Decision Fusion in Cooperative Adaptive Systems
Research on cooperative, adaptive intelligentsystems, involves studying, developing and evaluatingarchitectures and methods to solve complex problemsusing adaptive and cooperative systems. These systemsmay range from simple software modules (such as aclustering or a classification algorithm) to physicalsystems (such as autonomous robots, machines orsensors).The main characteristic of these systems is that theyare adaptive and cooperative. By adaptive, it is meantthat the systems have a learning ability that makes themadjust their behaviour or performance to cope withchanging situations. The systems are willing tocooperate together to solve complex problems or toachieve common goals.In pattern recognition, there are notable contributionson the use of multiple classifiers. The most dominantdecomposition model used is an ensemble of classifiers(identical structures) that are trained differently. Mostof the innovations are in the combining methods. Thereare weighting and voting approaches, probabilisticapproaches and approximate and fuzzy logicapproaches. In the area of sensor fusion, there havebeen some interesting ideas for fusing the data anddecisions of the sensors. However, most of thesecombining schemes are usually applied as a postprocessing stage.In this work are concerned with investigatingarchitectures and methods of aggregating decisions in amulti-classifier or multi-agent environment. Newarchitectures that allow active cooperation will bedeveloped. The classifiers (or agents) have to knowsome knowledge about others in the system. Differentforms of cooperation will be reported. In order for thesearchitectures to allow for dynamic decision fusion, theaggregation procedures have to have the flexibility toadapt to changes in the input and output and adjust toimprove on the final output. Changes will be learned bymeans of extracting features using feature detectors.Applications of these architectures to problems inclassification of data, distributed data mining andclustering will be illustrated.
Mohamed Kamel
Pattern Analysis and Machine Intelligence Lab.University of Waterloo Waterloo,Ontario,Canada,N2L 3G1
国际会议
厦门
英文
5
2004-05-26（万方平台首次上网日期，不代表论文的发表时间）