The Improvement And Application In Wavelet Neural Network Algorithm

Wavelet transform is a new mathematic theory and method which has been widely used in the world currently. Wavelet neural network(WNN), which is a kind of feed-forward neural network taking wavelet functions as its nerve cells' activations function, has also been used in various fields. Based on the research about study algorithms of wavelet network, parameters initialization and methods of structure optimization, this paper puts forward some improved algorithms. For the study algorithm, this paper contrapose the defects of study algorithm on the traditional wavelet neural network, introduces several familiar improved study algorithms and then presents a new study algorithm: Annealing Ants Colony Genetic Algorithm (AACGA). This algorithm introduces the stochastic disturbance ideal of simulated annealing algorithm into the crossover and the mutation of genetic algorithm so as to replace the original individuals with dynamically calculating probabilities of the crossover and the mutation. The dynamical probabilities make them have preference nature in these operations. The preference nature makes the algorithm have directivity during search process so that the algorithm has strong search capability in local region. And with occasional deflection, the algorithm can overcome the restriction of the local extremum to get away from the local region so as to realize global search. Therefore, the AACGA presented in this paper is a global search algorithm with strong local search capability. Wavelet networks using it as the study algorithm can hunt out the network parameters with which the network has the best fit effect to the samples. For the parameter parameters initialization, this paper presents a setting method. For the wavelet networks' initial parameters which links the initial parameters setting, the types of wavelet, the wavelet parameter of time-frequency and the samples of study together. For the structure optimization, this paper puts forward a design method of network's structure optimization based Ants Colony Genetic Algorithm–Rough Set (ACGA-RS) to the problem about big redundancy in the framework-based wavelet network. Firstly, it ascertains the elementary structure of wavelet network by analyzing the time-frequency, and base on the result, it deletes the redundant hidden layer node according the concept of Rough Set's attribute reduction to optimize the network's structure finally. In the end, we propose the WNN classifying device model, using the improving algorithms. And based on the model, we construct the classifying device of lymphoma cell and the classifying device of traffic image.