| 摘要 |
2066170 9202896 PCTABS00010 A modular neural network process control system and method with natural language configuration allows neural network functions to be easily configured and operated without programming expertise. Neural networks are provided as standard modules in a modular control system architecture. Each module has a storage area for holding module parameters and for holding a pointer to a neural network procedure. The procedure carries out the real time training and prediction functions of the neural network, and can be pointed to by multiple neural network modules. For neural network parameters, such as weight matrices, requiring large storage, the module storage also holds a pointer to a neural network parameter storage area. Partially defined neural network functions are presented to the user as a template, and the user completes the specification of the neural network functions by entering data into the template. Each type of partially defined function corresponds to a generic neural network procedure which is used to execute the module function. Or, neural network specifications can be accepted in a limited set of substantially natural language constructs, which are parsed into symbolic specifications. A neural network procedure is created by combining code fragments needed to implement the symbolic specifications.
|
