A0234
Title: Sparse approximation of kernel function estimators via genetic algorithm
Authors: Kiheiji Nishida - Kyoto Sangyo University (Japan) [presenting]
Abstract: A method is proposed for constructing multivariate kernel-based function estimators, specifically, both kernel density and regression estimators, using a genetic algorithm to obtain sparse representations. The method applies sparsification in two forms: Reducing the number of data points (data size sparsification) and reducing the number of input dimensions (feature sparsification). The algorithm generates multiple subsamples of user-specified fixed size through random sampling with replacement. Each subsample is treated as a chromosome, and each gene within it, in the terminology of genetic algorithms, represents either a data point or a variable, depending on the selected sparsification mode. Pairs of subsamples undergo genetic operations such as crossover, mutation, and direct inheritance, applied with predetermined probabilities. Fitness is evaluated based on performance in approximating the target density or regression function, and those with superior fitness are selected to survive and contribute to the next generation. Through repeated iterations, the algorithm evolves toward a compact yet accurate estimator. Simulation results show that the proposed method consistently outperforms well-known kernel-based estimators in accuracy and data condensation ratio.
