This codebase implements abstract classes and corresponding functions for Bayesian Optimal Experimental Design (BOED) in PyTorch.

Features

• BOED over discrete candidate designs with Nested Monte Carlo (NMC) Expected Information Gain (EIG) computations
• BOED over continuous design space by maximizing Prior Contrastive Estimation (PCE) (https://arxiv.org/abs/1911.00294)
• BOED by Deep Adaptive Design (DAD) (https://arxiv.org/abs/2103.02438)
• variational inference algorithm

Files in src

• data_utils: abtract Experimenter class for modelling experimental settings, abstract Design_Network class for DAD
• distributions: abstract prior and likelihood classes
• model_utils: functions for discrete BOED, continuous BOED with PCE, and DAD building on abstract Experimenter and distribution classes

Examples

We demonstrate the codebase on 2 different versions of 2d tissue slicing closely aligned to https://github.com/andrewcharlesjones/spatial-experimental-design:

• discrete: iteration over finite candidate designs with NMC
• continuous: weighting each point on the grid by its distance to the chosen slice and optimize PCE

Additionally, there is an experimental example with a simple 1d location finding inspired by https://arxiv.org/abs/2103.02438 and a 2d tissue model that samples by interpolation from an affine grid.

Additional Features

• constraining distribution params by simple transformations during optimization

Note of Caution

• Testing was limited (only performed on one example).
• Eventually, other variance reduction methods for the gradient estimation need to be implemented.