| 摘要 |
A trainable, state-sampled, network controller (TSSNC) or state-sampled controller (SSC) requires little information regarding a plant (as with neural networks), but can use what information is available (as in classical controllers), and provides a linear network (as for CMAC) improving calculation speeds. A form of a governing differential equation characterizing a plant may include parameters and their derivatives of various orders as variables combined in linear and nonlinear terms. Classical control theory, and a method such as a Fourier transform of governing equations, may provide 8a form of a control law, linear in certain weights or coefficients. Knowledge of coefficients is not required for either the form of the governing equations or the form of the control law. An optimization method may be used to train the SSC, defining a table of weights (contributions to coefficients) to be used in the matrix equation representing the control law the solution yielding a control output to the plant. Sampling plant outputs, during training, may be done at a selected spatial frequency in state space (each dimension a variable from the control law). Sampling is used to provide ideal interpolation of the weights over the entire range of interest. Minimum memory is used with maximum accuracy of interpolation, and any control/output value may be calculated as needed in real time by a minimal processor.
|
