Since Estimation of Distribution Algorithms (EDA) were proposed, manyattempts have been made to improve EDAs' performance in the context of globaloptimization. So far, the studies or applications of multivariate probabilisticmodel based continuous EDAs are still restricted to rather low dimensionalproblems (smaller than 100D). Traditional EDAs have difficulties in solvinghigher dimensional problems because of the curse of dimensionality and theirrapidly increasing computational cost. However, scaling up continuous EDAs forhigher dimensional optimization is still necessary, which is supported by thedistinctive feature of EDAs: Because a probabilistic model is explicitlyestimated, from the learnt model one can discover useful properties or featuresof the problem. Besides obtaining a good solution, understanding of the problemstructure can be of great benefit, especially for black box optimization. Wepropose a novel EDA framework with Model Complexity Control (EDA-MCC) to scaleup EDAs. By using Weakly dependent variable Identification (WI) and SubspaceModeling (SM), EDA-MCC shows significantly better performance than traditionalEDAs on high dimensional problems. Moreover, the computational cost and therequirement of large population sizes can be reduced in EDA-MCC. In addition tobeing able to find a good solution, EDA-MCC can also produce a useful problemstructure characterization. EDA-MCC is the first successful instance ofmultivariate model based EDAs that can be effectively applied a general classof up to 500D problems. It also outperforms some newly developed algorithmsdesigned specifically for large scale optimization. In order to understand thestrength and weakness of EDA-MCC, we have carried out extensive computationalstudies of EDA-MCC. Our results have revealed when EDA-MCC is likely tooutperform others on what kind of benchmark functions.
展开▼
