Ensemble structure of multiple local sensor fusion machine using evolutional pruning technique [an application to heading and rate of turn estimation]

Abstract

The paper presents the preliminary design of multiple-sensor fusion networks used to determine the rate of turn and heading information of an autonomous flying robot. Our approach is to implement an artificial neural network, in combination with an evolutional algorithm, such that the sensor fusion network design process can be decomposed into steps. In doing so, each of the local sensor fusion machine, radial basis function network, is first obtained by an independent training process based on the orthogonal least square algorithm. The final combined networks are then found via the technique called "evolutional ensemble averaging" (EEA). Rather than searching for an optimal combination, the EEA will use the best network combination among other candidates (in the sense that all the current requirements are satisfied) in term of the fitness function. The resulting network is then tested against those based on the other two techniques: the winner-take-all and the simple averaging ensemble method.