Resolving Inconsistencies in Shared Context Models using Multiagent Systems

Agents acting in physical space use perception in combination with their own world models and shared context models. The shared context models have to be adapted permanently to the conditions of the real world. If a measurement of an agent's sensor does not fit to the corresponding data in the shared context model the system contains an inconsistency. In this case it is necessary to decide whether the reason for the discrepancy is a change in the real world or a measurement error. A solution has to be based on only a limited number of measurements from different agents and a reduction of noise in the sensor data. The study reported in this paper evaluates procedures that combine a multitude of measurements to a single result that can be integrated in the shared context model. The statistically optimized procedure based on ratings of the participating agents is enhanced using scaled weighted arithmetic means which prevents the system from running into singularities caused by the feedback from the ratings. The method is combined with an additional preprocessing based on fuzzy clustering that detects aberrant measurements which can be excluded from further processing.