Research On Estimation Of Distribution Learning Algorithms

Machine learning is a scientific discipline that is concerned with the design and development of algorithms that allow computers to evolve behaviors based on empirical data, such as from sensor data or databases. A major focus of machine learning research is to automatically learn to recognize complex patterns and make intelligent decisions based on data. Estimation of distribution algorithm (abbr. EDA) is a relatively new branch of evolutionary algorithms. EDA replaces search operators with the estimation of the distribution of selected individuals+sampling from this distribution. The aim of EDA is to avoid the use of arbitrary operators (mutation, crossover) in favour of explicitly modelling and exploiting the distribution of promising individuals.This thesis is devoted to some learning algorithms and their theories and application studies based on EDA. The main achievements are summarized as follows:(1) A general learning framework based on EDA (abbr. FrEDL) is designed from the perspective of probability estimation based evolutionary computation. FrEDL consists of four steps, initialization, estimation, evolution and evaluation. The explicit probability basis about the FrEDL is analyzed. And the mathematical properties analysis of the implicit algebra and algebraic system pertained to FrEDL are provided.(2) A semi-supervised learning algorithm based on EDA (EDA-SSL) is presented. EDA-SSL uses a few data samples with class label to estimate class distributions of a mount of data instances without class labels. EDA-SSL is compared with several classification algorithms in error rates of classification and also with genetic algorithms (GA). The experimental and analytical results show EDA-SSL is better than or comparable with other algorithms in classification accuracy.(3) Unsupervised clustering learning algorithm based on EDA (EDA-USL) is designed to solve the analysis of data set without labels. EDA-USL is described and analyzed in detail by measurement methods of attributes and analysis methods of correlation between attributes. EDA-USL is verified on real-world data set and analyzed. The experimental results show that EDA-USL has highly stability and well performance in classification accuracy. (4) The capture problem among multi-agent is solved by EDA. The capture problem involves that some pursuers pursue several evaders through part of trajectory. The probabilistic evolutionary courses of multi-agent experiencing some competitions are analyzed in performances. The analysis shows that capture problem of multi-agent solved by EDA is better than other methods in several aspects.