GCG 0.4

design

working SCIP represents the original problem
read in problem is standard format (e.g., .lp-file, .mps-file) and corresponding blk-file
the master problem is represented by a relaxator, which replaces the LP-solving
separation is done by a separator included in the master SCIP, which calls the SCIPseparateSol() method in the original SCIP
branching is done simultaneously in the master and in the original SCIP
variables in the master are extreme points of conv(X) (and extreme rays, but functionality currently not implemented) -> convexification approach
many identical pricing problems -> fat too much effort for pricing, especially farkas pricing!
2 constraint handler assure that nodes in the master SCIP are linked to nodes in the original SCIP and vice versa
branching rules are implemented as branching rules in the master that have to define the additional callbacks specified in "type_branchgcg.h"

features

in blk-file constraints can be forced to be copied to the master problem; all other constraints are added to a block, if they only contain variables of this block
added possibility to use discretization approach, variables in the master then represent integer points in X, integrality in the master is demanded
identification of identical blocks -> break symmetrie in the master
2 branching rules implemented so far: branching on original variables (not possible if variable belongs to a block that has similar blocks in the problem), using pseudocost values to choose the variable to branch on ryan-foster type of branching for identical blocks with set partitioning structure
added GCGs own display plugin showing the number of LP iterations, of constraints, variables and cuts in the master