Dec. 20 2005 2:20PM

Dec. 20 2005 2:40PM

Dynamic decision making (DM) under uncertainty is a common framework for a range of problems addressed within the cybernetics domain. The Bayesian DM is the only known systematic design methodology. Similarly to others, it
suffers from the “curse of dimensionality”: the optimal solutions can rarely be implemented. Use of the distributed version – which concerns DM of multiple participants (MP) with limited ability to perceive, communicate, remember
and evaluate – seems to be the only viable remedy. The research running within the DAR center tries to develop the applicable version of distributed Bayesian DM. This requires to complement the standard Bayesian learning and
DM strategy design by new operations specific to the MP setup. Just merging of probabilistic models and distributions expressing DM aims seems to be needed. The usual incompatibility of merged distributions and apparent
arbitrariness of merging ways are addressed. The need for an implementable variant makes us focus on universal approximations based on finite probabilistic mixtures. This requires the feasible estimation of mixtures with dynamic
components and weights. Moreover, the inevitably limited analytical insight into properties of proposed solutions makes us to devote a substantial effort to creating MP DM simulation tools. The feedback from application partners is
obtained via the use of software that allow them to solve efficiently their problems. The presentation expands these points and prepares the contributions dealing with these aspects.

International Workshop on Data - Algorithms - Decision Making (Contact person: Martin Janžura)