A primal dual algorithm for contextual stochastic combinatorial optimization

In this talk we will first provide a comprehensive introduction to decision-focused learning (DFL). We will cover the main applications of DFL and the different paradigms found in the literature, namely learning by imitation, or by experience. Learning wih combinatorial layers presents several challenges, among them the computation of meaningful gradients for backpropagation. The Fenchel-Young loss serves as the main regularization tool to overcome this obstacle. In a second part, we will present our recent contribution, that is a primal dual algorithm to learn directly by experience. This algorithm uses a decomposition-coordination scheme, where the coordination step results in a classical imitation learning problem in Fenchel-Young loss. Our algorithm provides convergence guarantees and strong empirical performances.