Abstract
In this paper, a fuzzy competitive learning (FCL) paradigm adopting a principle of learn according to how well it wins is proposed, based upon which three existing competitive learning algorithms, namely, the unsupervised competitive learning, learning vector quantization, and frequency sensitive competitive learning, are fuzzified to form a class of FCL algorithms. Unlike the crisp competitive learning algorithms where only one neuron will win and learn at each competition, every neuron in the proposed FCL networks to a certain degree wins, depending on its distance to the input pattern, and learns the pattern accordingly. Thus, the concept of win has been formulated as a fuzzy set and the network outputs become the win memberships (in [0, 1]) of the competing neurons. Compared with the crisp competitive learning algorithms, the proposed fuzzy algorithms consist of various distinctive features such as i) converging more often to the desired solutions, or equivalently, reducing the likelihood of neuron underutilization that has long been a major shortcoming of crisp competitive learning; ii) better classification rate and generalization performance, especially in overlapping data sets: iii) providing confidence measure of the classification results. These features are demonstrated through numerical simulations of three data sets, including two artificially generated ones and a vowel recognition data set.
Original language | English |
---|---|
Pages (from-to) | 539-551 |
Number of pages | 13 |
Journal | Neural Networks |
Volume | 7 |
Issue number | 3 |
DOIs | |
Publication status | Published - 1 Jan 1994 |
Externally published | Yes |
Keywords
- Competitive learning
- Fuzzy c-means algorithm
- Fuzzy competitive learning
- Fuzzy sets
- Membership function
- Neuron underutilization
ASJC Scopus subject areas
- Cognitive Neuroscience
- Artificial Intelligence