A1371
Title: Generative adversarial network stopping for robust surrendering
Authors: Bowen Jia - City University of Macau (China) [presenting]
Abstract: An optimal stopping problem under model uncertainty is studied, where the underlying data-generating distribution is unknown. A distributionally robust framework is formulated by introducing an ambiguity set of probability measures around a reference model, capturing robustness against misspecification. To address the resulting high-dimensional and analytically intractable problem, a data-driven approach that combines generative adversarial networks (GAN) and deep neural networks (DNN) is proposed. The GAN approximates the worst-case distribution within the ambiguity set, while the DNN learns the optimal stopping strategy. Theoretical guarantees on the convergence of the proposed method under suitable conditions are provided. Numerical results demonstrate that the approach achieves robust performance and outperforms benchmark methods, particularly in high-dimensional settings where classical techniques are infeasible.
