This thesis is concerned with a general m-class, P-feature hierarchical pattern recognition problem. Many kinds of pattern recognition methods have been reported and it is well-known that the hierarchical pattern recognition method is more efficient than others in the view point of cost, time and so on . But, classifier designer doesn't know about what kind of hierarchical classifier is more efficient for his applications. so, in this thesis, some hierarchical pattern recognition algorithms are described and implemented using a computer. Five algorithms which are selected for comparative study are K-S distance method, permutation statistic method, mutual information method, projection method and Babes classifier based on tree structure. Three of them use a feature at each nonterminal node to partition the training patterns and the others use many features at each nonterminal node. The efficiency is compared in the view point of error rate, CPU time, memory size and so on. As a result, the algorithms which use many features are more efficient in recognition rate. But, in the new point of classification time and memory size, the algorithms which use a feature are more efficient. Finally, a new efficient classifier design algorithm combining the mutual information method and the projection method is proposed. Several experimental results shown the effectiveness of the proposed scheme.