Globally optimal vector quantizer design using stochastically competitive learning algorithm
Hao Bi, Guangguo Bi, Yimin Mao · 2002
This paper presents a learning scheme called stochastically competitive learning algorithm (SCLA) for globally optimal vector quantizer design. The SCLA incorporates the idea of stochastic relaxation into the on-line learning scheme of the Kohonen Learning Algorithm (KLA). The key of the SCLA is to replace the Euclidean winner rule with the stochastic competition such that at a given instant any codevector may be updated according to a probability related with its distance to the input. With computer simulations, the effectiveness of the SCLA has been demonstrated by comparing its performance with that of the GLA.>