Taehee Lee wrote that,
>But i have difficulty in finding the time phase model of SEM.
> Especially, the relationship of each variable in SEM which might be called
Spatial
> model of system is general problem, but the time lag for affecting with
each
> variable called Temporal model is very difficult problem.
>  If anyone knew the methods or theorical foundation for calculating that
problem, please inform me by Email.


 Very accurate predictions of NY Stock Exchange indexes and models
 of another different complex objects were obtained with the aid of
 twice-multilayered neuronet with active neurons, where each neuron
 is modelling GMDH algorithm. In this approach, which corresponds
 to the actions of human nervous system, the connections between
 several neurons are not fixed but change depending on
 the neurons themselves.

 The Group Method of Data Handling (GMDH) is self-organizing approach
 based on sorting-out of gradually complicated models and evaluation of
 them by external criterion on separate part of data sample. As input
 variables can be used any parameters, which can influence on the
 process. Computer is found structure of model and measure of
 influence of parameters on the output variable itself. That model is
 better that leads to the minimal value of external criterion. This
 inductive approach is different from commonly used deductive
 techniques or neural networks.

 The GMDH was developed for complex systems modeling, prediction,
 identification and approximation of multivariate processes, decision
 support after "what-if" scenario, diagnostics, pattern recognition and
 clusterization of data sample. It was proved, that for inaccurate,
 noisy or small data can be found best optimal simplified model,
 accuracy of which is higher and structure is simpler than structure of
 usual full physical model.

 Recent developments of the GMDH have led to neuronets with active
 neurons, which realize twice-multilayered structure: neurons are
 multilayered and they are connected into multilayered structure. This
 gives possibility to optimize the set of input variables at each
 layer, while the accuracy increases. The accuracy of forecasting,
 approximation or pattern recognition can be increased beyond the
 limits which are reached by neuronet with single neurons.

 Not only GMDH algorithms, but many modeling or pattern recognition
 algorithms can be used as active neurons. Its accuracy can be
 increased in two ways:
 - each output of algorithm (active neuron) generate new variable
   which can be used as a new factor in next layers of neuronet;
 - the set of input factors can be optimized at each layer. In usual
   once-multilayered NN the set of input variables can be chosen once
   only. The output variables of previous layers in such networks are
   very effective secondary inputs for the neurons of next layers.

 Neuronets with active neurons and basic GMDH algorithms was described in
   Selforganization of Neuronets with Active Neurons.
      Ivakhnenko,A.G.,Ivakhnenko,G.A.,Muller,J.A.  Pattern recognition and
      image analysis, 1994, vol.4, no.2;
   Self-Organization of Nets of Active Neurons.
      Ivakhnenko A.G., Muller J.-A.  SAMS, 1995, vol.20, pp.93-106.
 The GMDH theory and source code of some algorithms was also published in
     "Inductive Learning Algorithms for Complex System Modeling",
      Madala H.R. and Ivakhnenko A.G., 1994, ISBN: 0-8493-4438-7, CRC Press;
and
     "Self-organizing Methods in Modelling (Statistics: Textbooks and
      Monographs,vol.54)", Farlow, S.J. (ed.), 1984, ISBN: 0-8247-7161-3,
      Marcel Dekker Inc.

 This method is described now at  http://come.to/GMDH
 Regards,

 Gregory Ivakhnenko





--
National Institute for Strategic Studies
Kyiv, Ukraina






%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%