Online learning of ensemble learning by parallel boosting

摘要

We discuss the parallel boosting based on the on-line learning. In general, the weight vector of the learning machine is initialized by the independent identical random number. In this case, the correlation of the weight vectors of all the learning machines could be considered as the same becase of the similality of the time dependent equation of the weight vectors. In this paper, we analyze the time dependence of correlations of the weight vectors when they are not uniformly initialized. Moreover, the trajectories of the weight values for averaging are investigated. As the results, it is shown that the optimal values of the weights for parallel boostings are constant during the learning process, and the optimal weights depend on only the initial value of the correlation, of the weight vectors.