When Genetic Programming (GP) is applied to rule extraction from databases, the attributes of the data are often used for the terminal symbols. However, in the case of the database with a large number of attributes, the search space becomes vast because the size of the terminal set increases. As a result, the search performance declines. For improving the search performance, we propose new methods for dealing with the large-scale terminal set. In the methods, the terminal symbols are clustered based on the similarities of the attributes. In the beginning of search, by reducing the number of terminal symbols, the rough and rapid search is performed. In the latter stage of search, by using the original attributes for terminal symbols, the local search is performed. By comparison with the conventional GP, the proposed methods showed the faster evolutional speed and extracted more accurate classification rules.