| 摘要 |
This invention is a system and iterative non-learning method to determine optimal artificial neural network node and layer count, edge connection structure and transfer function for an artificial neural network. Optimality is indicated by the learning effort fo r the network being minimum and the generalization of the artificial neural network on provided data being maximum. A control and display subsystem receives a count of input and outpu t external interface nodes and associated node names from a user. Said subsystem also accepts end- conditions for the training of a series of artificial neural networks and establishes, together with a data delivery agent and data mapping agent, a relationship between variables of a data set and input and output network nodes. A network configuration agent and configuration agent controller create a series of artificial neural networks, each network in a series having a different internal nodal structure or transfer function than others in the series. A training agent trains each configured artificial neural network over an epoch of training, for each modified artificial neural network in a series. A data-logger records the training progress. An analyzer computes the improvement or reduction of training efficiency and ability of a particular network structure to generalize on a provided data set, compared to a previous structure. The control and display subsystem, analyzer and configuration and training controller subsequently determine a probable best structure of artificial neural network for a subsequent iteration of network creation and training testing.
|
