Method and device for learning a strategy and for implementing the strategy

A method for implementing a learned strategy which optimally adapts at least one parameter of an evolutionary algorithm, the evolutionary algorithm being a Covariance Matrix Adaptation Evolution Strategy (CMA-ES) or differential evolution algorithm, the strategy being learned by initializing the strategy, which ascertains a parameterization of the at least one parameter as a function of pieces of state information about a problem instance, and learning the strategy using reinforcement learning, the strategy being learned with a reward signal and from interactions of the CMA-ES algorithm (i) with the parameterization, the parameterization being determined using the strategy as a function of the pieces of state information and (ii) with the problem instance, wherein the parameterization is optimal for possible pieces of state information, the method comprising:

• determining the state information for the problem instance;
• ascertaining by the learned strategy a parameterization as a function of the state information, CMA-ES including the parameterization being applied for at least one iteration, wherein:
• the problem instance is a training method for machine learning, hyperparameters of the training method being optimized based on CMA-ES using the strategy,
• a deep neural network is trained using at least one of the optimized hyperparameters to produce a trained neural network,
• the trained neural network ascertains an output variable and controls an actuator that controls a physical operation of a technical system, wherein the output variable is used to ascertain a control variable for controlling the actuator.

A device for learning a strategy, which optimally adapts at least one parameter of an evolutionary algorithm, the evolutionary algorithm being a Covariance Matrix Adaptation Evolution Strategy (CMA-ES) or differential evolution algorithm, the device configured to:

• initialize the strategy, which ascertains a parameterization of the at least one parameter as a function of pieces of state information about a problem instance;
• and
• learn the strategy using reinforcement learning, the strategy being learned with a reward signal and from interactions of the CMA-ES algorithm (i) with the parameterization, the parameterization being determined using the strategy as a function of the pieces of state information and with the problem instance, wherein the parameterization is optimal for possible pieces of state information, wherein:
• the problem instance is a training method for machine learning, hyperparameters of the training method being optimized based on CMA-ES using the strategy,
• a deep neural network is trained using at least one of the optimized hyperparameters to produce a trained neural network,
• the trained neural network ascertains an output variable and controls an actuator that controls a physical operation of a technical system, wherein the output variable is used to ascertain a control variable for controlling the actuator.

A non-transitory machine-readable memory medium on which is stored a computer program for learning a strategy, which optimally adapts at least one parameter of an evolutionary algorithm, the evolutionary algorithm being a Covariance Matrix Adaptation Evolution Strategy (CMA-ES) or differential evolution algorithm, the computer program, when executed by a computer, causing the computer to perform the following steps:

• initializing the strategy, which ascertains a parameterization of the at least one parameter as a function of pieces of state information about a problem instance;
• and
• learning the strategy using reinforcement learning, the strategy being learned with a reward signal and from interactions of the CMA-ES algorithm (i) with the parameterization, the parameterization being determined using the strategy as a function of the pieces of state information and (ii) with the problem instance, wherein the parameterization is optimal for possible pieces of state information, wherein:
• the problem instance is a training method for machine learning, hyperparameters of the training method being optimized based on CMA-ES using the strategy,
• a deep neural network is trained using at least one of the optimized hyperparameters to produce a trained neural network,
• the trained neural network ascertains an output variable and controls an actuator that controls a physical operation of a technical system, wherein the output variable is used to ascertain a control variable for controlling the actuator.