Uncover our partner’s, Inria, paper, titled “QPLayer: efficient differentiation of convex quadratic optimization”.

The publication explores optimization layers in neural network architectures, focusing on convex Quadratic Programming (QP) layers for applications in learning, control, and robotics. The work leverages primal-dual augmented Lagrangian techniques to compute derivatives of both feasible and infeasible QPs, providing flexibility in the learning process. Experimental results demonstrate significant computational speed-ups compared to existing methods, with improved accuracy and numerical robustness. Additionally, the publication offers an open-source C++ software package called QPLayer for efficiently differentiating convex QPs which can be interfaced with modern learning frameworks.

Read the publication here.

Find AGIMUS publications here.