Public articles linked to the same research event.
arXiv The work introduces an autoregressive differentiable method for 0-1 integer programs: it fixes an arbitrary order of binary variables and trains a transformer to predict the next bit while remaining in the feasible set, first training on feasible incumbents from any solver to initialize inside the feasible set, then adding a Lagrangian penalty and further training with Gumbel-softmax activations on the relaxed objective to explore the feasible set; on non-convex quadratic knapsack instances it reports consistent improvement over state-of-the-art open-source solvers for dense problems up to 10,000 binary variables and empirically demonstrates a tunneling-like effect.
The work introduces an autoregressive differentiable method for 0-1 integer programs: it fixes an arbitrary order of binary variables and trains a transformer to predict the next bit while remaining in the feasible set, first training on feasible incumbents from any solver to initialize inside the feasible set, then adding a Lagrangian penalty and further training with Gumbel-softmax activations on the relaxed objective to explore the feasible set; on non-convex quadratic knapsack instances it reports consistent improvement over state-of-the-art open-source solvers for dense problems up to 10,000 binary variables and empirically demonstrates a tunneling-like effect.
The work introduces an autoregressive differentiable method for 0-1 integer programs: it fixes an arbitrary order of binary variables and trains a transformer to predict the next bit while remaining in the feasible set, first training on feasible incumbents from any solver to initialize inside the feasible set, then adding a Lagrangian penalty and further training with Gumbel-softmax activations on the relaxed objective to explore the feasible set; on non-convex quadratic knapsack instances it reports consistent improvement over state-of-the-art open-source solvers for dense problems up to 10,000 binary variables and empirically demonstrates a tunneling-like effect.
The work introduces an autoregressive differentiable method for 0-1 integer programs: it fixes an arbitrary order of binary variables and trains a transformer to predict the next bit while remaining in the feasible set, first training on feasible incumbents from any solver to initialize inside the feasible set, then adding a Lagrangian penalty and further training with Gumbel-softmax activations on the relaxed objective to explore the feasible set; on non-convex quadratic knapsack instances it reports consistent improvement over state-of-the-art open-source solvers for dense problems up to 10,000 binary variables and empirically demonstrates a tunneling-like effect.