Unsupervised learning is useful in exploratory data analysis, image segmentation and, with some added class knowledge, may be used for classification as well. In this paper we analyze a genetically guided algorithm (GGA) approach to optimization of certain clustering models. This approach can be directly applied to any clustering model which can be represented as a functional dependent upon a set of cluster centers (or point prototypes). The approach can be further generalized for models that require parameters other than the cluster centers.
In this paper the fuzzy and hard c-means (FCM/HCM respectively) functionals Jm (J1) are used as fitness functions [2,5]. This allows us to compare performance of the GGA with the conventional FCM/HCM algorithms and allows us to look at GGA optimization performance with similar, but different objective functions. Clustering algorithms such as FCM which use calculus-based optimization methods can be trapped by local extrema in the process of optimizing the clustering criterion. They are also very sensitive to initialization.
This paper provides answers to the following questions. Can a GA approach to clustering find extremum that won't be found with the iterative approach to minimizing the c-means functionals? Can a GA find the same extrema that an iterative version of FCM/HCM would or does FCM/HCM need to be run using the final cluster centers found by the GA as an initialization? Will the GA approach find the best or nearly best final partitions for a given data set, i.e. those partitions associated with the lowest Jm values?
In Section 2, we review the FCM/HCM algorithms with which comparisons will be made. In Section 3 the genetic guided clustering approach is presented. Section 4 contains a description of the three data sets employed in this work. Section 5 provides experimental results. Section 6 discusses time considerations and Section 7 summarizes our results.