Knowledge elicitation from extensive data files inevitably reduces the extracted information content. Results always serve to a subsequent, often dynamic, decision making. Its quality depends critically on the reduction made. This fact is rarely respected in the extensive set of methods for knowledge extraction, often, because of the "curse of dimensionality" connected with the methodologies that address the decision-making problem in its entirety. The proposed project will contribute to an improved solution of the above general problem: i) by solving general dynamic decision making via fully probabilistic methodology that describes both the subject and aims of decision making in probabilistic terms; ii) by designing approximation methodology allowing to solve practically a wide range decision making problems; iii) by verifying of the proposed algorithms on a non-trivial, economically significant application.