Research On Feature Selection Problem Based On Bat Algorithm

Feature selection is a technique that can effectively reduce the dimensionality of datasets to improve the efficiency of machine learning algorithms.In this thesis,by analyzing the shortcomings of the bat algorithm in the optimization process such as easy local convergence and poor global search ability,two improvement algorithms based on the bat algorithm are proposed to improve the feature selection performance.First,the feature selection objective is modeled by analyzing the feature selection problem.The commonly used conversion methods are analyzed,and the linear weighting method is chosen as a way to convert the multi-objective optimization problem into a single-objective optimization problem.The function constructed by this method will be used as the fitness function for subsequent studies on the bat algorithm.Next,the continuous space is discretized using the S-shaped function.An improved binary bat algorithm(COBBA)is proposed.By analyzing the disadvantages of the bat algorithm such as the lack of population diversity in the iterative process and the need for parameters to be tuned in advance and generated randomly,a chaotic sequence is chosen to generate a chaotic sequence using chaotic mapping to make the generated initial solution more diverse,while a set of initial solutions closer to the optimal solution is obtained by further optimizing the initial solution using the backward learning technique.Chebyshev chaotic sequences are used to improve the quality of the optimal solution by tuning the parameters.Simulation experiments are also conducted for COBBA on 32 classical datasets.The experimental results show that when COBBA is applied to solve the feature selection problem with more than half of the datasets,the convergence speed,average fitness value,and average classification accuracy of the algorithm are better than several other commonly used swarm intelligence algorithms.Thirdly,an improved bat algorithm(ELBBA)is proposed by using the MinMax function to discretize the continuous space.By analyzing the shortcomings of the unbalanced exploitation and exploration capabilities of the bat algorithm,the EPD mechanism is used to improve the local search capability,and the Lévy flight strategy is used to explore the surrounding possible solutions,which can effectively prevent the algorithm from falling into local optimal solutions.To further analyze the effectiveness of ELBBA,simulation experiments are conducted on 32 datasets to analyze the experimental results from several perspectives.The experimental results show that ELBBA has stronger optimal feature combination ability than several other commonly used swarm intelligence algorithms,and is more competitive in terms of average fitness value,average classification accuracy,and average number of selected features.Finally,a summary of the work is arranged in the thesis.Besides,the existing shortcomings of the improved algorithm are analyzed,and the prospect of further research is proposed.