The Research On Multi-Agent Based Pattern Recognition Frame APRF

The course of human's pattern recognition is divided into two cases: first case, if input pattern never be seen before, then feature extraction is need, so that the model of input pattern is built, in fact, pattern recognition is just modeling (or memory), and it is a down-top course; second case, if input pattern have been seen before, then classifying is completed by logic inference, in fact, the pattern recognition essentially is pattern emerge (inference), and it is a top-down course.The research of pattern modeling and emerge is independently in conventional pattern recognition. In this paper, we analyze the merit and default of down-top and top-down, and APRP (Agent-Based Pattern Recognition Frame) is pressed based on the theory and technology of Multi-agent, its main principle is reasoning but computer. Firstly, modeling by down-top, and then classifying by top-down, so that the course of pattern recognition accord to humans perceive.The main job are:(1) Bring up a new frame of pattern recognition APRP based on Multi- agent Systems (MAS), this frame merge into the methods of down-top and top-down. Pattern modeling solve "memory", and pattern emerge solve " inference ".(2) A model of dynamics cooperation DCMBR in MAS is presented based on roles, which is set up a bridge between pattern modeling and pattern emerge. Through the alliance, cooperation, negotiation and coordination in agents, so that system form complex model in structure, property and behavior, even though every agent's structure, property and rule of conduct is simple.(3) Bring up the concept of qualitative character. Because quantity feature can't reflect structure information in tradition, and there is not general way of feature extraction. Whereas qualitative feature is emerged by layer recursive, as sample increasing, qualitative feature tend to steady, it reflect whole structure information and detail information, too.(4) We will express pattern into AIM (agent influence map) based on knowledge, it is a memorial neural network and can realize to explain the phenomenon of transform from part to integration.(5) Bring up three pattern classifying models: pattern reasoning, character fusions and pattern associated based on the principle of emerge.