Study on replica extension of approximate probability calculation algorithms

点击翻译按钮获取中文摘要

Approximate algorithms for computing probabilities, such as belief propagation, sometimes offer remarkable performance for "easy" inference problems that come out in information and communication. However, when the profile of the target distribution is of the "many-valley" structure, they do not necessarily lead to the good performance. In this research project, we studied a scheme for improving the inference performance by introducing the concept of replica symmetry breaking to the approximate algorithms, and examined its possibilities and limitations based on various case studies.