As a specific functional organization of genes in prokaryotic genomes, operon contains a set of adjacent genes under the control of the corresponding regulatory signals, and is expressed as the transcript unit. It has been found that genes in an operon usually tend to have related functions, or belong to the same pathway in cell. Therefore the study of operon structure is significant to understand the gene functions and regulatory networks for prokaryotes. However with the current limitation of data acquisition of operons verified by experiments such as prokaryotic transcriptomics, computation methods to annotate the operons in a newly sequenced genome have so far been the major source of operon data, and will continue to be an important mission. Over the past decade, a set of computational approaches to operon prediction have been proposed, however mainly based on experimental operons as their training sets. Nevertheless the lack of experimental operon dataset has been the bottleneck of operon prediction. The authors employ an iterative self-learning algorithm which is independent of training set with known operon dataset. The algorithm develops based on a probabilistic model using features including gene distance, regulation signals of gene expression and functional annotation such as COG. The test result compared against the experimental operon data indicates that the algorithm can reach the best accuracy without any training set. Besides, this self-learning algorithm is superior to the algorithm trained on any species with known operons. Accordingly, the algorithm can be applied to any newly sequenced genome. Moreover, comparative analysis of bacteria and archaea enhances the knowledge of universal and genome specific features of operons.