Research And Implement Of A Computing Platform Based On Grid For Hybrid Neural Networks

Humanoid robot has become more and more intelligence, but it can not think and learn like human without strong computing power. Grid computing can supply much computing resource and storage resource to the humanoid robot. Artificial neural network is a complex network system that is comprised of many simple interconnected neurons. It imitates the physiological feature and part of function of human brain, but often just solves some special problems in that field. In order to imitate the complex function of brain, it's necessary to combine different kinds of neural networks to form a large-scale of hybrid neural network. The computing of neural network needs huge computing resource and storage resource, and much hugger when to form a large-scale of hybrid neural network. So according to the feature of neural network, it is a good entry point of applying grid computing to the humanoid robot.First, based on LabGrid V0.1 and referring to the architecture of grid computing, Globus Toolkit and Condor-G, this paper develops LabGrid V1.0 which is the basic operating environment for the subsequent experiments and application. And then the paper designs and implements the grid-based computing platform for hybrid neural networks (HNetCP) which uses the LabGrid as the grid middleware. HNetCP supplies a basic computing platform for the neural network computing and constructing the application of large-scale hybrid networks. Combining HNetCP with the research on the human behavior of sleeping learning, this paper implements an incremental learning system RILS to make the humanoid robot have the human-like ability of quick learning and sleeping learning. In the end, RILS is applied to the system of face recognition of humanoid robot.This paper firstly of puts forward and implements the grid-based computing platform for hybrid neural network, and designs a new hybrid neural network classifier system on this platform to have an experiment. The experimental results prove that the new classifier system has higher ratio of correct classification and it's feasible and effective to build a large-scale of hybrid neural network to imitate human's brain. Besides, this paper applies the incremental learning system with sleep ILS to the learning system of humanoid robot, and makes it run on the grid through the HNetCP, which is also innovative. Through experiments and application test, it proves that the incremental learning system is feasible, and makes the humanoid robot have the ability of incremental learning.