Empirical Inference of Numerical Information into Causal Strategy Models by Means of Artificial Intelligence - Distributed Artificial Intelligence, Agent Technology, and Collaborative Applications - page 310

Table 1. Prediction quality of linear regression vs. connectionist approximator

CFK MSE lin. MSE ANN significant

K SR 0.0048 0.0015 yes

K CG 0.0143 0.0056 yes

K SP 0.0022 0.0020 no

conclusion

This chapter proposes a new approach to validate causal strategy models consisting of nomologic mana-
gerial cause-and-effect hypotheses as a prerequisite for the approximation of the unknown underlying
functions: Therefore this chapter focuses on artificial neural network models due to their universal ap-
proximation properties which explicitly dispenses with a priori assessments about the functional form
of these associations. Therefore this chapter presents techniques for the separation causal function
kernels from causal models and their temporal disaggregation into multi layer perceptrons. The chapter
concludes with a functional review of the prototypical implementation based on these concepts. The
main findings show that the assumption of linear causal function for model selection purposes within
the stage of causal proof is admissible to a large extent. Merely the high noise level of one variable leads
to the acceptance of a spurious cause-and-effect hypothesis for an intermediary result. However, this
source of error will be one issue for future research: As Bartlett's significance test does not account for
the noise level of a dependent variable this technique has to be enhanced in order to provide improved
significance boundaries which are sensitive with respect to the predictive uncertainty.
In order to integrate this approach into a DSS it seems to be necessary to extend this approach for
analytic techniques based on this improved model base like simulation, what-if- or how-to-achieve-
analyses. These applications allow the quantitative anticipation and prediction of future impacts of
strategic scenarios.
Another field of interest is the transfer of the approach—as proposed here—to other scientific areas.
Basically, these techniques could be employed for every decision theoretic problem based on causal
networks such as economic theories, medical diagnostics, etc.

r eferences

Allen, R. (1964). Mathematical analysis for economists . New York: St. Martin's Press.

Bartlett, P. S. (1955). Stochastic processes . Cambridge, UK: Cambridge University Press.

Baum, E. B., & Haussler, D. (1988). What size net gives valid generalization? Neural Computation , 1 ,
151-160.

Brady, H. (2002, 16 July). Models of causal inference: Going beyond the neyman-rubin-holland theory.
Seattle (WA).

De Figueiredo, R. J. P. (1980). Implications and applications of Kolmogorov's superposition theorem.
IEEE Transactions of Autom. Control , 1227-1230.