Model Learning and Variance Control in Continuous EDAs Using PCA
Jun Liu, Hong‐Fei Teng · 2008
Estimation of Distribution Algorithms (EDAs) can be viewed as the outcome of the cooperation between evolutionary computation and probabilistic graphical models. In this paper, we review some continuous EDAs based on Gaussian network model and discuss their some known problems briefly. To prevent premature convergence and repair singular covariance matrix, we propose the PCA-EDA algorithm which integrates Principle Component Analysis (PCA) into continuous EDAs with the help of probabilistic PCA (PPCA), a probabilistic graphical model explaining PCA with latent variables. The model learning of PCA- EDA consists of principle components (PCs) selection and variables selection in each PC. Moreover variance control can be employed naturally and reliably. Experimental results support that presented algorithm can effectively avoid premature and singular problems.