CPLEX comes with a good tool to do tuning that allows the modeler to set which algorithm to use for the problems of a certain structure/type. But outside of the tuning tool are there any rules of thumb or relationship between the kind of problems and the best algorithm suiting them and that too just for LPs with just continuous variables? I am talking about primal, primal devex, dual, dual steep1, dual steep2 etc.. I am observing that the subproblem in my implementation of Benders Decomposition performs faster with primal simplex method in the beginning few iterations, and then towards later iterations dual simplex performs better. All that is changing in the sub problem is the right hand side of a subset of constraints. asked 18 Jan '18, 02:38 Naveen Divak... 
I'm not sure how reliable these "rules of thumb" are, and they're my understanding (not necessarily accurately remembered).
For MIPs, I only have two rules of thumb. The first is that CPLEX is usually smarter than I am. By "usually" I mean "almost always". The second is that the node log sometimes gives hints. If optimal or near optimal solutions show up early but the bound moves slowly, consider parameter settings that emphasize bound improvement. If it takes a long time to get a good solution, consider parameter settings that encourage CPLEX to dive more deeply or apply search heuristics more frequently to get better primal solutions. answered 23 Jan '18, 15:49 Paul Rubin ♦♦ Thanks for the answer. It matches with what I see when I choose different algorithms. For sure these are not hard and fast rules.
(23 Jan '18, 16:03)
Naveen Divak...
