17 instances was BB is the best thing to do - four strategies: preprocessing, cutting plane, sophisticated branching strategy, primal heuristics ##1 model preprocessing, algorithmic preprocessing - ![[Pasted image 20240306163849.png|100]] - milp vs combinatorial opt: A's structure (yes or no) ##2 valid ineq: uAx >= ub: "u" gomory cut tells us how to read this u from tableau (split cuts) ##3 branching - simplex is used by software as reoptimization if fast (feasibility; optimality is free) - two decisions: node (bfs, most promising) and variable (not well studies; variable value closer to .5) selection ##4 solving submilps; large neighborhood search - andrea don't see yet, cross learning (between different problem types (e.g. set covering and vrp) - for now just branch and bound for each - solving two lps for every variables for every branching - representing graph neural netwokrs (node on one side, constraints on the left) - branching on one variable (local info associated with var)