Real-time prediction in a stochastic domain via similarity-based data-mining

This paper introduces an application and a methodology to predict future states of a process under real-time requirements. The real-time functionality is achieved by creating a Bayesian tree via data-mining on agent-based simulations. The computationally expensive parts are handled in an offline phase, while the online phase is computationally cheap. In the offline phase the simulations are run and meaningful clusters of states are identified by use of virtual attributes. Then the transition probabilities between states of different clusters are organized in a Bayesian tree. Finally, in the online phase similarity measures are used again in order to classify query states into the clusters and to infer the probability of future states. The application domain is the support of military units during missions and maneuvers.