Rival penalized competitive learning, finite mixture, and multisets clustering

Rival penalized competitive learning (RPCL) can automatically select the number of clusters during learning by penalizing the rival in competition. The original adaptive RPCL algorithm is proposed for clusters of spherical shapes and its performance will degenerate considerably when the clusters are of complicated shapes. In this paper, the adaptive RPCL learning has been extended to solve this problem via the finite mixture modeling and multi-sets modeling, respectively. Moreover, two general competition types are suggested, referred to as Type A and Type B. The Type B RPCL with both the finite mixture modeling and multi-sets modeling includes the original RPCL as a special case. The experiments show that both Type A and Type B RPCL worked well and improved the original RPCL considerably for clusters of complicated shapes and strong overlapping. Moreover, the Type B RPCL is the best in automatic selection of correct number of clusters.