复值递归神经网络多平衡点的稳定性与吸引域.docVIP

  Stability and attractive basins of equilibria in complex-valued recurrent neural networks# Huang Yujiao, Wang Zhanshan** 5 10 15 20 25 30 35 (School of Information Science and Engineering, Northeastern University, ShenYang 110004) Abstract: This paper is concerned with multistability of complex-valued recurrent neural networks with a class nondecreasing piecewise linear activation function. By using decomposition of state space, sufficient conditions are established to ensure that n-dimensional complex-valued neural networks with a general class of activation function can have 9^n equilibria, 4^n of them are locally exponentially stable and others are unstable. Moreover, the attractive basins of equilibria are investigated. The results improve and extend the existing stability results in the literatures. One simulation example is given to illustrate the effectiveness of the results. Keywords: Multistability; complex-valued recurrent neural networks; activation function 0 Introduction In past decades, monostability and multistability of recurrent neural networks have been extensively studied due to its applications in image processing, pattern recognition, associative memories and many other fields [1-4.7-17]. In an associative memory neural network, the addressable memories or patterns are stored as stable equilibria or stable periodic orbits. Hence, it is necessary and meaningful that there exist multiple stable equilibria or periodic orbits [5,6]. References [7-11] investigated multiple stable equilibria or multiple stable periodic orbits of recurrent neural networks by a geometrical observation or decomposition of state space. In above references, a general n -dimensional neural networks with a class of activation functions have 3n equilibria or periodic orbits, and 2n of which are stable equilibria or stable periodic orbits. References [12-15] discussed multistability and multiperiodicity of recurrent neural networks with unsaturating piecewise linear tra