A Research On The Improvement Of Artificial Fish Algorithm And Its Optimization On BP Neural Network

Artificial neural network is a operation model with abstract system structure, work mode and operation model of biological neural network, of which BP neural network is most widely used. During actual applications, it is easy to fall into local extremum problem. Current solutions to this problem can be divided into two main categories: the first one is to improve BP algorithm principle; the second one is to use theories in other fields to optimize BP algorithm, such as artificial fish algorithm to optimize the BP neural network, while artificial fish algorithm still needs to be improved in its global optimization ability and optimization precision when applied in complex scenarios. Therefore, aiming at this deficiency, this paper improves artificial fish algorithm, and optimizes the BP neural network with improved artificial fish algorithm to improve its local minima’s ability to achieve correct output. In this paper, the main work is as follows:Firstly, in order to improve the optimization function of artificial fish algorithm, this paper proposes an artificial fish algorithm based on adaptive dynamic neighborhood structure(ADAFSA). In the basic artificial fish algorithm, visual and step length which are fixed, and the neighborhood structure of artificial fish is also determined by the visual, which affects the global search capability of the artificial fish and optimization precision. ADAFSA mainly has the following four improvements: using the way based on the distance and the number of iterations to build a more effective artificial fish neighborhood structure; setting the visual and step length as variables according to the adaptive neighborhood structure information; removing crowded degree factor and increasing the bulletin board and changing the artificial fish cluster, rear-end and foraging behavior; redesigning the algorithm steps and process. Simulation results show that the global optimization and optimization precision of ADAFSA are improved greatly.Secondly: this paper uses ADAFSA to optimize BP neural network(ADAFSA- BPNN). First, employing the global search ability of ADAFSA to adjust the initial BP neural network weights and thresholds to near global optimum value, then performing BP algorithm train the network weights and thresholds. Function fitting simulation experiments show that this optimized method improves the generalization performance of BP neural network.Thirdly: this paper applies the proposed ADAFSA-BPNN to the recognition of speech signal. First of all, designing ADAFSA-BPNN based speech recognition process; then, extracting and processing four types of music characteristic signal data; finally, performing recognition simulation experiments. The experiment results show that ADAFSA-BPNN performs better in speech signal recognition effect.