Research On Shuffled Frog Leaping Algorithm And Its Applications

Shuffled frog leaping algorithm(SFLA) is a kind of meta-heuristic algorithm which is based on cooperative search in subgroups. The algorithm has many advantages such as the simple concept, less control parameters, being easy to implement and ability to find the global optimum solution, etc. So it has broad implementation prospects. However, for some complex optimization problems, the standard SFLA still has some weaknesses such as the slow searching speed, trapping into the local optimum easily and getting a bad solution, etc. According to the deep research on SFLA, we try to improve the performance of the algorithm with different methods in this paper. And the improved algorithms are applied to function optimization, cluster analysis and GMM-based speaker recognition system, respectively. The main researches of this paper are listed as follows:(1) SFLA with an acceleration and aging mechanism(AAM-SFLA) is proposed for function optimization problems at first. At the searching prophase of SFLA, we make full use of the global optimum solution to speed up the algorithm. Then we take advantage of the aging phenomenon in nature, i.e., the weak leader will be replaced by other individuals. Thus, premature convergence caused by the misleading of a leader positioned at the local optimum position can be avoid. By comparing with other algorithms such as SFLA with comprehensive learning(SFLA-CL), particle swarm optimization(PSO), fast evolutionary programming(FEP) and SFLA on five standard benchmark functions which included 2 unimodal functions and 3 multimodal functions, it has been proved that AAM-SFLA can converge to global optimum more precisely, quickly and robustly.(2) Because the K-Means clustering algorithm is sensitive to initial clustering centers and easy to fall into local optimum, a clustering algorithm based on improved SFLA(ISFLA-KM) is proposed. Individual cognition and moving inertia of frog in SFLA are introduced into the new algorithm for enhanceing the global search. To accelerate the convergence, K-Means algorithm is performed after updating the frog position. The algorithm is experimentally validated on 2 artificial datasets and 4 real-life datasets from UCI database. According to the simulation results, the clustering effectiveness and clustering robustness of ISFLA-KM are improved by comparing with K-Means, SFLK, SFLA-KM, PSO and PSO-KM algorithm.(3) For the GMM parameters optimization problem in the speaker recognition system, an efficient shuffled collaborative optimization(SCO) algorithm is proposed. In the new algorithm, the population will be divided into three subgroups according to the grouping rules of standard SFLA. Each frog in subgroup makes use of the position-updated rule of PSO. Each subgroup is set to different parameters. They emphasize the capabilities of breadth search, deep search and comprehensive search respectively. Then shuffle all frogs to exchange information. To speed up the convergence, EM operation is performed after SCO algorithm. After training and testing with a set of speaker voice from TIMIT library, the experimental results show that the improved algorithm in this paper is better than the current existing algorithms and can reduce the speaker recognition error rate effectively.