Data Science and Applied Topology Seminar, CUNY Graduate Center
Time: Friday, September 12, 2025, 12:00-13:00 EDT
Location: GC 4421
Title: Decision Flow: Generative AI Meets Stochastic Optimal Control
Abstract: Decision Flow (*) is a novel attempt to use generative AI for sequential decision-making. The method provides a means of sampling from an analytically intractable probability distribution without the long-duration iterative refinement of Markov Chain Monte Carlo methods. Beginning with a prior distribution, an optimal transport problem is solved to evolve the prior to a posterior consistent with a target distribution over an underlying auto-regressive process, which represents decisions in a stochastic optimal control setting. We demonstrate that, if the prior reaches all possible trajectories, the posterior matches the target distribution exactly. Moreover, we illustrate that empirical convergence to the true distribution requires fewer samples than MCMC. Finally, we discuss an application to a mixed-integer optimization problem fundamental to electricity market clearing, and preview further opportunities for enhancing the fidelity, performance, and broad applicability of the method.
(*) Decision Flow originated in Chertkov et al. (2025) (arXiv:2503.14549).