Meta-heuristics like GAs are great tools. They are well-suited for certain jobs, and not so well-suited for others. For example: they perform poorly on search spaces where infeasibilty is common; they require a great deal of tuning and customization when there are intricate combinations of constraints; and as Matt pointed out, you have to provide your own bounds if you want a guarantee of how good the solution is. There is a growing academic literature that rigorously compares the performance of different algorithms on instances of the same problem, and predicts relative algorithmic performance based on statistical features of the instances. The answers those studies provide is almost always "it depends". There are no algorithms that always dominate, even on the same kind of problem. If you are interested in this topic, a good place to start is with the so-called "No Free Lunch" theorems. In practice, SAS/OR is very hard to beat. SAS provides a very broad collection of state of the art optimization algorithms, most of which are accessible from PROC OPTMODEL. In particular, even for problems for which GAs are well-suited, it is usually much less work in practice to encode the problem using simple expressions in OPTMODEL than it is to customize a GA (or Tabu Search, or GRASP, etc.) for that problem. The OPTMODEL code is also much more flexible and agile as the underlying business problem changes. Even implementing a GA is much easier to do in OPTMODEL than in most other environments and languages. With OPTMODEL you can do things like having your fitness function be itself a MILP or an NLP, or build a variety of hybrid algorithms that save months of development time and lots of money, and that are easy to read and maintain.
... View more