Holonic probabilistic agent merging algorithm

Data mining global models of complex domains is often difficult - a lot of methods are NP-hard as stated in D. M. Chickering et al. (1994) - and the resulting models either lack of details or are far too complex to work with them. In our case we were challenged with the task to model the shopping behavior of customers and the sales behaviour of items within a real supermarket store in order to forecast product sales figures on the basis of 'what-if'-scenarios. To forecast the scenarios we represent and simulate all entities of a store as holonic probabilistic agents. The knowledge bases of the agents consist of behavior networks which are data mined from real store data. For simulation we integrate on demand global coherences in our models without loosing the level of detail. This is realized by a process where the agents merge to a so called holon. We present an algorithm to merge the agents' behaviour networks where global coherences between the behaviors of different agents evolve as an effect of emergence.