The University of Edinburgh -
Division of Informatics
Forrest Hill & 80 South Bridge

MSc Thesis #9522

Title:Data Mining Using a Multi-Agent Self-Organizing Clustering System
Date: 1995
Abstract:There is a growing interest in extracting useful information from real-world data-bases. As the size and number of databases continue to grow at an increasingly accelerated pace, new and more powerful methods for mining these information spaces are needed. This thesis describes the implementation of an algorithm best classified as a self-organizing multi-agent system. Called the Multi-Agent Data Miner (MAD-Miner), this program partitions complex information sets into clusters so as to reveal the global structures hidden within. It is a refinement on an interactive system originally built by [DGF+91] and further developed by [LF94b]. The approach taken is modelled on the collective sorting abilities of ant colonies and involves simulation of simple local behavioural rules within an artificial environment. The result is an heuristic mapping of multidimensional data sets into a two dimensional environment. MADMiner successfully clusters complex data sets with up to as many as 36 dimensions and 871 elements. Clustering performance has been improved by introducing annealing to the system. Including heterogeneity amongst agents also improves clustering performance. Final testing of MADMiner has been carried out on data sets provided by the World Health Organization's (WHO) Tuberculosis Division.

[Search These Pages] [DAI Home Page] [Comment]