P-EDR: An algorithm for parallel implementation of Parzen density estimation from uncertain observations

We have developed a parallel version of a new algorithm for nonparametric density estimation when the input samples are not directly known, or they have some noise. The algorithm is an extension of the Parzen method for exact observations, but management of uncertainty implies heavy computational loads in terms of both calculus and storage. Therefore, a parallel version of the algorithm is more adequate to solve this extended problem in a practical time, specially for samples of medium-large sizes. Our parallel algorithm has been designed in an SPMD style and implemented in a message-passing parallel environment. An efficient treatment of the distribution of the main data structure among processors, together with a low communication cost scheme results in a high scalability of the algorithm. Preliminary performance evaluations in a cluster of workstations show excellent speed-up results