For general information on how to add your own primal heuristics to GCG, first check the SCIP documentation. However, one has to take into account some peculiarities when implementing heuristics that are included in the original SCIP instance, i.e. work on the original variables.
Many MIP heuristics make use of an LP feasible solution. In SCIP, such a solution is obtained by solving the LP relaxation. In GCG however, no LP relaxation is solved by default. A linearly feasible solution on the original variables comes from the GCG relaxator plug-in; it is a solution of the master LP that has been translated back into the original variables. To access it, one should use the method GCGrelaxGetCurrentOrigSol() in relax_gcg.h. Its fractional variables can be accessed by SCIPgetExternBranchCands() (rather than SCIPgetLPBranchCands() which is used in SCIP heuristics).
Note also that heuristics using LP solutions should use another timing than SCIP heuristics. Heuristics that are called after solving a node's relaxation typically have the timing SCIP_HEURTIMING_AFTERLPNODE. By default, no LPs are solved on the original problem. A heuristic relying on a linearly feasible solution should therefore have the timing SCIP_HEURTIMING_AFTERNODE to ensure that the heuristic is called at all. One then must ensure that the node's relaxation has indeed been solved to optimality and that the relaxation solution is valid. This can be done by placing
at the beginning of the HEUREXEC callback.
A common class of heuristics are diving heuristics; they solve LPs with modified bounds to perform a depth-first search on the Branch-and-bound tree. For this purpose, a probing mode and a diving mode have been implemented in SCIP, which can be invoked by SCIPstartProbing(scip) and SCIPstartDive(scip), respectively. In this mode, temporary bound changes on variables can be made, and modified LPs can be solved.
In GCG, a special probing mode has been implemented for the original instance. This mode serves for performing changes on the original instance but using the master LP instead of the original LP. It is invoked by GCGrelaxStartProbing() and terminated by GCGrelaxEndProbing() and features the methods GCGrelaxPerformProbing() and GCGrelaxPerformProbingWithPricing(), which will propagate any bound changes on the original instance to the extended instance and solve the resulting modified master LP, either without or with pricing new variables in. See e.g. heur_gcgcoefdiving.c for an example on how to use them.
The HEURCOPY callback is executed when a SCIP instance is copied, e.g. to solve a sub-SCIP. By defining this callback as
NULL the user disables the execution of the specified heuristic for all copied SCIP instances. This may deteriorate the performance of primal heuristics using sub-SCIPs.
For heuristics that are included in the original instance and make use of the extended instance as well (in particular, most of the heur_gcg* and heur_xp* plug-ins), this callback should be set to null. This is because sub-SCIPs are solved by SCIP rather than GCG and therefore do not know any master problem; including a GCG specific heuristic into them would cause errors.