As the combination of quantum computation theory and artificial neural network, quantum neural network is a novel theory. It has the strongpoint of quantum computation theory and artificial neural network. And it can make full use of the characteristics of quantum computation, such as the superposition of qubits, quantum parallel computation, and the entanglement of qubits, to overcome some inherence bugs of traditional artificial neural network. Most possibly, the quantum neural network will be a very important measure to deal with information in the future.The training algorithm and the structure of quantum neural network that based on multilevel activation function are presented in this dissertation. The quantum neural network model uses BP algorithm to modify the link weights and thresholds. So, some of the bugs of BP algorithm are avoidless for the quantum neural network. Aiming at this point, a linear superposition of arctangent function is introduced in as hide layer activation function, and error saturation prevention function is constructed to improve the convergence property of quantum neural network model. And quantum genetic algorithm is introduced to deal with the link weights and thresholds of the quantum neural network, searching for the best link weights and thresholds to initialize the network. In this way, the convergence property of quantum neural network is improved much more. The main tasks that have been done in this dissertation are shown as follows.(1) The characteristics of the quantum genetic algorithm are summarized. And the quantum gate is analyzed in detail. The policy of dynamic adjusting the quantum gate is used. The circumrotate angle of quantum gate is regarded as the variable of the current iterative generation. So, the convergence property is improved in this way. The quantum intercross and catastrophe operation are introduced in to improve the search ability of the algorithm.(2) The quantum neural network model based on multilevel activation function is analyzed and the detailed process of the quantum interval's training algorithm is given. A linear superposition of arctangent function is introduced in as hide layer activation function, and error saturation prevention function is constructed to avoid the phenomenon of error saturation in the training process and to improve the convergence property of quantum neural network model.(3) The quantum genetic and quantum neural network hybrid algorithm is proposed. The quantum genetic algorithm is used to optimize the initial link weights and thresholds of the quantum neural network model, the best link weights and thresholds are used to initialize the model. Simulation experiments on the improved quantum neural network model and the existing models are operated respectively. The validity and feasibility of the improved model is proved by the results. |