Neural Networks as Inference and Learning Engines

Abstract: This paper presents an abridgment of a neural network constructive methodology and applications with real data. The neural network can be considered as the learning core and inference engine of an expert system that produces either different network designs or simulations as output, its input being data sequences. Basically, it consists of additive structural learning, limiting it by a cross-validation technique.
Considerations about uncertainty treatment in neural networks are also presented, including uncertainty in data, in neuron activation, in outputs, and combination of several uncertainty sources.
Applications include three different sets of data, all of them related to the energy field. First, river streamflow estimation is discussed. Then CO₂ concentration prediction from gas injection rate is studied. Finally, the program learns to imitate a feedwater control system in a nuclear reactor. All tests show good results, as can be seen when compared with other standard methods.