Competitive and dynamic cooperative learning algorithm

Although rival penalized competitive learning (RPCL) and its variants can select cluster numbers automatically, they have some important limitations. Based on the semi-competitive learning mechanism of competitive and cooperative learning (CCL), this paper presents a new robust dynamic cooperative learning algorithm, or competitive and dynamic cooperative learning (CDCL), in which the learning rate of each seed point within a cooperative team can be adaptively adjusted. CDCL not only inherits the merits of CCL, RPCL and its variants, but also overcomes most of their shortcomings. It is insensitive to the initialization of seed points and applicable to heterogeneous clusters with an attractive and accurate convergence property. Experiments on Gaussian mixture clustering and color image segmentation have shown the efficacy of CDCL. Moreover, in complex cases its performance was far superior to CCL and some other variants of RPCL.