These questions are best placed here!
I would like to build up a NN which predicts the conditions of a machine after a certain time period. My idea was to build up kind of a grey box model since some of the functional connections between input and output are know (e.g. linear function between certain input and output parameters), but not their coefficients. For some connections also the direction (input A influences output in a positive way) is known.
Do you know if NN with such a grey box are possible/feasible? Is it possible to use MemBrain for such a method?
I'm not sure if I get your point:Devil wrote:Do you know if NN with such a grey box are possible/feasible? Is it possible to use MemBrain for such a method?
Do you want to model the behaviour between inputs and outputs on basis of prior knowledge and then still train this net with data?
Why would you try to include prior knowledge into the model?
Instead, you should leave it up to the net and the data to find relationships between inputs and corresponding outputs. Invoking apparent knowledge of the system restricts the solution in a way that normally is not desired.
However, to answer your question: I don't think that the modeling possibilities of MemBrain will allow very sophisticated transfer functions to be defined on basis of prio knowledge. Still, you may want to check out the mathematical neuron and link model in the MemBrain help file in order to get an idea of what you could construct.
I tried to set up the model already as black box. However, it was not really sucessful since the variation of the input parameter showed a wrong influence on the output. The main problem is, that I have only a limited amount of datasets available for training (around 30) and I've read in open literature that including a prior knowledge into a NN (thus creating a grey box) can increase the confidence of the network even with a small amount of training datasets... however, I haven't found a tool yet which is supporting NN in combination with a priori knowledge.