Research On Chaotic Neural Network And Its Applications

This thesis mainly addresses the existing chaotic neural network models and its applications. A varying initial value method is suggested to improve the performance of existing chaotic neural networks.Chaotic neural network is a new field in artificial intelligence. Transient chaotic neural network (TCNN) model is formed by introducing transient chaos into Hopfield neural network. The TCNN becomes a typical model in chaotic neural network research. At initial stage, The TCNN is chaotic, which is helpful for finding the global optimum solution by ergodic property of chaos. With the control parameter changing, the chaos in TCNN disappears gradually, and TCNN is degraded to common Hopfield neural network (HNN). Then, HNN would converge to possible global optimum solution by gradient descending mechanism. Due to the chaos ergodicity, TCNN reduces the probability of getting into local minimum. The longer the chaos is maintained, the larger probability of getting the global optimum solution is obtained. Therefore, enlarging chaos duration and boosting complexity of chaos becomes two start points to improve the performance of TCNN. Exponential annealing chaotic neural network (EACNN) and noise neural network (NCNN) are the direct examples of these efforts.It can be found that if the initial value of chaotic neural network is far away from the attracting basin of the global optimum solution in the sense of orbit,a very long time is needed for maintaining chaos in initial TCNN, so that the following HNN can find theglobal optimum solution. However, this means that a very long or incredible long time is needed, and the performance will degrade in this case. A varying initial value method is proposed to deal with such problem. If one initial value of TCNN cannot get global optimum within limited steps, a new initial value for TCNN is selected by chaos mapping to start the new search until the global optimization solution is gotten. The application simulation results show that a higher efficiency can be received by use of the proposed new model and fewer search steps are needed than the existing neural networks.Finally, the chaotic neural network is applied to the image segmentation in this thesis and some primary results are obtained.