Gravitation Search Training Algorithm for Asynchronous Distributed Multilayer Perceptron Model

More popular principles of neural networks distribution are trainset parallel processing, matrix operations parallel processing or parallel processing of neural networks collections. This paper describes architecture of asynchronous distributed system for training and simulating multilayer perceptron based on the actor model. Usual backpropagation training algorithms for multilayer perceptrons, such as gradient training, Levenberg-Marquardt training, RPROP training have a common disadvantage: all of them are partially synchronous. Global optimization algorithms (genetic training, annealing algorithm etc.) are more appropriate for distributed processing but most of them are slow and inefficient. Gravitation search algorithm is new global optimization procedure combining good convergence, efficiency and algorithm step dispensability. The paper develops asynchronous distributed modification of this algorithm and presents the results of experiments. The proposed architecture shows the performance increase for distributed systems with different environment parameters (high-performance cluster and local network with a slow interconnection bus).