This thesis considers a clustering problem for pattern recognition where clustering plays a role of furnishing information in a unsupervised learning approach to be used for designing a pattern classifier when any sufficient information on the associated data structure is not available.
For the problem two algorithms are exploited. One is a Tabu Search algorithm and the other one is a Tabu-Search-based algorithm. Because Tabu Search is not stick to a local optimum solution, the Tabu Search algorithm can obtain better solution than traditional algorithms, including the K-means algorithm, which stick to local optimum solutions. This Tabu Search characteristics are incorporated into exploiting a Tabu-Search-based algorithm by combining them with the functional procedures of operation packing and releasing for improving the overall effectiveness. Several similar algorithms including K-means algorithm, Simulated Annealing algorithm, Tabu Search algorithm, and Tabu-Search-based algorithm are then tested with numerical examples to compare them one another. The test results show that the Tabu-Search-based algorithm outperforms the other algorithms.