The Research Of Classifier Fusion Based On Multi-Agent System And Application In RoboCup2D

It is always an important problem to improve the performance of classifier, but there are so much difficult to figure out before get a type of better classifier. In traditional pattern recognition systems, only one or a little of characters are used in classifying. The pattern recognition system is difficult to achieve a good classifying result with large sorts of samples, in which noise is brought. Lately, it is found that classifier combination can improve accuracy of results. However, at present research on classifier combination mainly concentrated on selection of classifiers and fusion algorithm. Existing methods still have disadvantage on classifying, which get inaccurate results because feature extraction is not comprehensive. Additional, in a complex dynamic environment, there is little classifier algorithm to achieve recognition of dynamic objects.In this thesis, a multi-agent system (MAS) based classifier combination model is designed, which makes use of the characteristic of cooperation and resources sharing in MAS. This classifier combination model implements to classify in a complex dynamic environment, and it overcomes the shortcomings of low recognition rate resulting from that the information is not comprehensive. In this classifier combination model, role-based members of the classifier agent process the information collected firstly, and then transmit it to a fusion agent. The fusion agent chooses the important information according to the role of member agent. As a result, classifying result is gain from fusion agent with fusion algorithm. The model adopts the role-based management mode, so it reflects the agent's ability and autonomy. The role of each agent is not fixed, but also it is variable with environment and self-condition changing. Capability is the reference of role changing, because that capability represents the classifying performance of an agent. In addition, capability is also a reference for fusion agent to make a decision.In this thesis, CFBMAS algorithm is apply on misplay recognition in the debugging program. In Competition of RoboCup simulation league 2D, CFBMAS algorithm can recognize misplay action in passing, catching, intercepting and shooting behavior automatically. So we can propose a solution against our opponents. And the training efficiency is improved.In conclusion, the presentation in this paper is following:(1)The fusion framework of multi-classifier in star-model is designed;(2)Fusion algorithm based on MAS in dynamic environment is presented;(3)An application in RoboCup simulation league 2D for misplay recognition.