MATLAB: Neural networks for data prediction

data analyzingdata predictionDeep Learning Toolboxneural networkneural networks

As I am not advanced user of Matlab, maybe this is a stupid question for you.
Basically I have 6 experiments; each experiment has 9 parameters (x1,x2…x9) and one output value (Y).
Experiment matrix is following:
# x1 x2 x3 x4 x5 x6 x7 x8 x9 Y
All x parameters and Y parameter are experimental, and I want to use neural networks to build a relationship between all x and Y (e.g. formula that would consist of all x parameters, which would bring to Y parameter as close as possible.) And build a code, so that when I enter new x values, it would predict Y value.
I would appreciate any tips, how to get it done.
Thanks and best regards, Vitaly

Best Answer

  • [ I N ] = size(input) % [ 9 6 ]
    [ O N ] = size (target) % [ 1 6 ]
    Neq = prod(size(target)) % No. of EQUATIONS = N*O = 6
    % For a NN with I-H-O node topology, the No. of UNKNOWNS (weights and biases) is % Nw = (I+1)*H+(H+1)*O = I*H input layer weights, H*O output layer weights and H+O biases % Nw < Neq when H <= Hub where
    Hub = -1 + ceil( (Neq-O)/(I+O+1) )% = 0 (A Linear model)
    % However, for a linear NN with I-O node topology, the No. of UNKNOWNS (weights and biases) is % Nw = (I+1)*O % I*O = 9 weights + O = 1 bias = 10 > Neq = 6 equations (underdetermined system)
    % For a robust design Neq >> Nw (a VERY over determined system) is desired. Therefore, consider
    1. Much more data
    2. Fewer input variables ( e.g., help/doc STEPWISEFIT)
    3. BACKSLASH or PINV for an undetermined linear system
    3. Regularization (e.g., help/doc TRAINBR) .
    help fitnet
    doc fitnet
    Search Newsgroup and Answers
    greg fitnet
    Hope this helps
    Thank you for officially accepting my answer