Controlling Neural Nets

After six years of studying neural networks and their uses 90% are used for data mining purposes. I realized in the standard process of teaching nets not much emphasis has ben placed on the control aspect. For example say we have a kohonon net for classification in production. It's main use is to report to the user verification of specific requirements. Taking a different approach, we can train a net for a control system to perform various tasks, such as basic processes: jumping, running, evading, or coordinating movement with other nets. the tasks represent an output set for the net. How the various nets arrive at their output would be based on the linear, nonlinear and network composition systems. Too much emphasis has been placed on building better classifiers with little on no work on enhancing their capabilities.

The goal is not just to focus on teaching various patterns but to codify the patterns as defined commands. This command structure could be user controlled or dictated by the net's goal priorities. After basic commands can be accomplished more sophisticated command structures could be evolved to accomplish higher level goals. Such as search for a specific object and store it in a place where it can remember, and for example let the net build better storage systems for higher level memory functions or internal solution spaces. Both traditional OOP and Neural Net architectures could be developed into novel solutions. With the goal of increased intelligence and adaptation.

More enhanced solutions could be devised such as run fast and evade or organize two other nets to search for resources and share solutions from their solution space.