Using a priori knowledge to create probabilistic models for optimization

Recent studies have examined the effectiveness of using probabilistic models to guide the sample generation process for searching high dimensional spaces. Although the simplest models, which do not account for parameter interdependencies, often perform well on many problems, they may perform poorly when used on problems that have a high degree of interdependence between parameters. More complex dependency networks that can account for the interactions between parameters are required. However, building these networks may necessitate enormous amounts of sampling. In this paper, we demonstrate how a priori knowledge of parameter dependencies, even incomplete knowledge, can be incorporated to efficiently obtain accurate models that account for parameter interdependencies. This is achieved by effectively putting priors on the network structures that are created. These more accurate models yield improved results when used to guide the sample generation process for search and also when used to initialize the starting points of other search algorithms.