Inverse Optimization: An Efficient Learning Framework for Complex Behaviors

Abstract

We study a class of learning models known as inverse optimization (IO), where the goal is to replicate the behaviors of a decision-maker (i.e., optimizer) with an unknown objective function. We discuss recent developments in IO concerning convex training losses and optimization algorithms. The main message of this talk is that IO is a rich learning model that can capture complex, potentially discontinuous behaviors, while the training phase is still a tractable convex program. We motivate the discussion with applications from control (learning the MPC control law), transportation (2021 Amazon Routing Problem Challenge), and robotics (comparing with state-of-the-art methods, including deep neural networks, in MuJoCo environments).