Information Directed Sampling for Combinatorial Material Synthesis and Library Design

抄録

Combinatorial techniques have become more and more important in many areas of chemistry and chemical engineering research. It was suggested that simulated annealing can be used to improve the efficiency of sampling in combinatorial methods. However, without priori model estimates of fitness function, true importance sampling cannot be performed. In this case, the efficiency of annealing is only as good as random search. We suggested that a simple prediction model using currently available data can be constructed using a generalized regression neural network. An index of our uncertainty about a point in the search space can also be established using information entropy. An information free energy combined the two indices to direct the search so that importance sampling is performed. Two benchmark problems were used to model the optimization problem involved in combinatorial synthesis and library design. We showed that when importance sampling is performed, the combinatorial technique became much more effective. The improvement in efficiency over undirected methods is especially significant when the size of the problem becomes very large.