What is sometimes missing from the machine intelligence systems is that machine intelligence is an integrated part of the system design. Let we think of a complex IT system. It could be pretty much systematically designed in a way that certain parts are self-adaptable entities, meaning that they can be systematically trained and other parts are rather fixed. It is a general system design question as well which parts of the information system can be trained or for which parts of the information system make sense to be trained.
So as a basic system model let we imagine something object oriented system in which we have several objects or entities that practically contain some properties functions or constraints. As the most simple case let we imagine a self-adaptable entity as a set of {P1,P2, .. PO} general properties and a set of {F1,F2, .. FO} functions, in which each function maps some input parameters to some output ones based on the general properties. These functions are however not fix implemented but they are trained with some measured data. There might be a continuously learning designed, however it is probably a better system design to make a separation between different stages, like:
- performing phase: in which the system is working with static functions.
- learning phase: in which the system is adapting based on some measured or generated data.
- new performing phase: the result of the previous learning phase is used as static function, implying practically a new release of the system.
The functions themselves might follow different machine learning patterns, ranging from classical stochastic functions via artificial neural networks to more symbolic decision tree learning algorithms.
The model can be approved further to consider not just self-adaptable entities but self-adaptable connections between different entities or even inheritance between entities.
- performing phase: in which the system is working with static functions.
- learning phase: in which the system is adapting based on some measured or generated data.
- new performing phase: the result of the previous learning phase is used as static function, implying practically a new release of the system.
The functions themselves might follow different machine learning patterns, ranging from classical stochastic functions via artificial neural networks to more symbolic decision tree learning algorithms.
The model can be approved further to consider not just self-adaptable entities but self-adaptable connections between different entities or even inheritance between entities.
