Abstract
Generative modelling is a useful methodology for large-scale combinatorial problems involving many possible candidates. Such problems typically suffer from exponential scaling in complexity as the size of the system being studied is increased, and this impedes search space exploration and discovery. This work explores the parallel between the problem of generating high-value samples from a high-dimensional distribution and choosing a sparse set of grid points for high-dimensional integration using the Smolyak hierarchy. The proposed method, called Sparse Hierarchy Integration-Inspired Protocol (SHIP), is a Generative Flow Network (GFlowNet) based framework in which high-value higher-dimensional samples are generated by extending up from high-value lower-dimensional samples sequentially and iteratively. This protocol is tested on a model problem of approximating path-integral solutions to open quantum dynamics using Small Matrix Path Integral for reward evaluation. We show that the capacity of this approach to circumvent problems related to poor reward connectivity is drastically improved when action space representation is updated based on interpretable features identified in solutions at intermediate steps. These results serve to motivate the implementation of this protocol more broadly for other generative problems in physical and chemical systems.