship between control configuration and result and, using the evaluations, being successively updated in that the probability distribution of the task input parameter is conditioned on the evaluations; and
controlling the physical system according to the control configuration to carry out the further task, wherein the probability distribution of the task input parameter for the further task is conditioned on the evaluations in that probabilities for task input parameter values for which an output of the neural network is closer to results supplied by the evaluations is increased compared to probabilities for task input parameter values for which the output of the neural network is less close to the results supplied by the evaluations.