Empirical Inference of Numerical Information into Causal Strategy Models by Means of Artificial Intelligence - Distributed Artificial Intelligence, Agent Technology, and Collaborative Applications - page 309

Figure 8. Causal function approximator for K CG
series associated with each input and output nodes are split into a training and a validation set. If an
MLP incorporates correcting function approximators the latter have to be trained independently first.
After reaching a global error minimum for these correcting ANNs, their weights are kept fixed and
are included in the main function approximator as specified in this chapter including the special input
functions for first level input nodes as well as its inverse function for error backpropagation purposes.
The training of these MLPs leads to an approximation of the causal function determining the values
of the dependent variable of each causal function kernel. With the approximate knowledge of its func-
tional dependencies in the form of the trained ANN, the causal strategy models can be enhanced by a
prediction model which allows numerical analyses on future impacts of strategic scenarios. Since ANNs
are known for their universal approximation properties in the relevant literature, they are expected to
yield better approximation results than other techniques. A comparision with linear forecasting models
like multiple linear regression analysis shows a predominantly clear picture:
As it is shown in Table 1, the proposed approach enables decision makers to approximate a more
precise causal function compared to linear regression in all cases. As the latter does not account for
nonlinear functions, indirect effects, autocorrelation or noisy time series, it yields significantly8 8 higher
mean squared errors (MSE) for the validation set than the connectionist approach especially for these
problem classes. The only exception of this observation is CFK K SP , which is not included in any of the
problem classes: Only for these—strictly linear—functions, given the absence of other disturbances, it
is not possible to achieve significantly better predictive results using the connectionist approach.