Originally posted by: TobiasAchterberg
The random seed parameter affects quite a number of places in the code, in particular the MIP solver. For example, it has an influence on probing, on the LP perturbation, on primal heuristics, and on branching. Because the decisions in MIP are based on previous results, a single change (for example in probing or the optimal LP basis of the root node relaxation) will affect all of the remaining search process. Moreover, because CPLEX internally solves MIPs to find primal solutions, to separate cuts, and for other purposes, the random seed parameter will also implicitly affect those components since it is also used for the sub-MIP solves.
We are using this very strong impact of the random seed parameter on purpose. The fact that changing the random seed will change the solving path on almost any problem instance makes scientific experimentation easier, as modifying the random seed can be used in statistical experiments to distinguish random noise from significant effects imposed by a change in the algorithm, the model, or the parameter settings.
Tobias
#CPLEXOptimizers#DecisionOptimization