Design Amortization for Bayesian Optimal Experimental Design

Bayesian optimal experimental design is a sub-field of statistics focused on developing methods to make efficient use resources. Any potential evaluated in terms utility function, such as the (theoretically well-justified) expected information gain (EIG); unfortunately however, under most circumstances EIG intractable evaluate. In this work we build off successful variational approaches, which optimize parameterized model with respect bounds EIG. Past learning new from scratch for each considered. Here present novel neural architecture that allows experimenters single can estimate potentially infinitely many designs. To further improve computational efficiency, also propose train significantly cheaper-to-evaluate lower bound, and show empirically resulting provides an excellent guide more accurate, but expensive evaluate We demonstrate effectiveness our technique generalized linear models, class statistical models widely used analysis controlled experiments. Experiments method able greatly accuracy over existing approximation strategies, achieve these results far better sample efficiency.