A layered learning estimation of distribution algorithm
Yong Li, Qiang Yang, Xudong Gao, Zhenyu Lu, Jun Zhang · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2022
Though estimation of distribution algorithms (EDAs) have witnessed success in problem optimization, most of them suffer from sharp shrink of covariance, which may lead to premature convergence. To alleviate this issue, this paper proposes a layered learning estimation of distribution algorithm (LLEDA) by maintaining multiple probability distribution models. Specifically, LLEDA first separates the population into several layers based on fitness. Then, the mean position of each layer is computed. Subsequently, we let the estimated mean position in each lower layer randomly learn from the one of a randomly selected higher layer, so that the mean positions of lower layers could be promoted to be close to promising areas found in the current population. At last, the covariance of each layer is estimated based on the generated new mean position and the individuals in this layer. By this means, multiple probability models with high quality are maintained and then are used to sample promising and diversified offspring separately. Comparative experiments conducted on a widely used benchmark problem set demonstrate that the proposed LLEDA achieves competitive or even much better performance than several state-of-the-art and representative EDAs.