Neural Network inputs
in [Matlab]
|
Prev: parameter optimization for Simulink models with S functions
Next: xPC Target problem: Target disconnects during certain Host operations
From: Jehad Zughyer on 14 Apr 2010 17:02 I am working on constructing a neural network for function approximation. I appreciate if someone can help direct me to Matlab material about : 1: how to introduce input parameters, I have binary and continuous parameters, do I need to preprocess these inputs first? how would the network know that a 0 0 1 0 0 for example is a binary input vector and it won't try to approximate values of input between 0:1, for example 0.5. 2: my data will be introduced over time, I receive a data point ( input vector), so I would like the net to learn this new input in addition to keeping the old learned inputs, and then I need to make a decision, then I receive another data point.... So I will alternate between training and simulating, one input at a time. I will ask specific questions later if I can’t find the training material which explain the upper two points. Thanks, Jehad
From: Greg Heath on 14 Apr 2010 18:47
On Apr 14, 5:02 pm, "Jehad Zughyer" <ad6...(a)wayne.edu> wrote: > I am working on constructing a neural network for function approximation. I appreciate if someone can help direct me to Matlab material about : Are you using the NN Toolbox? If so, see online documentation and demos doc nnet help nnet and/or the online documentation and demos at the mathworks website. Also see my posts on pretraining advice and hidden node determination. Keywords greg heath pre training advice for newbies greg heath Neq Nw > 1: how to introduce input parameters, I have binary and continuous parameters, variables; not parameters. > do I need to preprocess these inputs first? Yes. See above. > how would the network know that a 0 0 1 0 0 for example is a binary input vector > and it won't try to approximate values of input between 0:1, for example 0.5. Why would you think a net would try to approximate input variable values? The training algorithm is presented with input vectors and target vectors any of the components can be binary or real. The most important thing is that there is a deterministic trend relationship between each input vector component and at least one output component. For information about variable scaling and coding see the comp.ai.neural-nets FAQ. > 2: my data will be introduced over time, I receive a data point ( input vector), so I would like the net to learn this new input in addition to keeping the old learned inputs, and then I need to make a decision, then I receive another data point.... So I will >alternate between training and simulating, one input at a time. Learning involves representatives of all possible input/output combinations. Therefore, given a trained net and a new training input/output pair to learn will require using a saved calibration set in addition to the new pair in order to make sure the net doesn't unlearn previous I/O associations. > I will ask specific questions later if I can’t find the training > material which explain the upper two points. Good Luck. Greg |