| 摘要 |
The invention relates to a method for computer-assisted learning of control and/or regulation of a technical system, operation of the technical system being characterized by states which the technical system can adopt during operation, and actions which are executed during operation of the technical system and transform a respective state of the technical system into a follow-on state. In the inventive method the statistical uncertainty of the training data used during learning is suitably taken into account when learning control of the technical system. This takes place as a result of the fact that the statistical uncertainty of a quality function, which models an optimal operation of the technical system, is determined by means of an uncertainty propagation and is incorporated during learning of an action-selecting rule. The inventive method is characterized in that the uncertainty propagation uses a covariance matrix in which non-diagonal elements are ignored. This approximation makes the method very computationally efficient and requires little memory space for its execution. Despite the approximation the method still delivers very good results when determining a suitable action-selecting rule. The inventive method can be used for learning control or regulation of any desired technical systems. In a variant the method is used for control or regulation of the operation of a turbine, in particular a gas turbine. In another variant the method is used for regulation or control of a wind power plant.
|
