 |
Semi-Amortized Models for Lagrangian Relaxation
1 : LIPN
Université Paris-Nord - Paris XIII
2 : CERMICS
Ecole Nationale des Ponts et Chausées
We present machine learning techniques to predict parameters of Lagrangian Relaxation. The solutions
of these methods can be used either as approximations of the solutions returned by iterative algorithms such as
subgradient descent and bundle method, or as informed starting points for such algorithms, saving many iterations.
We evaluate our proposition on instances of the Multi-Commodity Fixed-Charge Network Design Problem and show
the merits of our method.
