Computation of optimal transport and related hedging problems via penalization and neural networks (Q2020305)

The core idea in this article is to penalize the optimization problem in its dual formulation and reduce it to a finite dimensional one which corresponds to optimizing a neural network with smooth objective function. More precisely, the first step is to solve \[ \phi^m(f)=\inf _{\begin{array}{l} h\in H^m\\ h\ge f \end{array}} \int hd\mu_0 \] numerically, where \(H^m\) can be a set of neural networks with a fixed structure (but unspecified parameter values), and \(m\) measures the number of neurons per layer. To allow for a step-wise updating of the parameters (e.g. by gradient descent methods) for the space \(H^m\), the inequality constraint \(h\ge f\) is penalized. To this end, a reference probability measure \(\theta\) on the state space \(X\) is introduced. This leads to the penalized problem \[ \phi^m_{\theta,\beta}(f)=\inf _{h\in H^m}\{\int hd\mu_0+\int \beta (f-h)d\theta \}. \] Numerical examples from optimal transport, martingale optimal transport, portfolio optimization under uncertainty and generative adversarial networks that showcase the generality and effectiveness of the approach are given.